Multi-band radio frequency communication signal quality dynamic optimization method and intelligent water meter

Abstract

The present application relates to a multi-band radio frequency communication signal quality dynamic optimization method and an intelligent water meter. In view of the problem that the risk of water meter communication frequency band interruption is difficult to predict in advance in complex environment, a high sensitivity MEMS accelerometer integrated with a master control chip is used to collect three-dimensional micro-vibration signals of the water meter shell through low sampling rate, combined with time-frequency joint analysis to extract vibration modal characteristics, a graph neural network model is used to associate vibration state with radio frequency channel quality index, and real-time prediction of deterioration risk probability is realized. According to the environmental dynamic disturbance factor, the threshold value is adjusted and determined, combined with the radio frequency index step deterioration monitoring, the intelligent communication frequency band switching and context caching are triggered. The newly added vibration and radio frequency data are used for model adaptive optimization. The scheme can predict the channel interruption risk in advance, effectively improve the communication stability and adaptability of the intelligent water meter.

Description

Technical Field

[0001] This invention relates to the field of dynamic optimization of radio frequency communication signal quality and smart water meter communication management technology, and in particular to a method for dynamic optimization of multi-band radio frequency communication signal quality and a smart water meter. Background Technology

[0002] Currently, IoT terminals such as smart water meters widely employ multi-band radio frequency (RF) communication technology in practical deployments to improve the continuity and reliability of data transmission. To address issues such as complex pipe network environments and multi-path fading within buildings, mainstream solutions typically utilize real-time RF quality monitoring indicators such as RSSI (Received Signal Strength), SINR (Signal-to-Interference-Ratio), BER (Bit Error Rate), link retransmission rate, and packet loss rate. When the channel quality of the current operating frequency band is detected to be approaching a certain static or adaptive threshold, the RF front-end is triggered to switch to a backup frequency band or employ multiple concurrent links to ensure uninterrupted communication. Furthermore, some high-end IoT platforms have introduced trend prediction models based on historical RF data, reinforcement learning self-optimization strategies, cloud-based centralized collaborative frequency band scheduling, full-duplex link redundancy, and local multi-source data fusion methods to predict and avoid communication interruption risks in advance. The application of these technologies effectively improves the anti-interference capabilities and communication service stability of wireless terminals in industries such as water utilities and smart buildings.

[0003] However, the aforementioned existing technical solutions still face the following technical bottlenecks and defects in practical engineering applications. First, because the wireless channel state is highly dependent on the micro-disturbances of the deployment environment (such as water flow pulses in the pipeline, local vibrations of the pipe wall, micro-vibrations in buildings, etc.), relying solely on the quality indicators of the radio frequency itself for handover prediction suffers from insufficient sensitivity and response lag. This is mainly reflected in: the radio frequency signal itself is subject to interference from short-period noise and outliers, resulting in weak generalization ability of the prediction model; channel anomalies are often only perceived after the radio frequency monitoring indicators have deteriorated significantly, leading to unavoidable handover preparation and synchronization delays in the "detection-decision-execution" process; and in the case of delayed handover actions, it is easy to cause instantaneous link interruption or data frame loss. At the same time, some environmental perception methods based on multi-sensor fusion or relying on cloud-based augmented intelligence usually require additional configuration of external acquisition modules for vibration, temperature and humidity, and geographical location information, or obtaining handover control signals through cloud push. This not only increases the system hardware cost and energy consumption, but is also affected by factors such as external network latency and data synchronization security, making it difficult to fully meet the needs of extremely low power consumption and local autonomy in isolated environments such as smart water meters. Summary of the Invention

[0004] This application provides a method for dynamic optimization of multi-band radio frequency communication signal quality and a smart water meter, aiming to solve one of the problems or issues of the prior art mentioned in the background.

[0005] The multi-band radio frequency communication signal quality dynamic optimization method and smart water meter provided in this application specifically include: Dynamic optimization method for multi-band radio frequency communication signal quality: S1: During the idle period of the smart water meter communication, a high-sensitivity MEMS accelerometer integrated on the periphery of the main control chip is used to continuously collect the original signal of the three-dimensional axial micro-vibration of the water meter shell at an ultra-low sampling rate of less than 100Hz, and obtain the time-domain vibration data sequence including water flow pulsation, valve opening and closing and pipe wall resonance coupling excitation.

[0006] S2: Perform time-frequency joint analysis on the time-domain vibration data sequence, extract the dominant mode frequency, damping ratio and mode shape energy distribution parameters in the 0.5Hz to 20Hz frequency band, and generate a vibration mode fingerprint feature vector characterizing the physical response characteristics of the water meter.

[0007] S3: Based on the vibration mode fingerprint feature vector sequence of the same model of water meter in historical deployment scenarios within 30 to 120 seconds before the occurrence of different channel quality degradation events, the sequence is spatiotemporally aligned with the radio frequency signal received strength, signal-to-interference-plus-noise ratio and retransmission rate indicators of the corresponding time period, and a graph neural network topology structure is constructed with the vibration state of the water meter body as the node and the change of modal parameters in adjacent time windows as the edge weight.

[0008] S4: Input the real-time updated vibration mode fingerprint feature vector into the trained graph neural network topology, perform nonlinear mapping inference calculation, and output the channel degradation prediction result representing the probability value of the current communication frequency band interruption risk within the next 45-second time window.

[0009] S5: Dynamically adjust the risk judgment threshold based on the daily temperature gradient and water supply pressure fluctuation. If the channel degradation prediction result exceeds the dynamically adjusted risk judgment threshold three times in a row and the current radio frequency quality index does not fall below the baseline switching threshold, generate a frequency band switching pre-trigger command.

[0010] S6: In response to the frequency band switching pre-triggering command, perform target frequency band synchronization parameter acquisition and link layer context caching operations, complete the preloading and state maintenance of switching resources, and generate a pre-cached switching link context in an active state.

[0011] S7: During the pre-buffered switching link context hold period, monitor the instantaneous RF quality indicators in real time. If a step deterioration is detected with a single drop in RF signal received strength greater than or equal to 8dB or a sudden increase in retransmission rate greater than or equal to 40%, immediately activate the pre-buffered switching link context to perform physical layer frequency band switching.

[0012] S8: Record the evolution path of the vibration mode fingerprint feature vector and the actual radio frequency index change trajectory before and after this frequency band switching action, and feed it back to the historical deployment scenario database as new sample data to update the edge weight parameters of the graph neural network topology to achieve model adaptive optimization.

[0013] A smart water meter employs a multi-band radio frequency communication signal quality dynamic optimization method to optimize communication signal quality.

[0014] The multi-band radio frequency communication signal quality dynamic optimization method and smart water meter provided in this application have the following beneficial effects: (1) By introducing the micro-vibration signal of the water meter casing based on MEMS accelerometer as a forward proxy variable for channel quality evolution, an implicit correlation model between "vibration mode fingerprint" and future radio frequency interruption risk is constructed, which effectively overcomes the response lag problem caused by link switching relying on real-time RSSI threshold triggering or base station collaborative feedback in traditional wireless communication systems. Since physical processes such as water flow pulsation, valve action and pipeline resonance occur earlier than the significant degradation of channel performance, the low-frequency vibration characteristics (0.5–20Hz) they excite have strong time leading characteristics, which enables the system to predict potential link degradation trends before the communication quality has significantly deteriorated. Compared with existing technologies, the switching decision point is advanced by tens to hundreds of milliseconds, which significantly improves the foresight and accuracy of the switching timing judgment, avoids data packet loss and reconnection delay caused by sudden fading, and ensures the communication continuity and data reporting reliability of smart water meters in complex urban pipe network environments.

[0015] (2) A lightweight graph neural network is used to model the temporal vibration fingerprint. The change in modal parameters between adjacent time windows is used as dynamic edge weights. The nonlinear mapping relationship between the vibration evolution path at the device level and the channel degradation probability is learned, which solves the problem of poor generalization ability of the general prediction model in small sample and heterogeneous deployment scenarios. At the same time, the risk judgment threshold is adaptively adjusted by combining temperature gradient and water pressure fluctuation. A three-stage handover preparation mechanism of "preloading-caching-selective execution" is designed. The target frequency band context synchronization and parameter presetting are completed without actually initiating physical layer handover. The cache instruction is activated only when a momentary deterioration step is detected, realizing sub-second seamless handover. This mechanism balances handover agility and decision robustness without increasing conventional communication overhead, without requiring GPS positioning or map assistance, and without relying on information interaction on the base station side. It significantly reduces the false handover rate and resource waste, and improves the intelligent level of terminal autonomous decision-making and edge adaptability.

[0016] (3) The entire scheme takes the physical response of the terminal itself as the core sensing source, eliminating the dependence on complex architectures such as multi-source sensor fusion, reinforcement learning strategy optimization, or network layer resource scheduling, forming a lightweight, closed-loop, and locally-based prediction-caching-triggering paradigm that can be embedded in existing main control chips. Due to the high degree of device-specific characteristics and environment-dependent nature of vibration fingerprints, different water meters still exhibit distinguishable modal response differences under the same operating conditions, enhancing the robustness and interpretability of the model in cross-regional and cross-batch deployments, and is particularly suitable for autonomous health management of large-scale IoT terminals under conditions without continuous network feedback. This technical approach not only improves the dynamic adaptability of the communication link, but also provides a new idea for other constrained edge devices to achieve forward-looking network maintenance based on physical layer proxy signals, and has good scalability and engineering implementation value.

[0017] In summary, this solution, by mining the time-frequency characteristics of the vibration signal of the smart water meter, establishes an early warning mechanism for channel degradation, and integrates adaptive modeling and buffered switching control strategies, achieves highly sensitive, low-false-judgment, and fast-response wireless link resilience enhancement without additional communication overhead or external information dependence. This significantly improves the communication stability, switching efficiency, and system autonomy of IoT terminals in complex electromagnetic and physical environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for dynamic optimization of multi-band radio frequency communication signal quality and a smart water meter.

[0019] Figure 2 This is a flowchart of a method for dynamic optimization of multi-band radio frequency communication signal quality and a sub-flowchart for smart water meters.

[0020] Figure 3 This is another sub-flowchart of the method for dynamic optimization of multi-band radio frequency communication signal quality and smart water meters. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1 As shown, this application provides a method for dynamic optimization of multi-band radio frequency communication signal quality and a smart water meter, specifically including: Dynamic optimization method for multi-band radio frequency communication signal quality: S1: During the idle period of the smart water meter communication, a high-sensitivity MEMS accelerometer integrated on the periphery of the main control chip is used to continuously collect the original signal of the three-dimensional axial micro-vibration of the water meter shell at an ultra-low sampling rate of less than 100Hz, and obtain the time-domain vibration data sequence including water flow pulsation, valve opening and closing and pipe wall resonance coupling excitation.

[0024] S2: Perform time-frequency joint analysis on the time-domain vibration data sequence, extract the dominant mode frequency, damping ratio and mode shape energy distribution parameters in the 0.5Hz to 20Hz frequency band, and generate a vibration mode fingerprint feature vector characterizing the physical response characteristics of the water meter.

[0025] S3: Based on the vibration mode fingerprint feature vector sequence of the same model of water meter in historical deployment scenarios within 30 to 120 seconds before the occurrence of different channel quality degradation events, the sequence is spatiotemporally aligned with the radio frequency signal received strength, signal-to-interference-plus-noise ratio and retransmission rate indicators of the corresponding time period, and a graph neural network topology structure is constructed with the vibration state of the water meter body as the node and the change of modal parameters in adjacent time windows as the edge weight.

[0026] S4: Input the real-time updated vibration mode fingerprint feature vector into the trained graph neural network topology, perform nonlinear mapping inference calculation, and output the channel degradation prediction result representing the probability value of the current communication frequency band interruption risk within the next 45-second time window.

[0027] S5: Dynamically adjust the risk judgment threshold based on the daily temperature gradient and water supply pressure fluctuation. If the channel degradation prediction result exceeds the dynamically adjusted risk judgment threshold three times in a row and the current radio frequency quality index does not fall below the baseline switching threshold, generate a frequency band switching pre-trigger command.

[0028] S6: In response to the frequency band switching pre-triggering command, perform target frequency band synchronization parameter acquisition and link layer context caching operations, complete the preloading and state maintenance of switching resources, and generate a pre-cached switching link context in an active state.

[0029] S7: During the pre-buffered switching link context hold period, monitor the instantaneous RF quality indicators in real time. If a step deterioration is detected with a single drop in RF signal received strength greater than or equal to 8dB or a sudden increase in retransmission rate greater than or equal to 40%, immediately activate the pre-buffered switching link context to perform physical layer frequency band switching.

[0030] S8: Record the evolution path of the vibration mode fingerprint feature vector and the actual radio frequency index change trajectory before and after this frequency band switching action, and feed it back to the historical deployment scenario database as new sample data to update the edge weight parameters of the graph neural network topology to achieve model adaptive optimization.

[0031] Step S1: During the idle period of the smart water meter communication, a high-sensitivity MEMS accelerometer integrated on the periphery of the main control chip is used to continuously collect the original signals of the three-dimensional axial micro-vibration of the water meter casing at an ultra-low sampling rate of less than 100Hz, obtaining a time-domain vibration data sequence including water flow pulsation, valve opening and closing, and pipe wall resonant coupling excitation. Specifically, this includes: S1.1: Perform low-power wake-up control on the high-sensitivity MEMS accelerometer integrated on the periphery of the smart water meter main control chip to generate micro-vibration sensing execution instructions in a ready state, ensuring that the data acquisition process is started during the communication idle period.

[0032] During the idle period of the smart water meter communication, the input conditions are that the high-sensitivity MEMS accelerometer integrated on the main control chip is in a sleep state and the idle flag of the RF communication module is set. The execution target is the low-power wake-up control logic unit of the accelerometer. When the RF idle determination signal is true, a single pulse trigger signal is injected into the low-power wake-up control port of the accelerometer through register writing. The width of this signal is set to the minimum value that meets the hardware wake-up time constant, so as to achieve wake-up while ensuring energy efficiency. Based on the wake-up response readback delay parameter, a running status confirmation read is performed to verify that the hardware status bit of the accelerometer output has switched from the sleep state to the ready state. The sampling mode register is set using the peripheral bus controller to limit its operation to the three-dimensional axial full sampling mode during the idle period, and the sampling start delay is bound to the communication module unloading completion time to avoid the sampling signal from conflicting with the RF communication activity. By interacting with the clock management unit, the internal sampling clock division coefficient is adjusted to be lower than the corresponding value. A sampling frequency of Hz is used to meet the requirements of subsequent ultra-low sampling rate mechanical displacement data acquisition. The above hardware configuration results are encapsulated into micro-vibration sensing execution instructions. These instructions include core fields such as sampling mode, clock setting value, and working status flags, and are written to the accelerometer command buffer via the control bus to ensure that the data acquisition process can start instantly during communication idle periods. Through low-power wake-up control and sampling parameter configuration processing, the RF idle determination result of the previous step is transformed into a hardware-ready state that can execute sampling tasks at the physical layer, achieving seamless connection of the vibration data acquisition chain.

[0033] For example, in a smart water meter installed at the main water supply pipe of a floor, the MEMS accelerometer model is configured as XYZ-A1, and its static power consumption in sleep mode is less than μW. When the idle flag of the RF communication module is set to 1, the low-power wake-up pulse width is set to... The measured hardware wake-up time constant is ms. ms, ensuring pulse coverage of wake-up delay. After wake-up, the running status confirmation is read in... Completed within milliseconds, the status bit switches from 0 to 1, indicating readiness. The sampling mode register is set to full three-axis sampling, and a delayed binding communication offloading completion event is initiated to avoid radio frequency interference. The sampling clock divider is set to a fraction of the main frequency. The corresponding sampling frequency is Hz, satisfying the requirement of being below The Hz requirement. After the instruction is encapsulated and written to the command buffer, the average startup time for sampling during the communication idle period is measured to be [Hz value]. The power consumption of the entire chain is significantly reduced compared to conventional continuous sampling methods, and the data acquisition is free from radio frequency noise interference, providing a stable entry point for the accurate acquisition of subsequent mechanical displacement signals.

[0034] S1.2: Based on the micro-vibration sensing execution command, the mechanical displacement change of the water meter housing in the three-dimensional axis is discretized at an ultra-low sampling rate of less than 100Hz to obtain the original time-domain vibration data sequence containing water flow pulsation, valve opening and closing, and pipe wall resonance coupling excitation.

[0035] Based on the micro-vibration sensor execution command in the ready state, the axial sampling channel of the high-sensitivity MEMS accelerometer is called to synchronously sample the X, Y, and Z three-dimensional axial mechanical displacement changes of the water meter housing. This ensures that each axial sampling port works under the same time reference, so as to avoid phase distortion caused by timing misalignment between multi-axis data.

[0036] During the sampling process, the sampling period parameter is limited and the sampling frequency is fixed at a value lower than [the specified value]. The ultra-low sampling rate of Hz is used to reduce the proportion of high-frequency noise in the time-domain signal and reduce the power consumption load of the sampling task. The upper limit of this frequency is determined based on the length of the system idle cycle and the bandwidth requirements of the input signal.

[0037] At each sampling moment, the three-dimensional axial mechanical displacement change value is quantized and encoded. The analog voltage signal is converted into a digital quantity by an analog-to-digital converter (ADC), and the conversion result retains no less than [value missing]. The resolution is high enough to ensure that details of micro-vibrations are preserved.

[0038] The quantized displacement data streams of each axis are synchronized and aligned according to timestamps to form a discretized recording unit in the form of a three-dimensional vector. The data is then written into the storage cache in a time sequence to form a continuous original sample sequence, which is used to reflect the overall dynamic displacement characteristics excited by water flow pulsation, valve opening and closing, and pipe wall resonance coupling during communication idle periods.

[0039] A basic anomaly check is performed on the original discretized data in the cache to remove isolated outliers caused by instantaneous impacts or sensor start-up stabilization during the sampling process, ensuring the stability and integrity of the original time-domain vibration data sequence.

[0040] Through the above processing method, the sensor ready state in the previous step is transformed into a structured raw time-domain vibration data sequence, realizing the acquisition of micro-vibration signals of the water meter casing while taking into account both low power consumption and high-precision sampling.

[0041] For example, in a dynamically deployed smart water meter system, the three-axis sampling channels of the MEMS accelerometer are configured with sampling frequencies of [missing information]. Hz, ADC resolution is The total duration of the communication idle state within the sampling period is (bit, ). Seconds. During the acquisition process, the three-axis data are sampled synchronously under the same time base. After analog-to-digital conversion, they are converted into timestamp sequences in milliseconds and stored in a circular buffer. Each data unit is a continuous data unit. For the acquired discretized displacement vector sequence, an outlier removal operation is performed, with the detection threshold set when the displacement change on any axis exceeds the mean. A value exceeding one standard deviation is considered an anomaly and removed from the sequence. The above configuration achieved full fidelity in testing. The original time-domain vibration data sequence of Hz was successfully separated from the water flow pulsations in subsequent time-frequency analysis. Hz dominant mode frequency, caused by valve opening and closing Hz modal components and building micro-vibrations The Hz frequency peak significantly improves the accuracy of micro-vibration feature extraction.

[0042] S1.3: Perform power frequency interference filtering and baseline drift correction on the original time-domain vibration data sequence to eliminate the influence of environmental electromagnetic noise and sensor zero-point drift, and generate a pre-cleaned pure time-domain vibration data sequence.

[0043] S1.4: Based on the pre-cleaned pure time-domain vibration data sequence, sliding window truncation and overlapping splicing are performed to construct a set of micro-vibration signal frames with temporal continuity, providing standardized data units for subsequent time-frequency joint analysis.

[0044] S1.5: Perform amplitude normalization and dynamic range compression processing on the set of micro-vibration signal frames to unify the dimensional differences of vibration signals under different installation environments, and finally output a standardized time-domain vibration data sequence for subsequent feature extraction modules to use.

[0045] Step S2: Perform time-frequency joint analysis on the time-domain vibration data sequence to extract the dominant mode frequencies, damping ratios, and mode shape energy distribution parameters within the 0.5Hz to 20Hz frequency band, generating a vibration mode fingerprint feature vector characterizing the physical response characteristics of the water meter. Specifically, this includes: S2.1: Perform bandpass filtering preprocessing on the time-domain vibration data sequence to filter out DC drift components below 0.5 Hz and high-frequency mechanical noise interference above 20 Hz, thereby obtaining a clean broadband micro-vibration time-domain signal sequence.

[0046] Set the cutoff frequency parameter of the bandpass filter for the input standardized time-domain vibration data sequence, with the low cutoff frequency set to... Hertz is used to remove DC drift components, and the high cutoff frequency is set to Hertz is used to suppress high-frequency mechanical noise interference.

[0047] A second-order Butterworth bandpass filter is constructed based on the set frequency parameters. Its transfer function coefficients are calculated and the filter state vector is initialized for continuous data processing.

[0048] Forward filtering is applied to the input time-domain data sequence to transform the original vibration signal into a preliminary filtered output, eliminating low-frequency energy concentration and high-frequency sharp fluctuations.

[0049] The initial filtered output is subjected to reverse filtering, and the forward and reverse processes are combined to achieve zero phase distortion and preserve the true temporal characteristics of the vibration signal.

[0050] Amplitude consistency is checked on the forward and backward filtering results, and sample points that exceed the upper and lower limits of the dynamic range are proportionally scaled to ensure the physical rationality of the signal amplitude range.

[0051] By preprocessing with bandpass filtering, the time-domain vibration data sequence output from the previous step is transformed into a clean broadband micro-vibration time-domain signal sequence, thereby concentrating the spectral energy in... Expected technical effects in the Hertz range.

[0052] For example, in a smart water meter system installed in an underground pipeline, the length of the time-domain vibration data sequence after amplitude normalization is [missing information]. Number of sampling points, sampling rate Hertz. Set the low cutoff frequency of the bandpass filter. Hertz, high cutoff frequency Hertz, using the second-order Butterworth design method to solve for the coefficients, is used to perform forward filtering on the input filter of the acquired sequence. After forward filtering, inverse filtering is performed to remove phase distortion, and the output vibration signal spectrum is in Energy is significantly concentrated within the Hertz range, low-frequency DC drift components are completely suppressed, and the amplitude of high-frequency mechanical noise is significantly reduced, ensuring that the dominant mode frequency and damping ratio parameters extracted by subsequent time-frequency analysis have good stability and repeatability in complex underground environments.

[0053] S2.2: Perform short-time Fourier transform processing on the pure broadband micro-vibration time-domain signal sequence to map the time-domain waveform into a time-frequency two-dimensional energy distribution matrix, and obtain time-spectrum data containing information on water flow pulsation and pipe wall resonance coupling.

[0054] Frame length and frame shift parameters are set for the clean broadband micro-vibration time-domain signal sequence after bandpass filtering preprocessing. In each truncated frame, the Hanning window function is called to perform window weighting operation to reduce spectral leakage and maintain the integrity of modal information.

[0055] The weighted time-domain frame signal is transformed into a short-time Fourier transform using the discrete Fourier transform method, converting the time series into a complex spectral matrix distributed on the time and frequency axes.

[0056] The amplitude squares are calculated element by element for the complex spectrum matrix obtained by the transformation to obtain a two-dimensional time-frequency energy distribution matrix that characterizes the energy intensity of each frequency component.

[0057] Perform partitioning and indexing operations on the energy matrix along the frequency axis, retaining the corresponding energy distribution data within the 0.5Hz to 20Hz range to ensure that the analysis range is consistent with the target modal frequency bandwidth.

[0058] The extracted time-frequency sub-matrix is ​​calibrated for frequency resolution and time accuracy to generate time-spectrum data containing information on the coupling of water flow pulsation and pipe wall resonance.

[0059] This processing method transforms the broadband micro-vibration time-domain signal from the previous step into accurate time-spectrum data that can be used for peak search and subsequent modal analysis, thereby achieving a multidimensional expression of the signal's modal characteristics.

[0060] For example, on a DN20 smart water meter installed in an underground pipe network, the frame length of the broadband micro-vibration time-domain signal after bandpass filtering is set to... seconds, frame shift The window function uses a Hanning window, with a window length equal to the frame length. Within each frame... Point-based discrete Fourier transform, with frequency resolution of Hz. The energy matrix is ​​formed by summing the squares of the real and imaginary parts of the complex spectrum matrix. Retain by index. Hz to The energy distribution at Hz can be clearly observed in the obtained time-frequency spectrum. Hz to The continuous resonance peak at Hz caused by the pulsation of the water pump and at The sharp peaks near Hz, excited by tube wall resonance, can be directly located by subsequent peak searches, enabling the construction of input features for the channel state prediction model.

[0061] S2.3: Use the peak search method to detect local extreme points in the time spectrum data to identify the position of the resonant peak with the largest energy amplitude in the frequency band from 0.5 Hz to 20 Hz, and obtain the set of dominant mode frequency parameters characterizing the impact characteristics of pipeline fluid.

[0062] A local extremum search operation is performed on the time-frequency two-dimensional energy distribution matrix data generated by the short-time Fourier transform to locate the frequency components with peak energy amplitude characteristics in the 0.5 Hz to 20 Hz frequency band.

[0063] During the peak search process, the frequency axis data is traversed in segments. The energy amplitude sequence in each frequency segment is input into the threshold judgment function to remove frequency points with amplitudes lower than the set noise threshold, ensuring that extreme value detection is only for significant signal components.

[0064] The energy amplitude gradient change is calculated using the adjacent point difference method for the remaining frequency points. Points where the gradient changes from positive to negative are identified as candidate peak positions. A quadratic polynomial fitting operation is then performed near the candidate points to obtain the precise frequency of the peak position.

[0065] The peak normalization filtering method is used to compare the longitudinal amplitude of the fitting results, and secondary peaks with relative maximum amplitudes lower than a preset ratio are removed to form a stable and representative set of dominant resonance peaks.

[0066] The frequency value of each dominant resonance peak is recorded as the dominant mode frequency parameter and encapsulated as a parameter set output, providing standard input for subsequent damping ratio calculation and mode shape energy distribution analysis.

[0067] By using local extreme point detection and precise frequency fitting, the time-spectrum data from the previous step is transformed into a set of dominant modal frequency parameters characterizing the impact characteristics of pipeline fluid, thereby achieving high-precision extraction of the core frequency of vibration modes.

[0068] For example, in the vibration monitoring of a smart water meter with a rated flow rate of 15 cubic meters per hour, the time-domain signal after bandpass filtering undergoes a short-time Fourier transform to generate a 256×128 time-frequency matrix with a frequency resolution of 0.125 Hz. The frequency axis is segmented in 0.5 Hz increments, and the noise threshold is set to 0.02 times the global maximum value of the energy amplitude. In candidate peak identification, points with an amplitude difference exceeding 0.5 between adjacent frequency points and a gradient changing from positive to negative are selected as peak centers. Subsequently, a quadratic polynomial is used to fit the peak frequency within a range of ±0.125 Hz, with a fitting accuracy better than 0.01 Hz. The normalization filtering ratio is set to 0.6, and after removing low-amplitude peaks, the dominant mode frequency parameter set is output, including four frequency values: 0.875 Hz, 3.5 Hz, 7.25 Hz, and 12.75 Hz. Using these frequency values ​​for subsequent half-power bandwidth calculation can significantly improve the stability of damping ratio estimation. The robustness and consistency of this method for frequency extraction have been verified in multiple water meter installation environments. The output is a set of modal frequency parameters with high accuracy that is not affected by short-term environmental disturbances.

[0069] S2.4: Calculate the half-power bandwidth of the time-spectrum data based on the set of dominant mode frequency parameters to quantify the attenuation characteristics of each resonance peak and obtain a damping ratio parameter vector that reflects the constraint state of the water meter installation environment.

[0070] S2.5: Perform three-dimensional spatial energy integration based on the set of dominant modal frequency parameters and the damping ratio parameter vector to synthesize the relative contribution of each axial vibration component and generate the final vibration modal fingerprint feature vector characterizing the physical response characteristics of the water meter body.

[0071] Based on the dominant modal frequency parameter set and the damping ratio parameter vector, vibration components corresponding to X, Y, and Z along the three-dimensional axes are set as input objects for integration. For each axis, an energy spectrum function E(f) is constructed using its corresponding dominant modal frequency and damping ratio. This function is normalized with the dominant frequency as the center frequency and the damping ratio as the width factor. Integration is performed on the energy spectrum functions of different axes in the frequency domain, using a weighted integration method. The weights are determined based on the proportion of peak energy of each axis's modal frequency. The three-axis integration results are combined into a three-dimensional vector structure, and the relative contribution of each axis's vibration component is obtained through normalization. Feature concatenation is performed on this three-dimensional contribution vector, and it is combined with the original modal frequency parameter set and damping ratio parameter vector to form the final vibration modal fingerprint feature vector. Through the above three-dimensional spatial energy integration processing method, the frequency and damping parameter results of the previous step are transformed into quantifiable vibration component contribution indicators, realizing a complete characterization of the physical response characteristics of the water meter.

[0072] For example, in a DN20 smart water meter, the dominant modal frequency parameter set is 1.2Hz on the X-axis, 3.5Hz on the Y-axis, and 7.8Hz on the Z-axis, while the damping ratio parameter vector is 0.05 on the X-axis, 0.08 on the Y-axis, and 0.1 on the Z-axis. When constructing the energy spectrum function, a Gaussian energy spectrum function is generated with the modal frequencies as the center and the damping ratio as the width parameter. The frequency integration interval is set to 0.5Hz to 20Hz, and the weight for normalizing the peak energy to 1.0 is directly assigned based on the normalized amplitude. Calculations show that the integral value is 0.32 on the X-axis, 0.45 on the Y-axis, and 0.23 on the Z-axis, with normalized relative contributions of 0.32, 0.45, and 0.23 for each axis. The contribution vector is concatenated with the frequency and damping parameters to form a feature matrix [1.2, 0.05, 0.32; 3.5, 0.08, 0.45; 7.8, 0.1, 0.23], which is then used as the vibration modal fingerprint feature vector for subsequent graph neural network training. This process significantly improves the stability and discriminative power of vibration signals in the physical response characteristics characterization under complex installation environments.

[0073] like Figure 2 As shown, step S3 involves: Based on the vibration mode fingerprint feature vector sequence of the same model of water meter in historical deployment scenarios, occurring 30 to 120 seconds before different channel quality degradation events, spatiotemporally aligning it with the corresponding radio frequency signal received strength, signal-to-interference-plus-noise ratio (SINNR), and retransmission rate indicators for that time period. This constructs a graph neural network topology with the water meter's vibration state as nodes and the changes in modal parameters within adjacent time windows as edge weights. Specifically, this includes: S3.1: Based on the historical deployment scenario database, extract the vibration mode fingerprint feature vector sequence of the same model of smart water meter within the time window of 30 to 120 seconds before the occurrence of the channel quality degradation event, and perform timestamp standardization processing on the vibration mode fingerprint feature vector sequence to generate a historical vibration state time series dataset with a unified time reference.

[0074] Historical deployment scenarios refer to the collection of actual operation records of various smart water meter models under different physical installation environments (including different floors, different manhole locations, different regional pipe networks, and different water supply conditions). Each record includes the water meter model identifier, deployment location tag, vibration mode fingerprint feature vector sequence under continuous timestamps, and radio frequency signal received strength, signal-to-interference-plus-noise ratio, and retransmission rate indicators for the corresponding time period. The historical deployment scenario database is a structured storage unit used to store the above-mentioned multiple batches, multiple scenarios, and multiple model operation data. In actual use, historical data of the same model are selected from the database based on the current water meter model identifier for model training and prediction.

[0075] The vibration modal fingerprint feature vector is a multi-dimensional numerical combination composed of the dominant modal frequencies, damping ratios, and relative contributions of each axial vibration component of the water meter casing in a fixed order within the 0.5Hz to 20Hz frequency band. It is used to uniquely characterize the vibration response state of the water meter body at the current moment. Among them, the dominant modal frequencies reflect the speed of the main vibration of the water meter, the damping ratio reflects the rate of vibration decay, and the relative contributions of each axial vibration component reflect the distribution ratio of vibration energy in the X, Y, and Z directions.

[0076] The vibration modal fingerprint feature vector sequence is a data string formed by arranging the vibration modal fingerprint feature vectors collected under multiple consecutive time windows in chronological order. It is used to describe the evolution of the water meter vibration state over time.

[0077] Based on the operational log data of multiple batches of smart water meters stored in the historical deployment scenario database, using water meters of the same model as the filtering object, the search scope was limited to records marked with channel quality degradation events, and the time window parameter value was set between 30 and 120 seconds before the event trigger. For each retrieved record, the vibration mode fingerprint feature vector sequence within its corresponding time period was read to ensure that the sequence was continuous and without gaps within the time coverage. The vibration mode fingerprint feature vector sequence was processed by the timestamp standardization module. First, the format information and acquisition period parameter value of the original acquisition timestamp were read, and then all timestamps were converted into a unified absolute time base to eliminate time offsets caused by system clock drift or time zone differences between different acquisition batches. Time base alignment was performed on the converted absolute time series to map the time of each sampling point to a unified discrete time grid, and the interpolation calculation strategy of the vibration mode fingerprint feature vector was adjusted according to the discrete grid interval to complete the parameter values ​​of missing sampling points. Data version identifiers were generated for the interpolated feature vector sequence, and feature data under the same time base were uniformly incorporated into the storage structure of the historical vibration state time series dataset. Through the above data processing method, the time-frequency analysis results of the previous step are transformed into a historical vibration state time series dataset with a unified time reference, continuous and without missing data, and structured for model training, so as to achieve parameter consistency and time series comparability across deployment scenarios.

[0078] For example, in a batch of historical data from DN20 smart water meters deployed in an indoor pipe network, the database-marked channel quality degradation event trigger time is 14:32:10. The retrieval time window range is 30 to 120 seconds before the event, i.e., retrieving records with absolute times from 14:30:10 to 14:31:40. During this time period, the sampling period for the vibration modal fingerprint feature vector is 5 seconds, the original timestamp format is millisecond count, and there is a 200-millisecond system clock drift. After calling the standardization module, the original timestamps are uniformly converted to absolute times in UTC+8 format, and each feature vector is interpolated to the standard sampling grid at 5-second intervals. The interpolation method uses cubic spline interpolation. The two missing sampling points are used to obtain parameter values ​​after interpolation calculation. The dominant modal frequency smoothly transitions from 6.3Hz in the original data to 6.28Hz and 6.31Hz in the interpolated sequence. The interpolated feature vector sequence contains 19 consecutive sampling points, with all timestamps aligned to millisecond precision. In this example, the historical vibration state time series dataset generated after standardization and interpolation can be directly input into a lightweight graph neural network for training. During the validation phase, the model's channel degradation prediction accuracy for this batch of data was significantly improved, demonstrating the technical effectiveness of time reference unification and sequence continuity in cross-scenario model training.

[0079] S3.2: Using the time base of the historical vibration state time series dataset, perform spatiotemporal alignment operations on the radio frequency signal received strength, signal-to-interference-plus-noise ratio and retransmission rate indicators collected in the corresponding time period to construct a joint spatiotemporal correlation sample set containing multidimensional physical response parameters and multidimensional radio frequency quality parameters.

[0080] By utilizing the unified time reference of the historical vibration state time series dataset that has been timestamped and standardized, precise time alignment calculations are performed on the received radio frequency signal strength, signal-to-interference-plus-noise ratio, and retransmission rate indicators collected within the corresponding time period to ensure that the sampling points of different data sources are comparable at the same time node.

[0081] The received radio frequency signal strength sequence is interpolated and resampled according to a unified time base to eliminate the time deviation caused by different sampling frequencies and generate an equidistant sequence to meet the input requirements of subsequent joint modeling.

[0082] Synchronous interpolation and time window segmentation are performed on the signal-to-interference-plus-noise ratio (SNR) sequence to map the continuously changing SNR curve onto a discrete window that is consistent with the vibration mode fingerprint feature vector, thus forming a corresponding set of SNR window values.

[0083] Based on the retransmission rate index sequence, time offset correction and windowed statistics are performed on it. The average retransmission rate and burst characteristic values ​​within each time window are calculated according to a unified time base, so that they strictly correspond to the physical response parameters in the time domain.

[0084] A multi-dimensional data stitching strategy is adopted to merge the aligned vibration mode fingerprint feature vector with the received RF signal strength, signal-to-interference-plus-noise ratio and retransmission rate window sequence according to the time index to form a joint spatiotemporal correlation sample matrix, and to maintain the fixed position index of each parameter in the matrix, so as to facilitate the subsequent node and edge mapping of the graph neural network.

[0085] Through the above spatiotemporal alignment and matrix construction processing methods, the timestamp standardization results of the previous step are transformed into a joint spatiotemporal correlation sample set containing synchronous records of multidimensional physical response parameters and multidimensional radio frequency quality parameters, thereby achieving high-precision spatiotemporal mapping between vibration state and radio frequency quality.

[0086] For example, using historical data collected from a certain type of smart water meter deployed in an urban underground pipe network environment, the sampling frequency of the vibration modal fingerprint feature vector sequence is 10Hz, the sampling frequency of the radio frequency signal received intensity is 5Hz, the sampling frequency of the signal-to-interference-plus-noise ratio (SINR) is 2Hz, and the retransmission rate statistical period is 1 second. Cubic spline interpolation is applied to resample the signal received intensity sequence to 10Hz to match the vibration sequence. Linear interpolation is used to extend the SINR sequence to 10Hz, and the average retransmission rate and maximum spike value are calculated within each 1-second time window. In the constructed joint spatiotemporal correlation sample matrix, each time window contains three physical features: the dominant frequency of the vibration mode, the damping ratio, and the triaxial energy ratio, as well as four radio frequency features within the corresponding window: the mean RSSI, the mean SINR, the mean retransmission rate, and the spike value. Within this matrix, vibration and radio frequency indicators achieve precise correspondence in the time index. By inputting this sample set into the training module, the accuracy and stability of the model's prediction of channel degradation trends can be significantly improved.

[0087] S3.3: Based on the historical vibration state time series data in the joint spatiotemporal correlation sample set, the vibration mode fingerprint feature vector of each discrete time window is abstracted into an independent node in the graph topology to generate a set of nodes that characterize the physical response state of the water meter body at different times.

[0088] S3.4: For node pairs in adjacent time windows in the node set, calculate the change in modal parameters of the vibration modal fingerprint feature vector at the next time step relative to the vibration modal fingerprint feature vector at the previous time step, and quantify the change in modal parameters into the edge weight values ​​connecting the two nodes to generate a weighted edge set characterizing the dynamic characteristics of vibration evolution.

[0089] For adjacent time window node pairs in the node set, the vibration mode fingerprint feature vector representing the physical response state of the water meter body generated by the previous sub-step is received as input data to establish the correspondence between the node at the next time moment and the node at the previous time moment.

[0090] For each pair of nodes in the correspondence, the numerical components of the three dimensions of dominant modal frequency, damping ratio and mode shape energy distribution are extracted to form a parameter triplet.

[0091] By performing component-level subtraction on the triplet from the previous time step using difference operations, three difference parameters are obtained: change in modal frequency, change in damping ratio, and change in mode energy.

[0092] The three difference parameters are quantified in multiple dimensions using the weighted Euclidean norm. The difference of each dimension is multiplied by the corresponding weight coefficient, the squares are summed and the square root is taken to obtain the calculation result of the modal parameter change in a single scalar form. The weight coefficient is set according to the normalized value of the contribution of each modal parameter to the channel quality degradation prediction in the historical training samples.

[0093] The formula is as follows: in The modal frequency change This represents the change in damping ratio. This represents the change in mode energy. These are the weighting coefficients for each parameter.

[0094] The above calculation results are assigned to the edge weight values ​​of this node pair in the graph topology to form a weighted edge set that reflects the amplitude of vibration state evolution within adjacent time windows.

[0095] Through the above processing method, the node set of the previous step is transformed into structured data containing complete edge weight information, realizing the quantitative representation of the dynamic characteristics of vibration evolution in a lightweight graph neural network.

[0096] For example, in a set of historical data, the dominant mode frequency of the previous time window is Hz, the next time window is Hz; the damping ratio of the previous time window is The next time window is The previous time window mode energy was ×10 - ³J, the next time window is ×10 - ³ J. The weighting coefficients are set as follows: , , Calculate the differences between the parameters: Hz, , ×10 - ³ J. Multiply the difference by the weighting coefficient to obtain the weighted difference, sum the squares, and take the square root to obtain the change in modal parameters, which is approximately This value is assigned and stored as the edge weight of the node pair in the graph structure. In subsequent graph calculations, this edge weight effectively characterizes the dynamic change amplitude of the vibration state within this time window, and the prediction accuracy of the inference results of the prediction model with the participation of this weight data is significantly improved in predicting the risk of future communication frequency band interruptions.

[0097] S3.5: Integrate the node set and the weighted edge set to construct a lightweight graph neural network topology with the vibration state of the water meter body as nodes and the change in modal parameters within adjacent time windows as edge weights, so as to complete the initialization of the graph computing model used to output the channel degradation prediction results.

[0098] Based on the node set and weighted edge set output by the preceding sub-steps, an initialization object for the lightweight graph neural network topology is defined, and the node index of the vibration mode fingerprint feature vector corresponding to each historical time window is mapped to the row and column coordinate positions of the topology matrix.

[0099] All edge weights that have been quantized into values ​​are sparsely matrix-encoded and stored in the form of an adjacency matrix. Non-zero elements correspond to the modal parameter changes of adjacent time window node pairs, while zero elements correspond to node pairs with no direct evolutionary association.

[0100] A normalized Laplacian matrix is ​​constructed based on the adjacency matrix. The standardization of topological weights is achieved by multiplying the square root inverse of the node degree matrix with the adjacency matrix, as shown in the following formula: in For the normalized Laplace matrix, It is the identity matrix. Let the node degree matrix be... It is an adjacency matrix.

[0101] The normalized Laplacian matrix and the node feature matrix are paired and bound column-wise to form an initial data structure that can be input into the graph convolutional layer. Edge weight vectors are then embedded in this structure to achieve joint encoding of topology and features.

[0102] Parameterized weight initialization is performed on the initial data structure. The Xavier initialization method is used to generate a weight matrix that conforms to the topological size and feature dimensions to ensure that the model has a stable gradient distribution during the inference phase.

[0103] Through the above processing method, the node set and weighted edge set of the previous step are transformed into a lightweight graph neural network topology initialization model with strict mathematical definition and complete data structure, so as to realize an executable computation framework for subsequent channel degradation prediction.

[0104] For example, in a historical deployment scenario of a smart water meter, the node set contains vibration modal fingerprint feature vectors for 90 discrete time windows, each feature vector being a 12-dimensional floating-point number; in the weighted edge set, the modal parameter variation ranges from 0.05 to 1.2, the adjacency matrix is ​​a 90×90 sparse matrix, the proportion of non-zero elements is 0.18, and the diagonal elements of the degree matrix are the node connection values. The normalized Laplacian matrix is ​​used for calculation, and the identity matrix... The dimension is 90×90, and the square root inverse matrix elements are obtained through... Calculations show that, for example, when the degree matrix has 5 elements, its inverse square root is... The normalized Laplacian matrix has a uniform distribution of non-zero elements in a sparse storage state. Combined with the node feature matrix (90×12), it forms a graph convolution input data structure. The weight matrix (12×16-dimensional) is initialized using the Xavier method. In the validation phase, the model's forward propagation significantly improves the prediction stability of the probability value of future communication frequency band interruption risk. In the performance test, the gradient norm fluctuation amplitude is less than 0.005, indicating that the robustness of the initialized topology structure in prediction and inference under complex environments is improved.

[0105] like Figure 3 As shown, step S4 involves inputting the real-time updated vibration modal fingerprint feature vector into the trained graph neural network topology, performing nonlinear mapping inference calculations, and outputting a channel degradation prediction result representing the probability of interruption risk in the current communication frequency band within a 45-second time window. Specifically, this includes: S4.1: Perform node embedding encoding on the real-time updated vibration modal fingerprint feature vector and map it to the initial node hidden state vector in the lightweight graph neural network topology to establish a standard data interface between the physical response characteristics of the water meter and the graph model input space.

[0106] The initial conditions for input mapping of the vibration modal fingerprint feature vectors acquired in real time are defined, explicitly defining physical quantities such as three-dimensional axial modal frequencies, damping ratios, and mode shape energy distributions as the original carrier data for encoding. Based on the input specification of a lightweight graph neural network topology, the node embedding encoding module is invoked to perform numerical standardization on each modal parameter, eliminating dimensional differences and mapping them to a unified interval to ensure scale consistency of different physical response features in the encoding space. Using a parameter weight initialization strategy, the standardized modal features are assigned to corresponding graph nodes according to their node position indices, forming a multi-dimensional feature vector set at the node level. Affine transformations of the node feature vectors are performed using linear transformation operators and bias terms, projecting the original features onto a pre-defined hidden feature space. Considering the different sensitivities of different modal parameters to channel degradation trends, feature importance weighting is performed, enhancing the contribution of highly sensitive features in the hidden state through trainable weight coefficients. The encoded initial node hidden state vectors are normalized to ensure numerical stability and gradient propagation efficiency during subsequent convolutional message passing. By using this node embedding encoding method, a standard data interface is established between the graph topology structure constructed in the previous main step and the real-time vibration modal fingerprint feature vector, thereby achieving a high-fidelity mapping from physical response characteristics to the graph model input space.

[0107] For example, in a smart water meter deployment environment, the vibration modal fingerprint feature vector collected in real time includes three-dimensional axial dominant mode frequencies, respectively. Hz, Hz and Hz, the damping ratio parameter vectors are respectively , , The percentage of mode shape energy distribution is , , During node embedding encoding, each modal parameter is first normalized to the [0,1] interval according to its maximum value. For example, a frequency of 3.2Hz is normalized to... Then set the mapping matrix. Dimensions × Bias vector The initial hidden state vector is obtained by performing matrix multiplication on the zero vector. For example, the hidden state vector of the first node is calculated as follows: The result after encoding and mapping. Then, the feature importance weight coefficients are adjusted, for example, setting the frequency feature coefficients to... The characteristic coefficient of the damping ratio is set as The modal energy distribution characteristic coefficient is set as This approach significantly enhances highly sensitive features in the hidden space. After Z-score normalization, the mean of the node hidden state vector is zero and the standard deviation is one, effectively preventing gradient explosion or vanishing in subsequent graph convolutional layers. The final output initial node hidden state vector can significantly improve the accuracy and stability of channel degradation prediction.

[0108] S4.2: Based on the initial node hidden state vector and the preset edge weight parameters in the topology, perform graph convolution message passing operation, aggregate the vibration evolution path information carried by the modal parameter changes in adjacent time windows, and generate an intermediate layer node aggregation representation vector containing spatiotemporal correlation features.

[0109] For the initial node hidden state vector that has been mapped to the lightweight graph neural network topology, while maintaining the consistency of node feature dimensions, the pre-set edge weight parameters in the topology are used as message passing strength coefficients. Adjacency matrix multiplication is performed on each weighted edge connecting adjacent time window nodes to generate a neighbor node feature convergence matrix carrying modal parameter changes.

[0110] In the neighbor node feature aggregation matrix, the neighbor node feature vectors are weighted and summed using edge weight coefficients, and normalized according to node degree to ensure that the feature contribution of different time windows in the vibration evolution path remains relatively balanced during the aggregation process, eliminating feature bias caused by sparse or dense node connections.

[0111] A graph convolution kernel function is used to perform linear transformation and bias superposition on the normalized neighbor features, mapping the modal parameter changes to an intermediate feature space containing both temporal and spatial information, forming a preliminary aggregated representation vector that reflects the vibration state evolution process.

[0112] Spatiotemporal position encoding is introduced into the preliminary aggregated representation vector to encode the temporal window sequence information and the relative positional relationship of spatial nodes as additional feature components. These components are then fused into the aggregated representation vector through vector concatenation, thereby achieving a synchronous characterization of the temporal sequence and spatial correlation of the vibration evolution path.

[0113] Batch normalization and Dropout random deactivation are performed on the fused aggregated representation vector to suppress gradient instability and overfitting risks during model training, and finally output a stable intermediate layer node aggregated representation vector structure.

[0114] By using graph convolutional message passing and spatiotemporal location coding fusion processing, the initial node hidden state vector of the previous step is transformed into an intermediate layer node aggregation representation vector containing the modal parameter changes of adjacent time windows, temporal process and spatial correlation features, so as to realize the multi-dimensional feature aggregation required for channel degradation prediction.

[0115] For example, in the real-time prediction process of a certain type of smart water meter, the initial node hidden state vector dimension is set to 64, the edge weight parameters in the topology are taken from the training mean of historical samples, the adjacency matrix shape is 50×50, and the node degree ranges from 2 to 5. The neighbor node feature aggregation matrix obtained by performing adjacency matrix multiplication is normalized according to node degree and then input into the graph convolution kernel function. The convolution kernel weight matrix dimension is 64×128, and the bias vector dimension is 128. The 128-dimensional preliminary aggregation representation vector output by this linear transformation is introduced into position encoding. The time window encoding takes the form of alternating sine and cosine, and the spatial position encoding is assigned hierarchically according to the shortest path length of the node in the topology. After batch normalization processing, the Dropout inactivation rate is set to 0.2, and finally a 128-dimensional stable intermediate layer node aggregation representation vector is obtained. This vector is further used by a nonlinear activation function in subsequent steps to extract higher-order abstract features. The verification results show that the prediction model significantly improves the stability of calculating the probability value of channel interruption risk in the next 45 seconds under complex water supply pressure fluctuation conditions.

[0116] S4.3: Perform nonlinear activation function transformation on the aggregated representation vector of the intermediate layer nodes to extract high-order abstract features in the vibration state evolution trend and generate a high-dimensional latent space feature mapping vector that characterizes the dynamic change law of vibration mode fingerprint.

[0117] Multiple nonlinear activation function transformations are performed on the aggregated representation vector of intermediate layer nodes containing spatiotemporal correlation features, mapping it to a high-dimensional feature space to enhance the ability to capture complex patterns of water meter vibration state evolution trends.

[0118] Apply the hyperbolic tangent function transformation to the aggregated representation vector of each node, through The function compresses data in the interval [-1, 1], preserving details of small fluctuations while suppressing the effects of extreme values.

[0119] Applying a modified linear unit function transformation to the vectors after hyperbolic tangent transformation sets all negative values ​​to zero while retaining the linear growth property of positive values, ensures that the network maintains gradient stability in high activation regions.

[0120] The result of the modified linear unit transformation is processed using an exponential linear unit function, through... The function introduces smooth decay in the negative half-region, thereby improving the continuity of the characteristic distribution during the low-amplitude change phase of the vibration state.

[0121] Multiple sets of features obtained through different nonlinear function transformations are concatenated at the same index position to form a composite high-dimensional latent space feature mapping vector, which is used to characterize the nonlinear evolution of vibration modal fingerprints over time series.

[0122] By combining these nonlinear activation modes, the aggregated representation vector of intermediate layer nodes is transformed into a high-dimensional latent spatial feature mapping vector containing high-order spatiotemporal features, thereby enhancing the representation capability for complex dynamic changes.

[0123] For example, considering the vibration modal fingerprint feature vector collected from a certain model of smart water meter during its stable to fluctuating operating conditions, the intermediate layer node aggregate representation vector has a dimension of 64, and each component value is within the range of [-2.5, 3.1]. The vector is then processed element-wise. The function transformation yields compressed values ​​in the range [-0.99, 0.99]. Then, a ReLU transformation is performed, turning all negative values ​​to 0 while keeping positive values ​​unchanged. Finally, the ELU function is applied, resulting in exponentially decreasing values ​​in the negative region. For example, a value of -0.5, after ELU processing, becomes... The positive region remains linear. The three transformation results are concatenated by element index to form a 192-dimensional high-dimensional latent space feature mapping vector. In subsequent regression mapping, this vector is used to predict the channel quality change trend within the next 45 seconds. Test results show that this multi-activation combination significantly improves the model's prediction stability in complex deployment environments under the same training data, and the matching rate between the output risk probability value and the actual interruption event is greatly improved.

[0124] S4.4: Utilize the fully connected layer to perform dimensionality compression and regression mapping calculations on the high-dimensional latent space feature mapping vector, transforming the high-dimensional abstract features into single-dimensional scalar values, and generating an original risk score value that characterizes the possibility of channel quality degradation under the current communication environment.

[0125] For the high-dimensional latent space feature mapping vector obtained by nonlinear activation function transformation, the input dimension is set to... The target dimension of the mapping is Employing a fully connected layer weight matrix With bias vector Perform linear computation. Calculate the high-dimensional latent feature vector. Perform matrix multiplication with the weight matrix of the fully connected layer to generate an intermediate linear combination result. .

[0126] In the above intermediate results Based on this, superimpose the bias vector Perform a translation transformation to obtain the original output value before regression mapping. .

[0127] Input the original output values ​​before regression mapping into the regression calculation function. This function uses an identity mapping to maintain numerical continuity and outputs a single-dimensional scalar as the original risk score. .

[0128] Through matrix multiplication, bias superposition, and regression mapping function calculation of the fully connected layer, the high-dimensional latent space feature mapping vector is transformed into the original risk score value that characterizes the possibility of channel quality degradation under the current communication environment, thus realizing the direct mapping from features to risk score.

[0129] For example, in a smart water meter deployment scenario, the high-dimensional latent space feature mapping vector has a dimension of 256, and a fully connected layer weight matrix is ​​used. The dimension is set as Bias vector The value is set to a scalar of 0.05. In the actual calculation, the 256-dimensional eigenvectors are first multiplied with the weight matrix to obtain an intermediate linear combination result, such as... The value is 3.72. Then, a bias of 0.05 is added to... To obtain the original output value before regression mapping. =3.77. Because the regression function uses an identity mapping, the final original risk score is... The probability is 3.77. In this scenario, this score is mapped to the 0-1 range after subsequent probability normalization, resulting in a risk probability of 0.84. Verification shows that the actual channel degradation event corresponding to this probability value occurs precisely within the next 45 seconds, significantly improving prediction accuracy.

[0130] S4.5: Perform probability normalization on the original risk score value, map it to a continuous range from zero to one, and output the final channel degradation prediction result that represents the probability of interruption risk in the current communication frequency band within the next 45-second time window.

[0131] Step S5: Dynamically adjust the risk assessment threshold based on the daily temperature gradient and water supply pressure fluctuation. If the channel degradation prediction result exceeds the dynamically adjusted risk assessment threshold three times consecutively and the current radio frequency quality index has not fallen below the baseline switching threshold, generate a frequency band switching pre-trigger command. Specifically, this includes: S5.1: Obtain the daily temperature gradient data and water supply pressure fluctuation data as input conditions, and use the multivariate linear regression method to perform weighted fusion calculation on the environmental stress factor to generate the environmental dynamic disturbance coefficient characterizing the stability of the current deployment environment.

[0132] In this sub-step, the completed channel degradation prediction results are used as preliminary reference data, and the quantitative information of environmental physical disturbances at the deployment site is used as an adjustment factor. Daily temperature gradient data and water supply pressure fluctuation amplitude data are introduced as input conditions. A multivariate linear regression model is used to perform weighted fusion calculations of the environmental stress factor. The temperature gradient data is parsed into a unit-time temperature change rate index, and the water supply pressure fluctuation amplitude data is parsed into a unit-time pressure fluctuation amplitude index. Normalization is performed on both types of data to eliminate dimensional differences. The normalized temperature gradient index and pressure fluctuation index are used to construct multidimensional feature vectors according to preset weights, and the coefficient matrix of historical environmental disturbances and channel state samples is loaded into the regression model. Using the matrix multiplication method of multivariate linear regression, the feature vectors and coefficient matrices are multiplied to obtain the original fused disturbance values.

[0133] in, The original environmental disturbance value. For normalized temperature gradient index, This is a normalized pressure fluctuation indicator. and These are the weighting coefficients obtained from training with historical samples. The calculation results are smoothed and filtered to eliminate instantaneous abnormal fluctuations, and further, a continuous and stable sequence of environmental dynamic disturbance coefficients is generated using an exponentially weighted moving average method. Through these processing methods, the channel degradation prediction results from the previous step are numerically correlated with the current environmental stress conditions, providing a precise quantitative basis for environmental disturbances to dynamically adjust the subsequent risk assessment threshold, thus achieving consistency matching between pre-triggering conditions and the on-site environment.

[0134] For example, in a smart water meter deployment environment installed on a middle floor of a high-rise building, the normalized index value is 0.35 for a daily temperature gradient of 4.2 degrees Celsius / hour and 0.28 for a water supply pressure fluctuation of 12.5 kPa / hour. The weight coefficients obtained from historical training are then loaded. =1.15 and =0.87, forming an eigenvector and inputting it into the regression operation unit to obtain the original fused perturbation value. The value is 0.73. After being processed by a three-point moving average filter, this value stabilizes at 0.72, and the dynamic disturbance coefficient generated by the exponentially weighted moving average method is 0.70. This coefficient is used for the nonlinear mapping adjustment of the subsequent risk assessment threshold, ensuring that the decision logic of the frequency band switching pre-trigger command has significantly improved matching degree and reliability for the water meter environment of this floor.

[0135] S5.2: Based on the environmental dynamic disturbance coefficient, perform nonlinear mapping adjustment processing on the preset basic risk judgment threshold to output a dynamically adjusted risk judgment threshold that adapts to the current working conditions.

[0136] S5.3: Receive the channel degradation prediction results of three consecutive time windows as the input sequence, and use the sliding window counting method to count the number of consecutive times that the channel degradation prediction results exceed the dynamically adjusted risk judgment threshold, so as to generate a risk continuous satisfaction flag.

[0137] The channel degradation prediction results of three consecutive time windows are received as the input sequence. The probability values ​​of communication band interruption risk within the next forty-five seconds output by step S4.5 are called and arranged in chronological order to form a probability sequence of fixed length.

[0138] Using the dynamically adjusted risk judgment threshold as a comparison benchmark, threshold comparison processing is performed on each probability value in the input sequence to generate a corresponding over-limit flag sequence. A flag value of one indicates that the predicted value of the time window exceeds the limit, and a flag value of zero indicates that it does not exceed the limit.

[0139] Apply the sliding window counting method to the flag sequence, set the window length to three, sum the flag values ​​in each window, and output the cumulative number of times the limit is exceeded.

[0140] The cumulative number of overruns is compared with the window length. A risk satisfaction flag is generated using logical judgment operations. When the values ​​are equal, the risk satisfaction flag is set to one; when they are not equal, the flag is set to zero.

[0141] The generated risk continuously satisfies flag is output to step S5.4 for use by the current RF quality indicator comparison logic.

[0142] By combining the sliding window counting method with logical judgment, the channel degradation prediction result of the previous step is transformed into a binary flag signal that can be used to judge the persistence of the risk situation, thereby realizing the time continuity verification of the pre-triggering conditions for frequency band switching.

[0143] For example, in a smart water meter deployment environment, the channel degradation prediction results for three consecutive time windows are 0.78, 0.82, and 0.85, respectively, and the dynamically adjusted risk judgment threshold is 0.80. The three values ​​are arranged in chronological order and compared with the threshold to generate a flag sequence [0,1,1]. Using a sliding window algorithm with a window length of three, the cumulative number of exceedances is calculated as follows: The result is 2. Comparing the cumulative number of occurrences with the window length of 3, since 2 is not equal to 3, the risk continues to satisfy the flag value of zero. Under another environmental condition, if the prediction results are 0.82, 0.84, and 0.91, and the threshold remains 0.80, the generated flag sequence is [1,1,1], and the cumulative number of occurrences is calculated as follows: The result is 3, which is equal to the window length. Therefore, the risk continuous satisfaction flag value is one, which significantly improves the risk continuous satisfaction requirement in the frequency band switching pre-triggering condition.

[0144] S5.4: Read the current RF quality index value, perform a size comparison logic operation with the reference switching threshold to generate the current communication link availability status identifier.

[0145] Assuming the risk continues to meet the flag generation requirements, the current communication link's RF quality monitoring data is used as input. Real-time collected RF signal received strength, signal-to-interference-plus-noise ratio (SIR), and packet retransmission rate are loaded into the quality comparison control unit's input interface to establish a static reference value comparison benchmark with the baseline switching threshold. The received strength numerical analysis module is invoked to format the collected signal strength in physical units (dBm) and cache it in the quality comparison register. The threshold reading module is invoked to perform a register access operation on the baseline switching threshold, extracting the preset threshold value and outputting it to the comparison operator in dBm units. The RF signal received strength and the baseline switching threshold are input to the size comparison logic unit, which generates a link availability judgment result based on inequality rules. The comparison process can be represented as follows: in, The Boolean value result indicating the availability status of the communication link. This represents the current received radio frequency signal strength value. The threshold is used as a static reference value for switching. The judgment result is converted into a binary flag signal by the status encoding unit and stored in the link availability status register for direct use by subsequent logic AND gate decision modules. Through a chained processing method of quality analysis, threshold reading, comparison judgment and status encoding, a quantized interface is established between the risk continuously meeting the judgment conditions of the previous step and the current RF link availability status, realizing high-precision matching of availability judgment in the pre-trigger logic.

[0146] For example, in a smart water meter system deployed in an urban water supply network, the current measured value of the received radio frequency signal is... 64 dBm, signal-to-interference-plus-noise ratio of 14.2 dB, retransmission rate of 0.12, reference switching threshold set to 75 dBm. (This refers to the received signal strength value.) 64 is loaded into the quality comparison register, setting the reference switching threshold. 75 is loaded into the threshold register, and the comparison operator performs the determination process: Since the condition is met, the judgment result is true, and the link availability status flag is set to 1. This flag is stored in the link availability status register and is called by subsequent logic and decision units. Test results show that under this condition, the pre-trigger instruction generated in conjunction with the risk continuously satisfied flag can enter the handover preparation stage in advance when the channel degradation trend is obvious and the link is still available. After the handover resources are preloaded, it is activated in sub-seconds, and the communication stability is significantly improved.

[0147] S5.5: Based on the risk continuous satisfaction flag and the current communication link availability status identifier, perform a logical AND gate decision operation, and generate a frequency band switching pre-trigger instruction when both meet the preset conditions.

[0148] Based on the generated risk satisfaction flag and the current communication link availability status identifier, the input of the binary logic decision unit is loaded, and the risk satisfaction flag is assigned to the first input channel and the communication link availability status identifier is assigned to the second input channel.

[0149] Inside the logic decision unit, the two channel input signals are mapped to Boolean values ​​respectively. The risk continuously satisfied flag is true, indicating that the predicted risk has continuously exceeded the threshold. The communication link availability status flag is true, indicating that the radio frequency quality is still available.

[0150] Two Boolean truth values ​​are input into the AND logic unit, which performs a parallel input bitwise multiplication operation and outputs a logic high-level signal that satisfies the condition.

[0151] The logic high-level signal is subjected to gated filtering, and a sampling period consistency checker is used to ensure that the output trigger signal is only valid within the latest sampling period, so as to avoid false triggering caused by the risk of expiration.

[0152] The logic high-level signal that has passed the test is input into the frequency band switching pre-trigger instruction generation module. The instruction is formatted and encapsulated, and the current frequency band identifier, the target frequency band preset parameter index and the timestamp are added to generate the frequency band switching pre-trigger instruction used to drive the preloading of switching resources in the next step.

[0153] Through the above logic AND gate decision and signal encapsulation processing, the risk and availability determination results of the previous step are transformed into executable frequency band switching pre-trigger instructions, enabling early entry into the switching preparation stage when the channel degradation risk is high and the link is still available.

[0154] For example, in a dynamically deployed smart water meter application outdoors, the risk continuously satisfied flag is obtained by the sliding window counting method. A value of true indicates that the channel degradation probability has exceeded the dynamic adjustment threshold of 0.72 in the past three 10-second prediction periods. The current communication link availability status indicator is output by the radio frequency quality comparison module. A value of true indicates that the current RSSI is... 64dBm, higher than the baseline switching threshold 75dBm. The output is true when all inputs are true. After the high-level signal is confirmed as valid by the consistency checker, the instruction generation module encapsulates the frequency band switching pre-trigger instruction, appending the current frequency band ID=F12, the target frequency band ID=F8, and a timestamp of 1685023512. This instruction drives the frequency band synchronization parameter acquisition and link context caching module to perform a pre-loading operation in the next step. Actual deployment verification shows that, under these conditions, initiating handover preparation in advance can significantly improve the handover success rate and maintain data pass-through stability when communication degradation occurs.

[0155] Step S6: In response to the frequency band switching pre-trigger command, perform target frequency band synchronization parameter acquisition and link layer context caching operations, complete the preloading and state maintenance of switching resources, and generate a pre-cached switching link context in an inactive state. Specifically, this includes: S6.1: Based on the target frequency band identification information contained in the frequency band switching pre-trigger instruction, perform frequency synthesizer configuration parameter parsing processing on the radio frequency front-end control register to obtain a set of target frequency band synchronization parameters including center carrier frequency, channel bandwidth and spreading code sequence.

[0156] Based on the target frequency band identification information contained in the frequency band switching pre-trigger instruction, the identification value is loaded into the input of the RF front-end control register parsing module to determine the address range of the frequency synthesizer configuration register that needs to be parsed.

[0157] The original byte stream of the frequency synthesizer register output by the parsing module is processed by field separation according to the predetermined register mapping table to extract the encoded values ​​of the center carrier frequency field, channel bandwidth field and spreading code sequence field.

[0158] For the center carrier frequency field, the numerical decoding unit is invoked to convert its register encoded value into a frequency value in physical units.

[0159] For the channel bandwidth field, the bandwidth mapping table is called to match the register encoding index value to the corresponding bandwidth standard configuration, and the value in kilohertz (KHz) in physical units is output.

[0160] For the spreading code sequence field, bit stream reconstruction and error correction verification operations are performed to reconstruct the register encoded bit string into a complete spreading code sequence, and CRC verification is performed to ensure the integrity and correctness of the code sequence.

[0161] The decoded center carrier frequency value, channel bandwidth physical value, and spreading code sequence are integrated into a target frequency band synchronization parameter set and output to the subsequent link layer context buffer processing module.

[0162] By using register parsing and parameter decoding, the frequency band switching pre-trigger command generated in the previous step is converted into target frequency band synchronization parameter data that can be used for subsequent link resource preloading, thereby achieving accurate acquisition and consistency assurance of frequency, bandwidth and spreading code.

[0163] For example, in a smart water meter system operating in the UHF band, the target frequency band identifier value in the frequency band switching pre-trigger instruction is ID=F8. The parsing module locates the encoded value in the range of 0x1A to 0x1D in the frequency synthesizer register, where the center frequency field has a frequency division coefficient. =174, frequency dot coefficient =2, reference clock frequency =12.5 MHz, substituting into the calculation yields the center carrier frequency. =1087.5 MHz. The channel bandwidth field index value is 0x03, corresponding to a bandwidth value of 200 kHz in the mapping table. The spreading code sequence field register value is reassembled to obtain a 16-bit code string "1010011100110101", and the CRC check result is correct. The three parameters are integrated to form the target frequency band synchronization parameter set = {center frequency 1087.5 MHz, bandwidth 200 kHz, spreading code 1010011100110101}. The verification results of loading this set in the subsequent pre-buffered link context operation show that the target frequency band link establishment time is significantly shortened after the handover, and the communication stability is greatly improved.

[0164] S6.2: Use the target frequency band synchronization parameter set to perform snapshot capture processing on the medium access control layer state machine of the currently active communication link to extract the original data block of the current link layer context containing network address allocation, security key index and frame counter value.

[0165] Using the target frequency band synchronization parameter set output in step S6.1 as input, the center carrier frequency, channel bandwidth, and spreading code sequence parameters contained in this set are loaded into the parameter mapping unit of the link layer snapshot capture control module to establish a mapping table between radio frequency parameters and link layer state machine registers. The register read interface of the medium access control layer state machine is called, and register addresses related to the current communication session are accessed one by one according to the mapping table. A read operation is performed on the network address allocation register of the session management register group to extract the address allocation information of the current network node. The security key index register of the key index register group is accessed to obtain the security key index value of the current session and parse it into security parameters that can be loaded into the link layer after handover. The frame counter register is read to collect the frame counter value of the current session as a packet sequence state record of the link layer session. The network address allocation information, security key index value, and frame counter value read from the registers are combined according to a preset data structure to form a raw data block containing the complete state of the current link layer. Through the chained processing of register access and parameter extraction, the target frequency band synchronization parameter set of the previous step is transformed into a raw data block of link layer context that can be used for state transition, realizing the basic data preparation for subsequent context encapsulation and handover consistency.

[0166] For example, in a smart water meter system deployed in a municipal water supply network, the center carrier frequency parameter in the target frequency band synchronization parameter set is: MHz, channel bandwidth is kHz, spreading code sequence index value The register mapping table of the link-layer snapshot capture control module contains network address allocation register addresses. Security key index register address Frame count register address When performing a register read operation, from address... The network address allocation information read is the node ID. From address The security key index value read is From address The value of the frame counter read is The node ID, key index, and frame counter value are combined according to a preset format to obtain the raw data block of the link layer context. After being serialized, encapsulated, and validated for integrity in subsequent steps, this data block ensures a significantly improved stability of lossless transfer of link layer states and data pass-through when switching from the current communication frequency band to the target frequency band.

[0167] S6.3: Perform serialization encapsulation and integrity check code generation operations based on the original data block of the current link layer context to construct a standardized link layer context encapsulated data packet with anti-tampering characteristics, ensuring data consistency during state transition.

[0168] S6.4: Write the target frequency band synchronization parameter set and the standardized link layer context encapsulated data packet into the high-speed static random access memory reserved buffer inside the main control chip to complete the preloading and persistence of switching resources in the local storage medium.

[0169] S6.5: The status flag of the preloaded data in the reserved buffer of the high-speed static random access memory is processed to generate a pre-cache switching link context marked as ready and in an active state, which can be called by the subsequent step deterioration monitoring module in real time.

[0170] Using the target frequency band synchronization parameter set already written to the reserved buffer of high-speed static random access memory and the standardized link layer context encapsulated data packet as input objects, the status flag bit bit biting logic is executed on the buffer data, and the flag bit index is bound to the physical address of the buffer according to a preset mapping table to ensure that the index can directly point to the data unit to be activated when called in the future.

[0171] The status flag is assigned a logic high level signal and written to the status register. The hardware latching mechanism ensures that the flag remains unchanged while the data in the buffer is not overwritten by external write operations.

[0172] After the status register is updated, the consistency check module is called to perform cross-validation between the buffer data and the status flag bits, thereby confirming the integrity of the buffer data and verifying the correctness of the flag bits.

[0173] By combining the status flag and buffer index verification results, pre-cached switching link context metadata that is in a ready state and has an activation flag is generated. This metadata is written into the link context directory table for the step deterioration monitoring module to call in real time by index.

[0174] Add a timestamp field to the link context directory table to record the location of the flag, providing a basis for version comparison in subsequent calling stages and avoiding calling conflicts between different versions of link context.

[0175] By using status flag settings, register latching, data consistency verification, and directory management, the preloaded data from the previous step is transformed into an activated link context with immediate call capabilities, thus ensuring the secure preservation and rapid wake-up of resources during the switchover preparation phase.

[0176] For example, in a smart water meter deployment environment of a subway network, the physical address of the reserved buffer in the high-speed static random access memory is in the range of bytes 1024 to 1280. The target frequency band synchronization parameter set occupies the first 128 bytes, and the link layer context encapsulation data packet occupies the last 128 bytes. The status flag bit index is set to register address 0x1F, which points to the buffer start address 1024 in the mapping table. During the setting process, register 0x1F is assigned a logic high level, and the latch period is set to 50 milliseconds to ensure that the flag bit remains unchanged under short-term power supply fluctuations. The consistency check module reads the hash value of the buffer data and compares it with the checksum generated in the preloading stage. When the comparison results are consistent, a ready flag is generated. A new index entry ID=8 was added to the link context directory table, containing the physical address, flag index, and timestamp 1685023580. The step degradation monitoring module can directly call this ID when the signal strength differential sequence trigger condition is met. The corresponding buffer content is used to generate physical layer frequency band switching activation instructions. Field verification results show that this mechanism significantly reduces the interval between signal degradation and the start of the switching action, and effectively improves link stability.

[0177] Step S7: During the pre-buffered switching link context hold period, real-time monitoring of instantaneous RF quality indicators is performed. If a step-like degradation characteristic is detected, such as a single drop in RF signal received strength greater than or equal to 8dB or a sudden increase in retransmission rate greater than or equal to 40%, the pre-buffered switching link context is immediately activated to perform physical layer frequency band switching. Specifically, this includes: S7.1: Perform periodic sampling processing on the RF receive channel in the pre-buffered switching link context hold state to obtain the raw data stream of RF quality instantaneous indicators containing signal strength fluctuation information and retransmission count increment information.

[0178] S7.2: Perform sliding time window differential operation based on the original data stream of the instantaneous RF quality index to calculate the RF signal received strength differential sequence characterizing the rate of change of signal strength and the retransmission rate surge differential sequence characterizing the trend of retransmission frequency change.

[0179] Using the raw data stream of instantaneous RF quality indicators in a pre-buffered switching link context hold state as the input signal source, this data stream is divided into a set of sequentially arranged equal-length time windows according to a preset sampling period, forming a windowed data matrix with temporal continuity. For each RF signal received strength sample value sequence within the set of time windows, a difference operation is performed between the last sample value and the first sample value of the window. The difference values ​​of all time windows are arranged in chronological order to construct an RF signal received strength difference sequence characterizing the rate of change of signal strength. For the retransmission count increment sample value sequence within the set of time windows, a difference calculation is performed between the last sample value and the first sample value of the window. The retransmission difference values ​​of all time windows are arranged in chronological order to construct a retransmission rate surge difference sequence characterizing the trend of retransmission frequency change. Through the construction of the difference sequence, the raw data stream of instantaneous RF quality indicators output in the previous step is transformed into a difference sequence quantifying the rate of change, realizing the quantitative capture of instantaneous deterioration trends.

[0180] For example, in a certain deployment environment, the RF receiving channel continuously acquires RF signal strength and retransmission count increments at a sampling period of 0.5 seconds, forming a windowed data matrix with a length of 20 seconds, where each window contains 40 sampled values. When calculating the signal strength difference, the initial window value is -60dBm and the final value is -68dBm, resulting in a difference value of... dB indicates that the signal strength decrease rate within this window reaches the trigger threshold. When calculating the retransmission differential, the initial window value is 15 times, and the final value is 21 times, resulting in a differential value of [value missing]. The retransmission rate is significantly improved after the change in the value of the sudden increase. Combined with differential sequence analysis, the RF quality step deterioration trend corresponding to the window can be accurately determined and a flag bit can be output, providing reliable input data support for the subsequent step deterioration judgment logic.

[0181] S7.3: Using a preset step degradation judgment logic, perform threshold comparison analysis on the radio frequency signal received strength differential sequence and the retransmission rate surge differential sequence to generate a radio frequency quality step degradation flag bit that indicates whether the single drop amplitude is greater than or equal to eight decibels or whether the surge amplitude is greater than or equal to forty percent.

[0182] The input conditions include the RF signal received strength differential sequence and the retransmission rate surge differential sequence output from step S7.2, both obtained based on sliding time window differential operations and satisfying the sampling point index of a unified time reference. For the RF signal received strength differential sequence, the signal strength threshold comparison module in the step degradation judgment logic is called, and each differential value is compared with a preset single drop amplitude threshold to form a comparison operation unit. The sign determination and amplitude determination are combined to determine whether the differential value has reached the degradation condition. For the retransmission rate surge differential sequence, the retransmission frequency threshold comparison module in the step degradation judgment logic is called, and each differential value is compared with a preset surge amplitude threshold to form a comparison operation unit, and the same sign determination and amplitude determination processing is performed. The degradation judgment results of the signal strength differential sequence and the retransmission rate surge differential sequence are respectively generated into binary status identifiers, which serve as the basic data for subsequent logic input. The aforementioned binary status flags are input to a logic OR gate decision module. In this module, when any input status flag is true, the activation state of the step deterioration flag is output, ensuring that a handover decision is triggered when either signal strength or retransmission rate deteriorates. Through the above comparison and logic fusion processing, the differential sequence result from the previous step is transformed into an RF quality step deterioration flag that can directly drive the pre-buffered handover link context wake-up, achieving rapid identification and response to step deterioration characteristics.

[0183] For example, in a smart water meter deployment scenario, the sampling range of the signal strength differential sequence is -10.2 to +5.6, where the threshold for a single drop amplitude is set to... The sampling range of the differential sequence with a sudden increase in retransmission rate is -15 to +52, where the threshold for the sudden increase magnitude is set to... For the signal strength differential sequence, perform the following at each sampling point: and The comparison, if it satisfies If the sign is negative, a strength degradation flag value of 1 is generated; otherwise, 0 is generated. For differential sequences with a sudden increase in retransmission rate, each sampling point performs the following... and The comparison, if it satisfies If the sign is positive, a sudden deterioration flag value of 1 is generated; otherwise, 0 is generated. The above flags are input into a logical OR operation. In the resulting flag sequence, when either the intensity deterioration flag or the sudden deterioration flag is 1, a step deterioration flag activation state is output, thereby enabling rapid detection of significant single-incident RF quality degradation in this deployment environment. In this embodiment, the trigger accuracy of the deterioration flag is located when the signal strength decreases by -9.3 at sampling point 14 and the retransmission rate suddenly increases by +48 at sampling point 20, significantly improving the response speed and stability of frequency band switching triggering.

[0184] S7.4: Perform an immediate wake-up operation on the pre-buffered switching link context in the pending activation state according to the activation state of the RF quality step deterioration flag bit, so as to generate a physical layer band switching activation command containing target band synchronization parameters and complete link layer state information.

[0185] Based on the activation state of the RF quality step deterioration flag, read the target band synchronization parameters and complete link layer context structure data of the pre-buffered switching link context that are in the pending activation state.

[0186] The synchronization parameters of the target frequency band are subjected to consistency verification processing. The cached parameters are compared with the frequency band feature data collected in real time to ensure that there are no deviations in frequency, bandwidth and time slot configuration.

[0187] Integrity recovery operations are performed using session keys, buffer queue states, and retransmission records in the link layer context to generate a context mapping table that can be directly written to the RF front-end controller.

[0188] The target frequency band synchronization parameters that pass the verification are jointly encapsulated with the complete link layer context, and the two are combined with a unified data structure to form the activation command payload required for physical layer switching.

[0189] After the activation instruction payload is generated, the radio frequency control interface is called to generate the final physical layer frequency band switching activation instruction, which includes target frequency band synchronization parameters, link layer status information and execution flag bits.

[0190] Through the above-described step-by-step verification and encapsulation process, the step deterioration judgment result of the previous step is transformed into a physical layer activation command that can directly drive the RF front end to perform frequency band switching, ensuring the response accuracy and execution security of the switching process.

[0191] For example, in a smart water meter system deployed in a subway pipeline network environment, when the sliding time window differential analysis detects a single decrease in the received radio frequency signal strength, the magnitude of this decrease is... dB step characteristics and the retransmission rate sudden increase reaches When %, the flag is set to the active state. Retrieve the pre-buffered link context to confirm that the center frequency in the target band synchronization parameters is %. MHz, bandwidth kHz, the time slot configuration is completely consistent with the real-time monitoring data, and the number of unsent packets remaining in the buffer queue in the link layer context is The retransmission count is The center frequency, bandwidth, and time slot configuration are jointly encapsulated with the link layer buffer state. A physical layer frequency band switching activation command is generated through the RF control interface, and... Signal tuning and link information loading are completed within milliseconds. The data pass-through path is restored immediately after the handover. Verification results show that the handover process does not introduce additional communication delay and the link stability is significantly improved.

[0192] S7.5: In response to the physical layer frequency band switching activation command, control the radio frequency front-end module to perform carrier frequency reconstruction and time slot alignment processing to complete the physical layer frequency band switching action from the current communication frequency band to the target frequency band and restore the data pass-through path.

[0193] Step S8: Record the evolution path of the vibration mode fingerprint feature vector and the trajectory of the actual radio frequency index changes before and after this frequency band switching action, and feed this as new sample data back to the historical deployment scenario database to update the edge weight parameters of the graph neural network topology to achieve adaptive optimization of the model. Specifically, this includes: S8.1: Perform spatiotemporal alignment processing on the vibration modal fingerprint feature vector sequence within a preset time window before the frequency band switching trigger time and the vibration modal fingerprint feature vector sequence within a preset time window after the switching is completed, so as to construct a vibration modal fingerprint feature vector evolution path data block that characterizes the change in the body response caused by physical environment disturbance.

[0194] S8.2: Based on the vibration mode fingerprint feature vector evolution path data block, extract the real-time monitored radio frequency signal received intensity numerical sequence, signal-to-interference-plus-noise ratio numerical sequence and retransmission rate statistical sequence within the corresponding time period, and perform multi-source data timestamp matching operation to generate a dataset of actual radio frequency index change trajectories containing causal relationships.

[0195] S8.3: The vibration mode fingerprint feature vector evolution path data block and the actual radio frequency index change trajectory dataset are structured and encapsulated, and a success status label and environmental context metadata of this frequency band switching action are added to form a new sample data entry that meets the training format requirements.

[0196] S8.4: Write the newly added sample data entries into the dynamic growth sample pool of the historical deployment scenario database, and retrieve the historical stock samples of the same model of water meter under similar temperature gradient and water supply pressure fluctuation conditions in the database to construct a local sample subset for incremental learning.

[0197] S8.5: Using the co-occurrence frequency of the vibration mode fingerprint feature vector change and the actual radio frequency index degradation degree in the local sample subset, recalculate the edge weight parameters connecting adjacent time window nodes in the lightweight graph neural network topology to complete the adaptive optimization update of the model for the channel degradation mapping relationship under a specific deployment environment.

[0198] A smart water meter employs a multi-band radio frequency communication signal quality dynamic optimization method to optimize communication signal quality.

[0199] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0200] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0201] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic optimization method for multi-band radio frequency communication signal quality, specifically including: S1: Collect the original signal of three-dimensional axial micro-vibration of the smart water meter housing, and obtain the time-domain vibration data sequence including water flow pulsation, valve opening and closing and pipe wall resonance coupling excitation; S2: Based on the time-domain vibration data sequence, extract the dominant mode frequency, damping ratio and mode shape energy distribution parameters in the frequency band from 0.5Hz to 20Hz to generate a vibration mode fingerprint feature vector characterizing the physical response characteristics of the water meter body; S3: Constructing a graph neural network topology based on historical deployment scenarios; S4: Input the real-time updated vibration mode fingerprint feature vector into the graph neural network topology and output the channel degradation prediction result representing the probability value of the current communication frequency band interruption risk within the next 45-second time window; S5: Dynamically adjust the risk judgment threshold according to the daily temperature gradient and water supply pressure fluctuation. If the channel degradation prediction result exceeds the risk judgment threshold three times in a row and the current radio frequency quality index has not fallen below the baseline switching threshold, generate a frequency band switching pre-trigger command. S6: Respond to the frequency band switching pre-trigger command, perform target frequency band synchronization parameter acquisition and link layer context caching operations, complete the preloading and state maintenance of switching resources, and generate a pre-cached switching link context in an active state; S7: During the pre-buffered switching link context hold period, monitor the instantaneous RF quality indicators in real time. If a step deterioration is detected with a single drop in RF signal received strength greater than or equal to 8dB or a sudden increase in retransmission rate greater than or equal to 40%, activate the pre-buffered switching link context to perform physical layer frequency band switching action.

2. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, The graph neural network topology constructed based on historical deployment scenarios is specifically based on the vibration mode fingerprint feature vector sequence of the same model of water meter in historical deployment scenarios within 30 to 120 seconds before different channel quality degradation events, and spatiotemporally aligned with the radio frequency signal received strength, signal-to-interference-plus-noise ratio and retransmission rate indicators of the corresponding time period, to construct a graph neural network topology with the vibration state of the water meter body as nodes and the changes in modal parameters within adjacent time windows as edge weights.

3. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, The historical deployment scenarios refer to the collection of actual operation records of various models of smart water meters in different physical installation environments. Each record includes the water meter model identifier, deployment location label, vibration mode fingerprint feature vector sequence under continuous timestamps, and radio frequency signal received strength, signal-to-interference-plus-noise ratio and retransmission rate indicators for the corresponding time period.

4. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, Following S7, the following also includes: S8: Record the evolution path of the vibration mode fingerprint feature vector and the actual radio frequency index change trajectory before and after this frequency band switching action, and feed it back to the historical deployment scenario database as new sample data to update the edge weight parameters of the graph neural network topology to achieve model adaptive optimization.

5. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, Step S3 specifically includes: Based on the historical deployment scenario database, the vibration mode fingerprint feature vector sequence of the same model of smart water meter is extracted within the time window of 30 to 120 seconds before the occurrence of channel quality degradation event, and the vibration mode fingerprint feature vector sequence is timestamped to generate a historical vibration state time series dataset. Using the time base of historical vibration state time series dataset, the received intensity, signal-to-interference-plus-noise ratio and retransmission rate of radio frequency signals collected in the corresponding time period are spatiotemporally aligned to construct a joint spatiotemporally correlated sample set containing multidimensional physical response parameters and multidimensional radio frequency quality parameters. Based on the historical vibration state time series data in the joint spatiotemporal correlation sample set, the vibration mode fingerprint feature vector of each discrete time window is abstracted into an independent node in the graph topology, generating a set of nodes that characterize the physical response state of the water meter body at different times. Generate a weighted edge set based on the node set; By integrating the node set and the weighted edge set, a graph neural network topology is constructed with the vibration state of the water meter body as the node and the change in modal parameters within adjacent time windows as the edge weight.

6. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 5, characterized in that, The generation of a weighted edge set based on a node set specifically involves calculating the change in modal parameters of the vibration modal fingerprint feature vector at the next time step relative to the vibration modal fingerprint feature vector at the previous time step for node pairs in adjacent time windows in the node set, and quantifying the change in modal parameters into edge weight values ​​connecting the two nodes to generate a weighted edge set characterizing the dynamic characteristics of vibration evolution.

7. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, Step S4 specifically includes: The vibration modal fingerprint feature vector updated in real time is encoded by node embedding and mapped to the hidden state vector of the initial node in the graph neural network topology, thus establishing a standard data interface between the physical response characteristics of the water meter and the input space of the graph model. Based on the initial node hidden state vector and the preset edge weight parameters in the topology, the graph convolution message passing operation is performed to aggregate the vibration evolution path information carried by the modal parameter changes in adjacent time windows, and generate an intermediate layer node aggregation representation vector containing spatiotemporal correlation features. Generate high-dimensional latent space feature mapping vectors based on the aggregated representation vectors of intermediate layer nodes; Dimension compression and regression mapping calculations are performed on the high-dimensional latent space feature mapping vector to transform the high-dimensional abstract features into single-dimensional scalar values, generating an original risk score that characterizes the possibility of channel quality degradation under the current communication environment. The original risk score is probabilistically normalized and mapped to a continuous range from zero to one. The final channel degradation prediction result, which represents the probability of interruption of the current communication frequency band within the next 45-second time window, is output.

8. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 7, characterized in that, The process of generating a high-dimensional latent space feature mapping vector based on the intermediate layer node aggregation representation vector specifically involves performing a nonlinear activation function transformation on the intermediate layer node aggregation representation vector to extract high-order abstract features in the vibration state evolution trend and generate a high-dimensional latent space feature mapping vector that characterizes the dynamic change law of vibration mode fingerprint.

9. The method for dynamic optimization of multi-band radio frequency communication signal quality and the smart water meter according to claim 1, characterized in that, Step S5 specifically includes: The daily temperature gradient data and water supply pressure fluctuation data are used as input conditions. The environmental stress factor is weighted and fused using a multivariate linear regression method to generate an environmental dynamic disturbance coefficient that characterizes the stability of the current deployment environment. Based on the environmental dynamic disturbance coefficient, a nonlinear mapping adjustment process is performed on the basic risk judgment threshold to output a dynamically adjusted risk judgment threshold that adapts to the current working conditions. The channel degradation prediction results of three consecutive time windows are received as the input sequence. The sliding window counting method is used to count the number of consecutive times that the channel degradation prediction results exceed the dynamically adjusted risk judgment threshold, and a risk continuous satisfaction flag is generated. Read the current radio frequency quality index value, perform a size comparison logic operation with the benchmark switching threshold, and generate the current communication link availability status identifier; Based on the risk continuously satisfied flag and the current communication link availability status indicator, a logical AND gate decision operation is performed, and a frequency band switching pre-trigger instruction is generated when both meet the preset conditions.

10. A smart water meter, characterized in that, The method for dynamic optimization of multi-band radio frequency communication signal quality according to claim 1 is used to optimize the communication signal quality.

Images