A method for correcting temperature drift and bias of multiple sensors for a roadside berth

Abstract

The application relates to the technical field of intelligent parking and Internet of Things edge computing, and particularly discloses a multi-sensor temperature drift and bias correction method for a roadside parking space, which comprises the following steps: initializing multi-sensor collection and space baseline construction at a target roadside parking space, synchronously collecting multi-sensor time sequence data by each sensor, and establishing a space joint distribution model in an idle state; an improved fruit fly algorithm is introduced, a plurality of sensors are regarded as a group, global space bias drift caused by interference events is quickly identified through space smell concentration search, and real-time elimination is performed; the improved fruit fly algorithm is used to identify the global space bias drift, a sliding window principal component analysis is used to separate common mode interference, a double-layer reinforcement learning is used to compensate individual temperature drift, a convolutional neural network is used to identify a degradation mode, a time sequence difference learning is used to apply predictive compensation, a complete space environment disturbance elimination, time domain temperature drift calibration and equipment aging compensation closed loop are formed, and the equipment can maintain a high-precision baseline model and stable decision-making ability.

Description

Technical Field

[0001] This invention relates to the fields of smart parking and IoT edge computing technology, and in particular to a multi-sensor temperature drift and offset correction method for roadside parking spaces. Background Technology

[0002] In roadside parking scenarios, flip-up parking devices are exposed to the outdoor environment for extended periods, facing significant temperature changes, humidity fluctuations, and complex electromagnetic interference. Sensors are prone to temperature drift and bias deviation, leading to baseline shift and signal distortion, which in turn affects the accuracy of vehicle occupancy status identification. Furthermore, the response characteristics of sensing elements such as geomagnetic, ultrasonic, and millimeter-wave sensors differ at different temperatures. Without an effective correction mechanism, fixed thresholds will struggle to adapt to environmental changes, easily causing misjudgments and missed judgments.

[0003] For example, a data processing method for an accelerometer, as disclosed in Chinese Patent Publication No. CN121299166A, achieves high-fidelity denoising and dynamic compensation of acceleration signals by constructing an environmental noise model and implementing adaptive filtering and temperature drift correction, thereby improving the consistency of multi-channel measurements and the stability of system operation.

[0004] In the actual deployment and operation of roadside flip-type berth equipment, rainwater accumulation or snowmelt can cause abrupt changes in the spatial background field covering the entire berth area, leading to global consistency bias drift in the baselines of multiple sensors. This causes the occupancy decision based on the historical vacancy ground state to become instantly invalid. Moreover, temporal environmental factors such as high temperature exposure can also induce individual temperature drift in sensors such as geomagnetic sensors. The slowly changing baseline characteristics of these sensors are coupled with the signal characteristics of actual vehicle occupancy, resulting in an imbalance in the multimodal sensor fusion weights and a blurring of the decision boundary. At the same time, with the extension of the service life, the degradation bias of the sensor performance itself is further superimposed. The signal attenuation caused by decreased sensitivity is intertwined with environmental interference and temperature drift effects, causing the equipment to be in a sub-healthy state for a long time and produce intermittent missed alarms. Traditional correction methods are difficult to decouple and dynamically compensate for the above-mentioned spatial-temporal-degradation three-dimensional coupling bias. Therefore, a multi-sensor temperature drift and bias correction method for roadside berths is proposed to solve the above-mentioned problems. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a multi-sensor temperature drift and offset correction method for roadside parking spaces, which can effectively solve the problems involved in the prior art.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a multi-sensor temperature drift and offset correction method for roadside parking spaces, comprising the following steps:

[0007] Step 1: Initialize multi-sensor data acquisition and spatial baseline construction at the target roadside berth. Each sensor synchronously acquires multi-sensor time-series data including geomagnetism, ultrasound, millimeter wave, etc., and establishes a joint spatial distribution model in the idle state.

[0008] Step 2: Introduce an improved fruit fly algorithm, treating multiple sensors as a group. By searching for spatial odor concentration, quickly identify and remove global spatial bias drift caused by interference events such as water accumulation after rain in real time, thus achieving rapid identification of global environmental interference and real-time removal of abnormal data.

[0009] Step 3: Use sliding window principal component analysis to decompose the multi-sensor time-series data stream online, dynamically separate and compensate for the common-mode bias component caused by environmental abrupt changes, and achieve second-level self-calibration of the spatial baseline;

[0010] Step 4: Construct a two-layer reinforcement learning model. The first layer uses high-confidence occupancy pseudo-labels as reward signals to learn the individual temperature drift compensation function of each sensor at different temperatures to eliminate time-domain temperature drift interference. The second layer dynamically adjusts the multi-modal fusion weights based on the consistency and signal-to-noise ratio of the compensated signals, so that the decision boundary always tends to the reliable sensor combination in the current environment.

[0011] Step 5: Use the convolutional neural network in the image classification algorithm to encode the multi-sensor time series data into a time-frequency graph, identify the weak feature degradation patterns related to the health of the device, quantify the degree of sensor performance degradation, identify performance degradation at an early stage, and realize quantitative assessment of health.

[0012] Step 6: Establish a joint state machine that integrates health status, environmental status, and previous calibration results. Through time-series differential learning, apply predictive compensation biases to devices in different joint states to correct signal attenuation caused by performance degradation in advance, dynamically predict compensation, and extend the maintenance-free cycle of the devices.

[0013] Preferably, step 1 specifically includes:

[0014] When the berth is idle, control the multi-modal sensors, including geomagnetic, ultrasonic, and millimeter-wave sensors, to trigger the acquisition mode with a preset synchronous clock, acquire multi-sensor time-series data for no less than N consecutive sampling cycles, and perform timestamp calibration and resampling alignment on the data of each channel to ensure that the multi-sensor data can be compared and analyzed on a unified time reference.

[0015] Statistical features are extracted from the aligned multi-sensor time-series data, including the mean, variance, quantile, noise floor energy, and short-time fluctuation index of each sensor signal. A multi-dimensional feature vector is constructed in the idle state to fully characterize the intrinsic distribution characteristics of each sensor in the idle state.

[0016] The multidimensional feature vector is input into a multivariate Gaussian distribution fitter to calculate the joint probability density function of each sensor, forming a spatial joint distribution model containing the mean vector and covariance matrix. This model serves as the baseline for subsequent drift detection, establishing a statistical benchmark model for quantifying the correlation of multiple sensors.

[0017] Preferably, step 1 further includes:

[0018] During the construction of the joint spatial distribution model, ambient temperature data is collected synchronously and the correlation coefficient between each sensor signal and temperature is recorded. A temperature-baseline mapping table is established to distinguish between temperature drift and spatial baseline shift caused by sudden environmental changes, effectively decoupling temperature drift from environmental interference and improving the accuracy of baseline drift cause identification.

[0019] Periodic idle window scanning is performed on multi-sensor time series data. When the statistical characteristics of newly acquired data deviate from the baseline by more than a preset threshold, the baseline update mechanism is triggered. The mean vector and covariance matrix are updated recursively using exponential weighted moving average to achieve baseline adaptive tracking and avoid long-term drift accumulation caused by model solidification.

[0020] The updated spatial baseline results are stored in non-volatile memory, and the update timestamp and environmental label are recorded to form a spatial baseline library containing historical versions. This supports multi-version comparison and backtracking during subsequent drift analysis, constructs a traceable baseline evolution trajectory, and provides data support for degradation analysis and parameter rollback.

[0021] Preferably, step 2 specifically includes:

[0022] By treating multiple sensors such as geomagnetism, ultrasound, and millimeter waves as individuals within a fruit fly colony, and defining the real-time reading of each sensor as the spatial coordinates of that individual, a population distribution matrix in a multi-dimensional sensing space is constructed to achieve multi-sensor collaborative perception, laying a spatial mapping foundation for global bias drift identification.

[0023] Define an airspace odor concentration determination function, which maps the deviation of each sensor reading from the corresponding dimension in the idle baseline model at the current moment to the odor concentration value. Find the location of the individual with the highest concentration through group iterative search, identify the global bias drift direction and amplitude, accurately locate the bias drift direction and amplitude, and improve the sensitivity and accuracy of interference identification.

[0024] When it is detected that more than half of the individuals in the group simultaneously exhibit a large shift in the same direction for a duration exceeding a set threshold, it is determined to be a global spatial bias drift caused by interference events such as post-rain water accumulation. The current frame data is marked as abnormal and removed, and a rapid recalibration of the spatial baseline is triggered to promptly remove abnormal data, block interference propagation, and quickly restore the effectiveness of the baseline model.

[0025] Preferably, step 3 specifically includes:

[0026] Set a fixed-length sliding data window, center the multi-sensor time-series data stream within the window, calculate the covariance matrix and perform eigenvalue decomposition, extract the principal component vector and its corresponding eigenvalue, eliminate the DC component of the data, and make subsequent analysis focus on the fluctuation component that reflects the interference.

[0027] The principal component with the largest eigenvalue is identified as the common-mode bias component. The common-mode bias component corresponds to the same-direction drift component that occurs simultaneously in each sensor. The projection coefficient of the common-mode bias component is multiplied by the principal component vector to obtain the common-mode bias estimate. Then, the common-mode bias estimate is subtracted from the original multi-sensor time series data to achieve second-level self-calibration of the spatial baseline, real-time removal of global same-direction interference, restoration of baseline purity and ensuring decision reliability.

[0028] Preferably, the process of constructing the first layer of the two-layer reinforcement learning model in step 4 includes:

[0029] The high-confidence occupancy pseudo-labels generated by the flipping action of the roadside flipping berth equipment and the consistency constraints of multiple sensors are used as the reward signal for reinforcement learning to ensure that the reward signal is real and reliable and to avoid false triggers that lead to model mislearning. The state space is defined as the current reading of each sensor and the corresponding ambient temperature value, so that the model can fully learn the nonlinear mapping relationship between temperature and sensor output.

[0030] A temperature drift compensation function approximator based on a deep Q-network is constructed. Taking the state space as input, it outputs the optimal compensation bias of each sensor at the current temperature. By interacting with the environment, the network parameters are continuously updated to achieve adaptive optimization and fast convergence of the temperature drift compensation strategy.

[0031] The learned temperature drift compensation function is deployed at the edge, and the original data of sensors such as geomagnetism that are susceptible to temperature are compensated and corrected in real time according to the current temperature. This eliminates the coupling interference between time-domain temperature drift and real vehicle occupancy signals, effectively eliminating the coupling interference between time-domain temperature drift and real vehicle occupancy signals.

[0032] Preferably, the process of constructing the second layer of the two-layer reinforcement learning model in step 4 includes:

[0033] Based on the multi-sensor time-series data after temperature drift compensation, the signal consistency index of each sensor within a continuous time window is calculated, including the consistency rate of the decision results of each sensor and the correlation of signal fluctuations. This effectively quantifies the degree of multi-sensor collaboration and provides a reliable basis for weight adjustment.

[0034] A weighted reinforcement learning model is constructed, with the signal-to-noise ratio and consistency index of the compensated signal as the state, the fusion weights of each sensor as the action, and the final decision accuracy as the reward. The weight allocation strategy is dynamically optimized to achieve adaptive optimization of the fusion weights and improve the decision accuracy in complex environments.

[0035] The learned weight adjustment strategy is applied to multimodal fusion, so that the fusion decision boundary adaptively favors the sensor combination with high signal-to-noise ratio and strong consistency in the current environment, which improves the decision robustness in complex environments, ensures that the decision boundary always favors reliable sensors, and enhances the system's environmental adaptability.

[0036] Preferably, step 5 specifically includes:

[0037] The continuous acquisition window of multi-sensor time-series data is processed by time-frequency transformation. The one-dimensional time-series signal is converted into a two-dimensional time-frequency graph by continuous wavelet transform or short-time Fourier transform, which preserves the joint distribution characteristics of the signal in the time domain and frequency domain, and effectively enhances the visualization and separability of weak degradation features.

[0038] A convolutional neural network model is constructed, taking multi-channel time-frequency maps as input. Through convolutional layers, pooling layers and fully connected layers, weak feature degradation patterns related to device health are extracted, and the health scores of each sensor are output. This accurately quantifies the degree of sensor performance degradation and improves diagnostic consistency.

[0039] The health score output by the convolutional neural network is compared with a preset threshold to quantify the degree of sensor performance degradation, including the sensitivity decline index, the rise in noise floor, and the reduction ratio of effective detection distance. This enables early warning and graded assessment of performance degradation, supporting predictive maintenance decisions.

[0040] Preferably, step 5 further includes:

[0041] A time-series health tracking queue is constructed to record the historical health score change trajectory of each sensor. A long short-term memory network is used to predict the health degradation trend and output the expected health within a preset time window in the future, so as to realize the prediction of health status trend and provide a forward-looking decision basis for proactive maintenance.

[0042] By correlating the predicted health degradation trend with the original signal characteristics of the sensor, the dominant factors leading to performance degradation can be identified, including physical obstruction, aging of internal components, or abnormal power supply. This allows for precise location of the root cause of degradation and avoids the waste of maintenance costs caused by blindly replacing components.

[0043] Diagnostic codes are generated based on the identified dominant factors and stored in the local health log with timestamps. At the same time, summary information is reported to the cloud platform through the communication module to provide decision-making basis for predictive maintenance, build a traceable health record, and support the closed loop of cloud-based intelligent scheduling and predictive maintenance.

[0044] Preferably, step 6 specifically includes:

[0045] A joint state space is constructed that integrates health score, ambient temperature and humidity, water accumulation status indicators and previous calibration results. The current operating status of the equipment is mapped into a discrete joint state vector, realizing a full-dimensional digital representation of the equipment's operating conditions and providing a complete decision-making basis for refined compensation.

[0046] The temporal difference learning algorithm is adopted, with the joint state vector as input and the predictive compensation bias as output. The reward function is defined as the improvement of the decision accuracy after compensation. The optimal compensation strategy is obtained by offline training of the temporal difference learning model, and the global optimal compensation strategy is obtained by offline learning from historical data to ensure the effectiveness of the compensation decision.

[0047] The trained temporal difference learning model is deployed at the edge, and predictive compensation bias is applied dynamically in real time according to the current joint state. This corrects the signal attenuation caused by sensor performance degradation in advance, maintains the long-term stability of occupancy decisions, enables real-time querying and online adaptive fine-tuning at the edge, and maintains the optimal compensation effect in the long term.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. This method for multi-sensor temperature drift and bias correction in roadside parking spaces uses an improved fruit fly algorithm to identify global spatial bias drift, sliding window principal component analysis to separate common-mode interference, two-layer reinforcement learning to compensate for individual temperature drift, convolutional neural network to identify degradation patterns, and time-series difference learning to apply predictive compensation. This forms a complete closed loop of spatial environment disturbance elimination, time-domain temperature drift calibration, and equipment aging compensation, enabling the equipment to maintain a high-precision baseline model and stable decision-making capability even in complex outdoor environments.

[0050] 2. This multi-sensor temperature drift and bias correction method for roadside parking spaces uses a two-layer reinforcement learning model. The first layer uses high-confidence pseudo-labels as rewards to learn the individual temperature drift compensation function. The second layer dynamically adjusts the multi-modal fusion weights based on the signal-to-noise ratio and consistency of the compensated signal, making the decision boundary adaptively bias towards more reliable sensor combinations in the current environment. This effectively copes with sudden environmental changes such as post-rain water accumulation and high-temperature exposure, significantly improving the robustness and environmental adaptability of the fusion decision.

[0051] 3. This method for multi-sensor temperature drift and bias correction in roadside parking spaces utilizes a convolutional neural network to encode multi-sensor time-series data into a time-frequency graph, extracts weak feature degradation patterns related to health status, outputs health scores for each sensor, combines an LSTM network to predict future health status trends, and identifies dominant factors such as physical obstruction, device aging, and power supply anomalies through multi-dimensional correlation analysis, generating standardized diagnostic codes and automatically generating predictive maintenance work orders in the cloud, thus realizing the transformation from passive maintenance to predictive maintenance. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the workflow of a multi-sensor temperature drift and offset correction method for roadside parking spaces according to the present invention.

[0053] Figure 2 This is a schematic diagram illustrating the workflow of a multi-sensor temperature drift and offset correction method for roadside parking spaces according to the present invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0055] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a multi-sensor temperature drift and offset correction method for roadside parking spaces, comprising the following steps:

[0056] Step 1: Initialize multi-sensor data acquisition and spatial baseline construction at the target roadside berth. All sensors simultaneously acquire multi-sensor time-series data, including geomagnetic, ultrasonic, and millimeter-wave sensors, to establish a joint spatial distribution model in the idle state. This provides a high-precision baseline for subsequent drift detection and compensation. While the berth is idle, control the multi-modal sensors (including geomagnetic, ultrasonic, and millimeter-wave sensors) to trigger the acquisition mode with a preset synchronization clock, acquiring at least N consecutive sampling periods of multi-sensor time-series data. Timestamp and resampling alignment are performed on the data from each channel to ensure the accuracy of the multi-sensor data. Comparable analysis is performed on a unified time reference. Statistical features are extracted from the aligned multi-sensor time series data, including the mean, variance, quantile, noise floor energy, and short-term fluctuation index of each sensor signal. A multi-dimensional feature vector is constructed in the idle state to fully characterize the intrinsic distribution characteristics of each sensor in the idle state. The multi-dimensional feature vector is input into a multivariate Gaussian distribution fitter to calculate the joint probability density function of each sensor, forming a spatial joint distribution model containing the mean vector and covariance matrix. This model serves as the baseline for subsequent drift detection and establishes a statistical benchmark model for quantifying the correlation of multiple sensors.

[0057] It should be noted that after the roadside flap-type berth equipment completes its power-on self-test and confirms that the flap is in the initial horizontal position, it enters the idle state acquisition mode. The main control unit broadcasts a synchronous acquisition trigger signal to each sensor module via the CAN bus. The acquisition frequency is set to 20Hz for geomagnetic sampling rate, 10Hz for millimeter-wave radar, and 5Hz for ultrasonic sampling rate. At least 120 sampling cycles are continuously acquired to form an observation window of more than 2 seconds. Each sensor data frame is accompanied by a microsecond-level timestamp. After being corrected by the edge clock synchronization protocol, the heterogeneous sampling rate data is uniformly resampled to the 20Hz time reference axis using the cubic spline interpolation method. During the acquisition process, the ambient temperature sensor readings are recorded synchronously, and the motor static current is monitored to ensure that it remains at 0. Within the 12A to 0.18A idle baseline range, if abnormal current fluctuations are detected, the system is considered non-idle and the current acquisition is terminated. The synchronous acquisition mechanism ensures that subsequent statistical analysis is based on spatiotemporally consistent sensor data, avoiding spurious biases introduced by asynchronous sampling clocks. Sliding window statistical analysis is performed on the time-series data of each channel after resampling and alignment. The window length is set to 40 sampling points corresponding to a 2-second time window, and the sliding step size is 10 sampling points. Within each window, the mean and variance of the geomagnetic triaxial composite field strength, the second moment and quarter quantile of the ultrasonic echo energy integral, the short-time fluctuation index of the millimeter-wave radar range image energy, and the root mean square of the energy difference between adjacent frames are calculated. The noise floor energy is measured within the acquisition window. The signal minimum value under the triggered state is weighted and calculated using the 10th percentile. The geomagnetic noise floor threshold is set to 0.8 mG, the ultrasonic noise floor threshold to 3 mV, and the millimeter-wave noise floor energy threshold to 2 dB. The short-time fluctuation index is used to quantify the transient stability of the signal; when the geomagnetic fluctuation exceeds 1.2 mG for five consecutive frames, it is marked as a high-fluctuation segment. Finally, the above statistical features are combined into a 19-dimensional feature vector, including 6 dimensions of geomagnetic, 5 dimensions of ultrasonic, 5 dimensions of millimeter-wave, and 3 dimensions of ambient temperature, fully describing the multi-sensor intrinsic distribution characteristics under idle state. The 50 continuously acquired sliding window feature vectors are input into a multivariate Gaussian distribution fitter, and the expectation-maximization algorithm is used to estimate the 19-dimensional mean vector and covariance matrix, constructing the idle... The spatial joint distribution model under the condition is used. The off-diagonal elements of the covariance matrix quantify the correlation strength between geomagnetism and millimeter waves and ultrasound. After the model is built, the typical values ​​of each dimension in the mean vector are 48.2μT to 52.6μT for the total geomagnetic field, 4.2mV to 5.8mV for the ultrasound noise floor, and -72dBm to -68dBm for the millimeter wave energy reference. The model parameters, along with the acquisition timestamp, ambient temperature label and device ID, are written to the external Flash memory and backed up to the cloud to form a baseline library with version number. The baseline update adopts the exponential weighted moving average. The fusion coefficient between the newly acquired data and the historical baseline is set to 0.15 to ensure that the model can track slow drift and avoid sudden changes in the model due to a single interference.

[0058] Furthermore, step 1 also includes: during the construction of the joint spatial distribution model, synchronously collecting ambient temperature data and recording the correlation coefficients between each sensor signal and temperature, establishing a temperature-baseline mapping table to distinguish between temperature drift and spatial baseline shifts caused by environmental abrupt changes, effectively decoupling temperature drift from environmental interference, improving the accuracy of baseline drift cause identification, periodically scanning the idle window of multi-sensor time series data, and triggering a baseline update mechanism when the statistical characteristics of newly collected data deviate from the baseline by more than a preset threshold, using exponentially weighted moving average to recursively update the mean vector and covariance matrix to achieve baseline adaptive tracking, avoiding long-term drift accumulation caused by model solidification, storing the updated spatial baseline results in non-volatile memory, and recording the update timestamp and environmental label to form a spatial baseline library containing historical versions, supporting multi-version comparison and backtracking during subsequent drift analysis, constructing a traceable baseline evolution trajectory, and providing data support for degradation analysis and parameter rollback;

[0059] It should be noted that during the construction of the joint spatial distribution model, ambient temperature data was simultaneously collected, and the correlation coefficients between each sensor signal and temperature were recorded. Specifically, the edge control unit collected ambient temperature at a frequency of 1Hz using a digital temperature sensor integrated on the circuit board. For the geomagnetic sensor, idle baseline data at different temperature points were collected over a continuous 30-day natural temperature variation cycle. The Pearson correlation coefficient between the geomagnetic triaxial composite field strength and temperature was calculated, with typical correlation coefficient values ​​ranging from 0.23 to 0.37, indicating a mild positive correlation temperature drift characteristic of the geomagnetic field. The correlation coefficient between the ultrasonic echo energy integral value and temperature was -0.12 to -0.18, showing a weak negative correlation. The correlation coefficient between the millimeter-wave radar range image energy reference and temperature was less than 0.05, basically unaffected by temperature changes. Based on the above correlation analysis results, a temperature-baseline mapping table was established, with 5℃ intervals for each temperature range, storing the data for each temperature range. The typical baseline mean and variance of each sensor within the interval are used to distinguish between spatial baseline shifts caused by temperature drift and sudden environmental changes. When the temperature change exceeds 2°C and the deviation direction of multiple sensors is consistent with the predicted direction of the temperature-baseline mapping table, it is determined to be a temperature drift-dominated shift; otherwise, it is determined to be an environmental change-dominated shift. Further, the multi-sensor time-series data are periodically scanned in idle windows, with a scan cycle set to 6 hours. Each scan collects 120 sampling cycles to form a 2-second observation window, extracting a 19-dimensional feature vector and calculating the Mahalanobis distance with the baseline. The Mahalanobis distance threshold is set to 3.2, corresponding to the chi-square distribution critical value of the 99.5% confidence interval. When the Mahalanobis distance of three consecutive scans exceeds the threshold and the deviation direction is consistent, a baseline update mechanism is triggered. The update uses an exponentially weighted moving average algorithm. The fusion coefficient between the mean vector of the newly acquired data and the mean vector of the historical baseline is set to 0.15, and the fusion coefficient of the covariance matrix is ​​set to 0.1. Ensure the model can track slow drift while avoiding sudden model changes caused by single disturbances. During the update process, simultaneously calculate the matching degree between the newly acquired data and the corresponding temperature range in the temperature-baseline mapping table. If the matching degree is lower than 85%, it is marked as an environmental mutation event, and an environmental mutation flag is added to the updated model parameters. After the update is completed, store the new spatial baseline results in an external Flash memory with a capacity of 256MB. Use a cyclic overwrite method to retain the most recent 1000 baseline versions. Each storage record includes a 19-dimensional mean vector, a 19×19-dimensional covariance matrix in compressed storage format, an update timestamp, an environmental temperature label (current temperature value and temperature range), and an update trigger. The system sends a cause code (triggered by temperature drift, sudden environmental change, or periodic forced update) and a version number. The version number uses an incrementing integer format, with the high-order byte representing the year and week, and the low-order byte representing the number of updates in that week. Simultaneously, a baseline version summary is reported to the cloud platform via a 4G communication module. The cloud establishes a historical baseline database for each device, supporting retrieval and comparison by device ID, time range, version number, and other dimensions. When subsequent drift detection detects that the current data deviates from the latest baseline for more than 7 days, it automatically retrieves historical versions matching the temperature range from the historical baseline database for comparison. If a better match is found with a historical version, it is determined that the sensor performance has degraded rather than due to environmental changes, providing a basis for subsequent predictive maintenance.

[0060] Step 2 introduces an improved fruit fly algorithm, treating multiple sensors as a population. It rapidly identifies and removes global spatial bias drift caused by interference events such as post-rain puddles through spatial odor concentration search, achieving rapid identification of global environmental interference and real-time removal of abnormal data. Geomagnetic, ultrasonic, and millimeter-wave sensors are treated as individuals within the fruit fly population, with each sensor's real-time reading defined as its spatial coordinates. A population distribution matrix in a multi-dimensional sensing space is constructed, enabling multi-sensor collaborative sensing and laying the spatial mapping foundation for global bias drift identification. A spatial odor concentration determination function is defined, taking the current sensor readings as the basis for the determination. The deviation from the corresponding dimension in the idle baseline model is mapped to the odor concentration value. The location of the individual with the highest concentration is found through group iterative search. The direction and magnitude of global bias drift are identified, and the bias drift direction and magnitude are accurately located to improve the sensitivity and accuracy of interference identification. When more than half of the individuals in the group are detected to have a large deviation in the same direction at the same time and the duration exceeds the set threshold, it is determined to be a global spatial bias drift caused by interference events such as water accumulation after rain. The current frame data is marked as abnormal and removed. At the same time, the spatial baseline is quickly recalibrated to remove abnormal data in time, block the propagation of interference, and quickly restore the effectiveness of the baseline model.

[0061] It should be noted that, in order to identify and eliminate global interference caused by sudden changes in the airspace background field covering the entire berth area due to post-rain water accumulation or snowmelt, the multiple sensors are treated as individuals within a fruit fly colony. The real-time reading of each sensor is defined as the spatial coordinates of that individual. Specifically, the current value of the geomagnetic triaxial composite field strength (52.8 μT), the integral value of the ultrasonic echo energy (6.3 mV), and the millimeter-wave range image energy (-69 dBm) constitute a coordinate point in three-dimensional space. This point is compared with the mean vector [50.2 μT, 5.1 mV, -71 dBm] of the corresponding dimension in the idle baseline model to calculate the degree of deviation in each dimension. The airspace odor concentration determination function uses Mahalanobis distance to measure the deviation of the current reading from the baseline. The Mahalanobis distance calculation incorporates off-diagonal elements of the covariance matrix, fully considering the correlation characteristics between geomagnetism and millimeter-wave and ultrasonic waves. When, in five consecutive frames of data, the geomagnetic deviation exceeds 1.5 μT, the ultrasonic deviation exceeds 0.8 mV, and the millimeter-wave deviation exceeds 2.5 dB, and the deviation directions are consistent, the odor concentration value rapidly increases. The population iterative search algorithm identifies the location of the individual with the highest concentration, thereby accurately locating the direction and amplitude of the global bias drift. Based on the calculation of odor concentration and population location search, further criteria for determining global bias drift are set. When more than half of the sensors in the population, i.e., at least two sensors, simultaneously exhibit significant deviations in the same direction, and the duration exceeds a set threshold of 400 milliseconds (corresponding to eight samples), the odor concentration value rapidly increases. At frame rate, global spatial offset drift is identified as caused by interference events such as post-rain water accumulation or snowmelt. During the identification process, the consistency of the offset direction is simultaneously verified. Positive geomagnetic offsets exceeding 1.8 μT, ultrasonic positive offsets exceeding 1.2 mV, and millimeter-wave positive offsets exceeding 3.0 dB, all showing a positive increasing direction, are consistent with signal enhancement characteristics caused by water accumulation. Once identified as global offset drift, the 20 consecutive frames within the current observation window are marked as anomalous frames and directly removed from subsequent processing to prevent anomalous data from contaminating the idle baseline model and occupancy decision logic. After removal, a rapid spatial baseline recalibration process is immediately triggered. Recalibration uses the three most recent idle window data without anomaly markings, exponentially... The baseline model is updated using a weighted moving average method, with the fusion coefficient temporarily increased to 0.25 to accelerate convergence. After global bias drift removal, the device enters a rapid recalibration state to ensure the baseline model quickly returns to normal levels. During recalibration, the main control unit temporarily increases the acquisition frequency: geomagnetic sampling rate from 20Hz to 50Hz, millimeter wave from 10Hz to 20Hz, and ultrasonic from 5Hz to 10Hz. Sixty consecutive sampling cycles are collected to form a 1.2-second observation window for rapid reconstruction of idle baselines. The recalibration employs a layered update strategy: first, the mean vector is updated with the fusion coefficient for each dimension set to 0.25 to ensure new data quickly impacts the model; second, the covariance matrix is ​​updated with the fusion coefficient set to 0.15. To strike a balance between rapid response and maintaining the stability of the correlation structure, after recalibration, calculate the matching degree between the new model and the historical baseline of the current temperature range in the temperature-baseline mapping table. If the matching degree is higher than 90%, the recalibration is considered successful, the new model version number is incremented and stored in the Flash memory; if the matching degree is lower than 90%, trigger the secondary verification process, extend the observation window to 120 sampling periods and re-acquire data until the model is stable.

[0062] Step 3: Utilize sliding window principal component analysis to decompose the multi-sensor time-series data stream online, dynamically separate and compensate for common-mode bias components caused by sudden environmental changes, achieve second-level self-calibration of the spatial baseline, eliminate common-mode interference in seconds, and restore baseline purity. Set a fixed-length sliding data window, center the multi-sensor time-series data stream within the window, calculate the covariance matrix and perform eigenvalue decomposition, extract principal component vectors and corresponding eigenvalues, eliminate DC components in the data, and focus subsequent analysis on the fluctuation components reflecting interference. Identify the principal component with the largest eigenvalue as the common-mode bias component. The common-mode bias component corresponds to the same-direction drift component occurring simultaneously in each sensor. Multiply its projection coefficient by the principal component vector to obtain the common-mode bias estimate. Then, subtract the common-mode bias estimate from the original multi-sensor time-series data to achieve second-level self-calibration of the spatial baseline, real-time removal of global same-direction interference, and restoration of baseline purity to ensure decision reliability.

[0063] It should be noted that, in order to eliminate global common-mode bias interference caused by sudden environmental changes such as rainwater accumulation and snow melting, the edge-end main control unit sets a sliding data window with a fixed length of 40 sampling points (corresponding to a 2-second time window). The multi-sensor time-series data stream within the window is centralized. Centralization means subtracting the mean value of the corresponding sensor within the window from the current reading of each sensor to eliminate the DC component of the data, so that subsequent analysis focuses on the fluctuation component of the signal. Taking the geomagnetic triaxial composite field strength as an example, 40 data points are collected within the window at a sampling rate of 20Hz. After calculating their mean value, the mean value is subtracted from each sampling point to obtain the centralized geomagnetic sequence. Similarly, the integrated values ​​of millimeter-wave range image energy and ultrasonic echo energy are also centralized in the same way. After centralization, a 3×40-dimensional observation matrix is ​​constructed, where rows correspond to the three sensor channels (geomagnetic, millimeter-wave, and ultrasonic), and columns correspond to 40 sampling time points. After centralization, the main control unit calculates a 3×3-dimensional covariance matrix based on the observation matrix. The diagonal elements of this matrix reflect the fluctuation variance of each sensor, and the off-diagonal elements quantify the correlation strength between geomagnetic, millimeter-wave, and ultrasonic waves. Eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues ​​and their corresponding eigenvectors. After sorting the eigenvalues ​​from largest to smallest, the largest eigenvector is... The eigenvector corresponding to the value is the principal component direction, representing the dimension with the largest data fluctuation amplitude. In actual roadside environments, when water accumulation leads to enhanced geomagnetic field, increased ultrasonic echo energy, and elevated millimeter wave energy, the three sensors exhibit oscillations in the same direction. This common-mode component is precisely captured by the principal component corresponding to the largest eigenvalue. Taking measured data as an example, the largest eigenvalue is typically 0.42 to 0.58, much larger than the other two eigenvalues, indicating that the common-mode bias dominates the signal fluctuation in the current window. After identifying the principal component direction, the centered observation matrix is ​​projected onto this principal component vector to obtain the projection coefficient sequence. Then, the projection coefficients are multiplied by the principal component vector. The common-mode bias estimate is reconstructed. Taking a typical post-rain flooding event as an example, the principal component projection coefficients of geomagnetic, millimeter wave, and ultrasonic waves are 0.31, 0.28, and 0.25, respectively. After multiplying with the principal component vector [0.63, 0.58, 0.52], the common-mode bias at each sampling time is obtained. By subtracting the common-mode bias estimate point by point from the original centered data, the residual signal after deducting global interference can be obtained. In the residual signal, the geomagnetic fluctuation amplitude is reduced from the original 1.8μT to less than 0.3μT, the millimeter wave energy fluctuation is reduced from 3.2dB to less than 0.6dB, and the ultrasonic fluctuation is reduced from 1.5mV to less than 0.4mV.

[0064] Step 4: Construct a two-layer reinforcement learning model. The first layer uses high-confidence occupancy pseudo-labels as reward signals to learn the individual temperature drift compensation function of each sensor at different temperatures to eliminate time-domain temperature drift interference. The second layer dynamically adjusts the multi-modal fusion weights based on the consistency and signal-to-noise ratio of the compensated signals, so that the decision boundary always tends to the reliable sensor combination in the current environment. Adaptive temperature drift correction and weight optimization improve the robustness of decision in complex environments.

[0065] Step 5: Use the convolutional neural network in the image classification algorithm to encode the multi-sensor time series data into a time-frequency graph, identify the weak feature degradation patterns related to the health of the device, quantify the degree of sensor performance degradation, identify performance degradation at an early stage, and realize quantitative assessment of health.

[0066] Step 6: Establish a joint state machine that integrates health status, environmental status, and previous calibration results. Through time-series differential learning, apply predictive compensation biases to devices in different joint states to correct signal attenuation caused by performance degradation in advance, dynamically predict compensation, and extend the maintenance-free cycle of the devices.

[0067] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: the process of constructing the first layer of the two-layer reinforcement learning model in step 4 includes: using the flip-up action of the roadside flip-up parking equipment as the trigger and the high-confidence occupancy pseudo-label generated by the multi-sensor consistency constraint as the reward signal for reinforcement learning, ensuring that the reward signal is real and reliable, avoiding false triggers that lead to model mislearning, defining the state space as the current reading of each sensor and the corresponding ambient temperature value, so that the model can fully learn the nonlinear mapping relationship between temperature and sensor output, constructing a temperature drift compensation function approximator based on a deep Q network, taking the state space as input, outputting the optimal compensation bias of each sensor at the current temperature, continuously updating the network parameters through interaction with the environment, realizing the adaptive optimization and fast convergence of the temperature drift compensation strategy, deploying the learned temperature drift compensation function at the edge, and compensating and correcting the original data of sensors such as geomagnetism that are easily affected by temperature in real time according to the current temperature, eliminating the coupling interference between time-domain temperature drift and real vehicle occupancy signal, effectively eliminating the coupling interference between time-domain temperature drift and real vehicle occupancy signal;

[0068] It should be noted that in actual deployment, the roadside flap-type parking device generates high-confidence occupancy pseudo-labels through flap-action triggering and multi-sensor consistency constraints, serving as a reward signal for reinforcement learning. Specifically, when a vehicle enters the parking space, triggering the flap to descend, the main control unit detects that the motor current jumps from the idle state of 0.12A to a peak starting value of 0.45A and then falls back to a steady state of 0.20A. Combined with the IMU angle changing from 0° to -15° to confirm the flap-action is complete, the pseudo-label generation process is triggered. Simultaneously, through multi-sensor acquisition, the geomagnetic triaxial composite field strength increases from the idle baseline of 50.2μT to 54.8μT, the millimeter-wave range image energy increases from -71dBm to -65dBm, and the ultrasonic echo energy integral... The value increases from 5.1mV to 8.3mV, and all three values ​​remain stable for more than 1.5 seconds with consistent decision results, thus generating a high-confidence occupancy pseudo-label. This pseudo-label serves as a positive reward signal input to the reinforcement learning model for subsequent training and optimization of the temperature drift compensation function, ensuring the authenticity and reliability of the reward signal and avoiding false triggers that could lead to model mislearning. In the reinforcement learning framework, the state space is defined as the current real-time readings of each sensor and the corresponding ambient temperature values. Specifically, this includes the geomagnetic triaxial composite field strength, the integral value of ultrasonic echo energy, the energy of the millimeter-wave distance image, and the current ambient temperature collected by the temperature sensor. Taking the geomagnetic sensor as an example, its reading range is 48.2μT to 52.6μT, and its temperature range is -20℃ to 60℃, forming a continuous state. In the state space, a temperature drift compensation function approximator is constructed based on a deep Q-network. The network input layer is a 4-dimensional state vector, and the hidden layer uses a two-layer fully connected network with 128 neurons per layer and ReLU activation function. The output layer is the optimal compensation bias for each sensor at the current temperature. The geomagnetic compensation bias output range is -1.5μT to +2.0μT. The network collects empirical trajectories through interaction with the environment and uses an empirical playback mechanism to store the past 1000 interaction data. 64 empirical data are sampled in each batch to update the network parameters. The learning rate is set to 0.001, and the discount factor is set to 0.95 to ensure that the temperature drift compensation function converges quickly to the optimal strategy. The trained deep Q-network model is quantized and compressed before being deployed to the edge master unit. The model file... The size is controlled within 256KB to meet MCU resource constraints. After deployment, the main control unit reads ambient temperature sensor data at a frequency of 1Hz, queries the current temperature value in real time and inputs it into the temperature drift compensation network, and outputs the compensation bias of each sensor. Taking the geomagnetic sensor as an example, when the ambient temperature rises from 25℃ to 35℃, the compensation bias output by the network is +0.8μT. The main control unit subtracts this compensation from the original geomagnetic reading to obtain the correction value. After correction, the geomagnetic baseline is stabilized in the range of 50.1μT to 50.5μT. The fluctuation amplitude introduced by temperature drift is reduced from the original 1.2μT to less than 0.3μT. The corrected multi-sensor data enters the subsequent occupancy decision module, effectively eliminating the coupling interference between time-domain temperature drift and real vehicle occupancy signal.

[0069] Furthermore, the process of constructing the second layer of the two-layer reinforcement learning model in step 4 includes: based on the multi-sensor time-series data after temperature drift compensation, calculating the signal consistency index of each sensor within a continuous time window, including the consistency rate of the decision results of each sensor and the correlation of signal fluctuations, effectively quantifying the degree of multi-sensor collaboration, providing a reliable basis for weight adjustment, constructing a weight adjustment reinforcement learning model, taking the signal-to-noise ratio and consistency index of the compensated signal as the state, taking the fusion weight of each sensor as the action, and taking the final decision accuracy as the reward, dynamically optimizing the weight allocation strategy, realizing adaptive optimization of fusion weight, improving the decision accuracy in complex environments, applying the learned weight adjustment strategy to multimodal fusion, making the fusion decision boundary adaptively biased towards the sensor combination with high signal-to-noise ratio and strong consistency in the current environment, improving the decision robustness in complex environments, ensuring that the decision boundary always biases towards reliable sensors, and enhancing the system's environmental adaptability;

[0070] It should be noted that, based on the multi-sensor time-series data after temperature drift compensation, the edge control unit continuously calculates the signal consistency index of each sensor with a sliding window length of 2 seconds. Specifically, the consistency rate of the decision results of the geomagnetic, millimeter-wave, and ultrasonic channels is obtained by statistically analyzing the percentage of frames in which all three are simultaneously determined to be occupied or idle within the window. The typical consistency rate threshold is set at 85%. When the consistency rate exceeds this threshold, it indicates good multi-sensor coordination. The correlation of signal fluctuations is quantified using the Pearson correlation coefficient. The above consistency index and the signal-to-noise ratio of each channel together constitute the state space of reinforcement learning. In the signal-to-noise ratio calculation, the effective value of the geomagnetic signal is taken as the average value within the window. The noise floor is the noise energy of the idle baseline, with typical geomagnetic signal-to-noise ratios (SNRs) of 18 dB to 25 dB, ultrasonic signals of 12 dB to 18 dB, and millimeter-wave signals of 22 dB to 28 dB. The weighted reinforcement learning model uses the SNR and consistency index of the compensated signal as state inputs. The state vector has a 9-dimensional dimension, including the SNR of each of the three channels, pairwise correlation coefficients, and the consistency rate of the three channels. The action space is defined as the fusion weight coefficients of each sensor. The geomagnetic weight adjustment range is 0.2 to 0.5, ultrasonic weights are 0.1 to 0.4, and millimeter-wave weights are 0.3 to 0.6, with the sum of the three always equal to 1.0. The reward function is designed as an in-window occupancy judgment. The accuracy improvement was measured by comparing high-confidence pseudo-labels triggered by the flipping action. The model employs a deep deterministic policy gradient algorithm. Both the actor and critic networks are three-layer fully connected structures with 128 and 64 hidden layer neurons, respectively. The learning rate is set to 0.0001, the experience replay pool has a capacity of 2000 records, and 128 records are sampled in each batch to update the network parameters, ensuring stable convergence of the weight adjustment strategy in complex environments. After the trained weight adjustment strategy is deployed at the edge, the master control unit reads the current state every 1 second and outputs the optimal weight combination. Taking a post-rain water accumulation scene as an example, the millimeter-wave signal-to-noise ratio is maintained at 24d. B. The geomagnetic signal-to-noise ratio decreased to 14dB, and the ultrasonic signal-to-noise ratio was 11dB. The fusion weights of the model output were adjusted to 0.55 for millimeter wave, 0.25 for geomagnetic, and 0.20 for ultrasonic. This made the decision boundary adaptively biased towards the millimeter wave sensor with high signal-to-noise ratio and strong consistency in the current environment. Deployment and testing showed that the fusion decision accuracy after weight adjustment increased from 89.5% to 96.2%, and the false alarm rate decreased from 6.8% to 2.1%. This significantly enhanced the decision robustness of roadside parking equipment in complex environments. At the same time, the weight update process adopted an amplitude limiting strategy, with the single adjustment amplitude not exceeding 0.05, to avoid decision oscillation caused by drastic weight fluctuations.

[0071] Step 5 specifically includes: performing time-frequency transformation processing on the continuous acquisition window of multi-sensor time-series data, using continuous wavelet transform or short-time Fourier transform to convert the one-dimensional time-series signal into a two-dimensional time-frequency graph, preserving the joint distribution characteristics of the signal in the time and frequency domains, effectively enhancing the visualization and separability of weak degradation features, constructing a convolutional neural network model, using the multi-channel time-frequency graph as input, extracting weak feature degradation patterns related to equipment health through convolutional layers, pooling layers, and fully connected layers, outputting the health score of each sensor, accurately quantifying the degree of sensor performance degradation, improving diagnostic consistency, comparing the health score output by the convolutional neural network with a preset threshold, quantifying the degree of sensor performance degradation, including the sensitivity reduction index, the noise floor rise, and the effective detection distance attenuation ratio, realizing early warning and graded assessment of performance degradation, supporting predictive maintenance decisions;

[0072] It should be noted that in actual deployment applications, the edge control unit uses a 6-hour health monitoring cycle and performs time-frequency transformation processing on the continuous acquisition window of multi-sensor time-series data. Specifically, for the three channels of geomagnetic triaxial composite field strength, ultrasonic echo energy integral value, and millimeter-wave range image energy, 120 consecutive sampling points corresponding to a 6-second observation window are extracted for each channel. Continuous wavelet transform is used to convert the one-dimensional time-series signal into a two-dimensional time-frequency map. The wavelet basis function is Morlet wavelet, and the scale range is set to 1 to 64 to ensure coverage of the 0.5Hz to 20Hz frequency band. After transformation, each channel generates... A 64×120 dimensional time-frequency plot is generated, and the three channels are merged into a 3×64×120 dimensional multi-channel time-frequency feature plot, fully preserving the joint distribution characteristics of the signal in the time and frequency domains. During the construction of the time-frequency plot, the frequency resolution is set to 0.5Hz and the time resolution to 50ms, meeting the accuracy requirements for geomagnetic power frequency interference identification, millimeter-wave micro-motion feature extraction, and ultrasonic echo attenuation analysis. The transformed time-frequency plot data is normalized and stored in a circular buffer with a capacity of 50 groups for subsequent analysis by the convolutional neural network model. A lightweight convolutional neural network model is deployed at the edge, and the model output... The input layer receives a 3×64×120 dimensional multi-channel time-frequency map, which is then processed through three convolutional layers for feature extraction. The first convolutional layer has a kernel size of 3×3, 16 kernels, a stride of 1, and uses the same padding method with ReLU activation. The second convolutional layer has a kernel size of 3×3, 32 kernels, and a stride of 1. The third convolutional layer has a kernel size of 3×3, 64 kernels, and a stride of 1. Each convolutional layer is followed by a 2×2 max-pooling layer with a stride of 2, progressively reducing the feature map dimension. The feature maps extracted by the convolutional layers are flattened and then input into a two-layer fully connected network. The first fully connected layer... The first layer has 128 neurons, the second layer has 64 fully connected neurons, and the final output layer is a 3D vector, corresponding to the health scores of three sensors: geomagnetism, ultrasound, and millimeter wave. The score range is set from 0 to 100. During the model training phase, 12,000 sets of labeled time-frequency graph data were collected historically. The labels were marked by manual inspection records and fault logs. The training set and validation set were divided in an 8:2 ratio. The loss function was the mean squared error. The optimizer was Adam. The initial learning rate was set to 0.001. The training rounds were 50. The early stopping mechanism was set to terminate training when the verification loss did not decrease for 5 consecutive rounds.

[0073] The expression for the health score of a geomagnetic sensor is as follows:

[0074] ;

[0075] ;

[0076] In the formula: The health score of the geomagnetic sensor is given, with a lower score indicating more severe performance degradation. This represents the decrease in the amplitude of the geomagnetic response. This represents the amplitude of the geomagnetic response under baseline conditions. The current geomagnetic response amplitude is obtained from the historical average of the geomagnetic triaxial composite field strength under idle conditions. The threshold for judging the decrease in geomagnetic sensitivity is set at 1.2 μT. Only when the decrease exceeds this threshold will it be included in the health score deduction. This is the geomagnetic health attenuation coefficient, used to adjust the degree of influence of the decline on the health status, with a value of 0.85;

[0077] The expression for the health score of the ultrasonic sensor is as follows:

[0078] ;

[0079] ;

[0080] In the formula: Assess the health of ultrasonic sensors; This represents the rise in noise floor height. The 90th percentile of the integral value of the ultrasonic echo energy in the current idle state; Baseline noise energy; The threshold for judging ultrasonic noise rise is set to 2.5mV; This is the maximum allowable rise amplitude for ultrasonic noise. If this value is exceeded, the health status will drop to 0. The value is 10mV. The ultrasound health attenuation coefficient is set to 0.9.

[0081] The expression for the health score of a millimeter-wave sensor is as follows:

[0082] ;

[0083] ;

[0084] In the formula: Assess the health of millimeter-wave sensors; To effectively detect the attenuation ratio of the distance; Distance under baseline conditions The energy value at that location; This represents the energy value at the same distance in the current state. The threshold for determining the attenuation of millimeter-wave detection range is set to 3.5 dB; This is the maximum allowable attenuation value for millimeter waves. If this value is exceeded, the health status will drop to 0. The value is 15dB. The millimeter-wave health attenuation coefficient is set to 0.88.

[0085] The health score output by the convolutional neural network is compared with a preset threshold to quantitatively assess the degree of sensor performance degradation. The health threshold for the geomagnetic sensor is set at 75 points; a score below this threshold triggers a sensitivity degradation diagnosis. The sensitivity degradation index is calculated by comparing the geomagnetic response amplitude under current and baseline conditions. A typical degradation amplitude exceeding 1.2 μT is considered significant degradation. The health threshold for the ultrasonic sensor is set at 70 points. The noise floor rise amplitude is obtained by statistically analyzing the difference between the 90th percentile of the echo energy integral value under idle conditions and the baseline noise floor. A rise amplitude exceeding 2.5 mV is considered noise performance degradation. The health threshold for millimeter-wave sensors is set at 80 points. The effective detection distance attenuation ratio is obtained by analyzing the energy attenuation curve of distance image with distance. When the energy attenuation exceeds 3.5dB, it is determined that the detection capability has decreased. The health score and attenuation quantification results are packaged to generate a health log every 24 hours. The log includes the current health of each sensor, attenuation index, diagnosis timestamp and device ID. It is reported to the cloud platform through the 4G module. The cloud establishes the health change trend curve of each device. When the health is below the threshold three times in a row, a predictive maintenance work order is automatically generated to guide the on-site maintenance personnel to replace or calibrate the degraded sensor in a targeted manner.

[0086] Furthermore, step 5 also includes: constructing a time-series health tracking queue to record the historical health score change trajectory of each sensor; using a long short-term memory network to predict the health degradation trend; outputting the expected health within a preset time window in the future; realizing the prediction of health status trends; providing a forward-looking decision-making basis for proactive maintenance; correlating the health degradation trend prediction results with the original signal characteristics of the sensors to identify the dominant factors causing performance degradation, including physical obstruction, aging of internal components, or abnormal power supply; accurately locating the root cause of degradation; avoiding the waste of maintenance costs caused by blindly replacing components; generating diagnostic codes based on the identified dominant factors; storing them in the local health log with timestamps; and simultaneously reporting the summary information to the cloud platform through the communication module to provide a decision-making basis for predictive maintenance; constructing a traceable health record; and supporting the closed loop of cloud-based intelligent scheduling and predictive maintenance.

[0087] It should be noted that in the edge control unit, time-series health tracking queues are established for the three sensors: geomagnetic, ultrasonic, and millimeter-wave. The queue length is set to 30 historical health score data points, corresponding to 30 health monitoring cycles, or 180 hours, of health change trajectory. The control unit collects the health scores of each sensor at 6-hour intervals, inserts the new score at the end of the queue, and removes the oldest data point, maintaining a first-in-first-out dynamic update of the queue. Based on this time-series queue, a long short-term memory network prediction model is constructed. The network input layer receives a 30-step historical health sequence, the hidden layer is set as a two-layer LSTM structure, with 64 and 32 neurons in each layer, respectively, and the output layer is the expected health score for the next 5 monitoring cycles, or 30 hours. For the health sequence model training phase, 8000 sets of historically collected health trajectory data were used. The training and validation sets were divided in a 7:3 ratio. The mean squared error loss function was used, and the Adam optimizer was employed. The initial learning rate was set to 0.0005, and the training epochs were 100. An early stopping mechanism was implemented, terminating training if the validation loss did not decrease for 10 consecutive epochs to ensure the generalization ability and stability of the prediction model under limited resources at the edge. The expected health sequence for the next 30 hours output by the LSTM model was correlated with the original sensor signal characteristics in multiple dimensions to identify the dominant factors leading to performance degradation. For the geomagnetic sensor, the expected health decline trend was correlated with the diurnal fluctuation amplitude of the geomagnetic triaxial composite field strength, power frequency interference energy, and temperature drift. Correlation calculations were performed on the slope. When the diurnal fluctuation exceeded 1.5 μT and the expected health score was below 75, it was determined to be attenuation dominated by external physical obstruction. When the temperature drift slope exceeded 0.08 μT / ℃ and there were no significant obstruction characteristics, it was determined to be aging of internal components. For ultrasonic sensors, correlation analysis incorporated the 90th percentile of the echo energy integral value, the noise floor energy, and the jitter amplitude of the effective echo arrival time window. When the noise floor rise exceeded 2.5 mV and the effective echo jitter was greater than 3 ms, it was determined to be physical obstruction such as mud and sand cover. When the echo energy continued to attenuate and the noise floor did not change significantly, it was determined to be device aging. For millimeter-wave sensors, correlation analysis used the slope of the distance image energy attenuation curve with distance, Doppler spectral width, and virtual The false alarm rate is determined as follows: when the slope of the attenuation curve increases by more than 0.8 dB / m and the false alarm rate rises to more than 5%, it is determined that the power supply is abnormal and the transmission power is reduced; when the Doppler spectral width narrows by more than 15 Hz and the energy attenuation is greater than 3.5 dB, it is determined that the device is aging. Based on the dominant factors identified by the correlation analysis, the main control unit generates a standardized diagnostic code. The coding format uses a combination of three letters and four numbers. The first letter indicates the sensor type M / U / R, which corresponds to geomagnetic, ultrasonic, and millimeter wave, respectively. The second letter indicates the dominant attenuation factor O / A / P, which corresponds to physical blockage, device aging, and power supply abnormality, respectively. The last four numbers represent the combination code of diagnostic confidence and attenuation quantification value. The confidence level is determined by the correlation coefficient in the correlation analysis exceeding 0.The diagnostic code, along with the current health score, decay index, diagnostic timestamp, and device ID, is weighted and calculated using a feature count of 75, ranging from 0 to 100. Every 24 hours, a health log is generated and stored in local Flash memory, with a log capacity of 100 entries per cycle. Simultaneously, a summary information is reported to the cloud platform via a 4G module. This summary includes the device ID, diagnostic code, current health score, and decay trend prediction results. The cloud establishes health trend curves and diagnostic history records for each device. When the first and second digits of the diagnostic code reported by the same sensor are the same for three consecutive times, or when the expected health score is lower than the corresponding thresholds for each sensor (75 for geomagnetic, 70 for ultrasonic, and 80 for millimeter wave) three times consecutively, a predictive maintenance work order is automatically generated. The work order includes the device location, fault type, diagnostic confidence level, and suggested maintenance time window, guiding on-site maintenance personnel to carry the corresponding spare parts for targeted replacement or calibration, thus shifting from reactive maintenance to predictive maintenance.

[0088] Step 6 specifically includes: constructing a joint state space that integrates health score, ambient temperature and humidity, water accumulation status indicators, and previous calibration results; mapping the current operating state of the equipment to a discrete joint state vector; realizing a full-dimensional digital representation of the equipment's operating conditions; providing a complete decision-making basis for refined compensation; employing a temporal differential learning algorithm, using the joint state vector as input and predictive compensation bias as output; defining the reward function as the improvement in occupancy decision accuracy after compensation; obtaining the optimal compensation strategy through offline training of the temporal differential learning model; obtaining the globally optimal compensation strategy through offline learning of historical data; ensuring the effectiveness of compensation decisions; deploying the trained temporal differential learning model at the edge; dynamically applying predictive compensation bias in real time based on the current joint state; correcting signal attenuation caused by sensor performance degradation in advance; maintaining the long-term stability of occupancy decisions; realizing real-time query and online adaptive fine-tuning at the edge; and maintaining the optimal compensation effect in the long term.

[0089] It should be noted that in the edge control unit, a joint state space integrating multi-dimensional information is first constructed to comprehensively describe the current operating condition of the device. The joint state vector consists of four dimensions: the health score dimension quantifies the current health scores of the geomagnetic, ultrasonic, and millimeter-wave sensors into three levels: health status score ≥80 points, attention status 75 to 80 points, and warning status <75 points; the ambient temperature and humidity dimension divides the temperature into 16 intervals in 5°C increments, from -20°C to 60°C, and the humidity into 10 intervals in 10%RH increments, from 0%RH to 100%RH. The water accumulation status indicator is determined by combining the surface reflection energy of millimeter-wave radar and the attenuation characteristics of ultrasonic waves in the air, and is set into three states: no water accumulation, thin water accumulation with a depth of <5mm, and thick water accumulation with a depth of ≥5mm. The previous calibration result dimension records the average compensation bias of the three most recent calibrations, and is divided into five intervals with an interval of 0.3μT. The above four dimensions are combined to form a discretized joint state space, with a total of 21,600 states in total, which are 3×3×16×10×3×5. Each state corresponds to a unique joint state vector code. The main control unit reads the current value of each dimension at a 1-minute cycle and maps it to the corresponding state code. Offline training was performed using the Q-learning method in the temporal difference learning algorithm. The joint state vector was used as input, and the predictive compensation bias was used as output. The state space consisted of the aforementioned 21,600 discrete states. The action space was defined as the combination of compensation biases for the geomagnetic, ultrasonic, and millimeter-wave channels. The geomagnetic compensation range was set to -1.5μT to +2.0μT with a step size of 0.1μT (36 possible ranges); the ultrasonic compensation range was -2.0mV to +3.0mV with a step size of 0.2mV (26 possible ranges); and the millimeter-wave compensation range was -2.5dB to +3.5dB with a step size of 0.3dB (21 possible ranges). The total number of possible combinations in the action space is 36×26×21, totaling 19656. The reward function is designed to be the improvement in decision accuracy after applying compensation for two consecutive hours. The accuracy benchmark is the historical average accuracy without compensation. For every 1% increase in the improvement, the reward value increases by 10 points, and for every decrease, the corresponding points are deducted. During the training phase, 50,000 historically collected state transition trajectory data are used. Each trajectory includes the current state, the action to be performed, the next state, and the reward value. The learning rate is set to 0.01, the discount factor is set to 0.95, and the exploration rate of the ε-greedy strategy is initially set to 0.3 and then linearly decays to 0 with each training round.05. After 200 rounds of iterative training, the optimal compensation strategy Q-table is obtained. Each state-action pair in the table corresponds to an expected cumulative reward value. The trained Q-table is compressed and encoded, then deployed to the edge control unit. The storage space is controlled within 512KB to meet the MCU resource constraints. After deployment, the control unit performs the following operations every 1 minute: read the current joint state vector encoding, query the action with the highest expected cumulative reward value in the Q-table for that state, i.e., the optimal compensation bias combination, and apply the corresponding compensation amount to the original readings of geomagnetic, ultrasonic, and millimeter-wave signals respectively. When the geomagnetic compensation amount is +0.8μT, the original reading is subtracted from this value. When the ultrasonic compensation amount is -1.2mV, the original reading is subtracted from this value. The initial reading is added to this value. When the millimeter-wave compensation is +1.5dB, the original reading is subtracted from this value. The compensated data enters the occupancy decision process. Simultaneously, the main control unit statistically analyzes the accuracy of the decision after compensation every 6 hours. When the accuracy improvement is lower than expected, the Q-table online fine-tuning mechanism is triggered, using the most recent 1000 actual operating trajectory data points with a learning rate of 0.005 for incremental updates. This effectively corrects the signal attenuation caused by sensor performance degradation after applying predictive compensation. Specifically, the geomagnetic signal attenuation compensation accuracy reaches 0.1μT, the ultrasonic attenuation compensation accuracy reaches 0.2mV, and the millimeter-wave attenuation compensation accuracy reaches 0.3dB, extending the equipment's maintenance-free operation cycle.

[0090] The following is a detailed explanation of the workflow of a multi-sensor temperature drift and offset correction method for roadside parking spaces.

[0091] After the roadside flap berth equipment completes its power-on self-test and confirms that the flap is in the initial horizontal position, it enters the idle state acquisition mode. The main control unit broadcasts synchronous acquisition trigger signals to each sensor module through the CAN bus to acquire time-series data from multi-modal sensors such as geomagnetism, ultrasound, and millimeter waves. It also performs timestamp calibration and resampling alignment on the heterogeneous sampling rate data. The aligned data is then subjected to sliding window statistical analysis to extract statistical features such as mean, variance, quantiles, noise floor energy, and short-term fluctuation index, and constructs a multi-dimensional feature vector. Subsequently, a multivariate Gaussian distribution fitting is used to establish a spatial joint distribution model in the idle state, which serves as the baseline for subsequent drift detection. Ambient temperature data is collected synchronously to establish a temperature-baseline mapping table to distinguish between baseline shifts caused by temperature drift and environmental abrupt changes. The baseline is periodically scanned and updated to form a baseline library with version numbers and reported to the cloud.

[0092] When environmental changes, such as water accumulation after rain, cause global spatial bias drift, an improved fruit fly algorithm is introduced. This algorithm treats multiple sensors as individuals within a group and quickly identifies the direction and magnitude of global bias drift by searching for spatial odor concentration. After determining it as an interference event, abnormal frames are removed and rapid recalibration is triggered. At the same time, sliding window principal component analysis is used to decompose the multi-sensor time-series data stream online, extract the common-mode bias component and subtract it from the original data, achieving second-level self-calibration of the spatial baseline and eliminating global interference of environmental changes on occupancy decisions.

[0093] In terms of temporal temperature drift correction, a two-layer reinforcement learning model is constructed. The first layer uses the high-confidence occupancy pseudo-labels generated by the flip-board action trigger and the multi-sensor consistency constraint as the reward signal to learn the individual temperature drift compensation function of each sensor at different temperatures, and performs temperature drift correction on the original data in real time to eliminate the coupling interference between temporal temperature drift and vehicle occupancy signal. The second layer calculates the signal consistency index and signal-to-noise ratio of each sensor based on the temperature drift-compensated data, and dynamically adjusts the multimodal fusion weights with a deep deterministic strategy gradient algorithm to make the decision boundary adaptively favor the reliable sensor combination in the current environment, thereby improving the robustness of decision in complex environments.

[0094] To address the performance degradation of sensors during long-term service, continuous wavelet transform is performed on multi-sensor time-series data at a fixed period to generate multi-channel time-frequency maps. These maps are then input into a lightweight convolutional neural network to extract weak feature degradation patterns. The network outputs a health score for each sensor and quantifies the degree of degradation. Furthermore, a time-series health tracking queue is constructed, and a long short-term memory network is used to predict future health trends. Correlation analysis is then performed with the original signal features to identify dominant factors such as physical obstruction, device aging, and power supply anomalies. Standardized diagnostic codes are generated and reported to the cloud, providing a basis for predictive maintenance.

[0095] Finally, a joint state space integrating health status, ambient temperature and humidity, water accumulation status, and previous calibration results is established. The optimal compensation strategy Q-table is obtained through offline training using the Q-learning algorithm and deployed at the edge. The main control unit queries the optimal compensation bias corresponding to the current state in real time, applies predictive compensation to the raw readings of each sensor, and statistically analyzes the compensation effect at fixed intervals. When necessary, an online fine-tuning mechanism is triggered to correct signal attenuation caused by performance degradation in advance, maintain the long-term stability of occupancy decision, and form a complete closed loop from spatial environmental disturbance elimination to temporal temperature drift calibration and equipment aging compensation.

[0096] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-sensor temperature drift and offset correction method for roadside parking spaces, characterized in that, Includes the following steps: Step 1: Initialize multi-sensor data acquisition and construct airspace baseline at the target roadside berth. Each sensor synchronously acquires multi-sensor time-series data to establish a joint airspace distribution model in the idle state. Step 2: Introduce an improved fruit fly algorithm, treating multiple sensors as a group, and quickly identify and remove global spatial bias drift caused by interference events in real time through spatial odor concentration search. Step 3: Use sliding window principal component analysis to decompose the multi-sensor time-series data stream online, dynamically separate and compensate for the common-mode bias component caused by sudden environmental changes; Step 4: Construct a two-layer reinforcement learning model. The first layer uses high-confidence occupancy pseudo-labels as reward signals to learn the individual temperature drift compensation function of each sensor. The second layer dynamically adjusts the multimodal fusion weights based on the consistency and signal-to-noise ratio of the compensated signals. Step 5: Use the convolutional neural network in the image classification algorithm to encode the multi-sensor time series data into a time-frequency graph, identify the weak feature degradation patterns related to the health of the device, and quantify the degree of sensor performance degradation. Step 6: Establish a joint state machine that integrates health status, environmental status, and previous calibration results. Apply predictive compensation bias to devices in different joint states through temporal differential learning to correct signal attenuation caused by performance degradation in advance.

2. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: Step 1 specifically includes: When the berth is idle, the multimodal sensor is controlled to trigger the acquisition mode with a preset synchronous clock to acquire multi-sensor time-series data for no less than N consecutive sampling cycles, and the data of each channel is timestamped and resampled and aligned. Statistical features are extracted from the aligned multi-sensor time-series data, including the mean, variance, quantile, noise floor energy, and short-time fluctuation index of each sensor signal, and a multi-dimensional feature vector is constructed in the idle state. The multidimensional feature vector is input into a multivariable Gaussian distribution fitter to calculate the joint probability density function of each sensor, forming a spatial joint distribution model that includes the mean vector and the covariance matrix.

3. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 2, characterized in that: Step 1 further includes: During the construction of the joint spatial distribution model, ambient temperature data was collected simultaneously and the correlation coefficients between each sensor signal and temperature were recorded. A temperature-baseline mapping table was established to distinguish between temperature drift and spatial baseline shift caused by sudden environmental changes. Periodic idle window scanning is performed on multi-sensor time series data. When the statistical characteristics of newly acquired data deviate from the baseline by more than a preset threshold, the baseline update mechanism is triggered, and the mean vector and covariance matrix are updated recursively using exponential weighted moving average. The updated airspace baseline results are stored in non-volatile memory, and the update timestamp and environment label are recorded to form an airspace baseline library containing historical versions.

4. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: Step 2 specifically includes: The multiple sensors are treated as individuals in a fruit fly colony, and the real-time reading of each sensor is defined as the spatial coordinates of that individual, thus constructing a population distribution matrix in a multi-dimensional sensing space. Define an airspace odor concentration determination function, which maps the deviation of each sensor reading from the corresponding dimension in the idle baseline model at the current moment to the odor concentration value, and finds the location of the individual with the highest concentration through group iterative search, and identifies the global bias drift direction and amplitude. When it is detected that more than half of the individuals in the group simultaneously exhibit large deviations in the same direction for a duration exceeding a set threshold, it is determined to be a global spatial offset drift caused by an interference event. The current frame data is marked as abnormal and removed, and a rapid recalibration of the spatial baseline is triggered.

5. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: Step 3 specifically includes: Set a fixed-length sliding data window, center the multi-sensor time-series data stream within the window, calculate the covariance matrix and perform eigenvalue decomposition, and extract the principal component vector and its corresponding eigenvalue. The principal component with the largest eigenvalue is identified as the common-mode bias component. The common-mode bias component corresponds to the same-direction drift component that occurs simultaneously in all sensors. The projection coefficient of the common-mode bias component is multiplied by the principal component vector to obtain the common-mode bias estimate. The common-mode bias estimate is then subtracted from the original multi-sensor time series data.

6. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: The process of constructing the first layer of the two-layer reinforcement learning model in step 4 includes: The high-confidence occupancy pseudo-labels generated by the flipping action of the roadside flipping berth equipment and the consistency constraints of multiple sensors are used as the reward signal for reinforcement learning. The state space is defined as the current reading of each sensor and the corresponding ambient temperature value. A temperature drift compensation function approximator based on a deep Q-network is constructed. The state space is taken as input, and the optimal compensation bias of each sensor at the current temperature is output. The network parameters are continuously updated by interacting with the environment. The learned temperature drift compensation function is deployed at the edge to compensate and correct the raw sensor data that is susceptible to temperature in real time based on the current temperature, thereby eliminating the coupling interference between time-domain temperature drift and real vehicle occupancy signals.

7. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 6, characterized in that: The process of constructing the second layer of the two-layer reinforcement learning model in step 4 includes: Based on the multi-sensor time-series data after temperature drift compensation, the signal consistency index of each sensor within a continuous time window is calculated, including the consistency rate of the decision results of each sensor and the correlation of signal fluctuations. A weighted reinforcement learning model is constructed, with the signal-to-noise ratio and consistency index of the compensated signal as the state, the fusion weights of each sensor as the action, and the final decision accuracy as the reward, to dynamically optimize the weight allocation strategy. The learned weight adjustment strategy is applied to multimodal fusion, so that the fusion decision boundary adaptively favors sensor combinations with high signal-to-noise ratio and strong consistency in the current environment.

8. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: Step 5 specifically includes: The continuous acquisition window of multi-sensor time-series data is processed by time-frequency transformation. The one-dimensional time-series signal is converted into a two-dimensional time-frequency graph by continuous wavelet transform or short-time Fourier transform, preserving the joint distribution characteristics of the signal in the time and frequency domains. A convolutional neural network model is constructed, which takes a multi-channel time-frequency map as input, and extracts the weak feature degradation patterns related to device health through convolutional layers, pooling layers and fully connected layers, and outputs the health score of each sensor. The health score output by the convolutional neural network is compared with a preset threshold to quantify the degree of sensor performance degradation, including the sensitivity reduction index, the rise in noise floor, and the reduction ratio of effective detection distance.

9. A multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 8, characterized in that: Step 5 further includes: A time-series health tracking queue is constructed to record the historical health score change trajectory of each sensor. A long short-term memory network is used to predict the health degradation trend and output the expected health within a preset time window. The predicted health degradation trend results are correlated with the original signal characteristics of the sensor to identify the dominant factors leading to performance degradation, including physical obstruction, aging of internal components, or abnormal power supply. Diagnostic codes are generated based on the identified dominant factors and stored in the local health log with timestamps. At the same time, summary information is reported to the cloud platform through the communication module.

10. The multi-sensor temperature drift and offset correction method for roadside parking spaces according to claim 1, characterized in that: Step 6 specifically includes: Construct a joint state space that integrates health score, ambient temperature and humidity, water accumulation status indicator and previous calibration results, and map the current operating state of the equipment into a discretized joint state vector; The temporal difference learning algorithm is adopted, with the joint state vector as input and the predictive compensation bias as output. The reward function is defined as the improvement of the decision accuracy after compensation. The optimal compensation strategy is obtained by training the temporal difference learning model offline. The trained temporal difference learning model is deployed at the edge, and predictive compensation biases are applied dynamically in real time according to the current joint state to correct signal attenuation caused by sensor performance degradation in advance and maintain the long-term stability of occupancy decisions.

Images