A high-precision adaptive non-stop weighing system based on deep learning

Abstract

This invention relates to the field of dynamic weighing technology, specifically a high-precision adaptive non-stop weighing system based on deep learning. The system includes a front-end sensing unit, a multimodal weighing unit, an edge computing data processing unit, and a redundant data transmission unit. The multimodal weighing unit integrates arrays of quartz piezoelectric, piezoresistive, and fiber optic grating sensors. The edge computing unit utilizes a heterogeneous architecture and employs deep learning algorithms and attention mechanisms to dynamically allocate sensor weights based on vehicle speed to perform axle load calculation and error compensation. The front-end sensing unit performs multi-source sensing data fusion based on D-S evidence theory. By adopting the above scheme, this invention solves the problems of low accuracy and poor environmental adaptability under complex working conditions, achieving high-precision weighing across the entire speed domain and controlling the dynamic error within 0.5%.

Description

Technical Field

[0001] This invention relates to the field of dynamic weighing technology, and in particular to a high-precision adaptive non-stop weighing system based on deep learning. Background Technology

[0002] With the deepening construction of the globalized logistics system and the leapfrog development of the highway transportation industry, the safe operation and service life of highway infrastructure, as the main artery of the national economy, have become a core concern in the field of traffic management. Against this backdrop, the problem of overloaded and oversized freight vehicles has become increasingly prominent, causing irreversible physical damage to road surfaces and bridge structures. Furthermore, due to reduced vehicle braking performance and weakened handling stability, it significantly increases the probability of major traffic accidents. Related research data shows that when a truck's speed reaches 50 km / h and its total weight or axle load exceeds the limit by 30%, highway maintenance costs will surge by 20%. Under extreme overload conditions, such as when the axle load exceeds 100% of the standard axle load, the cumulative damage effect of a single trip on an asphalt road surface is equivalent to 256 repetitions of the standard axle load. For cement roads, this damage ratio is an astonishing 65,536 times. Therefore, developing and deploying intelligent non-stop weighing technology capable of high-precision, non-site enforcement has become a key path to regulate transportation order, protect transportation assets, and improve road safety.

[0003] In the current technological landscape, mainstream non-stop weighing solutions mainly fall into two categories: axle load scales and vehicle scales. From a design perspective, vehicle scales aim to acquire the full weight data of a vehicle through a large-scale load-bearing platform. While they offer high total weight measurement accuracy under quasi-static or extremely low-speed conditions, in dynamic traffic scenarios, factors such as vehicle center of gravity drift, high-frequency engine vibrations, and vertical jumps in the vehicle's suspension system generate complex alternating loads, significantly limiting the accuracy of axle load data analysis. In contrast, axle load scales focus on discrete weighing of each axle, designed to achieve higher traffic efficiency. However, in practical applications, due to the relatively simple sensor deployment, axle load scales are prone to axle count errors or tire identification inaccuracies when dealing with complex wheel and axle combinations, and their signal acquisition process is highly dependent on the stability of the vehicle's trajectory. If a vehicle decelerates non-linearly, changes lanes, or experiences minor road bumps, the raw signal captured by the sensor will generate severe noise fluctuations, leading to significant deviations in the total weight calculation. This not only directly causes subsequent toll disputes, but also induces severe traffic congestion due to repeated weighing or manual intervention.

[0004] A deeper analysis of the aforementioned technical shortcomings reveals a profound and less obvious technical contradiction in existing technologies when dealing with complex operating conditions: a mismatch between dynamic response bandwidth and intrinsic measurement accuracy. Specifically, traditional systems often rely on a single type of pressure sensor, whose physical response characteristics limit its linearity to a specific speed window. In high-speed traffic scenarios, the sensor must have an extremely high sampling frequency to capture instantaneous impact peaks, but high-frequency vibration noise severely erodes the signal-to-noise ratio of the useful signal. In low-speed or crawling conditions, the sensor requires extremely high sensitivity and quasi-static stability to overcome the effects of zero-point drift. Existing signal processing logic typically employs fixed-weight filtering algorithms, lacking adaptive adjustment mechanisms for real-time vehicle motion and environmental stress. This "one-size-fits-all" approach prevents the system from achieving dynamic closed-loop error compensation at the algorithm level when facing coupled factors such as full-speed-domain traffic, extreme weather, and road surface deformation. The root cause is that existing systems lack an intelligent logical architecture that can deeply integrate physical perception features with semantic environment information, which prevents them from fundamentally improving their limitations at the physical sensor level through digital signal processing.

[0005] In conclusion, although non-stop weighing technology has undergone a long period of evolution, balancing real-time performance with high precision, robustness with adaptability in complex, non-steady-state real-world traffic environments remains a bottleneck restricting the improvement of off-site enforcement efficiency. Dynamic response compensation at different driving speeds, deep decoupling of multi-sensor data, and reliable enhancement of the perception channel in harsh environments have become key technical challenges that urgently need to be overcome in the field of intelligent transportation systems. Constructing a high-precision adaptive non-stop weighing system based on a deep learning architecture that can perceive environmental changes in real time and dynamically optimize weighing decision logic is not only an urgent need to enhance the authority of traffic enforcement, but also an inevitable choice to promote the deep transformation of smart highways towards digitalization and intelligence. Summary of the Invention

[0006] To address the technical problems of existing non-stop weighing systems under complex working conditions, such as low dynamic response accuracy, poor environmental adaptability, inaccurate identification of axle number and wheel type, and insufficient data transmission reliability, this invention provides a high-precision adaptive non-stop weighing system based on deep learning.

[0007] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0008] A high-precision adaptive non-stop weighing system based on deep learning includes a front-end sensing unit, a multimodal weighing unit, an edge computing data processing unit, a redundant data transmission unit, a host computer management unit, and a remote operation and maintenance unit, deployed sequentially. These components—front-end sensing unit, multimodal weighing unit, edge computing data processing unit, redundant data transmission unit, host computer management unit, and remote operation and maintenance unit—construct a distributed communication network via industrial Ethernet to achieve real-time interaction and logical coordination of end-to-end data. Specifically, the front-end sensing unit and the multimodal weighing unit achieve millisecond-level clock alignment through a high-precision time synchronization module, ensuring strict consistency between vehicle motion characteristic data and dynamic weight acquisition data in the time domain. The edge computing data processing unit adopts a heterogeneous computing architecture and performs instruction-level optimization of the deep learning algorithm through a hardware acceleration module.

[0009] The front-end sensing unit is configured with a multi-source fusion sensing architecture, including infrared vehicle separators symmetrically arranged on both sides of the lane, dual inductive loop coil groups spaced apart front and rear, lidar sensors, microwave radar sensors, high-definition cameras, and speed sensors. Each sensing device performs physical layer data aggregation and preprocessing through a data fusion gateway. The infrared vehicle separators employ high-resolution through-beam infrared array sensors, integrating no fewer than 32 sets of infrared transmitting and receiving units to capture vehicle outline boundaries and generate high-level trigger signals. The dual inductive loop coil groups consist of a first and a second inductive loop coil spaced 2 meters apart, employing a multi-turn winding structure and equipped with a coil impedance self-compensation module to assist in vehicle positioning, axle trigger signal generation, and preliminary speed calibration. The lidar sensors are mounted on a gantry structure on the side of the lane, employing a 16-line laser scanning system with a scanning frequency set to 10Hz, to reconstruct the vehicle's three-dimensional geometric contours and extract axle spacing and tire quantity information. The microwave radar sensor and lidar sensor are coaxially deployed, operating at a frequency of 24GHz. In low-visibility environments such as rain, snow, fog, or sandstorms, they serve as redundant reinforcement for the lidar, providing continuous vehicle position and velocity vector data. The high-definition camera device uses a binocular vision camera with a single-channel pixel count of at least 5 megapixels and a frame rate of 25fps. It integrates an ISP image processing module to capture frontal and side images of the vehicle and license plate features. The speed sensor employs a laser-microwave composite speed measurement scheme, fusing two independent speed measurement data streams using a Kalman filter algorithm to ensure vehicle speed measurement accuracy within ±0.5km / h, covering a range of 0-80km / h. The time synchronization module uses a dual-mode GPS and BeiDou timing mechanism, achieving a timing accuracy on the order of 10ns.

[0010] The multimodal weighing unit includes an embedded weighing platform, a quartz piezoelectric sensor array, a piezoresistive sensor array, a fiber Bragg grating sensor array, a signal conditioning module, an analog-to-digital conversion module, and a weighing platform status monitoring module. The embedded weighing platform adopts a high-strength stainless steel and concrete composite structure, with dimensions of 6 meters × 3.75 meters. A buffer and shock-absorbing pad and a horizontal adjustment support are installed at the bottom, with a maximum static load capacity of 200 tons. The quartz piezoelectric sensor array is arranged in three rows along the longitudinal direction of the lane, with each row containing two symmetrically arranged sensors. Each sensor is 1.875 meters long, and the row spacing is 60 cm, used to capture high-frequency dynamic pressure signals. The piezoresistive sensor array is arranged in four groups along the transverse direction of the lane, with each group containing eight sensors, and the sensor spacing is 50 cm, used to collect quasi-static load signals during low-speed traffic. The fiber Bragg grating sensor array is embedded along the platform's diagonal and edge contours, monitoring the platform's physical deformation in real time by sensing stress and strain changes. The signal conditioning module integrates multiple charge amplifiers and instrumentation amplifiers. The gain adjustment range of the charge amplifiers is 1000-10000 times, and the gain adjustment range of the instrumentation amplifiers is 100-1000 times. The conditioning circuit also includes a second-order Butterworth low-pass filter and a notch filter, with the low-pass cutoff frequency adaptively adjusted between 10-100Hz to suppress 50Hz power frequency noise. The analog-to-digital conversion module uses a 24-bit high-precision ADC chip with a sampling rate of no less than 200kHz, supporting multi-channel synchronous sampling. The weighing platform status monitoring module includes tilt sensors, temperature sensors, and vibration sensors to acquire the platform's horizontal attitude, ambient temperature, and ambient background vibration.

[0011] The edge computing data processing unit adopts a heterogeneous computing architecture consisting of an ARM Cortex-A72 and an FPGA. The core controller uses an NXP i.MX8M Plus chip, and the hardware acceleration module uses a Xilinx Artix-7 FPGA chip. The ARM core performs system resource scheduling, data routing, and high-level logic operations, while the FPGA is responsible for parallel preprocessing and real-time trigger control of the underlying sensor signals. The controller connects to the data fusion gateway, analog-to-digital conversion module, and status monitoring module through dedicated physical interfaces. It performs axle load calculation, total overlap addition, axle type identification, and error closed-loop compensation through a built-in adaptive multimodal fusion weighing algorithm. The adaptive multimodal fusion weighing algorithm consists of a data preprocessing layer, a feature extraction layer, a fusion decision layer, and an error compensation layer.

[0012] In the adaptive multimodal fusion weighing algorithm, the data preprocessing layer uses a wavelet threshold denoising algorithm to denoise the signals from the three sensors, and combines environmental parameters provided by the state monitoring module to perform temperature compensation and vibration suppression. The feature extraction layer utilizes a convolutional neural network (CNN) to perform deep feature mining on the sensor signals. The CNN model contains 5 convolutional layers, 3 pooling layers, and 2 fully connected layers, outputting a 128-dimensional deep feature vector, and simultaneously extracting traditional time-frequency domain features such as signal peak value, integral area, and variance. The fusion decision layer introduces an attention mechanism based on the real-time vehicle speed feedback from the speed sensor. Dynamically allocate weights. Specifically, set a speed threshold. and :

[0013] when At that time, the characteristic weight of the piezoresistive sensor was assigned as 0.6, the weight of the quartz piezoelectric sensor as 0.3, and the weight of the fiber Bragg grating sensor as 0.1.

[0014] when At that time, the characteristic weight of the piezoresistive sensor was assigned as 0.4, the weight of the quartz piezoelectric sensor was assigned as 0.4, and the weight of the fiber Bragg grating sensor was assigned as 0.2.

[0015] when At that time, the weight of the piezoresistive sensor was assigned as 0.2, the weight of the quartz piezoelectric sensor as 0.6, and the weight of the fiber optic grating sensor as 0.2.

[0016] The fusion results are mapped through a fully connected layer to obtain the preliminary axle load. The error compensation layer is based on a support vector machine (SVM) model. The inputs include platform tilt angle, ambient temperature, vibration amplitude, vehicle speed, and wheel and axle geometric parameters. The output is the weighing deviation compensation value. Finally, the calibrated axle load and total weight data are output.

[0017] The redundant data transmission unit adopts a triple-mode redundancy architecture consisting of an RJ485 bus, industrial Ethernet, and 5G wireless communication. The RJ485 bus uses differential transmission mode with a rate of no less than 9600bps, supporting communication for local terminals in the field. The industrial Ethernet uses the gigabit standard, carrying TCP / IP and Modbus TCP protocols, for high-bandwidth data transmission between the edge computing unit and the host computer management unit. The 5G wireless communication module uses SA networking mode with a latency of less than 20ms, for remote operation and maintenance and remote data synchronization. The transmission unit has a built-in link monitoring module; when the main link packet loss rate or latency exceeds a preset threshold, the logic controller drives the physical layer switch to achieve seamless link switching. Simultaneously, the data encryption module uses the AES-256 algorithm to encrypt uplink data.

[0018] The host computer management unit is based on a collaborative architecture of industrial control computer and cloud platform, and includes a local data storage module, a visualization display module, an alarm and early warning module, and an enforcement management module. The local data storage module adopts a redundant SSD and HDD scheme, with a storage capacity of no less than 11TB, ensuring long-term traceability of vehicle passage records. The visualization display module renders vehicle axle load, license plate, vehicle speed, and system operating status in real time through a 27-inch high-definition touchscreen. The alarm and early warning module executes multi-level warnings based on the overload ratio and pushes alarm information to mobile terminals. The enforcement management module automatically generates violation notices and penalty tickets and interfaces with an external traffic enforcement database.

[0019] Furthermore, the data fusion gateway of the front-end sensing unit adopts a fusion algorithm based on DS evidence theory, the specific steps of which are as follows:

[0020] S11: Each sensing device collects raw data in parallel and removes outliers that exceed a reasonable threshold range;

[0021] S12: Extract vehicle position, speed, profile, and wheel axle geometric features;

[0022] S13: Construct the basic probability assignment function corresponding to each feature quantity ;

[0023] S14: The evidence combination rule is applied for fusion, and the calculation formula is as follows:

[0024] ;

[0025] in, , representing the coefficient of evidence conflict;

[0026] S15: Outputs the fused vehicle status information to achieve position recognition with an accuracy of 5cm.

[0027] Furthermore, the logic control flow of the adaptive multimodal fusion weighing algorithm is as follows:

[0028] S1: The edge computing data processing unit acquires the current vehicle speed output by the speed sensor in real time. and with preset switching threshold Compare;

[0029] S2: When When the system automatically switches to low-speed detection mode, it logically prioritizes calling the DC component signal of the piezoresistive sensor array to calculate the axle load, and uses the pulse leading edge of the quartz piezoelectric sensor for signal synchronization calibration.

[0030] S3: When When the system switches to high-speed dynamic mode, it prioritizes processing the high-frequency sampling signal of the quartz piezoelectric sensor array and uses the low-frequency trend term of the piezoresistive sensor to filter and compensate for the resonance interference generated by the road surface excitation.

[0031] S4: After calculating the axle weight of each axle, perform algebraic summation to obtain the total vehicle weight, and simultaneously analyze the number of axles and axle type based on the time axis distribution characteristics of the signal.

[0032] Furthermore, the remote operation and maintenance unit includes an equipment status monitoring module, a remote diagnostics module, a remote upgrade module, and an operation and maintenance management module. The equipment status monitoring module transmits system voltage, current, and the health status of each sensor node in real time via a 5G link. The remote diagnostics module has a built-in fault expert system that analyzes fault codes uploaded by the edge computing unit to locate the root cause. The remote upgrade module supports online updates of algorithm models based on differential technology. The operation and maintenance management module is responsible for the automated closed-loop management of maintenance logs.

[0033] Furthermore, the binocular vision camera is equipped with a YOLOv8 recognition algorithm optimized through transfer learning, the implementation process of which is as follows:

[0034] S21: Construct a diverse vehicle image dataset containing different lighting, precipitation, and occlusion conditions, and perform enhancement processing such as rotation, scaling, and Gaussian noise injection;

[0035] S22: Pre-train the YOLOv8 backbone network;

[0036] S23: Fine-tune the model using augmented datasets to optimize the convergence point of the loss function;

[0037] S24: Deploy the optimized weight parameters to the camera ISP hardware to achieve real-time identification of license plates, vehicle models and loading status, with a license plate recognition accuracy of 99.8%.

[0038] In a preferred embodiment of the present invention, the individual sensors of the quartz piezoelectric sensor array use high-purity quartz crystals as sensitive elements, with a linearity error of less than 1.0% and an operating temperature range covering -40℃ to +80℃.

[0039] In a preferred embodiment of the present invention, the fiber optic grating sensor array adopts a quasi-distributed deployment scheme and uses wavelength division multiplexing technology to achieve synchronous demodulation of signals from multiple sensor nodes in a single fiber channel, with a wavelength resolution better than 1 pm.

[0040] In a preferred embodiment of the present invention, the edge computing data processing unit and the multimodal weighing unit use LVDS differential signals for high-speed raw data transmission to enhance the system's anti-electromagnetic interference performance.

[0041] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0042] 1. Significant improvement in weighing accuracy. By introducing a three-mode sensor cross-deployment scheme of quartz piezoelectric, piezoresistive and fiber optic grating, and combining it with a deep learning adaptive fusion algorithm based on attention mechanism, this system completely solves the accuracy bottleneck of traditional single sensor solutions in the full speed range (0-80km / h), and the dynamic weighing error is successfully controlled within 0.5%.

[0043] 2. The system possesses extremely strong physical robustness and environmental adaptability. By fusing lidar, microwave radar, and visual perception data through DS evidence theory, interference from re-triggering logic caused by severe weather is effectively eliminated. At the same time, through a multi-factor error compensation model, the influence of temperature, tilt angle, and vibration on the re-triggering eigensignal is mitigated, and the MTBF index is improved to over 10,000 hours.

[0044] 3. It achieves a high degree of automation and intelligent control. The system deeply integrates edge computing and cloud management, from vehicle entry recognition and multimodal data fusion weighing to the solidification of evidence for violation enforcement, the entire process requires no manual intervention, significantly improving the administrative enforcement efficiency of traffic management departments.

[0045] 4. Excellent engineering compatibility and traffic support capabilities. The system supports non-stop continuous passage weighing at speeds of 0-80km / h, and both the physical and protocol layers follow standardized designs, enabling seamless integration and data exchange with various existing intelligent traffic enforcement platforms, thus possessing extremely high value for promotion and application. Attached Figure Description

[0046] Figure 1 This is a block diagram showing the overall composition of the non-stop weighing system of the present invention;

[0047] Figure 2 This is a schematic diagram of the physical deployment of the front-end sensing unit and the multimodal weighing unit of the present invention;

[0048] Figure 3 This is a schematic diagram showing the arrangement of sensors in the multimodal weighing unit of the present invention;

[0049] Figure 4 This is a logical hierarchy diagram of the adaptive multimodal fusion weighing algorithm of the present invention;

[0050] Figure 5 This is a flowchart illustrating the data fusion algorithm in the front-end sensing unit of the present invention. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0052] A high-precision adaptive non-stop weighing system based on deep learning is designed with industrial-grade high reliability in mind. It utilizes an industrial Ethernet network to construct a distributed, high-bandwidth, low-latency communication network. The system, from the physical layer to the logical layer, sequentially deploys a front-end sensing unit, a multimodal weighing unit, an edge computing data processing unit, a redundant data transmission unit, a host computer management unit, and a remote operation and maintenance unit. Millisecond-level clock alignment is achieved between these units through a high-precision time synchronization module, specifically employing a dual-mode GPS and BeiDou timing mechanism with an accuracy on the order of 10ns. This clock synchronization mechanism ensures strict consistency in the time domain between the motion feature data captured by the front-end sensing unit and the dynamic weight data collected by the multimodal weighing unit when a vehicle passes through the weighing area, providing a highly reliable raw data alignment benchmark for subsequent deep fusion algorithms.

[0053] The front-end sensing unit, serving as the system's input, employs a multi-source fusion sensing architecture. Its physical layer hardware is integrated into the gantry structure and surrounding support facilities above the lane. Infrared vehicle separators are symmetrically positioned on both sides of the lane, integrating 32 sets of high-resolution through-beam infrared transmitting and receiving units. These units are vertically arranged with a 10mm spacing, capable of capturing signal transitions caused by vehicles blocking the light path and generating high-level trigger signals in real time, thus accurately defining the vehicle's logical boundaries. Simultaneously, a dual-loop inductive coil group is buried beneath the road surface, consisting of a first and second inductive coil spaced 2 meters apart. To eliminate the influence of environmental humidity and metallic interference on the resonant signal, each coil is equipped with a coil impedance self-compensation module, dynamically adjusting the resonant frequency to maintain consistent detection sensitivity, primarily used for assisting vehicle positioning and initial speed calibration.

[0054] In the advanced perception layer of the front-end perception unit, a lidar sensor is installed at the center of the crossbeam of the gantry structure. Employing a 16-line laser scanning system with a scanning frequency set at 10Hz, it can perform a 360-degree spatial scan of passing vehicles, thereby reconstructing the vehicle's three-dimensional geometric contours. The point cloud data output by the lidar sensor is used to extract key physical information such as the vehicle's wheel-axle distance, number of tires, and vehicle height. To address the issue of decreased lidar detection accuracy in adverse weather conditions, a microwave radar sensor operating at 24GHz is coaxially deployed within the system. Utilizing the Doppler effect, the microwave radar sensor provides continuous vehicle position and velocity vector data in rain, snow, fog, or dust storms, serving as redundancy reinforcement for the lidar.

[0055] Furthermore, the high-definition camera equipment adopts a binocular vision camera architecture with a single-channel pixel count of 5 megapixels and a frame rate of 25fps, and integrates an ISP image processing module. This camera not only captures high-definition images of the vehicle's front and sides but also loads a YOLOv8 recognition algorithm optimized through transfer learning. The algorithm's implementation first involves constructing a diverse vehicle image dataset containing different lighting conditions, precipitation, and occlusion, and performing enhancement processing such as rotation, scaling, and Gaussian noise injection. Subsequently, the YOLOv8 backbone network is pre-trained, and the model is fine-tuned using the enhanced dataset to optimize the convergence point of the loss function. Finally, the optimized weight parameters are deployed to the camera's ISP hardware to achieve real-time discrimination of license plates, vehicle models, and loading status. Its measured license plate recognition accuracy reaches 99.8%. The speed sensor employs a laser-microwave composite speed measurement scheme, fusing two independent speed measurement data streams through a Kalman filter algorithm to ensure that the vehicle speed measurement accuracy remains stable within ±0.5 km / h in the 0-80 km / h range.

[0056] All physical layer data collected by the front-end sensing units converges to the data fusion gateway. This gateway runs a fusion algorithm based on the DS evidence theory. Specifically, after each sensing device collects raw data in parallel, outliers exceeding a reasonable threshold range are first removed, followed by the extraction of vehicle position, speed, contour, and wheel axle geometric features. For each feature, the algorithm constructs a corresponding basic probability assignment function. When applying the evidence combination rule for fusion, the calculation formula is as follows: ,in This represents the coefficient of evidence conflict. Through this fusion processing, the system can output high-confidence vehicle status information, achieving vehicle position recognition with an accuracy of 5cm.

[0057] The multimodal weighing unit is the core execution layer of this system, with its core being an embedded weighing platform. This platform employs a composite structure of high-strength stainless steel and concrete, measuring 6 meters × 3.75 meters, effectively covering the width of a standard lane. To suppress transient impact vibrations generated when vehicles enter, a buffer and shock-absorbing pad layer composed of nitrile rubber and spring steel is installed at the bottom of the platform, along with adjustable horizontal supports, ensuring a static load-bearing capacity of 200 tons.

[0058] The multimodal weighing unit integrates three sensor arrays with different physical mechanisms. Quartz piezoelectric sensor arrays are arranged longitudinally along the lane, with each row containing two 1.875-meter-long sensor units, and the spacing between rows is strictly controlled at 60cm. These quartz piezoelectric sensors utilize their high-frequency response characteristics to specifically capture high-frequency pressure signals when vehicles dynamically pass by. Piezoresistive sensor arrays are arranged laterally along the lane in four groups, each group consisting of eight sensors spaced 50cm apart, primarily used to collect load signals under low-speed or quasi-static conditions. The fiber optic grating sensor array employs a quasi-distributed deployment scheme, using wavelength division multiplexing (WDM) technology to achieve synchronous demodulation of signals from multiple nodes within a single fiber optic channel. Its wavelength resolution is better than 1pm, and it is mainly used to sense minute physical deformations of the weighing platform.

[0059] In terms of signal processing, the signal conditioning module integrates a high-precision charge amplifier and an instrumentation amplifier. The charge amplifier provides a gain adjustment range of 1000 to 10000 times for the weak current signal from the quartz piezoelectric sensor; the instrumentation amplifier provides a gain of 100 to 1000 times for the piezoresistive signal. To eliminate interference from the power system, a second-order Butterworth low-pass filter and a notch filter are also designed in the conditioning circuit, whose low-pass cutoff frequency can be adaptively adjusted between 10Hz and 100Hz according to the current vehicle speed. The analog-to-digital conversion module uses a 24-bit high-precision ADC chip with a sampling rate set to 200kHz, supporting multi-channel synchronous sampling to ensure waveform distortion-free operation. In addition, the weighing platform status monitoring module includes tilt sensors, temperature sensors, and vibration sensors to acquire real-time data on the platform's horizontal attitude, ambient temperature, and background noise.

[0060] The edge computing data processing unit adopts a heterogeneous computing architecture consisting of an ARM Cortex-A72 and an FPGA. The core controller uses an NXP i.MX8M Plus chip, and the hardware acceleration module uses a Xilinx Artix-7 FPGA. In this architecture, the FPGA is responsible for the parallel pipeline preprocessing and real-time trigger control of the underlying sensor signals, while the ARM core is responsible for high-level system resource scheduling and the operation of the adaptive multimodal fusion weighing algorithm. The algorithm consists of a data preprocessing layer, a feature extraction layer, a fusion decision layer, and an error compensation layer. During algorithm execution, the data preprocessing layer uses a wavelet threshold denoising algorithm to denoise the three sensor signals and performs temperature compensation in conjunction with environmental parameters provided by the status monitoring module.

[0061] The feature extraction layer utilizes a convolutional neural network (CNN) to extract deep features from the signal. This CNN model contains 5 convolutional layers, 3 pooling layers, and 2 fully connected layers, capable of outputting a 128-dimensional feature vector. The fusion decision layer introduces an attention mechanism based on the real-time vehicle speed feedback from the speed sensor. Dynamically allocate weights. Specifically, set a speed threshold. and ;

[0062] when At that time, the characteristic weights of the piezoresistive sensor were assigned as 0.6, the quartz piezoelectric sensor as 0.3, and the fiber Bragg grating sensor as 0.1.

[0063] when At that time, the weights of the three factors were 0.4, 0.4, and 0.2, respectively.

[0064] when At that time, the piezoresistive weight decreased to 0.2, the quartz piezoelectric weight increased to 0.6, and the fiber grating remained at 0.2;

[0065] This weighting strategy fully leverages the sensitivity advantages of different sensors across various speed ranges. The initially generated axle load data is further fed into an error compensation layer, which is based on a support vector machine (SVM) model. This layer takes platform tilt angle, temperature, vibration amplitude, and vehicle geometry parameters as inputs and outputs the final calibrated axle load and total weight.

[0066] The logical control of the adaptive multimodal fusion weighing algorithm follows this process: First, the edge computing data processing unit acquires the real-time vehicle speed and compares it with a preset threshold. When the vehicle is at a low speed ( When the vehicle passes through, the system automatically enters low-speed detection mode, logically prioritizing the DC component signal of the piezoresistive sensor array and using the pulse leading edge of the quartz piezoelectric signal for synchronization. When the vehicle is traveling at high speed... When the signal passes through, the system switches to high-speed dynamic mode, prioritizing the processing of high-frequency sampling signals from the quartz piezoelectric sensor array and using the low-frequency trend term of the piezoresistive sensor to filter for resonance interference on the weighing platform. The system calculates the weight of each axis, performs algebraic accumulation, and analyzes the number of axes based on the signal's time axis distribution.

[0067] Redundant data transmission units ensure secure and reliable data transmission. The system employs a tri-mode redundancy architecture consisting of an RJ485 bus, industrial Ethernet, and 5G wireless communication. The RJ485 bus handles basic communication for local terminals; the industrial Ethernet carries gigabit-rate TCP / IP data streams, connecting edge computing units and the host computer; the 5G wireless module uses SA networking, providing an ultra-low latency channel of less than 20ms for remote operation and maintenance. The link monitoring module monitors the packet loss rate in real time, triggering seamless physical layer switching once the threshold is exceeded. Simultaneously, the data encryption module uses the AES-256 algorithm to encrypt all uplink sensitive data.

[0068] The host computer management unit is deployed on an industrial control computer and equipped with a redundant storage system of no less than 11TB. The visualization display module renders vehicle passage information in real time through a high-definition touchscreen, including axle load, total weight, license plate, and violation status. The law enforcement management module can automatically generate law enforcement documents and connect to external databases. The remote operation and maintenance unit integrates equipment status monitoring, expert fault diagnosis, online algorithm upgrades based on differential technology, and operation and maintenance log management, which greatly reduces the system's maintenance costs.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-precision adaptive non-stop weighing system based on deep learning, characterized in that, It includes a front-end sensing unit, a multimodal weighing unit, an edge computing data processing unit, a redundant data transmission unit, a host computer management unit, and a remote operation and maintenance unit, which are deployed in sequence. The front-end sensing unit, the multimodal weighing unit, the edge computing data processing unit, the redundant data transmission unit, the host computer management unit, and the remote operation and maintenance unit construct a distributed communication network through industrial Ethernet to realize real-time interaction and logical coordination of data across the entire link. The system also includes a high-precision time synchronization module. The front-end sensing unit and the multimodal weighing unit achieve millisecond-level clock alignment through the high-precision time synchronization module to ensure strict consistency between vehicle motion characteristic data and dynamic weight acquisition data in the time domain. The front-end sensing unit is configured with a multi-source fusion sensing architecture to extract the vehicle's physical contours, motion vectors, and trigger signals; the multimodal weighing unit uses a cross-deployment of sensor arrays with various physical mechanisms to capture dynamic axle load signals and platform environmental status across the entire speed domain; the edge computing data processing unit adopts a heterogeneous computing architecture and incorporates an adaptive multimodal fusion weighing algorithm to perform deep feature extraction and error compensation on the collected multidimensional data; the redundant data transmission unit is configured with a multimodal redundancy architecture to perform encrypted data transmission and link status monitoring.

2. The high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The front-end sensing unit includes infrared vehicle separators symmetrically arranged on both sides of the lane, dual ground loop coil groups arranged at intervals in front and behind, lidar sensors, microwave radar sensors, high-definition camera equipment, speed sensors, and data fusion gateways. The infrared vehicle separator uses a through-beam infrared array sensor, which integrates no less than 32 sets of infrared transmitting and receiving units to capture the vehicle outline boundary and generate a high-level trigger signal. The dual inductive coil group consists of a first inductive coil and a second inductive coil spaced 2 meters apart, and each coil is equipped with a coil impedance self-compensation module to assist in vehicle positioning, axle trigger signal generation, and preliminary speed calibration. The lidar sensor is mounted on the gantry structure on the side of the lane and uses a 16-line laser scanning system to reconstruct the three-dimensional geometric contour of the vehicle and extract information on wheel axle spacing and tire quantity. The microwave radar sensor and the lidar sensor are coaxially arranged and operate at a frequency of 24 GHz. As a redundancy reinforcement for the lidar, it provides continuous vehicle position and velocity vector data. The high-definition camera device uses a binocular vision camera and integrates an ISP image processing module to capture multi-dimensional image features of the vehicle. The speed sensor adopts a laser-microwave composite speed measurement scheme and fuses two independent speed measurement data through a Kalman filter algorithm. The data fusion gateway runs a fusion algorithm based on DS evidence theory. Its execution steps include: parallel acquisition of raw data from each sensing device and removal of outliers; extraction of vehicle position, speed, contour, and wheel axle geometric features; and construction of basic probability assignment functions corresponding to each feature. The application includes a conflict of evidence coefficient. The evidence combination rules are fused to output the fused vehicle status information, achieving position recognition with an accuracy of 5cm.

3. The high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The multimodal weighing unit includes an embedded weighing platform, a quartz piezoelectric sensor array, a piezoresistive sensor array, a fiber optic grating sensor array, a signal conditioning module, an analog-to-digital conversion module, and a weighing platform status monitoring module. The embedded weighing platform adopts a high-strength stainless steel and concrete composite structure. The bottom of the platform is equipped with a buffer and shock-absorbing pad and a horizontal adjustment support. The maximum static load capacity is 200t. The quartz piezoelectric sensor array is arranged in no less than 3 rows along the longitudinal direction of the lane. Each row contains 2 quartz piezoelectric sensors with a single unit length of 1.875 meters and a row spacing of 60 cm, which are used to capture high-frequency dynamic pressure signals. The piezoresistive sensor array is arranged in four groups along the lane, each group containing eight sensors, for collecting quasi-static load signals during low-speed passage. The fiber grating sensor array adopts a quasi-distributed deployment scheme, and is embedded and installed along the diagonal and edge contours of the embedded weighing platform, and senses the physical deformation of the platform through wavelength division multiplexing technology. The weighing platform status monitoring module includes a tilt sensor, a temperature sensor, and a vibration sensor, used to acquire the platform's horizontal attitude, ambient temperature, and ambient background vibration.

4. The high-precision adaptive non-stop weighing system based on deep learning according to claim 3, characterized in that, The signal conditioning module integrates a multi-channel charge amplifier and an instrumentation amplifier. The gain adjustment range of the charge amplifier is 1000-10000 times, and the gain adjustment range of the instrumentation amplifier is 100-1000 times. The signal conditioning module also includes a second-order Butterworth low-pass filter and a notch filter. Its low-pass cutoff frequency is configured to adaptively adjust between 10-100Hz according to the real-time vehicle speed to suppress 50Hz power frequency noise. The analog-to-digital conversion module uses a 24-bit high-precision ADC chip with a sampling rate of no less than 200kHz, and supports multi-channel synchronous sampling of the signals from the quartz piezoelectric sensor array and the piezoresistive sensor array.

5. The high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The edge computing data processing unit adopts a heterogeneous computing architecture consisting of ARM and FPGA, with the core controller using the NXPi.MX8M Plus chip and the hardware acceleration module using the Xilinx Artix-7 FPGA chip. The FPGA chip is responsible for the parallel preprocessing and real-time trigger control of the underlying sensor signals. The FPGA chip is equipped with an independent signal extraction IP core. By monitoring the slope change rate of the quartz piezoelectric signal in real time, the full-speed sampling mode is activated when the slope change rate exceeds the preset dynamic threshold value, and the sampling window is moved forward by 50ms. The ARM chip performs system resource scheduling, data routing, and logical operations of the adaptive multimodal fusion weighing algorithm; the edge computing data processing unit and the multimodal weighing unit transmit data through LVDS differential signals.

6. The high-precision adaptive non-stop weighing system based on deep learning according to claim 3, characterized in that, The adaptive multimodal fusion weighing algorithm consists of a data preprocessing layer, a feature extraction layer, a fusion decision layer, and an error compensation layer. The data preprocessing layer uses a wavelet threshold denoising algorithm to denoise the three sensor signals, and combines the environmental parameters provided by the weighing platform status monitoring module to perform temperature compensation and vibration suppression. The feature extraction layer uses a convolutional neural network (CNN) to perform deep feature mining on the sensor signal. The CNN model contains 5 convolutional layers, 3 pooling layers and 2 fully connected layers to output a 128-dimensional deep feature vector and simultaneously extract the time-frequency domain features of signal peak value, integral area and variance. The fusion decision layer introduces an attention mechanism based on the real-time vehicle speed feedback from the speed sensor. Dynamic weight allocation, specifically including: Set the first speed threshold Second speed threshold ; when At that time, the characteristic weight of the piezoresistive sensor was assigned as 0.6, the weight of the quartz piezoelectric sensor as 0.3, and the weight of the fiber Bragg grating sensor as 0.

1. when At that time, the characteristic weights of the piezoresistive sensor, the quartz piezoelectric sensor, and the fiber Bragg grating sensor were assigned as 0.4, 0.4, and 0.2, respectively. when At that time, the weights of the piezoresistive sensor feature were assigned as 0.2, the weights of the quartz piezoelectric sensor and the fiber optic grating sensor were assigned as 0.6, and the weights of the fiber optic grating sensor were assigned as 0.

2. The fusion result was mapped through a fully connected layer to obtain the initial axis weight.

7. The high-precision adaptive non-stop weighing system based on deep learning according to claim 6, characterized in that, The error compensation layer is based on the support vector machine (SVM) model. It takes the tilt angle of the embedded weighing platform, ambient temperature, vibration amplitude, vehicle speed and wheel and axle geometric parameters as input items and outputs the weighing deviation compensation value to obtain the calibrated axle weight and total weight data. The logic control flow of the adaptive multimodal fusion weighing algorithm is as follows: S1: Edge computing data processing unit obtains current vehicle speed and with preset switching threshold contrast; S2: When When the system switches to low-speed detection mode, it prioritizes the use of the DC component signal of the piezoresistive sensor array to calculate the axle load and uses the pulse leading edge of the quartz piezoelectric sensor for signal synchronization calibration. S3: When When the system switches to high-speed dynamic mode, it prioritizes processing the high-frequency sampling signal of the quartz piezoelectric sensor array and uses the low-frequency trend term of the piezoresistive sensor to filter and compensate for the resonance interference of the refractive plateau. S4: After calculating the axle weight of each axle, perform algebraic summation to obtain the total vehicle weight, and analyze the number of axles and axle type based on the signal time axis distribution.

8. A high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The high-definition camera device is loaded with a YOLOv8 recognition algorithm optimized through transfer learning. The implementation process is as follows: constructing a vehicle image dataset containing different lighting, precipitation, and occlusion conditions and performing data augmentation processing; pre-training the YOLOv8 backbone network; and fine-tuning the model using the augmented dataset to optimize the convergence point of the loss function. The optimized weight parameters are deployed to the camera's ISP hardware to enable real-time identification of license plates, vehicle models, and loading status.

9. A high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The redundant data transmission unit adopts a three-mode redundancy architecture consisting of RJ485 bus, industrial Ethernet and 5G wireless communication; the 5G wireless communication module adopts SA networking mode with a latency of less than 20ms. The redundant data transmission unit has a built-in link monitoring module and a data encryption module. When the packet loss rate or delay of the main link exceeds a preset threshold, the link monitoring module drives the physical layer switch to achieve seamless link switching. The data encryption module uses the AES-256 algorithm to encrypt the uplink data.

10. A high-precision adaptive non-stop weighing system based on deep learning according to claim 1, characterized in that, The host computer management unit includes a local data storage module, a visualization display module, an alarm and early warning module, and an enforcement management module; the visualization display module renders vehicle axle load, license plate, vehicle speed, and system operating status in real time through a high-definition touch screen; The remote operation and maintenance unit includes an equipment status monitoring module, a remote diagnostic module, a remote upgrade module, and an operation and maintenance management module. The equipment status monitoring module transmits system electrical parameters and sensor health status in real time via a 5G link. The remote diagnostic module has a built-in fault expert system for root cause localization through fault codes. The remote upgrade module supports online updates of algorithm models based on differential technology.

Images