Large freight vehicle-mounted weighing method and system based on time sequence large model

Abstract

The invention belongs to the technical field of large freight vehicle weighing, and discloses a large freight vehicle-mounted weighing method and system based on a time sequence large model, and the method comprises the steps: 1, fusing multi-source data collected by a plurality of sensors through employing a timestamp alignment algorithm, and generating and storing time sequence coded data with spatial-temporal characteristics; 2, preprocessing the original time sequence data, and constructing a multi-modal training data set; 3, performing pre-training and fine tuning by using the data set to construct a time sequence large model based on a Transform architecture; 4, inputting newly collected load data with time sequence characteristics into the time sequence large model to carry out real-time dynamic weighing calculation; 5, constructing an error correction hybrid system to accurately correct the weighing result; and 6, outputting an accurate weighing result with a time sequence characteristic in real time, and carrying out reliability prediction on a load change trend.

Description

Technical Field

[0001] The present invention belongs to the technical field of large-piece freight vehicle weighing, and in particular relates to a large-piece freight vehicle on-board weighing method and system based on a time series large model. Background Art

[0002] Oversized freight refers to the process of loading, weighing, and transporting cargo that exceeds conventional dimensions, weight, or volume. In this process, a good weighing method is crucial for saving weighing time and improving transportation efficiency. Currently, traditional solutions for weighing oversized freight vehicles face the following three technical obstacles: (1) insufficient spatiotemporal alignment accuracy of multiple sensors, resulting in spatiotemporal feature drift; (2) poor model generalization under dynamic interference conditions; (3) a conflict between limited hardware resources and real-time computing; and (4) a lack of analysis and prediction of load changes, making it difficult to provide early warning of overloading risks.

[0003] While some existing time series models offer advantages in time series analysis, their ability to extract spatiotemporal features from multiple sensors is limited, and a complete closed-loop system encompassing data acquisition, feature fusion, model training, error correction, and trend prediction has yet to be established. Therefore, a method and system for onboard weighing of large-scale freight vehicles based on large-scale time series models is urgently needed to improve the real-time, accuracy, and reliability of such weighing. Summary of the Invention

[0004] To address the challenges of the existing technology, the present invention provides a method and system for onboard weighing of large freight vehicles based on a large time series model. This method overcomes these challenges by addressing feature distortion caused by spatiotemporal drift of multiple sensors through FPGA hardware synchronization and spatiotemporal convolution. It also addresses the poor model generalization under dynamic conditions through the design of causal attention and online transfer learning mechanisms. It also addresses the conflict between hardware resources and real-time computational performance through the use of model distillation and structured pruning strategies. Furthermore, through time series analysis and uncertainty assessment of vehicle motion states, an ARIMA-Prophet combined prediction model is constructed to address weight change trend prediction and early warning. Furthermore, the present invention's technical approach forms a complete closed-loop system consisting of "data acquisition - feature fusion - model training - error correction - trend prediction." Furthermore, the constructed error correction hybrid system accurately corrects weighing results, effectively reducing weighing errors and improving the reliability of the results, enabling real-time, high-precision weighing of freight vehicles.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for weighing large freight vehicles based on a time series large model, the method comprising:

[0007] Step 1: Use a timestamp alignment algorithm to fuse the multi-source data collected by multiple sensors to generate and store time series coded data with spatiotemporal characteristics. Multiple sensors are installed at various locations on large freight vehicles.

[0008] Step 2: Preprocess the time series encoding data with spatiotemporal characteristics to construct a multimodal training dataset;

[0009] Step 3: Use the multimodal training dataset for pre-training and fine-tuning to build a large time series model based on the Transformer architecture;

[0010] Step 4: Input the collected load data with time series characteristics into the time series model based on the Transformer architecture to perform real-time dynamic weighing calculations;

[0011] Step 5: Construct an error correction hybrid system to correct the weighing results;

[0012] Step 6: Output the weighing results with time series characteristics in real time and make predictions on the load change trend.

[0013] Preferably, the step 1 comprises:

[0014] Step 1.1, collecting dynamic load information at each location through a distributed pressure sensor array;

[0015] Step 1.2: Use the nanosecond timestamp alignment algorithm based on the FPGA hardware time synchronization trigger circuit to fuse multi-source data and generate time series load data with spatiotemporal characteristics.

[0016] Step 1.3: Store the time series payload data in the dual-channel ring buffer memory of the edge computing node, and use the DMA mechanism to achieve non-blocking data transmission with the pre-processing module;

[0017] The sensor array in step 1.1 refers to at least 2M pressure sensors covering the support point area of ​​the suspension system of the large freight vehicle, with at least 2 sensors per support point, and 0.5 sensors / m2 in the non-support area. 2 The density distribution form of the adjacent sensors d and the maximum adjacent axle spacing L of the vehicle should satisfy the constraint condition d≤0.3L, where M is the number of vehicle suspension support points and L is the maximum horizontal distance between adjacent axle centers;

[0018] The technical basis for FPGA selection in step 1.2 is: FPGA clock frequency is configured in the range of 500MHz-1GHz, FPGA clock frequency f FPGA ≥2f to meet the nanosecond synchronization requirement, and the timestamp alignment error δ satisfies:

[0019]

[0020] Where f is the sensor sampling frequency and f≥1kHz;

[0021] In step 1.3, the dual-channel ring buffer memory adopts a ping-pong cache architecture, alternately performs data writing and reading operations by setting a double buffer, and cooperates with a DMA chain transfer mechanism;

[0022] The dual-channel ring buffer memory capacity is configured as 2×N×B×T, where N is the total number of sensors, B is the data bandwidth of a single sensor, and T is the maximum non-blocking transmission time.

[0023] Preferably, the step 2 comprises:

[0024] Improved SURE threshold denoising based on wavelet packet decomposition, conversion of one-dimensional time series data into two-dimensional feature maps through Gram angular field transform, time series anomaly detection and waveform reconstruction, and dynamic reconstruction error threshold ε are set as:

[0025] ε=μ+3σ;

[0026] Where μ is the mean reconstruction error of the training set and σ is the standard deviation;

[0027] The construction of the multimodal training dataset is achieved by integrating the load distribution feature map and the vehicle motion state vector through feature-level fusion, and using the attention mechanism to achieve cross-modal feature interaction.

[0028] Preferably, in step 3, the time series large model based on the Transformer architecture includes: a spatiotemporal convolutional coding layer, a multi-head causal attention mechanism, a dynamic weight allocation module, and a model distillation;

[0029] The spatiotemporal convolutional coding layer uses a two-dimensional dilated causal convolution kernel to extract local spatiotemporal features, and the dilation coefficient d is d=2 n-1 Exponential growth, where n is the number of layers and the convolution kernel size satisfies the formula k = 2d + 1;

[0030] The multi-head causal attention mechanism uses a sliding window attention with a sliding window of 64 and 8 attention heads to limit the computational complexity;

[0031] The dynamic weight allocation module implements the importance assessment of sensor channels through a learnable gating network; the learnable gating network consists of two fully connected layers + Softmax activation;

[0032] The model distillation adopts a teacher-student architecture. The teacher model is a 12-layer Transformer with a hidden layer dimension of 512 and 8 attention heads. The student model adopts a 4-layer structured pruned network with a hidden layer dimension of 256. The top 50% important attention heads are retained in the Transformer layer. Knowledge transfer is achieved through KL divergence loss, and the pruning standard is the L1 norm of the attention head weights.

[0033] Preferably, in step 5, the error correction hybrid system uses an adaptive Kalman filter for dynamic error compensation at the front end, and sets the process noise covariance matrix Q t Dynamic update mechanism; the back-end application correction network adopts online transfer learning strategy to update network parameters; the error correction amount is superimposed on the model output through residual connection design to achieve error correction of the model output result;

[0034] Among them, the process noise covariance matrix Q t The dynamic update mechanism is:

[0035] Q t =αQ t-1 +(1-α)·diag(ξ t ·ξ t T );

[0036] Among them, α = 0.95 is the smoothing factor, ξ t is the residual vector at time t;

[0037] The state vector of the adaptive Kalman filter is x = [w,w t ,a] T , where w is the load, w t is the load change rate, a is the vehicle acceleration;

[0038] The back-end correction network adopts a three-layer gated recurrent unit structure. The input is the Kalman residual. The input dimension is consistent with the Kalman residual dimension. The number of hidden units does not exceed 32, and the time step is 10.

[0039] Online transfer learning uses an elastic weight solidification algorithm and sets the Fisher information matrix threshold to 10 after cross-validation. -4 Freeze parameters;

[0040] The update period of the front-end adaptive Kalman filter is 10ms, and the update period of the back-end correction network parameters is 100ms, establishing a 1:10 cascade update timing chain.

[0041] Preferably, the step 6 comprises:

[0042] Step 6.1: Output the time series weighing results in real time, and use the sliding window mechanism to smooth the results;

[0043] Step 6.2, analyze the time series based on STL decomposition;

[0044] Step 6.3: Based on the analysis results, combined with the dynamic uncertainty assessment of the vehicle's motion state, an ARIMA-Prophet combined prediction model is constructed to predict the weight change trend and provide early warning;

[0045] In step 6.1, an exponentially weighted sliding with a window length of 128 is used to smooth the results.

[0046] In step 6.3, the Monte Carlo dropout method was used to evaluate uncertainty, with a dropout rate of 0.2 and 100 samplings to calculate the confidence interval.

[0047] Step 6.3 builds an ARIMA-Prophet combined forecasting model, which includes: the ARIMA model is responsible for short-term trend forecasting (i.e., <5s), and the Prophet model is responsible for extracting periodic components. The output is fused through weighted averaging, and the weight coefficient β satisfies:

[0048]

[0049] Among them, t p is the prediction time, τ=5s is the ARIMA effective window, when t p >τ, the Prophet output is fully adopted;

[0050] Step 6.3: When the preset conditions are met simultaneously, the three-level warning mechanism will be triggered:

[0051] (1) Actual load > rated load × 90%;

[0052] (2) predicted weight change rate > 5% / s;

[0053] (3) Monte Carlo dropout evaluation confidence interval P ≥ 95%;

[0054] The three-level warning mechanism is: local alarm prompt, overload status upload, and on-board ECU speed limit control.

[0055] The present invention also provides a large freight vehicle weighing system based on a time series large model, the system is used to implement the above method, the system includes: a data acquisition module, a preprocessing module, a model calculation module, and a correction output module;

[0056] The data acquisition module is used to collect dynamic load data of large freight vehicles, so that the data has temporal and spatial characteristics and is stored;

[0057] The preprocessing module is used to process data with spatiotemporal characteristics and construct a multimodal training dataset;

[0058] The model calculation module is used to perform real-time dynamic weighing calculations on the collected load data with time series characteristics by applying the time series large model built based on the Transformer architecture;

[0059] The correction output module is used to perform error correction on the weighing calculation results of the model, output the weighing results in real time, and perform weight change trend prediction and early warning.

[0060] Preferably, the data acquisition module includes: an acquisition unit, a fusion unit and a storage unit;

[0061] The acquisition unit is used for collecting dynamic load data information at various positions of large freight vehicles by using a distributed pressure sensor array;

[0062] The fusion unit is used to fuse multi-source data by using the timestamp alignment algorithm to generate time series load data with spatiotemporal characteristics;

[0063] The storage unit is used to store the time series payload data in the dual-channel ring buffer memory of the edge computing node, and realize non-blocking data transmission with the preprocessing module through the DMA mechanism.

[0064] Preferably, the pre-processing module includes: a denoising unit, a transformation unit and a detection and reconstruction unit;

[0065] The denoising unit is used to remove the influence of noise on the original time series data by using the improved SURE threshold based on wavelet packet decomposition;

[0066] The transformation unit is used to convert one-dimensional time series data into a two-dimensional feature map through GAF;

[0067] The detection and reconstruction unit is used to detect abnormal time series data and reconstruct waveforms.

[0068] Preferably, the correction output module includes: a filtering unit, a correction unit, a connection unit, an output unit, a prediction unit and an early warning unit;

[0069] The filtering unit is used to perform dynamic error compensation on the weighing calculation results of the model using adaptive Kalman filtering;

[0070] The correction unit is used to apply the correction network to the weighing calculation result of the model to perform error correction;

[0071] The connection unit is used to superimpose the error correction amount and the model output through the residual connection design to achieve error correction of the model output result;

[0072] The output unit is used to output time series weighing results in real time;

[0073] The prediction unit is used to predict the weight change trend by constructing an ARIMA-Prophet combined prediction model;

[0074] The early warning unit is used to trigger a three-level early warning mechanism including local alarm prompts, overload status upload and on-board ECU speed limit control.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] The present invention correlates multi-sensor data with spatiotemporal characteristics in the time dimension and gives the model prediction capability based on time series analysis, which helps the dynamic weighing system become more intelligent and provides users with a scientific and efficient technical means based on model prediction.

[0077] The present invention can effectively solve the following problems: the problem of spatiotemporal feature distortion caused by spatiotemporal drift of multi-source sensors; the problem of poor model generalization ability under unstable working conditions; the contradiction between hardware resources and real-time computing;

[0078] The time series large model constructed by the present invention can continuously accept new weight data and update the weighing results in real time, and the application of the error correction hybrid system effectively reduces the weighing error and improves the reliability of the results;

[0079] The present invention establishes a complete vehicle-mounted weighing system based on a large time series model, providing important technical support and significant practical application value for industries such as weighing large freight vehicles, logistics warehousing, and transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0081] Figure 1 This is a flow chart of a method for onboard weighing of large freight vehicles based on a time series large model according to an embodiment of the present invention;

[0082] Figure 2 Schematic diagram of the actual weighing process of a method for onboard weighing of large freight vehicles based on a time series large model according to an embodiment of the present invention;

[0083] Figure 3 The present invention is a flowchart of the actual weighing process of a method for weighing large freight vehicles based on a time series large model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0084] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0085] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0086] Example 1

[0087] like Figure 1 As shown, the present invention provides a method for weighing large freight vehicles based on a time series large model, comprising the following steps:

[0088] Step 1: Use the timestamp alignment algorithm to fuse the multi-source data collected by multiple sensors to generate and store time series load data with spatiotemporal characteristics;

[0089] Step 2: Preprocess the original time series data and construct a multimodal training dataset;

[0090] Step 3: Use the dataset for pre-training and fine-tuning to build a large time series model based on the Transformer architecture;

[0091] Step 4: Input the latest collected load data with time series characteristics into the time series model to perform real-time dynamic weighing calculation;

[0092] Step 5: Construct an error correction hybrid system to accurately correct the weighing results;

[0093] Step 6: Output accurate weighing results with time series characteristics in real time and make reliability predictions on load change trends.

[0094] In this embodiment, step 1 specifically includes the following steps:

[0095] Step 1.1, collecting dynamic load information at each location through a distributed pressure sensor array;

[0096] Step 1.2: Use the nanosecond timestamp alignment algorithm based on the FPGA hardware time synchronization trigger circuit to fuse multi-source data and generate time series load data with spatiotemporal characteristics.

[0097] In step 1.3, the time series load data is stored in the dual-channel ring buffer memory of the edge computing node, and non-blocking data transmission with the preprocessing module is achieved through the DMA (Direct Memory Access) mechanism.

[0098] As a preferred technical solution, the various positions in step 1.1 refer to the various support point areas of the suspension system of the large freight vehicle and non-support areas such as the bottom of the vehicle plate.

[0099] As a preferred technical solution, the sensor array in step 1.1 refers to at least 2M pressure sensors covering the support point area of ​​the suspension system of large freight vehicles, with at least 2 sensors per support point and 0.5 sensors / m2 in the non-support area. 2 The density distribution form is shown in Figure 2. Furthermore, the distance between adjacent sensors (d) and the maximum distance between adjacent axles (L) must satisfy the constraint d≤0.3L. Here, M is the number of suspension support points on the vehicle (for a six-axle vehicle, M=12); L is the maximum horizontal distance between adjacent axle centers.

[0100] As a preferred technical solution, the technical basis for FPGA selection in step 1.2 is: FPGA clock frequency is configured in the range of 500MHz-1GHz, FPGA clock frequency f FPGA ≥2f to meet the nanosecond synchronization requirements. The timestamp alignment error δ must meet the following formula requirements

[0101]

[0102] Where f is the sensor sampling frequency and f≥1kHz.

[0103] As a preferred technical solution, the timestamp alignment algorithm in step 1.2 is a technology that uses hardware-level synchronization to eliminate timing deviations caused by clock differences, transmission delays, etc. during multi-sensor data acquisition. Its core is to establish a unified global time base to synchronize the sampling times of all sensors within nanosecond accuracy. It includes the following steps:

[0104] Generate hardware-level time base through FPGA clock tree design and synchronous trigger pulse;

[0105] Compensate for sensor signal transmission delays and implement dynamic drift correction for each sensor clock deviation to complete timestamp compensation calculations;

[0106] Cubic spline interpolation is used for non-uniformly sampled data to unify the sensor data into a standard time grid to achieve data resampling alignment.

[0107] As a preferred technical solution, the sensor signal transmission delay compensation is designed to eliminate the transmission delay caused by hardware differences such as PCB routing and interface chips. By using FPGA synchronous triggering to send standard pulses to all sensors, the time difference T between the pulse sending and the return is recorded. i , to calculate the one-way transmission delay Δt i , the calculation formula is:

[0108]

[0109] where t fpga This is the internal processing delay of the FPGA and can be determined by hardware logic timing calibration.

[0110] Create a sensor delay parameter table and set Δt i Stored in the FPGA on-chip register, it is updated every 100ms through a calibration pulse to adapt to delay changes caused by environmental factors such as temperature and vibration. oi To perform delay compensation, the calculation formula is:

[0111] t oi =t ni -Δt i

[0112] where t ni The raw timestamp generated by the FPGA using the global clock when acquiring data from multiple sensors.

[0113] Finally, all sensor data are processed in parallel to ensure that the delay compensation is completed within two FPGA clock cycles to realize the sensor signal transmission delay compensation calculation.

[0114] As a preferred technical solution, the one-way transmission delay Δt i The calculation formula can still be expressed as:

[0115]

[0116] Among them L i is the PCB trace length from the sensor to the FPGA, v p is the temperature-related signal propagation speed, t ic is the inherent delay of the interface chip, t pcb is the additional delay for vias and connectors.

[0117] As a preferred technical solution, compared with the sensor signal transmission delay compensation method, most existing solutions use a fixed delay compensation value, ignoring the differences in PCB trace lengths, interface chip delay fluctuations, and changes in signal propagation speed caused by temperature, which will lead to larger actual compensation errors.

[0118] As a preferred technical solution, the dynamic drift correction of the clock deviation of each sensor is used to solve the clock asynchrony problem caused by crystal oscillator frequency deviation and temperature drift of multiple sensors. By calculating the clock frequency deviation, constructing the drift function and performing timestamp correction calculation, the dynamic drift correction of the clock deviation of each sensor can be completed. Using the FPGA global clock as a reference, the clock signal of each sensor is phase detected and the clock frequency deviation Δf is calculated. i , whose expression is:

[0119] Δf i =f FPGA -f i

[0120] where f i is the sensor's local clock frequency.

[0121] A first-order linear drift function is established to describe the sensor clock offset Δφ i (t), its expression is:

[0122] Δφ i (t) = Δφ i (0)+Δf i ·t

[0123] where Δφ i (0) is the initial phase deviation, which can be initialized by a synchronous trigger pulse.

[0124] Perform clock deviation correction on the actual timestamp of multi-sensor collected data. The calculation formula is:

[0125] t′ oi =t FPGA -Δφ i (t i )

[0126] where t FPGA FPGA global timestamp.

[0127] As a preferred technical solution, the use of cubic spline interpolation to achieve data resampling alignment for non-uniformly sampled data aims to unify non-uniformly sampled data caused by sampling frequency fluctuations and inconsistent transmission delays onto a global standard time grid, providing a unified format input for subsequent large-scale time series model calculations. For non-uniformly sampled data, mirror extension and exponential decay are used to suppress boundary effects. By defining a standard time grid and constructing a cubic spline interpolation function, the corresponding value is calculated using the interpolation function at each standard time point to generate a uniformly sampled sequence, ensuring that all sensor data is aligned to the standard time grid T.

[0128] According to the system's highest sampling frequency f maxTo define the global standard time grid T = {t0, t1, ..., t n}, and the time interval For the non-uniform sampling data of each sensor, a cubic spline interpolation function is constructed, and its expression is:

[0129] S i (t) = a i +b i (tt i )+c i (tt i ) 2 +d i (tt i ) 3 t∈[t i ,t i+1 ]

[0130] The following conditions must be met to construct the interpolation function:

[0131]

[0132] The interpolation function coefficient a can be solved by the natural spline boundary condition, that is, the second-order derivatives at the beginning and end are equal to 0 i 、b i 、c i d i .

[0133] As a preferred technical solution, the multi-source data fusion in step 1.2 involves generating unified time-series load data with spatiotemporal characteristics from the spatiotemporal data collected by the distributed pressure sensor array through temporal encoding, spatial encoding, cross-modal fusion, and physical constraint optimization. Multi-source data fusion is achieved by separately temporally and spatially encoding the sensor information, then performing cross-modal attention fusion, and finally optimizing the decision-making process based on physical constraints.

[0134] As a preferred technical solution, the sensor information is temporally and spatially encoded, wherein the time encoding is based on the process of using the timestamp alignment algorithm to establish a unified global time base so that the sampling moments of all sensors are synchronized, and a one-dimensional time series feature vector X containing the timestamp is generated. t,w =[t0,t1,...,t n ;w0,w1,...,w n ], where t i is the standardized time, w i is the corresponding load data.

[0135] Spatial coding collects the direct load of the suspension system in the support area and the indirect transmission signal of the load distribution in the non-support area, and assigns a three-dimensional coordinate (x c ,y c ,z c ), where x c 、y c Corresponding to the bottom plane position of the vehicle plate, z c The vehicle panel is divided into a two-dimensional grid corresponding to the sensor layout. Each grid node is associated with the adjacent sensor. The spatial adjacency matrix A is constructed to represent the spatial correlation of the sensors. Finally, the support point or non-support point area attribute label is added to each sensor to generate a unique hot encoding Region. c ∈{0,1}, as the prior knowledge of spatial features.

[0136] As a preferred technical solution, the coordinate system is established with the origin set at the geometric center of the vehicle, the x-axis along the vehicle length, the y-axis along the vehicle width, and the z-axis vertically upward.

[0137] As a preferred technical solution, the cross-modal fusion aims to fuse time series features and spatial position features through the attention mechanism to capture the dependencies between different sensors in the time and space dimensions. t,w The position coordinates (x c ,y c ), Region tag Region c Splice to form the spatiotemporal feature vector X of a single sensor t,w,c =[X t,w ;x c ;y c ;Region c ]. Sliding window causal attention is used to capture the temporal dependency of single sensor load data. The adjacency matrix A is used to calculate the attention weights of adjacent sensors to determine the load transfer relationship between the support point and the non-support area. Multi-head attention heads are designed to process temporal and spatial features respectively. Finally, a linear transformation matrix is ​​designed to map the spatiotemporal feature vectors of multiple sensors to a unified dimension and output the fused feature vector X. f , to achieve multimodal feature fusion.

[0138] As a preferred technical solution, the physical constraint-based decision optimization aims to ensure that the multimodal feature fusion results comply with the conservation law of vehicle load and eliminate abnormal data. By defining physical constraints and constructing a constrained optimization objective function, the Lagrange multiplier method is used to solve the constrained optimization problem. The objective function is optimized using the gradient descent method, and the multimodal feature fusion parameters are adjusted to ensure that the results meet the physical constraints. Residual analysis is performed on the fused data. If the difference between the fused data of a single sensor and the original signal exceeds 3σ (σ is the standard deviation of the residual of the training set), the hardware layer is triggered to remeasure the transmission delay and clock offset of the sensor, update the timestamp compensation parameters, and perform cubic spline interpolation repair on the abnormal points. The abnormal data is corrected using the data of neighboring sensors.

[0139] As an optimal technical solution, the physical constraints are defined as follows: ensuring that the total load collected by all sensors is equal to the actual load of the vehicle, avoiding abnormal data that violates the law of conservation of mass, and defining load conservation constraints; ensuring that the load distribution of each support point of the vehicle suspension system conforms to the principle of mechanical equilibrium, avoiding unreasonable load mutations at adjacent support points, and defining support point balance constraints; the load distribution in the non-support area should show a gradual trend with the distance from the support point, avoiding sudden changes without physical basis, and defining non-support area gradual constraints.

[0140] As a preferred technical solution, the constrained optimization objective function is to convert the inequality constraint into a penalty term through the Lagrange multiplier method, and construct a comprehensive objective function including data fitting term, load conservation term, equilibrium constraint term, and non-equilibrium constraint term, and its expression is:

[0141] η=η r +λ1η c +λ2η b +λ3η s

[0142] where η r is the data fitting term, η c is the load conservation term, η b is the equilibrium constraint term, η s is the non-equilibrium constraint term, λ1, λ2, λ3 are function coefficients, and λ1>λ2>λ3.

[0143] As a preferred technical solution, the dual-channel ring buffer memory in step 1.3 adopts a ping-pong cache architecture, alternately performs data writing and reading operations by setting a double buffer, and cooperates with a DMA chain transmission mechanism to ensure data continuity.

[0144] As a preferred technical solution, the dual-channel ring buffer memory capacity is configured as 2×N×B×T, where N is the total number of sensors, B is the data bandwidth of a single sensor, and T is the maximum unblocked transmission time (T≥100ms). For example, when the total number of sensors N=28, the data bandwidth of a single sensor B=100KB / s, and the maximum unblocked transmission time T=100ms, the cache capacity is 2×28×100KB / s×0.1s=560KB. In this embodiment, the data preprocessing process in step 2 specifically includes improved SURE (Stein Unbiased Risk Estimation) threshold denoising based on wavelet packet decomposition, conversion of one-dimensional time series data into two-dimensional feature maps through Gramian Angular Field (GAF), and time series anomaly detection and waveform reconstruction.

[0145] As a preferred technical solution, the improved SURE threshold denoising method based on wavelet packet decomposition introduces an adaptive adjustment factor and constructs an asymmetric nonlinear threshold function, allowing different subbands to use independent thresholds for denoising. While minimizing the SURE risk, it dynamically balances noise suppression with signal detail preservation, especially adapting to the different characteristics of high-energy signals in support areas and low-energy signals in non-support areas.

[0146] As a preferred technical solution, compared with the improved SURE threshold denoising method based on wavelet packet decomposition, the traditional SURE denoising is based on Stein's unbiased risk estimation and uniformly shrinks the wavelet coefficients through a global threshold. There are two shortcomings: first, the same threshold calculation method is used for all sub-bands, and the differences in signal strength and noise characteristics between different sub-bands are not considered, which affects the denoising effect; second, the traditional soft threshold will cause a constant deviation in the denoised coefficients, while the hard threshold will introduce artifacts due to discontinuity.

[0147] As a preferred technical solution, the improved SURE threshold denoising based on wavelet packet decomposition specifically refers to converting the original time series signal (a sequence of load measurement values ​​of a single sensor at continuous time points) w t Perform j-layer wavelet packet decomposition to generate 2 j sub-bands, each sub-band corresponds to a different frequency range. Get the wavelet packet coefficients b of each sub-band l,k , l is the number of subband layers, k is the subband index under the current layer, and the median b′ of the subband coefficient with stronger noise resistance is adopted l,k To estimate the noise standard deviation ψ. For each sub-band, the sub-band energy adjustment factor λ is introduced l , allowing different sub-bands to use independent thresholds T l , construct an improved SURE risk function, whose expression is:

[0148]

[0149] Solve T for each subband independently l =argminS(T l ,λ l ), find the optimal threshold through one-dimensional search, replacing the single calculation of the traditional global threshold.

[0150] As a preferred technical solution, the adjustment factor λ l By sub-band energy E l Dynamically determined, the expressions are:

[0151]

[0152] Where min(E) and max(E) are the minimum and maximum subband energies in all subbands, respectively.

[0153] As a preferred technical solution, the improved threshold function is divided into three regions, which solves the problem of constant deviation of traditional soft threshold and discontinuity of hard threshold. Its expression is:

[0154]

[0155] As a preferred technical solution, the Gram angular field transform (GAT) is a method for converting one-dimensional time series data into a two-dimensional image. By mapping the sequence's values ​​to angular relationships using trigonometric identities, this method preserves the dynamic characteristics of the time series data and facilitates subsequent processing of large time series models. By normalizing the data and mapping the angles and constructing a Gram angular field matrix, a two-dimensional feature map can be generated, completing the Gram angular field transform.

[0156] As a preferred technical solution, the normalized data is the original time series signal w t Linear normalization is performed to the interval [-1,1] to obtain Its expression is:

[0157]

[0158] Where min(w) and max(w) represent the minimum and maximum values ​​of the one-dimensional time series data of a single sensor, respectively.

[0159] As a preferred technical solution, the angle mapping is to use the inverse cosine function to transform the normalized data Mapped to angle θ t , whose expression is:

[0160]

[0161] As an optimal technical solution, the construction of the Gram angle field matrix is ​​to construct N×N order Gram angle and matrix G respectively.s and the Gram angle difference matrix G d . Where the matrix G s It is used to reflect the trend characteristics of time series data. Its elements are defined as the cosine value of the sum of the angles of the time series points. The matrix G d It is used to capture the periodic characteristics of time series data. Its elements are defined as the sine value of the angle difference of the time series points. Its expressions are:

[0162] G s (i,j)=cos(θ i +θ j )

[0163] G d (i,j)=sin(θ i -θ j )

[0164] As a preferred technical solution, generating a two-dimensional feature map is to convert G s and G d As two independent channels to form a N×N×2 dual-channel feature map, or take G s With G d The mean of the matrix and generates a single-channel feature map. Finally, the matrix value is normalized to [0, 255] and the feature map is converted into a two-dimensional grayscale image or pseudo-color image to facilitate the subsequent large time series model to extract spatial features.

[0165] The two-dimensional feature map generated in this paper is a dual-channel feature map, which can reflect the trend and periodicity of load changes respectively.

[0166] As a preferred technical solution, the time series anomaly detection and waveform reconstruction method aims to detect abnormal data in the denoised signal, ensure signal integrity through dynamic thresholding and interpolation repair, and provide reliable time series data for subsequent model training and real-time weighing. Reconstructing the time domain waveform of the original signal using the processed data includes the following steps:

[0167] The denoised wavelet packet coefficients are transformed through inverse wavelet packet transform to reconstruct the signal, and the residual sequence between the reconstructed signal and the original signal is calculated;

[0168] Set the dynamic reconstruction error threshold and form an abnormal area by locating abnormal points;

[0169] Anomalies are repaired through cubic spline interpolation, and the repaired signal is secondary smoothed using a sliding window to suppress residual high-frequency noise.

[0170] As a preferred technical solution, the denoised wavelet packet coefficients are transformed through inverse wavelet packet to reconstruct the signal, and the residual sequence between the reconstructed signal and the original signal is calculated, mainly by using the denoised wavelet packet coefficients b″ l,kPerform inverse transform to reconstruct the time domain signal Its expression is:

[0171]

[0172] In order to quantify the information loss or abnormal disturbance in the denoising process, the residual r between the reconstructed signal and the original signal is calculated, which is expressed as:

[0173]

[0174] As a preferred technical solution, the dynamic reconstruction error threshold is set to form an abnormal area by locating the abnormal point, and the dynamic reconstruction error threshold ε is set to:

[0175] ε=μ+3σ

[0176] Among them, μ is the mean reconstruction error of the training set, and σ is the standard deviation.

[0177] If | r t If |>ε, the point is considered an outlier. By traversing the residual sequence, all points exceeding the threshold and their five adjacent points before and after are marked to form an outlier region.

[0178] As a preferred technical solution, the anomaly repair is performed by cubic spline interpolation, and the repaired signal is smoothed twice using a sliding window to suppress residual high-frequency noise. Specifically, for the point in the abnormal area, a cubic spline function S is constructed using at least 5 normal points before and after. i (t) is used to repair the abnormal points, and the exponentially weighted moving average with a window length of 16 is used to perform secondary smoothing on the repaired signal to suppress the residual high-frequency noise.

[0179] In this embodiment, the construction of the multimodal training dataset in step 2 is achieved by integrating the load distribution feature map and the vehicle motion state vector through feature-level fusion, and using the attention mechanism to achieve cross-modal feature interaction.

[0180] As an optimal technical solution, the feature-level fusion is to fuse the spatiotemporal characteristics of the load distribution collected by the sensor (a two-dimensional feature map generated by GAF transformation, which implies the trend and periodicity of load changes) with the dynamic characteristics of the vehicle's motion state (one-dimensional vectors such as acceleration and angular velocity) to form a multi-modal system containing the "load distribution-motion state" coupling relationship, solving the problem that a single mode is difficult to distinguish complex working conditions.

[0181] The load distribution feature map encodes the spatial layout of the sensor support points and non-support areas, as well as the time series pattern, while the motion state vector encodes the vehicle dynamics parameters. The fusion of the two can form a three-dimensional feature tensor of "space-time-dynamics", which is suitable for the spatiotemporal attention mechanism of the Transformer architecture and can also improve the time series model's understanding of complex working conditions.

[0182] As an optimal technical solution, the process of constructing a multimodal training data set is, specifically, to generate a dual-channel two-dimensional feature map by performing Gram angular sum field (GAS) and angular difference field (GAD) transformation on the one-dimensional load time series data of each sensor, and to add regional attribute encoding to each pixel (the support point area is 1, the non-support area is 0) to form a load distribution feature map (GAS+GAD+spatial label); obtain acceleration, angular velocity and other information through the on-board IMU (Inertial Measurement Unit) to construct the vehicle motion state vector, and use the FPGA hardware synchronization mechanism described above to strictly align the timestamps of the motion state data and the load data to eliminate the interference of asynchronous noise on cross-modal association; perform global average pooling on the two-dimensional load feature map to compress it into a one-dimensional load feature vector, linearly project the load feature vector to 64 dimensions, and map the motion state vector to 64 dimensions through a fully connected layer, and then splice the load vector with unified dimensions and the motion state vector to form the initial multimodal feature F r ; Use the load feature as the query vector Q and the motion state as the key-value pair (K, V) to design the attention module and calculate the attention weight α t ; Add the attention output to the original splicing feature through the residual connection, and generate the fusion feature F after layer normalization f ; Use the static weighing calibration value as label w t , combined with the multimodal fusion feature F after timestamp alignment f Forming training samples (F f ,w t ) to cover typical working conditions such as no-load, full-load, acceleration, and turning; data enhancement is performed by adding random masks to the load feature map or adding Gaussian noise to the motion vector; and the data is divided into training set, validation set, and test set in an 8:1:1 ratio to ensure balanced samples of each working condition.

[0183] As a preferred technical solution, the attention weight α t The calculation formula is:

[0184]

[0185] where d k The key vector dimension is used to capture the influence of different motion parameters on load characteristics through 8-head parallel calculation.

[0186] As an optimal technical solution, the cross-modal interaction enables the model to capture the causal relationship between vehicle motion and load distribution changes, improve the model's ability to identify actual load changes during dynamic weighing, and suppress interference such as vibration and noise.

[0187] In this embodiment, step 3 of constructing a large time series model based on the Transformer architecture includes: a spatiotemporal convolutional coding layer, a multi-head causal attention mechanism, a dynamic weight allocation module, and model distillation.

[0188] As a preferred technical solution, the spatiotemporal convolutional coding layer uses a two-dimensional dilated causal convolution kernel to extract local spatiotemporal features, and its dilation coefficient d is d=2 n-1 (n is the number of layers) grows exponentially, and the convolution kernel size satisfies the formula k=2d+1.

[0189] As a preferred technical solution, the multi-head causal attention mechanism uses a sliding window attention with a sliding window of 64 and an attention head number of 8 to limit the computational complexity;

[0190] As a preferred technical solution, the dynamic weight allocation module uses a learnable gating network (including two fully connected layers + Softmax activation) to achieve the importance evaluation of sensor channels;

[0191] As an optimal technical solution, the model distillation adopts a teacher-student architecture, where the teacher model is a 12-layer Transformer with a hidden layer dimension of 512 and 8 attention heads; the student model adopts a 4-layer structured pruned network with a hidden layer dimension of 256. The top 50% important attention heads are retained in the Transformer layer (the number of attention heads is 4), and knowledge transfer is achieved through KL divergence loss. The pruning standard is the L1 norm of the attention head weights.

[0192] In this embodiment, step 4 inputs the newly collected load data with time series characteristics into the time series large model for real-time dynamic weighing calculation. Specifically, based on the Transformer architecture-based time series model constructed in step 3, multiple sensors are used to collect real-time vehicle load data. The multi-source data collected by the multiple sensors are fused through the timestamp alignment algorithm in step 1 to generate time series load data with spatiotemporal characteristics. The time series data is then preprocessed in step 2. Finally, the processed real-time load vehicle data is input into the time series large model, and the model is used to perform weighing calculations and output the results synchronously. Since the constructed time series large model has a fast response speed, reaching the microsecond level, the entire process from multi-sensor data collection to the model output of the weighing results takes a very short time, achieving the real-time effect of dynamic weighing calculation.

[0193] In this embodiment, the step 5 constructs an error correction hybrid system, in which the front end uses an adaptive Kalman filter to perform dynamic error compensation and sets the process noise covariance matrix Q t A dynamic update mechanism is proposed; the back-end application correction network is used to update the network parameters using an online transfer learning strategy; the error correction amount is superimposed on the model output through the residual connection design to achieve error correction of the model output results.

[0194] As a preferred technical solution, the process noise covariance matrix Q t The dynamic update mechanism is:

[0195] Q t =αQ t-1 +(1-α)·diag(ξ t ·ξ t T );

[0196] Where α=0.95 is the smoothing factor, ξ t is the residual vector at time t.

[0197] As a preferred technical solution, the adaptive Kalman filter state vector is x=[w,w t ,a] T , where w is the load, w t is the load change rate, and a is the vehicle acceleration.

[0198] As an optimal technical solution, the back-end correction network adopts a three-layer gated recurrent unit structure, the input is the Kalman residual, the input dimension is consistent with the Kalman residual dimension, the number of hidden layer units does not exceed 32, and the time step is 10.

[0199] As a preferred technical solution, the online transfer learning adopts the elastic weight solidification algorithm, and sets the Fisher information matrix threshold to 10 after cross-validation. -4 Freeze the parameters.

[0200] As a preferred technical solution, the update period of the front-end adaptive Kalman filter is 10ms, and the update period of the back-end correction network parameters is 100ms, establishing a 1:10 cascade update timing chain.

[0201] In this embodiment, step 6 specifically includes the following steps:

[0202] Step 6.1: Output the time series weighing results in real time and use the sliding window mechanism to smooth the results. The two are executed synchronously:

[0203] Step 6.2, time series analysis based on STL decomposition;

[0204] In step 6.3, the dynamic uncertainty assessment of the vehicle's motion state is combined with the construction of an ARIMA-Prophet combined prediction model to predict and warn the weight change trend.

[0205] As a preferred technical solution, step 6.1 uses exponential weighted sliding with a window length of 128 to smooth the results.

[0206] As a preferred technical solution, the time series analysis based on STL decomposition in step 6.2 is a method of performing multiple iterative smoothing of the time series through local weighted regression to gradually separate the trend component, seasonal component and residual component. Its expression is:

[0207] Y t =T t +S t +R t ;

[0208] Where Y t Original sequence, T t is the trend term, S t is the seasonal term, R t is the residual term.

[0209] As a preferred technical solution, the uncertainty assessment in step 6.3 adopts the Monte Carlo dropout method, sets the dropout rate to 0.2, and performs 100 samplings to calculate the confidence interval.

[0210] As a preferred technical solution, step 6.3 constructs an ARIMA-Prophet combined forecasting model, in which the ARIMA model is responsible for short-term (<5s) trend forecasting, and the Prophet model is responsible for periodic component extraction. The outputs of the two are fused by weighted averaging, and the weight coefficient β satisfies the following formula:

[0211]

[0212] where t p is the prediction time, τ=5s is the ARIMA effective window, when t p When >τ, Prophet output is fully adopted.

[0213] As a preferred technical solution, step 6.3 will trigger the three-level warning mechanism when the following conditions are met at the same time:

[0214] (1) Actual load > rated load × 90%;

[0215] (2) predicted weight change rate > 5% / s;

[0216] (3) Monte Carlo dropout evaluation confidence interval P ≥ 95%.

[0217] As a preferred technical solution, the three-level warning mechanism is respectively: local alarm prompt, overload status upload, and vehicle-mounted ECU speed limit control.

[0218] The present invention can output the weighing results of large freight vehicles in real time, keenly capture the characteristics of vehicle load changes, and make reliable predictions of the load change trends of large freight vehicles, so as to facilitate early warning of overloading risks.

[0219] Example 2

[0220] This embodiment takes the actual weighing of a six-axle heavy freight vehicle on a highway as an example. Figure 2 This is a schematic diagram of the actual weighing process of a large-scale freight vehicle weighing method based on a time series large model, including a data collector with a data acquisition module, a six-axle large-scale freight vehicle, a highway, an on-board server including a data processing module and an edge computing module, an on-board central controller including a control module, an on-board display, a cloud server, etc. Figure 3 The figure is a flow chart of the actual weighing process of a large freight vehicle weighing method based on a time series large model.

[0221] Among them, the data acquisition module of the data collector is used to collect the weighing data measured by various sensors distributed on the six-axle large freight vehicle; the data processing module of the on-board server is mainly used to perform a series of processing on the collected weighing data, so that the edge computing module can use it directly later; the edge computing module of the server mainly uses the time series analysis large model to independently analyze and calculate the processed weighing data; the control module of the on-board central controller is mainly used to coordinate and control the operations of other modules, realize data exchange between models, and monitor the operating status of the entire system in real time, and make timely adjustments as needed to ensure the normal operation of the system; the on-board display is mainly used to display the weighing calculation results and real-time prediction of weight change trends; the cloud server is mainly used when the vehicle triggers the third-level early warning mechanism, and the vehicle overload information will be automatically uploaded to the server, recording the basic information and overload situation of the overloaded vehicle to facilitate subsequent management of the supervision platform.

[0222] During the actual weighing measurement process, when a six-axle heavy-duty freight truck with a rated load of 40t travels on a highway at a speed range of 0-80km / h, the data collector uses 30 pressure sensors to collect vehicle load and other related data. Under the control module of the central controller, the collected weighing data is input into the server. After being processed by the server data processing module and calculated by the edge computing module, the weighing calculation results and the predicted results of the weight change trend are transmitted back to the display. When all of the following conditions are met at the same time, the three-level warning mechanism of local alarm prompt, overload information upload, and on-board ECU speed limit control will be triggered: (1) the real-time load exceeds 36t; (2) the predicted weight change rate exceeds 5% / s; (3) the Monte Carlo dropout evaluation confidence interval is not less than 95%. If the vehicle triggers the three-level warning mechanism, the vehicle's overload information will be automatically uploaded to the cloud server for subsequent management by the supervision platform.

[0223] After actual measurement, compared with the traditional method, the method of this embodiment has achieved the following effects:

[0224]

[0225] Example 3

[0226] The present invention also provides a large freight vehicle weighing system based on a time series large model, comprising: a data acquisition module, a preprocessing module, a model calculation module, and a correction output module;

[0227] The data acquisition module is used to collect dynamic load data of large freight vehicles, make the data have temporal and spatial characteristics and store them;

[0228] The preprocessing module is used to process data with spatiotemporal features and construct a multimodal training dataset for subsequent model pre-training;

[0229] The model calculation module is used to perform real-time dynamic weighing calculations on the latest collected load data with time series characteristics by applying the constructed Transformer time series model;

[0230] The correction output module is used to correct errors in the model's weighing calculation results, output accurate weighing results in real time, and predict and warn weight change trends.

[0231] In this embodiment, the data acquisition module includes: an acquisition unit, a fusion unit and a storage unit;

[0232] The acquisition unit is used for collecting dynamic load data information at various positions of large freight vehicles by using a distributed pressure sensor array;

[0233] The fusion unit is used to fuse multi-source data by using the timestamp alignment algorithm to generate time series load data with spatiotemporal characteristics;

[0234] The storage unit is used to store the time series payload data in the dual-channel ring buffer memory of the edge computing node, and realize non-blocking data transmission with the preprocessing module through the DMA mechanism.

[0235] The positions in the collection unit refer to the supporting point areas of the suspension system of the large freight vehicle and non-supporting areas such as the bottom of the vehicle plate.

[0236] The sensor array in the acquisition unit refers to the distribution of multiple sensors covering the support point area and non-support area of ​​the suspension system of large freight vehicles. Each suspension system supports no less than 2 sensors, and the non-support area has a density of 0.5 sensors per square meter.

[0237] The multi-source data fusion in the fusion unit is achieved by temporally and spatially encoding the sensor information respectively, then performing cross-modal attention fusion, and finally optimizing the decision based on physical constraints to achieve the fusion of multi-source data.

[0238] In this embodiment, the pre-processing module includes: a denoising unit, a transformation unit and a detection and reconstruction unit;

[0239] The denoising unit is used to remove the influence of noise on the original time series data by using the improved SURE threshold based on wavelet packet decomposition;

[0240] The transformation unit is used to convert one-dimensional time series data into a two-dimensional feature map through GAF;

[0241] The detection and reconstruction unit is used to detect abnormal time series data and reconstruct waveforms.

[0242] In this embodiment, the model calculation module is used to perform real-time weighing calculations on the latest collected load data with time series characteristics by applying the constructed Transformer time series model.

[0243] In this embodiment, the correction output module includes: a filtering unit, a correction unit, a connection unit, an output unit, a prediction unit and an early warning unit;

[0244] The filtering unit is used to perform dynamic error compensation on the weighing calculation results of the model using adaptive Kalman filtering;

[0245] The correction unit is used to apply the correction network to the weighing calculation result of the model to perform error correction;

[0246] The connection unit is used to superimpose the error correction amount and the model output through the residual connection design to achieve error correction of the model output result.

[0247] The output unit is used to output time series weighing results in real time;

[0248] The prediction unit is used to predict the weight change trend by constructing an ARIMA-Prophet combined prediction model;

[0249] The early warning unit is used to trigger a three-level early warning mechanism including local alarm prompts, overload status upload and on-board ECU speed limit control.

[0250] The warning unit must meet all of the following conditions:

[0251] (1) Actual load > rated load × 90%;

[0252] (2) predicted weight change rate > 5% / s;

[0253] (3) Monte Carlo dropout evaluation confidence interval ≥ 95%.

[0254] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for weighing large freight vehicles based on a time series large model, characterized in that: The method comprises: Step 1: Use a timestamp alignment algorithm to fuse the multi-source data collected by multiple sensors to generate and store time series coded data with spatiotemporal characteristics. Multiple sensors are installed at various locations on large freight vehicles. Step 2: Preprocess the time series encoding data with spatiotemporal characteristics to construct a multimodal training dataset; Step 3: Use the multimodal training dataset for pre-training and fine-tuning to build a large time series model based on the Transformer architecture; Step 4: Input the collected load data with time series characteristics into the time series model based on the Transformer architecture to perform real-time dynamic weighing calculations; Step 5: Construct an error correction hybrid system to correct the weighing results; Step 6: Output the weighing results with time series characteristics in real time and make predictions on the load change trend.

2. The method according to claim 1, characterized in that The step 1 comprises: Step 1.1, collecting dynamic load information at each location through a distributed pressure sensor array; Step 1.2: Use the nanosecond timestamp alignment algorithm based on the FPGA hardware time synchronization trigger circuit to fuse multi-source data and generate time series load data with spatiotemporal characteristics. Step 1.3: Store the time series payload data in the dual-channel ring buffer memory of the edge computing node, and use the DMA mechanism to achieve non-blocking data transmission with the pre-processing module; The sensor array in step 1.1 refers to at least 2M pressure sensors covering the support point area of ​​the suspension system of the large freight vehicle, with at least 2 sensors per support point, and 0.5 sensors / m2 in the non-support area. 2 The density distribution form of the adjacent sensors d and the maximum adjacent axle spacing L of the vehicle should satisfy the constraint condition d≤0.3L, where M is the number of vehicle suspension support points and L is the maximum horizontal distance between adjacent axle centers; The technical basis for FPGA selection in step 1.2 is: FPGA clock frequency is configured in the range of 500MHz-1GHz, FPGA clock frequency f FPGA ≥2f to meet the nanosecond synchronization requirement, and the timestamp alignment error δ satisfies: Where f is the sensor sampling frequency and f≥1kHz; In step 1.3, the dual-channel ring buffer memory adopts a ping-pong cache architecture, alternately performs data writing and reading operations by setting a double buffer, and cooperates with a DMA chain transfer mechanism; The dual-channel ring buffer memory capacity is configured as 2×N×B×T, where N is the total number of sensors, B is the data bandwidth of a single sensor, and T is the maximum non-blocking transmission time.

3. The method according to claim 1, characterized in that The step 2 comprises: Improved SURE threshold denoising based on wavelet packet decomposition, conversion of one-dimensional time series data into two-dimensional feature maps through Gram angular field transform, time series anomaly detection and waveform reconstruction, and dynamic reconstruction error threshold ε are set as: ε=μ+3σ; Where μ is the mean reconstruction error of the training set and σ is the standard deviation; The construction of the multimodal training dataset is achieved by integrating the load distribution feature map and the vehicle motion state vector through feature-level fusion, and using the attention mechanism to achieve cross-modal feature interaction.

4. The method according to claim 1, wherein In step 3, the time series model based on the Transformer architecture includes: a spatiotemporal convolutional coding layer, a multi-head causal attention mechanism, a dynamic weight allocation module, and a model distillation; The spatiotemporal convolutional coding layer uses a two-dimensional dilated causal convolution kernel to extract local spatiotemporal features, and the dilation coefficient d is d=2 n-1 Exponential growth, where n is the number of layers and the convolution kernel size satisfies the formula k = 2d + 1; The multi-head causal attention mechanism uses a sliding window attention with a sliding window of 64 and 8 attention heads to limit the computational complexity; The dynamic weight allocation module implements the importance assessment of sensor channels through a learnable gating network; the learnable gating network consists of two fully connected layers + Softmax activation; The model distillation adopts a teacher-student architecture. The teacher model is a 12-layer Transformer with a hidden layer dimension of 512 and 8 attention heads. The student model adopts a 4-layer structured pruned network with a hidden layer dimension of 256. The top 50% important attention heads are retained in the Transformer layer. Knowledge transfer is achieved through KL divergence loss, and the pruning standard is the L1 norm of the attention head weights.

5. The method according to claim 1, wherein In step 5, the error correction hybrid system uses an adaptive Kalman filter to perform dynamic error compensation at the front end, setting the process noise covariance matrix Q t Dynamic update mechanism; the back-end application correction network adopts online transfer learning strategy to update network parameters; the error correction amount is superimposed on the model output through residual connection design to achieve error correction of the model output result; Among them, the process noise covariance matrix Q t The dynamic update mechanism is: Q t =αQ t-1 +(1-a)·diag(ξ t ·x t T ); Among them, α = 0.95 is the smoothing factor, ξ t is the residual vector at time t; The state vector of the adaptive Kalman filter is x = [w,w t ,a] T , where w is the load, w t is the load change rate, a is the vehicle acceleration; The back-end correction network adopts a three-layer gated recurrent unit structure. The input is the Kalman residual. The input dimension is consistent with the Kalman residual dimension. The number of hidden units does not exceed 32, and the time step is 10. Online transfer learning uses an elastic weight solidification algorithm and sets the Fisher information matrix threshold to 10 after cross-validation. -4 Freeze parameters; The update period of the front-end adaptive Kalman filter is 10ms, and the update period of the back-end correction network parameters is 100ms, establishing a 1:10 cascade update timing chain.

6. The method according to claim 1, wherein The step 6 comprises: Step 6.1: Output the time series weighing results in real time, and use the sliding window mechanism to smooth the results; Step 6.2, analyze the time series based on STL decomposition; Step 6.3: Based on the analysis results, combined with the dynamic uncertainty assessment of the vehicle's motion state, an ARIMA-Prophet combined prediction model is constructed to predict the weight change trend and provide early warning; In step 6.1, an exponentially weighted sliding with a window length of 128 is used to smooth the results. In step 6.3, the Monte Carlo dropout method was used to evaluate uncertainty, with a dropout rate of 0.2 and 100 samplings to calculate the confidence interval. Step 6.3 builds an ARIMA-Prophet combined forecasting model, which includes: the ARIMA model is responsible for short-term trend forecasting (i.e., <5s), and the Prophet model is responsible for extracting periodic components. The output is fused through weighted averaging, and the weight coefficient β satisfies: Among them, t p is the prediction time, τ=5s is the ARIMA effective window, when t p >τ, the Prophet output is fully adopted; Step 6.3: When the preset conditions are met simultaneously, the three-level warning mechanism will be triggered: (1) Actual load > rated load × 90%; (2) predicted weight change rate > 5% / s; (3) Monte Carlo dropout evaluation confidence interval P ≥ 95%; The three-level warning mechanism is: local alarm prompt, overload status upload, and on-board ECU speed limit control.

7. A large freight vehicle weighing system based on a time series large model, the system is used to implement the method according to any one of claims 1 to 6, characterized in that: The system includes: a data acquisition module, a preprocessing module, a model calculation module, and a correction output module; The data acquisition module is used to collect dynamic load data of large freight vehicles, so that the data has temporal and spatial characteristics and is stored; The preprocessing module is used to process data with spatiotemporal characteristics and construct a multimodal training dataset; The model calculation module is used to perform real-time dynamic weighing calculations on the collected load data with time series characteristics by applying the time series large model built based on the Transformer architecture; The correction output module is used to perform error correction on the weighing calculation results of the model, output the weighing results in real time, and perform weight change trend prediction and early warning.

8. The system according to claim 7, characterized in that The data acquisition module includes: an acquisition unit, a fusion unit and a storage unit; The acquisition unit is used for collecting dynamic load data information at various positions of large freight vehicles by using a distributed pressure sensor array; The fusion unit is used to fuse multi-source data by using the timestamp alignment algorithm to generate time series load data with spatiotemporal characteristics; The storage unit is used to store the time series payload data in the dual-channel ring buffer memory of the edge computing node, and realize non-blocking data transmission with the preprocessing module through the DMA mechanism.

9. The system according to claim 7, wherein: The preprocessing module includes: denoising unit, transformation unit and detection and reconstruction unit; The denoising unit is used to remove the influence of noise on the original time series data by using the improved SURE threshold based on wavelet packet decomposition; The transformation unit is used to convert one-dimensional time series data into a two-dimensional feature map through GAF; The detection and reconstruction unit is used to detect abnormal time series data and reconstruct waveforms.

10. The system according to claim 7, wherein: The correction output module includes: a filtering unit, a correction unit, a connection unit, an output unit, a prediction unit and an early warning unit; The filtering unit is used to perform dynamic error compensation on the weighing calculation results of the model using adaptive Kalman filtering; The correction unit is used to apply the correction network to the weighing calculation result of the model to perform error correction; The connection unit is used to superimpose the error correction amount and the model output through the residual connection design to achieve error correction of the model output result; The output unit is used to output time series weighing results in real time; The prediction unit is used to predict the weight change trend by constructing an ARIMA-Prophet combined prediction model; The early warning unit is used to trigger a three-level early warning mechanism including local alarm prompts, overload status upload and on-board ECU speed limit control.

Images