Vehicle weight and type detection method and system based on piezoelectric sensor

Abstract

The invention discloses a vehicle weight and type detection method and system based on a piezoelectric sensor, and relates to the field of vehicle detection.The method comprises the steps that firstly, piezoelectric signals generated when a vehicle passes through are synchronously collected and preliminarily conditioned; inputting the signal into an edge computing node, extracting time domain, frequency domain and space domain features by using self-supervised learning, and identifying a vehicle driving state through an improved random forest algorithm; carrying out secondary conditioning on the signal according to the state; and finally, aligning the multi-channel signal time sequence deviation by adopting dynamic time warping during high-speed driving, and outputting a detection result in combination with a multi-task Bayesian regression model. In slow driving, a blind source separation method is adopted to separate signals, and vehicle types and weight estimation values are synchronously output through a multi-task deep learning model, so that the problem that vehicle weight and types cannot be accurately distinguished due to the fact that vehicle dynamic information in high-speed driving or crowded conditions cannot be accurately processed in the prior art is solved. According to the invention, high-precision detection of the vehicle weight and type is realized.

Description

Technical Field

[0001] The present invention relates to the field of vehicle detection, and in particular to a vehicle weight and type detection method and system based on a piezoelectric sensor. Background Art

[0002] With the acceleration of urbanization and the continued growth of vehicle ownership, the development of intelligent transportation systems has become a key approach to improving traffic efficiency, ensuring traffic safety, and promoting energy conservation and emission reduction. Within this development, accurate, real-time detection of vehicles on the road, particularly the precise identification of vehicle weight and type, has become a key component in overload management, road maintenance planning, and traffic flow optimization. Traditional vehicle detection methods, such as manual observation and ground-sensing coils, have numerous limitations in terms of accuracy, real-time performance, information dimensionality, and maintenance costs. Against this backdrop, piezoelectric sensor technology, with its unique advantages, provides strong technical support for the precise perception of vehicle dynamic information.

[0003] However, the weight and type detection results of piezoelectric sensors are affected by the vehicle's state while it's in motion. For example, at high speeds, the contact time between the tire and the sensor is short, making existing detection methods unable to accurately capture and process the dynamic information of high-speed vehicles, thus affecting detection accuracy. Furthermore, in crowded traffic, where the distances between vehicles are small, existing detection methods are unable to accurately distinguish the sensor signals of adjacent vehicles, resulting in inaccurate detection results and an inability to distinguish the weight and type of adjacent vehicles.

[0004] Therefore, a vehicle weight and type detection method and system based on a piezoelectric sensor are needed to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention discloses a vehicle weight and type detection method and system based on piezoelectric sensors. Through the combination of multi-channel synchronous acquisition, self-supervised learning, an improved random forest algorithm, dynamic signal conditioning, dynamic time warping, blind source separation and a multi-task deep learning model, high-precision detection of vehicle weight and type is achieved, and high-precision vehicle weight estimation and type classification in high-speed and low-speed scenarios are achieved.

[0006] The present invention adopts the following technical solutions:

[0007] A method for detecting vehicle weight and type based on a piezoelectric sensor comprises the following steps:

[0008] Step 1: collecting piezoelectric signals generated by a vehicle passing by using a multi-channel piezoelectric sensor array, and performing preliminary signal conditioning on the collected signals using a signal conditioner. The multi-channel piezoelectric sensor array achieves synchronous collection of multi-channel signals through a synchronous trigger mechanism;

[0009] Step 2: Input the preliminarily conditioned signal to the edge computing node for processing. The edge computing node uses a self-supervised learning method to extract the time domain, frequency domain, and spatial domain features of the input signal, and uses an improved random forest algorithm to analyze the extracted time domain, frequency domain, and spatial domain features to identify the vehicle's driving state, which includes high-speed driving state and slow driving state.

[0010] Step 3: Perform secondary signal conditioning on the collected signal based on the identified vehicle driving state;

[0011] Step 4: Perform vehicle weight and vehicle type detection on the secondary conditioned signal based on the identified vehicle driving status. When the vehicle is driving at high speed, the dynamic time warping method is used to align the timing deviation of the multi-channel signal, and the vehicle type and weight detection results are output based on the multi-task Bayesian regression model. When the vehicle is driving slowly, the blind source separation method is used to separate the piezoelectric signal components of a single vehicle, and the multi-task deep learning model synchronously outputs the vehicle type classification results and weight estimation values.

[0012] Furthermore, in step one, the signal conditioner includes a filtering unit, a gain amplification unit, an analog-to-digital conversion unit, and an environmental parameter compensation unit. The filtering unit uses a bandpass filter to eliminate ground vibration noise and electromagnetic noise. The gain amplification unit dynamically adjusts the gain according to the signal strength. The analog-to-digital conversion unit uses an analog-to-digital converter to convert the amplified analog piezoelectric signal into a digital piezoelectric signal. The environmental parameter compensation unit offsets the temperature drift effect of the piezoelectric signal through an NTC thermistor. The synchronous trigger mechanism generates a timestamp through an internal clock to mark the piezoelectric signal collected by each sensor in the multi-channel piezoelectric sensor array.

[0013] Furthermore, in step 2, the self-supervised learning method obtains the time difference sequence of the piezoelectric signals of the multi-axle wheels by calculating the duration of the single tire signal, thereby converting the piezoelectric signal into a time series feature matrix, and generates a time-frequency feature matrix by performing a short-time Fourier transform on the multi-channel piezoelectric signal. Then, a spatial position feature matrix is ​​constructed according to the physical layout of the multi-channel piezoelectric sensor array, and a vehicle geometric model is fitted based on the multi-point pressure distribution of the piezoelectric signal to calculate the characteristic wheelbase.

[0014] Furthermore, in step 2, the working method of the improved random forest algorithm includes the following steps:

[0015] S201, inputting the extracted time domain, frequency domain, and spatial domain features into an improved random forest algorithm, and dividing the historical input feature data into a training set and a test set, wherein the training set is used to train the improved random forest algorithm, and the test set is used to evaluate the performance of the algorithm;

[0016] S202. For each decision tree, a dynamic feature selection mechanism and a sliding window are used to select the optimal splitting feature from the input features. The dynamic feature selection mechanism calculates local statistics of the input time domain features through windowing, dynamically evaluates the importance of the time domain features, and dynamically adjusts the weights of the frequency domain features based on the distribution of the main vibration frequencies. The importance ranking of the spatial domain features is then dynamically adjusted using the covariance matrix of the multi-channel piezoelectric sensor array.

[0017] S203, when splitting a node in each decision tree, defining a splitting criterion based on the time domain, frequency domain, and spatial domain characteristics of the piezoelectric signal, and evaluating the node splitting effects at different time scales to select the optimal splitting point to split into left and right child nodes;

[0018] S204, repeating S202 and S203 until the stop condition is met;

[0019] S205. Construct multiple decision trees to form a random forest. Use the training set generated by self-sampling to train each decision tree. Use an incremental learning strategy to automatically trigger algorithm updates. During the training process of each decision tree, dynamically adjust the sample weights of the training set, and record the output results and feature importance.

[0020] S206. For the input time domain, frequency domain and space domain features, each decision tree outputs a prediction result, and the category with the most output results among all decision trees is selected as the final result through confidence-weighted voting.

[0021] Furthermore, in step three, for the high-speed state of the vehicle, a lightweight convolutional neural network is used to extract transient features; for the slow state of the vehicle, a three-layer wavelet packet decomposition is used to extract steady-state features.

[0022] Furthermore, in step four, the dynamic time warping method takes the piezoelectric signal of each sensor channel as a time series, and calculates the distance matrix between different piezoelectric signals, and uses a dynamic programming method to find the optimal path in the distance matrix. The optimal path is the path with the smallest sum of distances from the starting point of the distance matrix to the end point of the distance matrix, and the piezoelectric signals of different sensor channels are stretched or compressed according to the optimal path to achieve time series alignment. The multi-task Bayesian regression model includes an input layer, a hidden layer and an output layer. The input layer is used to receive the aligned multi-channel piezoelectric signal, the hidden layer is used to extract and transform the signal, and the output layer is used to output the prediction results of vehicle type and weight. The multi-task Bayesian regression model updates parameters according to the Bayesian theorem to minimize the prediction error.

[0023] Furthermore, in step four, the blind source separation method uses a linear transformation matrix to transform the piezoelectric mixed signal into a source signal, and realizes the separation of the mixed signal by maximizing the statistical independence of the source signal. The multi-task deep learning model uses a convolutional neural network and a long short-term memory network to model the dynamic characteristics of the vehicle, and simultaneously learns the two tasks of vehicle type classification and weight estimation by sharing some network layers. The convolutional neural network receives the single-channel signal after blind source separation, and uses one-dimensional convolution to capture the local time-frequency pattern, and then uses maximum pooling to reduce the temporal resolution. The long short-term memory network dynamically weights the feature contributions of different time steps by introducing a self-attention layer.

[0024] Furthermore, a vehicle weight and type detection system based on a piezoelectric sensor comprises:

[0025] A multi-channel piezoelectric sensor array is used to collect piezoelectric signals generated by vehicles passing by, and a synchronous trigger mechanism is used to ensure synchronous signal collection;

[0026] The signal conditioner is used to perform preliminary and secondary conditioning on the original signals collected by the sensor and transmit the conditioned signals to the edge computing node for further processing;

[0027] The edge computing node uses a self-supervised learning method to extract the time, frequency, and spatial domain features of the signal, and uses an improved random forest algorithm to analyze the extracted features to identify the vehicle's driving status;

[0028] The vehicle detection module uses dynamic time warping to align the timing deviations of multi-channel signals when the vehicle is traveling at high speeds, and outputs vehicle type and weight detection results based on a multi-task Bayesian regression model. When the vehicle is traveling slowly, blind source separation is used to separate the piezoelectric signal components of a single vehicle, and a multi-task deep learning model simultaneously outputs vehicle type classification results and weight estimates.

[0029] The data display module displays the detection process data in real time through the display device and sends the data to the cloud for back-end analysis and storage;

[0030] The communication module is used for communication between different modules within the system and data interaction with external systems.

[0031] The beneficial effects of the present invention are:

[0032] 1. The present invention adopts a multi-channel piezoelectric sensor array to comprehensively and accurately collect the piezoelectric signals generated when a vehicle passes, and adopts a synchronous trigger mechanism to realize the synchronous collection of multi-channel signals, avoiding data deviations and errors caused by inconsistent signal collection time, and improving detection accuracy in dynamic environments.

[0033] 2. The present invention uses edge computing nodes for real-time signal processing. The edge computing nodes introduce self-supervised learning methods to extract time domain, frequency domain and spatial domain features, reducing the need for manual labels, enabling the system to adapt more flexibly to different types of vehicles and road conditions, and using an improved random forest algorithm to analyze the extracted features to identify the vehicle's driving status, whether it is high-speed driving or slow driving, thereby improving the accuracy of state recognition.

[0034] 3. The present invention performs secondary signal conditioning on the collected signals based on the identified vehicle driving state. The piezoelectric signals of vehicles in different driving states have different characteristics. Targeted secondary signal conditioning can further optimize signal quality and improve the accuracy of subsequent detection.

[0035] 4. For high-speed vehicles, this invention uses dynamic time warping to align the timing deviations of multi-channel signals. It then outputs vehicle type and weight detection results based on a multi-task Bayesian regression model. This dynamic time warping method effectively addresses the timing discrepancies between multi-channel signals, ensuring signal consistency across the temporal dimension. The multi-task Bayesian regression model can simultaneously handle both vehicle type and weight detection tasks, leveraging inter-task correlations to improve detection accuracy and quantify uncertainty.

[0036] 5. This invention uses blind source separation to separate the piezoelectric signal components of individual vehicles for slow-moving vehicles. A multi-task deep learning model then simultaneously outputs vehicle classification and weight estimation results. Blind source separation can separate mixed signals into individual source signals without knowing the specific form of the source signals, thereby more accurately extracting the signal characteristics of individual vehicles. The multi-task deep learning model possesses powerful learning capabilities and can automatically learn complex feature representations from large amounts of data, improving the accuracy of vehicle classification and weight estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0038] Figure 2 Schematic diagram of the process of the improved random forest algorithm in the present invention;

[0039] Figure 3 This is the overall system architecture diagram of the present invention. DETAILED DESCRIPTION

[0040] The following is a combination of the embodiments of the present invention Figure 1 To the attached Figure 3The technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0041] The embodiment of the present invention discloses a vehicle weight and type detection method based on a piezoelectric sensor, as shown in the attached Figure 1 , including the following steps:

[0042] Step 1: Signal acquisition and preliminary signal conditioning

[0043] A multi-channel piezoelectric sensor array installed on the road generates a piezoelectric effect on the sensors as vehicles move, capturing vibration signals. These signals can reflect characteristics such as vehicle weight, speed, and type. A synchronized triggering mechanism ensures that all sensors in the multi-channel sensor array begin collecting signals at the same time, thus ensuring synchronization and timing consistency of the multi-channel signals. The collected raw signals undergo preliminary processing in a signal conditioner, which typically includes noise suppression, signal amplification, and filtering to ensure signal quality and usability.

[0044] PVDF (polyvinylidene fluoride) piezoelectric thin film sensors offer high sensitivity and a wide frequency response of 0.1Hz-10kHz, making them suitable for dynamic load detection. 8-16 sensors are linearly arranged at 50cm intervals below the lane, covering the wheel contact area. FPGA-based hardware synchronization is implemented, with a global clock signal ensuring a multi-channel sampling phase difference of less than 1μs. A sampling rate of ≥10kHz and a resolution of ≥16 bits are required to capture the transient impact characteristics of tire contact.

[0045] Step 2: Feature extraction and vehicle driving state recognition

[0046] After preliminary conditioning, the signal is transmitted to the edge computing node. The edge computing node can process the signal locally, reducing data transmission latency and improving real-time performance. Signal features are extracted using self-supervised learning methods. Signal characteristics in the time, frequency, and spatial domains provide critical information for subsequent analysis. Time domain features are used to analyze how the signal changes over time, such as the impact force changes when a vehicle passes. Frequency domain features analyze the signal spectrum using methods such as Fourier transform to obtain the frequency characteristics of vehicle vibration. Due to the layout of the multi-channel sensor array, spatial domain analysis can be used to obtain the vehicle's motion trajectory and the spatial distribution characteristics of the piezoelectric signal.

[0047] The extracted features are analyzed using an improved random forest algorithm to identify the vehicle's driving state. Common driving states include: High-speed driving: The vehicle passes quickly, resulting in significant vibration and impact, typically when the vehicle passes through the detection area at speeds ≥ 30 km / h; Slow-speed driving: The vehicle generates less vibration at low speeds, with a different feature distribution, and a speed range: Typically, the vehicle passes through the detection area at speeds < 30 km / h.

[0048] Step 3: Secondary Signal Conditioning

[0049] The signal undergoes secondary signal conditioning based on the identified driving state. High-speed state: Enhances high-frequency vibration characteristics (such as tire-road impact) and suppresses interference from wind speed and vehicles in adjacent lanes. Slow state: Emphasizes engine vibration characteristics and removes interference from pedestrians, small animals, etc.

[0050] Step 4: Vehicle weight and type detection

[0051] The vehicle weight and vehicle type are detected on the secondary conditioned signal based on the identified vehicle driving status. When the vehicle is driving at high speed, the dynamic time warping method is used to align the timing deviation of the multi-channel signals, and the vehicle type and weight detection results are output based on the multi-task Bayesian regression model. When the vehicle is driving slowly, the blind source separation method is used to separate the piezoelectric signal components of a single vehicle, and the multi-task deep learning model synchronously outputs the vehicle type classification results and weight estimation values.

[0052] Detection at high speed:

[0053] When a vehicle is traveling at high speeds, multi-channel signals may experience timing deviations due to signal propagation delays. Dynamic Time Warping (DTW) can align these timing deviations to ensure signal consistency. A multi-task Bayesian regression model uses multi-task learning to predict vehicle type and weight based on aligned signal inputs using a Bayesian regression model. Bayesian regression can provide uncertainty assessment and share information across different tasks.

[0054] Detection in slow driving state:

[0055] Blind Source Separation: For slow-moving vehicles, signals are complex and may be affected by interference from other vehicles. Using blind source separation, we can isolate the piezoelectric signal components from individual vehicles, ensuring that each vehicle's signal can be analyzed independently. A multi-task deep learning model: Based on the separated signals, a deep learning model is used to generate simultaneous outputs for vehicle type classification and weight estimation. Deep learning models can capture complex nonlinear relationships, providing more accurate detection results.

[0056] This method combines piezoelectric sensor arrays, signal conditioning, edge computing, self-supervised learning, dynamic time warping, blind source separation, and deep learning technologies to accurately identify vehicle type and weight under different driving conditions. This efficient and intelligent detection method has high practical value and development prospects in fields such as intelligent transportation and road monitoring.

[0057] In step 1, the signal conditioner includes a filtering unit, a gain amplifying unit, an analog-to-digital conversion unit, and an environmental parameter compensation unit.

[0058] The filtering unit uses a bandpass filter to eliminate ground vibration noise (low-frequency interference, such as 5-20Hz) and electromagnetic noise (high-frequency interference, such as above 500Hz). The bandpass filter allows signals within a specific frequency range (such as 20Hz-200Hz) to pass through, while filtering out both low-frequency and high-frequency noise. Implementation: An active filter is constructed using an RC circuit or operational amplifier, and processed using a digital filtering algorithm.

[0059] The gain amplifier unit dynamically adjusts the gain based on signal strength, adjusting the amplification factor according to the input signal strength to avoid signal saturation or distortion. Workflow: Detect the input signal's amplitude or RMS value; calculate the optimal gain through a feedback circuit or digital algorithm; and control the variable gain amplifier (VGA) or programmable gain amplifier (PGA). Application scenario: The piezoelectric sensor output signal may fluctuate significantly due to factors such as distance and material.

[0060] The analog-to-digital conversion unit uses an analog-to-digital converter to convert the amplified analog piezoelectric signal into a digital piezoelectric signal for subsequent digital processing. Key parameters: Resolution (e.g., 16 bits): Determines quantization accuracy (for example, with a 3.3V reference voltage, the minimum resolution voltage is 3.3V / 2^16 ≈ 50μV). Sampling rate: Must meet the Nyquist sampling theorem (typically at least twice the highest signal frequency). Implementation: Select a low-noise, high-resolution ADC chip (e.g., the ADS1255).

[0061] The environmental parameter compensation unit uses an NTC thermistor to offset the temperature drift of the piezoelectric signal. The compensation principle is: the NTC is connected in parallel or series with the sensor to form a voltage divider circuit. As the temperature changes, the NTC resistance changes, and the voltage divider ratio changes, thereby compensating for the temperature drift of the sensor output. Digital compensation is performed using a lookup table or polynomial fitting. A synchronous trigger mechanism uses an internal clock to generate a timestamp to mark the piezoelectric signal collected by each sensor in the multi-channel piezoelectric sensor array, ensuring data synchronization. The internal clock generates synchronization pulses, triggering all ADC sampling simultaneously. A high-precision timestamp (e.g., microsecond level) is added to each channel's data, aligning the time during subsequent processing.

[0062] The output end of the filtering unit is connected to the input end of the gain amplifying unit, the output end of the gain amplifying unit is connected to the input end of the analog-to-digital conversion unit, and the output end of the analog-to-digital conversion unit is connected to the input end of the environmental parameter compensation unit.

[0063] In step 2, the self-supervised learning method calculates the duration of a single tire signal to obtain a time difference sequence of the piezoelectric signals from multiple axles, converting the piezoelectric signals into a time series feature matrix. For each tire, the signal strength varies over different time periods. By calculating the duration of each tire signal (i.e., the time width of the signal waveform), a time difference sequence (i.e., the time difference between the piezoelectric signal from each tire and the other tires) is obtained. This time difference sequence can reflect the dynamic characteristics of the vehicle, such as vehicle speed, acceleration, and even wheel weight distribution. These time differences serve as key temporal features in subsequent analysis. Using the time difference sequence of the piezoelectric signals, the signals are converted into a feature matrix. By extracting time domain features such as the signal's mean, standard deviation, peak value, and waveform length, a matrix describing the signals is formed. The time series feature matrix not only contains the time domain signal features of each sensor but also includes correlation information between different tires, such as phase and amplitude differences in the piezoelectric signals.

[0064] A time-frequency feature matrix is ​​generated by performing a short-time Fourier transform (STFT) on the multi-channel piezoelectric signal. The short-time Fourier transform (STFT) converts the signal from the time domain to the time-frequency domain, thereby simultaneously acquiring both the time and frequency information of the signal. By performing STFT processing on the multi-channel piezoelectric signal, a time-frequency feature matrix for each sensor signal can be obtained. This matrix contains the signal's spectral information at different time periods and can reveal the characteristic frequencies and corresponding physical conditions of different tires. For example, high-frequency components may be associated with the vibration characteristics of the wheel, while low-frequency components may reflect the overall dynamic characteristics of the vehicle.

[0065] A spatial position feature matrix is ​​constructed based on the physical layout of the multi-channel piezoelectric sensor array. This matrix takes into account the actual physical layout of the piezoelectric sensor array (e.g., sensor location and distribution). The spatial position feature matrix for each sensor reflects its location and arrangement in physical space. By correlating the sensor's time-frequency characteristics with its spatial location, it is possible to analyze the pressure distribution between different tires and its relationship to the vehicle's geometry. For example, differences in the piezoelectric signals between the front and rear axles may reflect differences in wheelbase between the wheels. A vehicle geometric model is fitted based on the multi-point pressure distribution of the piezoelectric signals to calculate the characteristic wheelbase. The piezoelectric signals provide information about the pressure distribution of the vehicle tires when they are in contact with the ground. By fitting the piezoelectric signals from multiple sensors, a pressure profile for the entire wheel can be obtained. These pressure profiles reflect the pressure variations between different tires on different surfaces. Based on these pressure profiles, the vehicle's geometric characteristics, particularly parameters such as wheelbase and track width, can be fitted. By combining the known geometric model with sensor data, the vehicle's characteristic wheelbase (i.e., the distance between the front and rear axles) can be calculated.

[0066] The characteristic wheelbase is a key parameter derived by fitting a vehicle's geometric model to sensor data. It can be calculated based on the pressure difference between the front and rear wheels, the time difference sequence of the sensor signals, their time-frequency characteristics, and their spatial positional characteristics. Accurate calculation of the characteristic wheelbase helps determine the vehicle's type, weight, and driving state.

[0067] The core of the entire process is to extract time, frequency, and spatial domain features from the multi-channel signals acquired by the piezoelectric sensor array using self-supervised learning methods. This is then combined with physical models to model and calculate the vehicle's geometric characteristics. Ultimately, based on this characteristic information, the vehicle's weight, type, and other geometric parameters can be accurately detected.

[0068] In step 2, the working method of the improved random forest algorithm is as shown in the attached Figure 2 As shown, the following steps are included:

[0069] S201: Input features and dataset segmentation

[0070] Time, frequency, and spatial domain features are input into a modified random forest algorithm, and the historical input feature data is divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate model performance. First, features are extracted through data preprocessing (for example, time domain features can be the signal's mean and standard deviation, and frequency domain features can be obtained using Fourier transforms). The data is then partitioned, for example, using 80% of the data for training and 20% for testing.

[0071] S202: Select the optimal split feature

[0072] A dynamic feature selection mechanism and a sliding window are used to select the optimal splitting features. For time domain features, their importance is dynamically evaluated and the frequency domain feature weights are adjusted based on the dominant frequency distribution. The weights of spatial domain features are adjusted using the covariance matrix. A sliding window (such as a fixed-length window in a time series signal) can be applied to calculate local statistics of time domain features. For frequency domain features, their importance weights are dynamically adjusted based on the dominant frequency of the signal (for example, using the fast Fourier transform (FFT) to calculate the spectrum). The weights of spatial domain features can be adjusted by calculating the correlation between different sensors using the covariance matrix.

[0073] S203: Node Splitting Criteria and Multi-Scale Evaluation

[0074] When splitting each decision tree node, a splitting criterion is defined based on the characteristics of the piezoelectric signal, and the effectiveness of the node split at different time scales is evaluated. The effectiveness of each candidate split is evaluated using features based on the time, frequency, and spatial domains (such as information gain or the Gini index). The tree's classification ability can be improved by dynamically adjusting the splitting criterion, taking into account the importance of features within different time windows.

[0075] S204: Circuit Training and Splits

[0076] Repeat S202 and S203 until a stopping condition is met (e.g., a predetermined tree depth or number of split points is reached, or the tree splitting effect is maximized). The stopping condition can be set to the maximum tree depth or when the information gain is less than a certain threshold. After each split, the importance of each feature is re-evaluated, and the splitting strategy is dynamically adjusted based on the new evaluation results.

[0077] S205: Random Forest Training and Incremental Learning

[0078] Multiple decision trees are constructed to form a random forest. Bootstrap sampling is used to generate training sets for each decision tree. Incremental learning is employed to trigger algorithm updates and dynamically adjust the sample weights within the training set. Bootstrap sampling is used to generate the training set for each decision tree, with each tree using different training data. Incremental learning allows for continuous adjustment of sample weights during model training, allowing the model to adapt to new data distributions or changes.

[0079] S206: Final result output

[0080] Each decision tree outputs a prediction, and the final result is selected through confidence-weighted voting. Each decision tree outputs a class label and its corresponding confidence score. The final classification result is obtained by combining the confidence scores of each tree through voting. Trees with higher confidence scores have a greater influence on the final decision.

[0081] In step three, different feature extraction methods are used for secondary signal conditioning according to different operating states of the vehicle.

[0082] At high speeds, a lightweight convolutional neural network is used to extract transient features. At high speeds, transient features (such as acceleration, deceleration, and sharp turns) become more pronounced, making traditional feature extraction methods ineffective. Therefore, a lightweight convolutional neural network (CNN) can more efficiently extract these transient features. Sensor data from a high-speed vehicle may include acceleration, steering angle, and speed. To increase speed and reduce computational burden, fewer convolutional layers and parameters can be used while maintaining strong feature extraction capabilities. This approach can include replacing traditional convolutions with depthwise separable convolutions to reduce model complexity. Skip connections can be used to avoid information loss and accelerate model convergence. A trained convolutional neural network is then used to extract transient features. These features may include acceleration patterns, turn amplitude, and speed changes. Training is performed using high-speed data and evaluation is performed using cross-validation.

[0083] When the vehicle is moving slowly, three-layer wavelet packet decomposition is used to extract steady-state features. When a vehicle is moving slowly, dynamic changes are minimal, and steady-state features (such as those associated with steady low-speed driving) are more prominent. In this case, wavelet packet decomposition can effectively capture both low-frequency and high-frequency components of the signal and extract steady-state features. Data from the vehicle is collected while it is moving slowly, focusing on low-frequency features such as vibration signals at steady speeds and steering wheel angle changes. Appropriate wavelet basis functions (such as Daubechies wavelets or Symlet wavelets) are selected. Three-layer wavelet packet decomposition is performed to decompose the signal into different frequency bands. Each decomposition layer separates the signal into low-frequency and high-frequency components, which facilitates the extraction of steady-state features. Features are extracted from the decomposed signal at each layer, such as statistical features such as energy, mean, and variance, to reflect steady-state features. Correlation analysis is used to select the most representative features of the vehicle's slow-moving state, reducing redundant information.

[0084] In step four, the dynamic time warping (DTW) method treats the piezoelectric signals from each sensor channel as time series. These signals originate from different sensors and may be acquired at different time points, resulting in different lengths and sampling rates. A distance matrix is ​​then calculated between the different piezoelectric signals. By calculating the distance matrix between different sensor channels, the DTW method can measure the similarity between these time series. Common distance metrics include Euclidean distance or Manhattan distance. DTW uses a dynamic programming algorithm to find an optimal path in the distance matrix. This path is the path from the starting point to the end point of the matrix, minimizing the sum of all distances along this path. This means that aligning the two signals along this path will make them as similar as possible. Based on the found optimal path, the signals are dynamically adjusted (stretched or compressed) to align the signals from different channels on the time axis. This way, even if the sampling points and times of the signals are inconsistent, the DTW method will align the signals, facilitating subsequent analysis.

[0085] The multi-task Bayesian regression model consists of an input layer, hidden layers, and an output layer. The input layer receives aligned multi-channel piezoelectric signals as input to the neural network model. Signals from each sensor channel are processed in the input layer. The hidden layers extract and transform the signals. These features may include frequency domain information and time domain statistical characteristics, which are nonlinearly mapped using the activation functions of the hidden layer neurons. The output layer outputs predictions for vehicle type and weight. This is the key to multi-task learning. The model simultaneously handles multiple tasks (predicting vehicle type and weight), so there are two different neurons in the output layer, one for each task. The multi-task Bayesian regression model updates parameters according to Bayes' theorem to minimize prediction error. The Bayesian regression model uses Bayes' theorem for parameter updates. Through an iterative optimization process, the model continuously adjusts its parameters to minimize prediction error. The advantage of Bayesian regression is that it can handle uncertainty and provide a probability distribution for each parameter.

[0086] The aligned signals are fed into a multi-task Bayesian regression model. The model is trained using historical data to update the Bayesian regression parameters. Once the model is trained, new piezoelectric signal data can be input, and the model will output the corresponding vehicle type and weight predictions. This method combines the advantages of time series signal processing (DTW) and Bayesian regression, making it particularly suitable for accurate vehicle detection and analysis of multi-channel sensor data in practical applications.

[0087] In step 4, the blind source separation method uses a linear transformation matrix to transform the piezoelectric mixed signal into a source signal, and extracts the source signal from the piezoelectric mixed signal, which is usually a plurality of independent source signals in the mixed signal. A linear transformation matrix (for example, independent component analysis ICA) is used to transform the mixed signal and convert it into a source signal. The key to this method is to achieve separation of the mixed signal by maximizing the statistical independence of the source signals, which means separating the signals that are as unrelated as possible from the mixed signal.

[0088] The multi-task deep learning model uses a combination of convolutional neural networks and long short-term memory networks to model the dynamic characteristics of vehicles, and simultaneously learns the two tasks of vehicle type classification and weight estimation by sharing some network layers.

[0089] The convolutional neural network receives a single-channel signal that has undergone blind source separation, representing the output of the piezoelectric sensor. Through one-dimensional convolution operations, the CNN is able to capture local time-frequency patterns in the signal and extract features for subsequent analysis. A maximum pooling operation is used to reduce the temporal resolution and redundant information while retaining important features, enhancing the model's computational efficiency. A long short-term memory (LSTM) network is used to capture the temporal information and dynamic characteristics of the signal. To enhance the performance of the LSTM network, a self-attention layer is introduced that dynamically weights the feature contribution of each time step. In this way, the model can better handle long sequences of data and identify time steps that are important for vehicle type classification and weight estimation.

[0090] By sharing convolutional and LSTM layers, the model can simultaneously learn the dynamic characteristics of the vehicle while completing multiple tasks (such as classification and regression), avoiding the high computational overhead of training a separate model for each task. The collaborative training of the two tasks helps the model better extract features with generalization capabilities, improving performance.

[0091] The implementation steps of this method include processing raw sensor signals using blind source separation techniques and performing multi-task learning with a deep learning model. Convolutional neural networks reduce temporal resolution by processing time-frequency patterns and pooling, while LSTMs enhance the modeling capabilities of temporal information through a self-attention mechanism, effectively completing vehicle type classification and weight estimation tasks. This design not only improves feature extraction but also enhances model efficiency through shared network layers.

[0092] A vehicle weight and type detection system based on piezoelectric sensors, as shown in the attached Figure 3 Shown, including:

[0093] A multi-channel piezoelectric sensor array collects piezoelectric signals generated by passing vehicles, ensuring that the signals are responsive to the physical changes caused by the vehicles. Multiple piezoelectric sensors are installed in an array and arranged on the road surface. A synchronized triggering mechanism ensures that each sensor acquires signals synchronously, avoiding errors caused by signal delays. Appropriate sensor sensitivity and response frequency are selected to ensure that signals from different vehicles at different speeds can be captured.

[0094] The signal conditioner is used to perform preliminary and secondary conditioning on the collected raw signals, including denoising, amplification, filtering, etc., to ensure good signal quality and transmit them to the edge computing node. A low-noise amplifier (LNA) is used to amplify the sensor signal, and a digital filter is used to remove irrelevant noise and high-frequency interference to ensure signal clarity in the frequency domain. An analog-to-digital converter (ADC) is used to convert the analog signal into a digital signal to provide accurate input for subsequent processing.

[0095] Edge computing nodes use self-supervised learning methods to extract the time, frequency, and spatial domain features of the signal, and use an improved random forest algorithm to analyze the features and identify the vehicle's driving status. The self-supervised learning method extracts meaningful features from the raw data without supervision by designing a suitable objective function. Time domain features: Such as the amplitude and waveform shape of the signal. Frequency domain features: The spectral characteristics of the signal are extracted through the fast Fourier transform (FFT). Spatial domain features: The spatial information in the multi-sensor array can help identify the movement trajectories of different vehicles. The use of the improved random forest algorithm can better cope with the nonlinear relationship between features and improve the accuracy of vehicle recognition.

[0096] The vehicle detection module uses dynamic time warping (DTW) to align the timing deviation of signals when the vehicle is traveling at high speed. When the vehicle is traveling slowly, it uses blind source separation (BSS) to separate the signals of individual vehicles and synchronously outputs vehicle type and weight information through a multi-task learning model.

[0097] At high speeds: A dynamic time warping algorithm is used to correct timing deviations in sensor signals, ensuring that multi-channel signals are aligned on the time axis. An improved random forest algorithm is used to analyze the aligned signals to identify the vehicle type and estimate its weight.

[0098] Slow-moving state: Blind source separation methods (such as independent component analysis (ICA)) are used to separate multiple overlapping vehicle signals. A multi-task deep learning model is used to simultaneously classify vehicle type and estimate weight, ensuring accurate and real-time recognition.

[0099] The data display module displays real-time inspection data on a display device and sends the data to the cloud for further analysis and storage. It displays real-time inspection data, such as vehicle transit time, type, and weight, using visualization tools like charts and text boxes to clearly present the data. Through a data interface with the cloud platform, inspection results are uploaded to the cloud for subsequent big data analysis, statistics, and long-term data storage.

[0100] The communication module enables the coordinated operation of internal modules and exchanges data with external systems. Wireless communication methods (such as Wi-Fi and LoRa) are used to achieve communication between different modules within the system. Data is exchanged with external systems through standard interfaces (such as REST API and MQTT), ensuring real-time data transmission and updates, ensuring communication security, and preventing data leakage or tampering.

[0101] The system integrates multiple key technologies, including a multi-channel piezoelectric sensor array, signal conditioning, edge computing, feature extraction, vehicle type and weight detection, and data display and communication modules. It effectively detects and analyzes vehicle type and weight, adapting to vehicle detection requirements at varying speeds. Self-supervised learning and multi-task deep learning methods enhance the system's recognition accuracy and real-time performance.

[0102] In this system, the connections between the various parts can be achieved through:

[0103] The multi-channel piezoelectric sensor array is responsible for collecting the piezoelectric signals generated by passing vehicles. Each sensor is connected to a signal conditioner, which transmits the collected raw signal to the signal conditioner via analog signal transmission.

[0104] The signal conditioner performs preliminary conditioning on the signal (such as amplification, filtering, etc.) and may perform secondary conditioning (such as noise removal, frequency smoothing, etc.) to ensure that the signal quality is suitable for subsequent processing.

[0105] The signal processed by the signal conditioner is transmitted to the edge computing node via a data bus or wired / wireless communication.

[0106] The edge computing nodes will further analyze and extract features from these signals, extract time domain, frequency domain, and spatial domain features through self-supervised learning, and use an improved random forest algorithm for analysis to identify the vehicle's driving status.

[0107] Based on the analysis results of the edge computing node, after receiving the relevant data, the vehicle detection module applies the dynamic time warping (DTW) method to align the timing deviations of multi-channel signals during high-speed driving to ensure the temporal consistency of the signals.

[0108] For slow-moving vehicles, the vehicle detection module uses a blind source separation method to separate the piezoelectric signal components of a single vehicle, and simultaneously outputs the vehicle type classification results and weight estimation values ​​based on a multi-task deep learning model.

[0109] The detection results of the vehicle detection module (such as vehicle type, weight, etc.) are sent to the data display module through the built-in communication interface.

[0110] The data display module displays the data results of the vehicle detection process in real time, including vehicle type, weight detection results, etc., and provides a visual interface for users to view.

[0111] The display device is used to display the system's detection data in real time and send this data to the cloud for subsequent analysis and storage.

[0112] Data is transmitted wirelessly or wired through the communication module, ensuring that the test results can be uploaded to the cloud in a timely manner for further processing, storage and analysis.

[0113] The communication module not only connects the different modules within the system (such as signal conditioners, edge computing nodes, vehicle detection modules, and data display modules), but also interacts with external systems. This can be achieved through Wi-Fi, Bluetooth, 4G / 5G networks, or other IoT communication protocols to achieve real-time data transmission and sharing.

[0114] Overall connection structure: sensor array → signal conditioner → edge computing node → vehicle detection module → data display module. The signal conditioner, edge computing node, and vehicle detection module exchange and synchronize data through the communication module. The data display module displays the data to the user and sends the data to the cloud through the communication module.

[0115] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is clearly not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A vehicle weight and type detection method based on a piezoelectric sensor, characterized in that: The following steps are involved: Step 1: collecting piezoelectric signals generated by a vehicle passing by using a multi-channel piezoelectric sensor array, and performing preliminary signal conditioning on the collected signals using a signal conditioner. The multi-channel piezoelectric sensor array achieves synchronous collection of multi-channel signals through a synchronous trigger mechanism; Step 2: Input the preliminarily conditioned signal to the edge computing node for processing. The edge computing node uses a self-supervised learning method to extract the time domain, frequency domain, and spatial domain features of the input signal, and uses an improved random forest algorithm to analyze the extracted time domain, frequency domain, and spatial domain features to identify the vehicle's driving state, which includes high-speed driving state and slow driving state. Step 3: Perform secondary signal conditioning on the collected signal based on the identified vehicle driving state; Step 4: Perform vehicle weight and vehicle type detection on the secondary conditioned signal based on the identified vehicle driving status. When the vehicle is driving at high speed, the dynamic time warping method is used to align the timing deviation of the multi-channel signal, and the vehicle type and weight detection results are output based on the multi-task Bayesian regression model. When the vehicle is driving slowly, the blind source separation method is used to separate the piezoelectric signal components of a single vehicle, and the multi-task deep learning model synchronously outputs the vehicle type classification results and weight estimation values.

2. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In step one, the signal conditioner includes a filtering unit, a gain amplification unit, an analog-to-digital conversion unit, and an environmental parameter compensation unit. The filtering unit uses a bandpass filter to eliminate ground vibration noise and electromagnetic noise. The gain amplification unit dynamically adjusts the gain according to the signal strength. The analog-to-digital conversion unit uses an analog-to-digital converter to convert the amplified analog piezoelectric signal into a digital piezoelectric signal. The environmental parameter compensation unit offsets the temperature drift effect of the piezoelectric signal through an NTC thermistor. The synchronous trigger mechanism generates a timestamp through an internal clock to mark the piezoelectric signal collected by each sensor in the multi-channel piezoelectric sensor array.

3. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In the step 2, the self-supervised learning method obtains the time difference sequence of the piezoelectric signals of the multi-axle wheels by calculating the duration of the single tire signal, thereby converting the piezoelectric signal into a time series feature matrix, and generates a time-frequency feature matrix by performing a short-time Fourier transform on the multi-channel piezoelectric signal. Then, a spatial position feature matrix is ​​constructed according to the physical layout of the multi-channel piezoelectric sensor array, and a vehicle geometric model is fitted based on the multi-point pressure distribution of the piezoelectric signal to calculate the characteristic wheelbase.

4. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In step 2, the working method of the improved random forest algorithm includes the following steps: S201, inputting the extracted time domain, frequency domain, and spatial domain features into an improved random forest algorithm, and dividing the historical input feature data into a training set and a test set, wherein the training set is used to train the improved random forest algorithm, and the test set is used to evaluate the performance of the algorithm; S202. For each decision tree, a dynamic feature selection mechanism and a sliding window are used to select the optimal splitting feature from the input features. The dynamic feature selection mechanism calculates local statistics of the input time domain features through windowing, dynamically evaluates the importance of the time domain features, and dynamically adjusts the weights of the frequency domain features based on the distribution of the main vibration frequencies. The importance ranking of the spatial domain features is then dynamically adjusted using the covariance matrix of the multi-channel piezoelectric sensor array. S203, when splitting a node in each decision tree, defining a splitting criterion based on the time domain, frequency domain, and spatial domain characteristics of the piezoelectric signal, and evaluating the node splitting effects at different time scales to select the optimal splitting point to split into left and right child nodes; S204, repeating S202 and S203 until the stop condition is met; S205. Construct multiple decision trees to form a random forest. Use the training set generated by self-sampling to train each decision tree. Use an incremental learning strategy to automatically trigger algorithm updates. During the training process of each decision tree, dynamically adjust the sample weights of the training set, and record the output results and feature importance. S206. For the input time domain, frequency domain and space domain features, each decision tree outputs a prediction result, and the category with the most output results among all decision trees is selected as the final result through confidence-weighted voting.

5. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In the step three, for the high-speed state of the vehicle, a lightweight convolutional neural network is used to extract transient features; for the slow state of the vehicle, a three-layer wavelet packet decomposition is used to extract steady-state features.

6. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In step four, the dynamic time warping method takes the piezoelectric signal of each sensor channel as a time series, calculates the distance matrix between different piezoelectric signals, and uses a dynamic programming method to find the optimal path in the distance matrix. The optimal path is the path with the smallest sum of distances from the starting point of the distance matrix to the end point of the distance matrix. The piezoelectric signals of different sensor channels are stretched or compressed according to the optimal path to achieve time series alignment. The multi-task Bayesian regression model includes an input layer, a hidden layer and an output layer. The input layer is used to receive the aligned multi-channel piezoelectric signal, the hidden layer is used to extract and transform the signal, and the output layer is used to output the prediction results of vehicle type and weight. The multi-task Bayesian regression model updates parameters according to the Bayesian theorem to minimize the prediction error.

7. The vehicle weight and type detection method based on a piezoelectric sensor according to claim 1, characterized in that: In step four, the blind source separation method uses a linear transformation matrix to transform the piezoelectric mixed signal into a source signal, and realizes the separation of the mixed signal by maximizing the statistical independence of the source signal. The multi-task deep learning model uses a convolutional neural network and a long short-term memory network to model the dynamic characteristics of the vehicle, and simultaneously learns the two tasks of vehicle type classification and weight estimation by sharing some network layers. The convolutional neural network receives the single-channel signal after blind source separation, and uses one-dimensional convolution to capture the local time-frequency pattern, and then uses maximum pooling to reduce the temporal resolution. The long short-term memory network dynamically weights the feature contributions of different time steps by introducing a self-attention layer.

8. A vehicle weight and type detection system based on a piezoelectric sensor, characterized in that: A vehicle weight and type detection method based on a piezoelectric sensor as described in any one of claims 1 to 7, comprising: A multi-channel piezoelectric sensor array is used to collect piezoelectric signals generated by vehicles passing by, and a synchronous trigger mechanism is used to ensure synchronous signal collection; The signal conditioner is used to perform preliminary and secondary conditioning on the original signals collected by the sensor and transmit the conditioned signals to the edge computing node for further processing; The edge computing node uses a self-supervised learning method to extract the time, frequency, and spatial domain features of the signal, and uses an improved random forest algorithm to analyze the extracted features to identify the vehicle's driving status; The vehicle detection module uses dynamic time warping to align the timing deviations of multi-channel signals when the vehicle is traveling at high speeds, and outputs vehicle type and weight detection results based on a multi-task Bayesian regression model. When the vehicle is traveling slowly, blind source separation is used to separate the piezoelectric signal components of a single vehicle, and a multi-task deep learning model simultaneously outputs vehicle type classification results and weight estimates. The data display module displays the detection process data in real time through the display device and sends the data to the cloud for back-end analysis and storage.

Images