A dynamic weighing method and system

Abstract

This invention relates to a dynamic weighing method and system, belonging to the field of measurement technology, and solves the problems of vibration interference and temperature drift in existing methods. The method includes: acquiring vibration and temperature signals, wherein shallow feature signals of the vibration signals are obtained using wavelet packet feature extraction; establishing a neural network and training the neural network using historical vibration and temperature signals to generate a neural network model; inputting the current shallow feature signal and current temperature signal into the neural network model to dynamically adjust the dynamic filter parameters; and generating a dynamic filter based on the dynamic filter parameters, and inputting a pressure signal into the dynamic filter to obtain a dynamic weighing result. After the signal collected by the pressure sensor is passed through an adaptive filter with variable filter parameters, an accurate weighing result that overcomes vibration interference and has temperature compensation is finally obtained, thereby accelerating the dynamic weighing speed.

Description

Technical Field

[0001] This invention relates to the field of measurement technology, and in particular to a dynamic weighing method and system. Background Technology

[0002] With the improvement of modern people's quality of life and pursuit of efficiency, the application scenarios of dynamic weighing are increasing. For example, to achieve non-stop toll collection for trucks at highway entrances, non-stop load detection, i.e., dynamic measurement of vehicle weight, is essential. Another example is in the express delivery industry, where modern sorting centers are all fully automated assembly lines, which relies heavily on dynamic measurement of package weight. Using static measurement would significantly reduce sorting efficiency. Dynamic weighing is more difficult and costly to implement than static weighing. During dynamic weighing, numerous interference factors exist, such as conveyor belt vibration and environmental interference. Therefore, special precision sensors are needed to address these issues. However, these special dynamic weighing sensors are very expensive. Therefore, designing a method that can overcome vibration and environmental interference through software algorithms is crucial.

[0003] Currently, there are three main methods to overcome dynamic weighing interference: one is to improve the mechanical structure and sensor structure, another is to achieve this through conventional filter structures, and the third is to achieve this through filter circuits. Summary of the Invention

[0004] Based on the above analysis, the embodiments of the present invention aim to provide a dynamic weighing method and system to solve the problems of vibration interference and temperature drift in existing methods.

[0005] On one hand, embodiments of the present invention provide a dynamic weighing method, comprising: acquiring vibration signals and temperature signals, wherein shallow feature signals of the vibration signals are obtained using wavelet packet feature extraction; establishing a neural network and training the neural network using historical vibration signals and historical temperature signals to generate a neural network model; inputting the current shallow feature signals and the current temperature signals into the neural network model to dynamically adjust the dynamic filter parameters; and generating and acquiring a dynamic filter based on the dynamic filter parameters, and inputting a pressure signal into the dynamic filter to obtain a dynamic weighing result.

[0006] The beneficial effects of the above technical solution are as follows: The filter parameters are dynamically adjusted using a neural network model, where the filter parameters change with the signals from the vibration and temperature sensors to remove interference from dynamically changing vibration signals and temperature drift. After the signal collected by the pressure sensor is passed through an adaptive filter with variable filter parameters, an accurate weighing result that overcomes vibration interference and includes temperature compensation is finally obtained, thereby accelerating the dynamic weighing speed.

[0007] Based on a further improvement of the above method, the filter parameters change with the changes in the current vibration signal acquired in real time by the vibration sensor; and the filter parameters change with the changes in the current temperature signal acquired in real time by the temperature sensor.

[0008] Further improvements to the above method, obtaining the shallow feature signal of the vibration signal using wavelet packet feature extraction further include: sampling and performing wavelet packet transformation on the historical vibration signal using a positive coefficient wavelet filter to obtain tree-structured wavelet packet coefficients; selecting the optimal basis using an information entropy cost function; normalizing the optimal basis and using a reconstruction algorithm to obtain a reconstructed signal; and calculating the corresponding energy value signal based on the reconstructed signal, wherein the energy value signal is the shallow feature signal.

[0009] Based on a further improvement of the above method, the neural network model is a parallel 3-layer neural network model. The process of establishing the neural network and training it using historical vibration signals and historical temperature signals to generate the neural network model further includes: dividing the data of the historical shallow feature signals and the historical temperature signals into a training set and a test set; training the parallel 3-layer neural network according to the training set to generate the parallel 3-layer neural network model; and inputting the test set into the parallel 3-layer neural network model to obtain dynamic filter parameters.

[0010] Further improvements to the above method, training a parallel 3-layer neural network based on the training set to obtain the parallel 3-layer neural network model further include: inputting historical shallow feature signals into the vibration signal branch of the parallel 3-layer neural network model and inputting historical temperature signals into the temperature signal branch of the parallel 3-layer neural network model; merging the outputs of the vibration signal branch and the temperature signal branch into a multi-dimensional vector after passing them through a pooling layer; using the multi-dimensional vector as the input of the third convolutional layer; and sequentially inputting the output of the third convolutional layer into two fully connected layers and an activation function to obtain the parallel 3-layer neural network model.

[0011] A further improvement to the above method, which involves generating a dynamic filter based on the dynamic filter parameters and inputting the pressure signal into the dynamic filter to obtain a dynamic weighing result, further includes: generating a finite-length impulse response filter based on the dynamic filter parameters; and inputting the pressure signal collected in real time by the pressure sensor into the finite-length impulse response filter for filtering to generate a dynamic weight value in real time.

[0012] A further improvement to the above method includes, during the training of the parallel 3-layer neural network based on the training set: calculating the error between the dynamic weight value output by the finite-length impulse response filter and the desired weight value; adjusting each weight parameter according to the error using stochastic gradient descent; and when the error is not within the error requirement range, repeatedly training the parallel 3-layer neural network until the error is within the error requirement range, and storing the weight parameters and the dynamic weight value.

[0013] Based on a further improvement of the above method, the vibration sensor is placed on the surface of the conveyor belt or the weighing platform; the pressure sensor and the temperature sensor are combined into a sensor pair, wherein multiple sensor pairs are placed on the bottom surface of the conveyor belt or the weighing platform.

[0014] On the other hand, embodiments of the present invention provide a dynamic weighing system, comprising: a signal acquisition module for acquiring vibration signals, temperature signals, and pressure signals; a feature pre-extraction module for obtaining shallow feature signals of the vibration signals using wavelet packet feature extraction; a neural network training module for establishing a neural network and training the neural network using historical vibration signals and historical temperature signals to generate a neural network model; a dynamic filter parameter generation module for inputting the current shallow feature signals and the current temperature signals into the neural network model to dynamically adjust the dynamic filter parameters; and a dynamic filter for generating and obtaining a dynamic filter based on the dynamic filter parameters, and inputting the pressure signal into the dynamic filter to obtain a dynamic weighing result.

[0015] Based on further improvements to the above system, the filter parameters change with the vibration signal acquired in real time by the vibration sensor; and the filter parameters change with the temperature signal acquired in real time by the temperature sensor.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0017] 1. This invention enables high-precision dynamic weighing in complex environments. Due to the application of a neural network model in this method, the system has a certain degree of autonomous learning ability. Under complex and ever-changing conditions, it can calculate the acquired data through this method, thereby quickly obtaining more accurate measurement results.

[0018] 2. The filter parameters are dynamically adjusted using a neural network model, where the filter parameters change with the signals from the vibration sensor and temperature sensor.

[0019] 3. After the signal collected by the pressure sensor is filtered, an accurate weighing result with temperature compensation and vibration interference is finally obtained, thereby accelerating the dynamic weighing speed.

[0020] This invention enables high-precision dynamic weighing in complex environments. Due to the application of a neural network model in this method, the system has a certain degree of autonomous learning ability. Under complex and ever-changing conditions, it can calculate the acquired data through this method, thereby quickly obtaining relatively accurate measurement results.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0023] Figure 1 This is a flowchart of a dynamic weighing method according to an embodiment of the present invention.

[0024] Figure 2 This is a flowchart of a high-precision dynamic weighing method according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a parallel neural network structure according to an embodiment of the present invention.

[0026] Figure 4 This is a block diagram of a dynamic filter according to an embodiment of the present invention.

[0027] Figure 5 This is a block diagram of a dynamic weighing system according to an embodiment of the present invention. Detailed Implementation

[0028] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which constitute a part of the present invention and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0029] A specific embodiment of the present invention discloses a dynamic weighing method. (See reference...) Figure 1The dynamic weighing method includes: in step S102, acquiring vibration signals and temperature signals, wherein shallow feature signals of the vibration signals are obtained using wavelet packet feature extraction; in step S104, establishing a neural network and training the neural network using historical vibration signals and historical temperature signals to generate a neural network model; in step S106, inputting the current shallow feature signals and current temperature signals into the neural network model to dynamically adjust the dynamic filter parameters; and in step S108, generating and acquiring a dynamic filter based on the dynamic filter parameters, and inputting the pressure signal into the dynamic filter to obtain a dynamic weighing result.

[0030] Compared with existing technologies, the dynamic weighing method provided in this embodiment dynamically adjusts filter parameters through a neural network model. These filter parameters change with the signals from the vibration and temperature sensors to remove interference from dynamically changing vibration signals and temperature drift. After the signal collected by the pressure sensor is passed through an adaptive filter with variable filter parameters, an accurate weighing result that overcomes vibration interference and includes temperature compensation is obtained, thereby accelerating the dynamic weighing speed.

[0031] The following text will refer to Figures 1 to 3 The steps of the dynamic weighing method according to an embodiment of the present invention will be described in detail.

[0032] Vibration sensors are placed on the surface of the conveyor belt or weighing platform; pressure sensors and temperature sensors are combined into sensor pairs, wherein multiple sensor pairs are placed on the bottom surface of the conveyor belt or weighing platform.

[0033] refer to Figure 1 and Figure 2 In step S102, vibration signals and temperature signals are acquired, wherein shallow feature signals of the vibration signals are obtained using wavelet packet feature extraction. Specifically, obtaining shallow feature signals of the vibration signals using wavelet packet feature extraction further includes: sampling and performing wavelet packet transformation on historical vibration signals using a positive coefficient wavelet filter to obtain tree-structured wavelet packet coefficients; selecting the optimal basis (e.g., the optimal basis is 8) using an information entropy cost function; normalizing the optimal basis and using a reconstruction algorithm to obtain the reconstructed signal; and calculating the corresponding energy value signal based on the reconstructed signal, wherein the energy value signal is the shallow feature signal.

[0034] In step S104, a neural network is established and trained using historical vibration and temperature signals to generate a neural network model. For example, refer to... Figure 3The neural network model is a parallel 3-layer neural network model. The process of establishing the neural network and training it using historical vibration and temperature signals to generate the neural network model further includes: dividing the historical shallow feature signals and historical temperature signals into training and testing sets; training the parallel 3-layer neural network based on the training set to generate the parallel 3-layer neural network model; and inputting the testing set into the parallel 3-layer neural network model to obtain dynamic filter parameters. The process of training the parallel 3-layer neural network based on the training set to obtain the parallel 3-layer neural network model further includes: inputting historical shallow feature signals into the vibration signal branch of the parallel 3-layer neural network model and inputting historical temperature signals into the temperature signal branch of the parallel 3-layer neural network model; merging the outputs of the vibration signal branch and the temperature signal branch into a multi-dimensional vector after passing them through pooling layers; using the multi-dimensional vector as the input to the third convolutional layer; and sequentially inputting the output of the third convolutional layer into two fully connected layers and an activation function to obtain the parallel 3-layer neural network model.

[0035] In step S106, the current shallow feature signal and the current temperature signal are input into the neural network model to dynamically adjust the dynamic filter parameters. The filter parameters change with the changes in the current vibration signal acquired in real time by the vibration sensor; and the filter parameters change with the changes in the current temperature signal acquired in real time by the temperature sensor.

[0036] In step S108, a dynamic filter is generated based on the dynamic filter parameters, and the pressure signal is input into the dynamic filter to obtain a dynamic weighing result. Specifically, generating and obtaining a dynamic filter based on the dynamic filter parameters and inputting the pressure signal into the dynamic filter to obtain a dynamic weighing result further includes: generating a finite-length impulse response filter based on the dynamic filter parameters; and inputting the pressure signal collected in real time by the pressure sensor into the finite-length impulse response filter for filtering to generate a dynamic weight value in real time. Before using the dynamic filter and training the neural network device, the dynamic filter is adjusted. Specifically, during the training of the parallel 3-layer neural network based on the training set, the process further includes: calculating the error between the dynamic weight value output by the finite-length impulse response filter and the desired weight value; adjusting each weight parameter according to the error using stochastic gradient descent; and when the error is not within the required error range, repeatedly training the parallel 3-layer neural network until the error is within the required error range, and storing the weight parameters and the dynamic weight value. When the required error range decreases, the training time for the parallel 3-layer neural network increases, and the accuracy of the dynamic filter improves. When the required error range increases, the training time for the parallel 3-layer neural network decreases, and the accuracy of the dynamic filter decreases.

[0037] Another specific embodiment of the present invention discloses a dynamic weighing system. (See reference...) Figure 5According to an embodiment of the present invention, the dynamic weighing system includes: a signal acquisition module 502, a feature pre-extraction module 504, a neural network training module 506, a dynamic filter parameter generation module 508, and a dynamic filter 510.

[0038] The signal acquisition module 502 is used to acquire vibration signals, temperature signals, and pressure signals, including a vibration sensor for acquiring vibration signals, a temperature sensor for acquiring temperature signals, and a pressure sensor for acquiring pressure signals. The feature pre-extraction module 504 is used to obtain shallow feature signals of the vibration signals using wavelet packet feature extraction. The neural network training module 506 is used to build a neural network and train the neural network using historical vibration signals and historical temperature signals to generate a neural network model. The dynamic filter parameter generation module 508 is used to input the current shallow feature signals and the current temperature signals into the neural network model to dynamically adjust the dynamic filter parameters. Specifically, the filter parameters change with the vibration signals acquired in real time by the vibration sensor; and the filter parameters change with the temperature signals acquired in real time by the temperature sensor. The dynamic filter 510 is used to generate a dynamic filter based on the dynamic filter parameters and input the pressure signal into the dynamic filter to obtain dynamic weighing results.

[0039] The following text will refer to Figures 1 to 3 The dynamic weighing method according to embodiments of the present invention will be described in detail with specific examples.

[0040] In view of this, the first aspect of this application provides a high-precision dynamic weighing method based on neural networks, comprising:

[0041] refer to Figure 2 Data is acquired using vibration and temperature sensors. The analog signals are sampled and discretized using an A / D converter to generate a training set. A positive coefficient wavelet filter is used to sample and perform wavelet packet transform on the acquired pressure and vibration sensor signals, resulting in a tree-structured wavelet packet coefficient. An information entropy cost function is used to select the optimal basis. The selected basis is normalized, and a reconstruction algorithm is applied to obtain the reconstructed signal. The energy values ​​of the reconstructed signal are calculated, and these energy values ​​are used as the extracted feature signals, also known as feature vectors.

[0042] Specifically, to obtain the datasets needed for training and testing, a pressure sensor can be used to collect pressure signals to obtain p(t), a vibration sensor can be used to collect vibration signals to obtain z(t), and a temperature sensor can be used to collect temperature signals to obtain t(t). After passing these signals through an analog-to-digital converter, the corresponding discrete-time sequences are obtained, namely the pressure signal sequence p(t), the vibration signal sequence z(n), and the temperature signal sequence t(n), which are convenient for computer processing.

[0043] The vibration signal acquired by the vibration sensor is decomposed into three layers of wavelet packets, and a real coefficient filter is used. n} n∈Z , {g n} n∈Z g n =(-1) n h 1-n A set of wavelet functions defined by μ0, μ1, h, and g is as follows:

[0044]

[0045] When n = 0,

[0046] t is the time series (t is n of z(n)), and k is the time series shift factor.

[0047] Define by recursion The determined wavelet packet.

[0048]

[0049]

[0050]

[0051] In this context, the subscript n represents the wavelet layer.

[0052] After performing a 3-level wavelet packet decomposition, the coefficients of 8 sub-bands (i.e., basis) are obtained, which are respectively

[0053] The coefficients obtained from wavelet decomposition are reconstructed using the following formula:

[0054]

[0055] The sub-band signal is represented as The total signal S can be expressed as

[0056] By reconstructing the signal, the corresponding energy signal is calculated as follows:

[0057] The resulting vector is the extracted feature vector P.

[0058] The shallow features P extracted by wavelets are used as input, and a parallel 3-layer neural network is used to extract the hyperparameters of the adaptive filter.

[0059] The aforementioned feature vectors are input into the first convolutional layer of a parallel 3-layer convolutional neural network. The output of the first convolutional layer, after passing through an activation function, serves as the input to the second convolutional layer. Similarly, the temperature sensor signal is input into the first convolutional layer, and its output, after passing through an activation function, serves as the input to the second convolutional layer. The outputs of the second convolutional layers containing the vibration signal and temperature signals, after passing through pooling layers, are merged into a single multidimensional vector. This multidimensional vector serves as the input to the third convolutional layer. After passing through the convolutional layer and activation function, the output of the third convolutional layer passes through two fully connected layers and another activation function to obtain the trained parameters, which are then used as the parameters for the finite-length impulse response filter. The pressure sensor signal is passed through the filter, and the error between the filter output and the actual weight value is calculated. Stochastic gradient descent is used to adjust the weight parameters. If the error requirement is not met, the above steps are repeated for further training. When the error requirement is met, the corresponding result is output, and the corresponding weight parameter values ​​and filter parameters are saved. After the pressure sensor signal is passed through a filter, the error between the final output and the expected value is calculated. Gradient descent is used to adjust the weight parameters. If the error requirement is not met, the above steps are repeated to continue training. When the error requirement is met, the corresponding result is output and the corresponding weight parameter values ​​are saved. In the convolutional layers, the elements in the weight vector are set to 0 with a certain probability of deactivation to enhance the network's generalization ability. The activation function in this network is y(x) = max(0,x).

[0060] Use the PyTorch deep learning framework to build a neural network model.

[0061] refer to Figure 2 and Figure 3 The first convolutional layer expands the feature vector extracted in step one to obtain T. 1×10 =[0|P|0], N 1×10 =[0|Q|0][T 1×10 N 1×10 As the input data for the first convolutional layer, 32 sets of 1×3 convolutional kernels w1 (the kernels are randomly generated) and feature vector T are used. 1×10 and eigenvector N 1×10 Convolutions are performed separately with the feature vector T. 1×10 The vector obtained after the convolution operation is With eigenvector N 1×10 The vector obtained after the convolution operation is For ease of representation, and W 1T and W 1N The m columns of elements corresponding to the vector. The output of the first layer is... and

[0062] The second convolutional layer uses the output data of the first layer as its input data, and employs 64 sets of 1×3 convolutional kernels w2. in The operation is the same as above. The output of the second layer is... and in and Both are 1×8 vectors.

[0063] The pooling layer takes the output of the second layer as input. (For ease of representation) and They represent and The m-th element of the vector.

[0064]

[0065]

[0066]

[0067] W3 = [W 3T |W 3N Generate a 2×4×64 matrix.

[0068] The fourth layer is a convolutional layer. It takes the output W3 of the pooling layer as input and performs convolutions with W3 using 64 2×4 convolutional kernels w4. For ease of representation, W... 3(x,y,z) This represents the element with coordinates (x, y, z) in the 3D matrix W3. The output of this layer is represented as...

[0069]

[0070] Layer 5, fully connected layer I. The weights of fully connected layer I consist of 32 1×64 vectors, denoted by f1. The input is the output W4 of the fourth convolutional layer, and the output of this layer is...

[0071] Layer 6, Fully Connected Layer II. The weights of Fully Connected Layer II are 16 vectors of 32 each, denoted as f2. The input is the output F1 of the fifth convolutional layer. The output of this layer is... F2 is used as the coefficient of the filter.

[0072] refer to Figure 2 and Figure 4 The signal collected by the pressure sensor is filtered to obtain an accurate weighing result that overcomes vibration interference and has temperature compensation, thereby accelerating the dynamic weighing speed.

[0073] Considering that the algorithm can be implemented on distributed computing devices such as FPGAs in the future, the filter adopts a non-recursive finite-length unit impulse response filter structure. This filter structure is convenient for implementing a pipelined structure on FPGAs, which can accelerate the parallel processing speed of filtering.

[0074] The output of the signal after passing through the adaptive filter is: In this expression, n is the time series index, N is the filter order, X represents the input signal sequence, and F2 is the weight vector in the adaptive filter. This weight vector F2 is... Figure 4 The numbers shown are b0, b1, b2, and b. N .

[0075] After applying mean filtering to the output signal, the final output result is:

[0076] Once the final result is obtained, it can be stored in the backend database along with the label of the measured object through the transmission system or displayed in real time in the visualization system.

[0077] A dynamic weighing system for rapid dynamic measurement includes: a signal acquisition module for acquiring pressure sensor signals, vibration signals, and temperature signals to generate training sets and real-time measurement data; a feature pre-extraction module for obtaining shallow features of the signals to accelerate training, using a positive coefficient wavelet filter to sample and perform wavelet packet transform on the acquired pressure sensor and vibration sensor signals to obtain tree-structured wavelet packet coefficients; an information entropy cost function to select the optimal basis; normalization of the selected basis and a reconstruction algorithm to obtain the reconstructed signal; and calculation of the corresponding energy value from the reconstructed signal; and a neural network training module that inputs the features extracted from the vibration signal through wavelet packet features into the first convolutional layer, and the output of the first convolutional layer is passed through a function layer as the input to the second convolutional layer. Similarly, the temperature sensor signal is input into the first convolutional layer, and the output of the first convolutional layer is passed through a function layer as the input to the second convolutional layer. The outputs of the second convolutional layer of the vibration signal layer and the second convolutional layer of the temperature signal layer, after passing through an activation function, are used as the input vector for a pooling layer. These outputs are then combined into a multi-dimensional vector, which is used as the input to the third convolutional layer. After passing through the convolutional layer and an activation function, the output of the third convolutional layer passes through two fully connected layers and another activation function to obtain the final result. The elements of the weight vectors in the convolutional layers are set to 0 with a certain probability of deactivation to enhance the network's generalization ability. The activation function in this network is y(x) = max(0,x). The filtering module uses the results of the neural network training module as parameters and the data obtained from the signal acquisition device and feature pre-extraction module, passing it through a finite impulse response filter to obtain the final result. The data transmission module transmits the output of the filtering module to a database for storage and subsequent processing.

[0078] The system acquires relevant data through a signal acquisition module, uses the data acquired by the signal acquisition module as input to a feature pre-extraction module to obtain shallow features of the signal, uses the extracted shallow features as input to a neural network training module, obtains filter parameters after training, inputs the data from the temperature sensor and pressure sensor into the filter with adjusted parameters, and obtains the final weighing result after filtering.

[0079] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0080] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic weighing method, characterized in that, The method comprises: obtaining a vibration signal and a temperature signal, wherein a shallow feature signal of the vibration signal is obtained by using a wavelet packet feature extraction method; establishing a neural network and training the neural network by using historical vibration signals and historical temperature signals to generate a neural network model; inputting a current shallow feature signal and a current temperature signal into the neural network model to dynamically adjust a dynamic filter parameter, wherein the filter parameter changes with a change in a current vibration signal collected by a vibration sensor in real time, and the filter parameter changes with a change in a current temperature signal collected by a temperature sensor in real time; and obtaining an adaptive dynamic filter with a variable filter parameter based on the dynamic filter parameter, and inputting a pressure signal into the adaptive dynamic filter to obtain an accurate dynamic weighing result that overcomes vibration interference and has temperature compensation, wherein obtaining an adaptive dynamic filter based on the dynamic filter parameter and inputting a pressure signal into the adaptive dynamic filter to obtain a dynamic weighing result further comprises: generating a finite impulse response filter based on the dynamic filter parameter; and inputting a pressure signal collected by a pressure sensor in real time into the finite impulse response filter for filtering to generate a dynamic weight value in real time.

2. The dynamic weighing method according to claim 1, characterized in that, The method of obtaining a shallow feature signal of the vibration signal by using a wavelet packet feature extraction method further comprises: sampling the historical vibration signal, performing wavelet packet transformation, and obtaining wavelet packet coefficients in a tree structure by using a positive coefficient wavelet filter; selecting an optimal basis by using an information entropy cost function; performing normalization processing on the optimal basis and obtaining a reconstruction signal by using a reconstruction algorithm; calculating a corresponding energy value signal based on the reconstruction signal, wherein the energy value signal is the shallow feature signal.

3. The dynamic weighing method of claim 1, wherein, The neural network model is a parallel 3-layer neural network model, and the method of establishing a neural network and training the neural network by using historical vibration signals and historical temperature signals to generate a neural network model further comprises: dividing data of the historical shallow feature signal and the historical temperature signal into a training set and a test set; and training a parallel 3-layer neural network according to the training set to generate the parallel 3-layer neural network model; inputting the test set into the parallel 3-layer neural network model to obtain a dynamic filter parameter.

4. The dynamic weighing method according to claim 3, characterized in that, The method of training a parallel 3-layer neural network according to the training set to obtain the parallel 3-layer neural network model further comprises: inputting a historical shallow feature signal into a vibration signal branch of the parallel 3-layer neural network model and inputting a historical temperature signal into a temperature signal branch of the parallel 3-layer neural network model; merging outputs of the vibration signal branch and the temperature signal branch into a multi-dimensional vector after each output passes through a pooling layer; inputting the multi-dimensional vector as an input of a third convolutional layer; and sequentially inputting an output of the third convolutional layer into two fully connected layers and an activation function to obtain the parallel 3-layer neural network model. The method further comprises, in the process of training the parallel 3-layer neural network according to the training set:

5. The dynamic weighing method according to claim 4, characterized in that, ​ calculating an error between a dynamic weight value output by the finite impulse response filter and an expected weight value; adjusting each weight parameter in a manner of random gradient descent according to the error; and repeating training of the parallel 3-layer neural network when the error is not within a range of error requirement until the error is within the range of error requirement, and storing the weight parameter and the dynamic weight value.

6. The dynamic weighing method of claim 1, wherein, a vibration sensor is arranged on a surface of a conveyor belt or a weighing platform; the pressure sensor and the temperature sensor are arranged as a sensor pair, and a plurality of sensor pairs are arranged on a bottom surface of the conveyor belt or the weighing platform.

7. A dynamic weighing system, characterized in that comprising: a signal acquisition module, configured to acquire a vibration signal, a temperature signal and a pressure signal; a feature pre-extraction module, configured to obtain a shallow feature signal of the vibration signal by using a wavelet packet feature extraction method; a neural network training module, configured to establish a neural network and train the neural network by using historical vibration signals and historical temperature signals to generate a neural network model; a dynamic filter parameter generation module, configured to input a current shallow feature signal and a current temperature signal into the neural network model to dynamically adjust a dynamic filter parameter, wherein the filter parameter changes with a change in a vibration signal acquired by a vibration sensor in real time, and the filter parameter changes with a change in a temperature signal acquired by a temperature sensor in real time; and an adaptive dynamic filter, configured to obtain an adaptive dynamic filter with a variable filter parameter based on the dynamic filter parameter, and input a pressure signal into the adaptive dynamic filter to obtain a dynamic weighing result, wherein obtaining an adaptive dynamic filter based on the dynamic filter parameter and inputting a pressure signal into the adaptive dynamic filter to obtain a dynamic weighing result further comprises: generating a finite impulse response filter based on the dynamic filter parameter; and inputting a pressure signal acquired by a pressure sensor in real time into the finite impulse response filter for filtering to generate a dynamic weight value in real time.

Images