CN102506983A

CN102506983A - Weighing error automatic compensation method of vehicle scale

Info

Publication number: CN102506983A
Application number: CN2011103351631A
Authority: CN
Inventors: 林海军; 滕召胜; 杨进宝; 汪鲁才; 李仲阳; 谭旗; 迟海; 刘让周; 郑丹
Original assignee: Hunan Normal University
Current assignee: Hunan Normal University
Priority date: 2011-10-31
Filing date: 2011-10-31
Publication date: 2012-06-20

Abstract

The invention discloses a high-accuracy weighing and error automatic compensation method for a truck scale, which includes three aspects: compound neural network construction, neural network off-line training and truck scale online weighing. First, construct three sub-neural networks based on prior knowledge; then use standard weights of different tonnages and weighing signal acquisition circuits to train samples, transmit them to the host computer, and use training software to complete the off-line training of the three sub-neural networks to obtain the corresponding parameter, and download it to the lower computer to prepare for the on-line measurement of the truck scale; when the truck scale is on-line to measure, the lower computer first obtains the weighing signal vector as the input of the three sub-neural networks through the weighing signal acquisition circuit, and calculates the weight of each sub-neural network. The output of the network; roughly estimate the weight of the load to be measured, and automatically obtain the output weight of the composite neural network; weighted and fused the output of the sub-neural network to obtain the final weighing result, and at the same time complete the weighing error compensation of the truck scale. The invention can realize the automatic compensation of the unbalanced load error and the linearity error of the truck scale, and greatly improves the accuracy of the weighing result.

Description

Automatic Compensation Method for Truck Scale Weighing Error

技术领域 technical field

汽车衡称重误差自动补偿方法，本发明涉及一种汽车衡高准确度称重与误差补偿的方法，具体地说，涉及一种利用神经网络方法进行汽车衡高准确称重与误差补偿的方法，属于称重系统检测与信息处理领域，也可应用于其它多传感器系统。Auto-compensation method for truck scale weighing error, the present invention relates to a method for high-accuracy weighing and error compensation of truck scale, in particular, relates to a method for high-accuracy weighing and error compensation of truck scale by using neural network method The invention belongs to the field of weighing system detection and information processing, and can also be applied to other multi-sensor systems.

技术背景 technical background

汽车衡(包括地上衡、地中衡等)作为衡器的重要分支，称量范围由几吨到几百吨甚至上千吨，是厂矿、商家等用于大宗货物计量的主要称重设备，其广泛应用于仓储贸易、交通运输、工矿企业等部门，随着工业生产与交通运输的发展，汽车衡的需求量将越来越大。现有汽车衡除计量功能外，还具有监测、管理等多种功能，其对企业的现代化管理具有重要意义。As an important branch of weighing instruments, truck scales (including ground scales, ground scales, etc.) have a weighing range from a few tons to hundreds of tons or even thousands of tons. It is widely used in warehousing trade, transportation, industrial and mining enterprises and other departments. With the development of industrial production and transportation, the demand for truck scales will increase. In addition to the measurement function, the existing truck scale also has multiple functions such as monitoring and management, which is of great significance to the modern management of enterprises.

由于价格等原因，目前模拟式电子汽车衡占据汽车衡市场的主导地位。现有模拟式电子汽车衡主要由承重传力机构(秤体)、模拟称重传感器、称重显示仪表三大主部件组成，由此可完成汽车衡基本的称重功能，其利用传感器输出信号累加方式，在模拟接线盒中将各路称重传感器的输出信号集中累加，获得一个与被测载荷质量成比例的电压信号，经信号调理、A/D转换后，由单片机处理获得称重结果，送显示、通信，完成被测载荷的称重。汽车衡称重结果的准确度受偏载误差与线性度误差等影响。偏载误差是由于汽车衡受各种非线性因素的影响，被测载荷处于汽车衡承载面上不同位置时，称量结果不一致而产生的误差；线性度误差是由于各路称重传感器的特性不一致，导致汽车衡的输入与输出并非理想的线性关系，从而产生的称重误差。现有汽车衡的偏载误差与线性度误差补偿过程是分开的，偏载误差补偿方法有：(1)通过人工反复调节汽车衡接线盒中电阻器，改变每路传感器通道增益，补偿偏载误差，这种方法人工操作繁琐，工作效率低，补偿效果差；(2)利用位置传感器检测载荷加载在承载器上的位置，根据位置信息实现偏载误差补偿和称重融合，该方法适合小承载器的电子秤，对于大承载器的汽车衡，需要进一步改进，同时该方法需要增加位置传感器，增加了汽车衡的成本，且不易于工程实现；(3)采用多元线性回归方法，即利用标准砝码加载在汽车衡各受力支点上(即称重传感器)，得到一组线性方程，通过求解该方程获得各称重传感器通道的增益系数，从而实现偏载误差补偿，该方法没有考虑汽车衡各种非线性因素影响，补偿效果不明显；(4)采用数字称重传感器实现偏载误差补偿，该方法没有考虑汽车衡各种非线性因素影响，补偿效果差，同时数字称重传感器成本高。现有汽车衡线性度误差补偿是在偏载误差补偿完成后，利用如下方法完成：首先利用不同重量的标准砝码依次加载在汽车衡秤体上，获得目标称重结果与实际称重结果，然后将实际称重结果倍乘修正系数，使之等于目标称重结果。这种方法是在基于汽车衡输入-输出为线性关系的基础上的，与实际相差较大，因此补偿效果较差。Due to price and other reasons, analog electronic truck scales currently occupy a dominant position in the truck scale market. The existing analog electronic truck scale is mainly composed of three main components: the load-bearing force transmission mechanism (scale body), the analog load cell, and the weighing display instrument, which can complete the basic weighing function of the truck scale. It uses the sensor output signal In the accumulative mode, the output signals of each load cell are accumulated in the analog junction box to obtain a voltage signal proportional to the mass of the measured load. After signal conditioning and A/D conversion, the weighing result is processed by the single-chip microcomputer. , send display, communication, and complete the weighing of the measured load. The accuracy of the weighing results of the truck scale is affected by the partial load error and linearity error. The partial load error is due to the influence of various nonlinear factors on the truck scale. When the measured load is at different positions on the truck scale’s bearing surface, the weighing results are inconsistent; the linearity error is due to the characteristics of each load cell. Inconsistency leads to the non-ideal linear relationship between the input and output of the truck scale, resulting in weighing errors. The offset load error and linearity error compensation process of existing truck scales are separated, and the offset load error compensation methods include: (1) By manually adjusting the resistor in the junction box of the truck scale repeatedly, changing the channel gain of each sensor, and compensating for the offset load error, this method is cumbersome manual operation, low work efficiency, and poor compensation effect; (2) Use the position sensor to detect the position of the load loaded on the carrier, and realize offset load error compensation and weighing fusion according to the position information. This method is suitable for small The electronic scale of the loader needs to be further improved for the truck scale with a large loader. At the same time, this method needs to increase the position sensor, which increases the cost of the truck scale, and is not easy to implement in engineering; (3) using the multiple linear regression method, that is, using The standard weights are loaded on each force fulcrum of the truck scale (that is, the load cell), and a set of linear equations is obtained. By solving the equation, the gain coefficient of each load cell channel is obtained, so as to realize the compensation of the unbalanced load error. This method does not consider The compensation effect is not obvious due to the influence of various nonlinear factors of the truck scale; (4) The compensation of the partial load error is realized by using a digital load cell. This method does not consider the influence of various nonlinear factors of the truck scale, and the compensation effect is poor. high cost. The linearity error compensation of the existing truck scale is completed by the following method after the eccentric load error compensation is completed: firstly, the standard weights of different weights are loaded on the truck scale body in sequence to obtain the target weighing result and the actual weighing result, Then multiply the actual weighing result by the correction factor to make it equal to the target weighing result. This method is based on the linear relationship between the input and output of the truck scale, which is quite different from the actual situation, so the compensation effect is poor.

发明内容 Contents of the invention

为克服已有技术的不足，本发明的目的在于提供一种基于复合神经网络的汽车衡高准确度称重与误差补偿方法，该方法能够自动、准确地完成汽车衡称重误差补偿，提高称重结果准确度。In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a high-accuracy weighing and error compensation method for truck scales based on a composite neural network, which can automatically and accurately complete the weighing error compensation for truck scales, and improve the weighing efficiency of truck scales. Accuracy of results.

受承载器的刚度与强度、汽车衡加工与安装过程中产生的内应力与机械形变及尺寸误差、称重传感器灵敏度的分散性等因素影响，同一被测载荷加载在承载器不同位置时，称重结果不一致，即称重结果与加载点位置有关，具有偏载误差。汽车衡具有多路称重传感器，各路称重传感器的输入-输出特性、灵敏度等不一致，导致汽车衡的输入与输出并非线性关系，从而产生线性度误差。汽车衡检定的国家标准《JJG539-97数字指示秤检定规程》规定，汽车衡不同秤量段的最大允许误差不同，因此测量不同载荷时，所要求准确度是不一样的。Affected by factors such as the stiffness and strength of the carrier, the internal stress and mechanical deformation and dimensional error generated during the processing and installation of the truck scale, and the dispersion of the sensitivity of the load cell, when the same measured load is loaded on different positions of the carrier, the weighing The weighing results are inconsistent, that is, the weighing results are related to the position of the loading point, and there is an eccentric load error. The truck scale has multiple load cells, and the input-output characteristics and sensitivities of each load cell are inconsistent, resulting in a non-linear relationship between the input and output of the truck scale, resulting in linearity errors. The national standard "JJG539-97 Digital Indicating Scale Verification Regulations" for truck scale verification stipulates that the maximum allowable error of different weighing sections of truck scales is different, so when measuring different loads, the required accuracy is different.

基于以上分析，本发明提出的汽车衡称重误差自动补偿方法，利用复合神经网络逼近汽车衡输入-输出之间的非线性函数关系，从而完成汽车衡高准确度称重与误差补偿。这种方法包括复合神经网络构造、神经网络离线训练与汽车衡在线称量三个方面，具体如下：Based on the above analysis, the automatic compensation method for the weighing error of the truck scale proposed by the present invention uses a compound neural network to approximate the nonlinear function relationship between the input and output of the truck scale, thereby completing the high-accuracy weighing and error compensation of the truck scale. This method includes three aspects: composite neural network construction, neural network offline training and truck scale online weighing, as follows:

(1)复合神经网络构造(1) Composite neural network structure

根据汽车衡检定的国家标准《JJG539-97数字指示秤检定规程》规定，汽车衡不同秤量段(一般分为3个秤量段)的最大允许误差E_pmax不同，即According to the national standard "JJG539-97 Digital Indicating Scale Verification Regulations" for the verification of truck scales, the maximum allowable error E _pmax of different weighing sections (generally divided into 3 weighing sections) of truck scales is different, that is

${E E.}_{p p max max} = = \{\begin{matrix} 0.5 0.5 e e & M m \leq \leq 500500 e e \\ 1.0 1.0 e e & 500500 e e < < M m \leq \leq 20002000 e e \\ 1.5 1.5 e e & 20002000 e e < < M m \leq \leq {M m}_{max max} \end{matrix} - - - - - - ((11))$

式中，M为汽车衡的秤量；M_max为汽车衡最大秤量；e为汽车衡检定分度值。由式(1)可知，汽车衡不同的秤量段，其最大允许误差不同。因此根据这一误差特征，构造3个子神经网络(子神经网络1、子神经网络2与子神经网络3)，每个子神经网络负责相应秤量段的误差补偿，各子神经网络结构相同，均为一个N输入1输出的三层网络，子神经网络可以是BP神经网络、径向基函数神经网络(RBFNN)或其它前向神经网络。In the formula, M is the weighing capacity of the truck scale; M _max is the maximum weighing capacity of the truck scale; e is the calibration division value of the truck scale. It can be seen from formula (1) that the maximum allowable error is different for different weighing sections of the truck scale. Therefore, according to this error feature, three sub-neural networks (sub-neural network 1, sub-neural network 2, and sub-neural network 3) are constructed, and each sub-neural network is responsible for the error compensation of the corresponding weighing section. The structure of each sub-neural network is the same. A three-layer network with N inputs and one output, and the sub-neural network can be a BP neural network, a radial basis function neural network (RBFNN) or other feed-forward neural networks.

(2)神经网络离线训练(2) Neural network offline training

神经网络离线训练包括子神经网络1、子神经网络2和子神经网络3的训练。利用不同吨位的标准砝码(设标准砝码的重量为y_j，j＝1，2，...，K)，分别加载在汽车衡承载面上的不同位置，系统下位机通过N路相互独立的称重信号采集电路，获得N路称重传感器输出信号S_i(i＝1，2，...，N)，经数据预处理后，获得称重信号向量X_j(j＝1，2，...，K)，并构成训练样本(X_j，y_j)，即The offline training of the neural network includes the training of sub-neural network 1, sub-neural network 2 and sub-neural network 3. Use standard weights of different tonnages (set the weight of the standard weights to be y _j , j=1, 2, ..., K), and load them on different positions on the load-bearing surface of the truck scale, and the lower computers of the system communicate with each other through N channels. The independent weighing signal acquisition circuit obtains N load cell output signals S _i (i=1, 2, ..., N), and after data preprocessing, obtains the weighing signal vector X _j (j=1, 2,...,K), and constitute the training samples (X _j , y _j ), namely

$(({X x}_{j j},, {y the y}_{j j})) = = (\begin{matrix} {x x}_{1111} & {x x}_{21 twenty one} & . . . . . . & {x x}_{N N 11} & {y the y}_{11} \\ {x x}_{1212} & {x x}_{22 twenty two} & . . . . . . & {x x}_{N N 22} & {y the y}_{22} \\ . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . \\ {x x}_{11 j j} & {x x}_{22 j j} & . . . . . . & {x x}_{Nj Nj} & {y the y}_{j j} \\ . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . \\ . . & . . & . . & . . & . . \\ {x x}_{11 K K} & {x x}_{22 K K} & . . . . . . & {x x}_{NK NK} & {y the y}_{K K} \end{matrix}) - - - - - - ((22))$

下位机通过RS232或其它串行通信接口，将训练样本传送至上位机(即PC机)，利用上位机的训练软件分别对子神经网络1、子神经网络2和子神经网络3进行离线训练。训练完成后获得各子神经网络参数，并通过串行通信接口将这些参数下载至下位机，为汽车衡在线测量作准备。The lower computer transmits the training samples to the upper computer (ie PC) through RS232 or other serial communication interfaces, and uses the training software of the upper computer to perform offline training on sub-neural network 1, sub-neural network 2 and sub-neural network 3 respectively. After the training is completed, the parameters of each sub-neural network are obtained, and these parameters are downloaded to the lower computer through the serial communication interface to prepare for the on-line measurement of the truck scale.

称重数据预处理采用采用一种数字滤波方法完成，该方法连续采样Q个称重传感器输出电压，并对其按从小到大顺序排列得到S_im，去掉前、后各M个采样值，对其余Q-2M个S_im进行均值滤波，即The preprocessing of weighing data is completed by a digital filtering method. This method continuously samples the output voltages of Q load cells, and arranges them in ascending order to obtain S _im . The remaining Q-2M _Sims are average filtered, namely

${x x}_{i i} = = \frac{11}{Q Q - - 22 M m} {Σ Σ}_{m m = = M m + + 11}^{Q Q - - M m} {S S}_{im im} - - - - - - ((33))$

汽车衡只有在第一次安装与年检时需要复合神经网络训练，平时称重时无需进行训练。Truck scales only need compound neural network training during the first installation and annual inspection, and no training is required during normal weighing.

(3)汽车衡在线称量与误差补偿(3) Truck scale online weighing and error compensation

子神经网络构造与离线训练完成后，汽车衡即可进行在线称量与误差补偿，其步骤如下：After the sub-neural network construction and offline training are completed, the truck scale can perform online weighing and error compensation. The steps are as follows:

(a)下位机首先通过N路相互独立的称重信号采集电路，获得N路称重传感器输出信号，经式(3)所示的数据预处理后，获得N路称重信号向量X，为在线称量做准备。(a) The lower computer first obtains the output signals of N load cells through N independent weighing signal acquisition circuits, and obtains the N load signal vector X after the data preprocessing shown in formula (3), which is Prepare for online weighing.

(b)将称重信号向量X作为子神经网络1、2、3的输入，计算各子神经网络的输出y_i(i＝1，2，3)。(b) Take the weighing signal vector X as the input of the sub-neural networks 1, 2, and 3, and calculate the output y _i (i=1, 2, 3) of each sub-neural network.

(c)利用称重信号向量X，粗略估计被测载荷的重量，并根据国家标准《JJG539-97数字指示秤检定规程》的规定，通过权值调节器自适应获得复合神经网络输出权值W。设w₁、w₂、w₃为W的分量，则有(c) Use the weighing signal vector X to roughly estimate the weight of the load to be measured, and according to the national standard "JJG539-97 Digital Indicator Scale Verification Regulations", obtain the output weight W of the composite neural network through the self-adaptation of the weight regulator . Let w ₁ , w ₂ , and w ₃ be the components of W, then we have

${w w}_{11} = = \{\begin{matrix} 11 & if if & {\overset{~ ~}{y the y}}_{i i} \leq \leq 500500 e e \\ 00 & if if & 500500 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq 20002000 e e \\ 00 & if if & 20002000 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq {M m}_{max max} \end{matrix} - - - - - - ((44))$

${w w}_{22} = = \{\begin{matrix} 00 & if if & {\overset{~ ~}{y the y}}_{i i} \leq \leq 500500 e e \\ 11 & if if & 500500 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq 20002000 e e \\ 00 & if if & 20002000 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq {M m}_{max max} \end{matrix} - - - - - - ((55))$

${w w}_{33} = = \{\begin{matrix} 00 & if if & {\overset{~ ~}{y the y}}_{i i} \leq \leq 500500 e e \\ 00 & if if & 500500 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq 20002000 e e \\ 11 & if if & 20002000 e e < < {\overset{~ ~}{y the y}}_{i i} \leq \leq {M m}_{max max} \end{matrix} - - - - - - ((66))$

$\{\begin{matrix} W W = = (\begin{matrix} {w w}_{11} \\ {w w}_{22} \\ {w w}_{33} \end{matrix}) \\ {Σ Σ}_{j j = = 11}^{33} {w w}_{j j} = = 11 \end{matrix} - - - - - - ((77))$

被测载荷的粗略估计通过称重结果估计器完成，其可采用与子神经网络相同的结构与训练算法，也可采用其它结构和算法，但由于估计结果误差较大，不能作为最终的称重结果值。The rough estimation of the measured load is completed by the weighing result estimator, which can use the same structure and training algorithm as the sub-neural network, or other structures and algorithms, but due to the large error of the estimation result, it cannot be used as the final weighing result value.

(d)汽车衡称重结果获取。根据各个子神经网络的输出y_i和复合神经网络输出权值，将子神经网络的输出加权融合，得到最终的称重结果，并同时完成汽车衡称重误差补偿，即(d) Acquisition of truck scale weighing results. According to the output yi of each sub-neural network and the output weight of the composite neural network, the output of the sub-neural network is weighted and fused to obtain _the final weighing result, and at the same time complete the weighing error compensation of the truck scale, that is

$y the y = = {w w}_{11} {y the y}_{i i} + + {w w}_{22} {y the y}_{22} + + {w w}_{33} {y the y}_{33} = = {Σ Σ}_{j j = = 11}^{33} {w w}_{j j} {y the y}_{j j} - - - - - - ((88))$

本发明的N路称重传感器输出信号，采用N路相互独立的信号采集电路，并利用式(3)所示的数据预处理方法获得，其原理如附图2所示，包括称重信号调理、模数转换(A/D转换)等。The N-way load cell output signal of the present invention adopts N-way independent signal acquisition circuits, and utilizes the data preprocessing method shown in formula (3) to obtain, its principle is as shown in Figure 2, including weighing signal conditioning , Analog-to-digital conversion (A/D conversion), etc.

本发明的下位机为高性能单片机、DSP处理器或其它嵌入式系统设备，与其它电路构成主控板，并安装在汽车衡秤体下方，便于汽车衡在线测量。The lower computer of the present invention is a high-performance single-chip microcomputer, a DSP processor or other embedded system equipment, constitutes a main control board with other circuits, and is installed under the scale body of the truck scale to facilitate on-line measurement of the truck scale.

本发明的上位机为PC机，用于离线训练汽车衡误差补偿模型，并进行参数优化。The upper computer of the present invention is a PC, which is used for off-line training of the truck scale error compensation model and parameter optimization.

本发明的主控板必须做好防潮、防电磁干扰、防雷击处理，以免影响被测载荷称重结果的准确性。The main control board of the present invention must be protected against moisture, electromagnetic interference, and lightning, so as not to affect the accuracy of the weighing result of the measured load.

本发明与已有技术相比有以下优点：本发明可以实现汽车衡偏载误差与线性度误差的自动补偿，大大提高了称重结果的准确度。Compared with the prior art, the present invention has the following advantages: the present invention can realize the automatic compensation of the unbalanced load error and the linearity error of the truck scale, and greatly improves the accuracy of the weighing result.

附图说明 Description of drawings

图1是本发明的汽车衡复合神经网络训练流程框图。Fig. 1 is a block diagram of the training process of the compound neural network of the truck scale of the present invention.

图2是本发明的汽车衡在线称量流程框图。Fig. 2 is a block diagram of the on-line weighing process of the truck scale of the present invention.

图3是本发明的复合神经网络结构图。Fig. 3 is a structural diagram of the composite neural network of the present invention.

图4是本发明的汽车衡主控制板原理框图。Fig. 4 is a functional block diagram of the main control panel of the truck scale of the present invention.

图5是本发明的称重误差补偿效果图，其中(a)为补偿前、后的称重结果对比图，(b)为补偿前、后的称重结果误差曲线对比图。Fig. 5 is a weighing error compensation effect diagram of the present invention, wherein (a) is a comparison diagram of weighing results before and after compensation, and (b) is a comparison diagram of error curves of weighing results before and after compensation.

其中，1、第一路称重传感器，2、第一路调理电路，3、第一路模数转换电路，4、第二路称重传感器，5、第二路调理电路，6、第二路模数转换电路，7、第N路称重传感器，8、第N路调理电路，9、第N路模数转换电路，10、下位机，11、显示电路，12、键盘电路、13上位机(PC机)。Among them, 1. The first load cell, 2. The first conditioning circuit, 3. The first analog-to-digital conversion circuit, 4. The second load cell, 5. The second conditioning circuit, 6. The second Analog-to-digital conversion circuit, 7. Nth load cell, 8. Nth conditioning circuit, 9. Nth analog-to-digital conversion circuit, 10. Lower computer, 11. Display circuit, 12. Keyboard circuit, 13. Upper position machine (PC machine).

具体实施方式 Detailed ways

本发明提出了一种汽车衡称重误差自动补偿方法。以下结合附图1～4作详述，但不作为本发明的限定。The invention proposes an automatic compensation method for the weighing error of the truck scale. The following will be described in detail in conjunction with accompanying drawings 1 to 4, but not as a limitation of the present invention.

实施例一Embodiment one

设定本实施例中，汽车衡有8路称重传感器(N＝8)，量程为40吨，每路称重传感器的最大容量为20吨，分度数为4000，检定分度值e和实际分度值d均为10kg。采用TI公司的高性能单片机MSP430F449作为下位机10。It is assumed that in this embodiment, the truck scale has 8 load cells (N=8), the measuring range is 40 tons, the maximum capacity of each load cell is 20 tons, the number of divisions is 4000, and the verification division value e and the actual The division value d is 10kg. The high-performance single-chip microcomputer MSP430F449 of TI Company is used as the lower computer 10 .

(1)复合神经网络构造。在本实施例中，子神经网络采用径向基函数神经网络(RBFNN)，RBFNN的基函数采用高斯函数。根据式(1)所示方法，构造3个子神经网络，以子神经网络1为例，此时的子神经网络为一个8输入1输出的网络，因此其输出为(1) Compound neural network structure. In this embodiment, the sub-neural network adopts a radial basis function neural network (RBFNN), and the basis function of the RBFNN adopts a Gaussian function. According to the method shown in formula (1), three sub-neural networks are constructed. Taking sub-neural network 1 as an example, the sub-neural network at this time is a network with 8 inputs and 1 output, so its output is

${y the y}_{11} = = {b b}_{11} + + {Σ Σ}_{j j = = 11}^{m m 11} {h h}_{11,, j j} {w w}_{11,, j j} = = {Σ Σ}_{j j = = 00}^{m m 11} {h h}_{11,, j j} {w w}_{11,, j j} = = {W W}_{11}^{T T} {H h}_{11} - - - - - - ((99))$

式中，隐层神经元的个数m₁由实验确定为9；W₁为子RBFNN的权矢量，W₁＝(w_1，0，w_1，1，w_1，2，…，w_1，m)^T，其中w_n，0＝1；H₁为径向基函数矢量，H₁＝(h_1，0，h_1，1，h_1，2，…，h_1，m)¹，其中h_1，0＝b₁；b₁为输出层偏置值，In the formula, the number m ₁ of neurons in the hidden layer is determined to be 9 by experiments; W ₁ is the weight vector of sub-RBFNN, W ₁ = (w _1,0 ,w _1,1 ,w _1,2 ,...,w _{1 , m} ) ^T , where w _n,0 =1; H ₁ is the radial basis function vector, H ₁ =(h _1,0 ,h _1,1 ,h _1,2 ,…,h _1,m ) ¹ , where h _1,0 = b ₁ ; b ₁ is the output layer bias value,

${h h}_{11,, j j} = = exp exp ((- - \frac{{| | | | X x - - {C C}_{11,, j j} | | | |}^{22}}{{22 r r}_{11,, j j}^{22}})) j j = = 1,2 1,2,, . . . . . . {m m}_{11} - - - - - - ((1010))$

式中，X为输入矢量，X＝(x₁，x₂…，x₈)^T；r_1，j为第j个节点的扩展常数；C_1，j为第j个节点的中心矢量，C_1，j＝(c_11，j，c_12，j，…，c_18，j)^T，||·||为欧几里德距离。对于子RBFNN2、子RBFNN3有同样的分析。In the formula, X is the input vector, X=(x ₁ , x ₂ ..., x ₈ ) ^T ; r _1,j is the expansion constant of the jth node; C _1,j is the center vector of the jth node, C _{1, j} = (c _{11, j} , c _{12, j} , ..., c _{18, j} ) ^T , where ||·|| is the Euclidean distance. The same analysis is carried out for sub-RBFNN2 and sub-RBFNN3.

(2)复合神经网络训练。利用不同吨位的标准砝码(如0.5吨、1吨、3吨、6吨、12吨、24吨、36吨等)加载在汽车衡承载面上的不同位置，系统通过称重传感器、调理电路、模数转换电路和下位机10，采集140组8路称重传感器信号，经式(3)所示的数据预处理方法(Q＝50，M＝5)，获得140组训练样本与测试样本，其中84组用于各子RBFNN训练，56组用于各子RBFNN测试。下位机10通过串行通信接口将这些样本传送至上位机13，上位机13利用训练软件进行子神经网络离线训练，各子RBFNN的隐层神经元个数均为12。网络结束训练后，上位机13将各子RBFNN的参数(如扩展常数R_n、中心矢量C_n、权值矩阵W_n和输出层偏置b_n)下载到下位机10中，为汽车衡在线称重与误差补偿作准备。汽车衡复合神经网络训练方法如图1所示。(2) Composite neural network training. Using standard weights of different tonnages (such as 0.5 tons, 1 ton, 3 tons, 6 tons, 12 tons, 24 tons, 36 tons, etc.) , analog-to-digital conversion circuit and lower computer 10, gather 140 groups of 8-way load cell signals, through the data preprocessing method (Q=50, M=5) shown in formula (3), obtain 140 groups of training samples and test samples , of which 84 groups are used for each sub-RBFNN training, and 56 groups are used for each sub-RBFNN test. The lower computer 10 transmits these samples to the upper computer 13 through the serial communication interface, and the upper computer 13 uses the training software to perform offline training of the sub-neural network, and the number of hidden layer neurons of each sub-RBFNN is 12. After the network training is finished, the upper computer 13 downloads the parameters of each sub-RBFNN (such as the expansion constant R _n , the center vector C _n , the weight matrix W _n and the output layer bias b _n ) to the lower computer 10, and provides an online Prepare for weighing and error compensation. The training method of the compound neural network of the truck scale is shown in Figure 1.

(3)汽车衡在线称重与误差补偿。汽车衡在线称重与误差补偿时，系统通过称重传感器、调理电路、模数转换电路和下位机10，采集8路称重传感器输出信号，经式(3)所示的数据预处理方法(Q＝50，M＝5)，获得称重信号向量X，粗略估计被测载荷的重量，并根据国家标准《JJG539-97数字指示秤检定规程》的规定，自适应获得复合神经网络输出权值W；同时将称重信号向量X作为子RBF神经网络1、2、3的输入，根据保存在下位机10中的子RBFNN1、子RBFNN2、子RBFNN3的参数(如R_n、C_n、W_n和b_n)，分别计算出RBFNN1、子RBFNN2、子RBFNN3的输出y₁、y₂、y₃；最后利用式(8)完成汽车衡在线称量和称重误差补偿，获得被测载荷的称重结果。汽车衡在线称重方法如图2所示。利用汽车衡检定的国家标准《JJG539-97数字指示秤检定规程》，对采用该方法的汽车衡进行现场检定，表1偏载误差检定结果，为表2为线性度误差检定结果。(3) Truck scale online weighing and error compensation. During the on-line weighing and error compensation of the truck scale, the system collects 8 load cell output signals through the load cell, the conditioning circuit, the analog-to-digital conversion circuit and the lower computer 10, and the data preprocessing method shown in formula (3) ( Q=50, M=5), obtain the weighing signal vector X, roughly estimate the weight of the measured load, and according to the provisions of the national standard "JJG539-97 Digital Indicator Scale Verification Regulations", adaptively obtain the output weight of the composite neural network W; simultaneously weigh signal vector X as the input of sub-RBF neural network 1,2,3, according to the parameter (as R _n , C _n , W _n of sub-RBFNN1, sub-RBFNN2, sub-RBFNN3 stored in lower computer 10 and b _n ), respectively calculate the output y ₁ , y ₂ , and y ₃ of RBFNN1, sub-RBFNN2, and sub-RBFNN3; finally, use formula (8) to complete the on-line weighing and weighing error compensation of the truck scale, and obtain the scale of the measured load Focus on results. The truck scale online weighing method is shown in Figure 2. Using the national standard "JJG539-97 Digital Indicating Scale Verification Regulations" for the verification of truck scales, the truck scales using this method are verified on site. Table 1 shows the results of the eccentric load error verification, and Table 2 shows the results of the linearity error verification.

表1偏载误差检定结果Table 1 Verification results of eccentric load error

表2线性度误差检定结果Table 2 Linearity Error Verification Results

表1中，1#表示1号称重传感器所在位置的加载区域，其它的有相同的意义。由表1、2可以看出，采用这种方法的汽车衡偏载误差远小于国家标准规定的允许误差，线性度误差同样小于国家标准规定的允许误差，误差补偿效果明显。In Table 1, 1# indicates the loading area where the No. 1 load cell is located, and the others have the same meaning. It can be seen from Tables 1 and 2 that the unbalanced load error of the truck scale using this method is much smaller than the allowable error specified by the national standard, and the linearity error is also smaller than the allowable error specified by the national standard, and the error compensation effect is obvious.

实施例2Example 2

设定本实施例中，汽车衡有6路称重传感器(N＝6)，量程为40吨，每路称重传感器的最大容量为20吨，分度数为4000，检定分度值e和实际分度值d均为10kg。采用TI公司的DSP TMS320VC5502作为下位机10。Set in this embodiment, the truck scale has 6 load cells (N=6), the measuring range is 40 tons, the maximum capacity of each load cell is 20 tons, the division number is 4000, the verification division value e and the actual The division value d is 10kg. The DSP TMS320VC5502 of TI Company is adopted as the lower computer 10 .

(1)复合神经网络构造。在本实施例中，子神经网络采用BP神经网络(BPNN)，BPNN的隐层激励函数f₁采用S形函数，输出层激励函数f₂采用线性函数，以子神经网络1为例，此时的子神经网络为一个6输入1输出的网络，因此其输出为(1) Compound neural network structure. In this embodiment, the sub-neural network adopts BP neural network (BPNN), the hidden layer activation function _f1 of BPNN adopts an S-shaped function, and the output layer activation function _f2 adopts a linear function. Taking sub-neural network 1 as an example, at this time The sub-neural network of is a network with 6 inputs and 1 output, so its output is

${y the y}_{11} = = {W W}^{22} {F f}_{11} + + {b b}^{22} = = \frac{{W W}^{22}}{11 + + {e e}^{- - (({W W}^{11} X x + + {b b}^{11}))}} + + {b b}^{22}$

$= = {Σ Σ}_{j j = = 11}^{m m 11} ((\frac{{w w}_{}^{j j}}{11 + + {e e}^{- - (({Σ Σ}_{i i = = 11}^{N N} {w w}_{ij ij}^{11} {x x}_{i i} + + {b b}_{i i}^{11}))}} + + {b b}^{22})) - - - - - - ((1111))$

式中，隐层神经元的个数m₁由实验确定为5；W¹、W²分别为BPNN输入层到隐层、隐层到输出层的权值矩阵；b¹、b²分别为隐层的偏置值向量和输出层的偏置值。F¹为隐层函数矢量(即S形函数输出)。对于子BPNN2、子BPNN3有同样的分析。In the formula, the number m ₁ of neurons in the hidden layer is determined to be 5 by experiments; W ¹ and W ² are the weight matrixes from the input layer to the hidden layer and from the hidden layer to the output layer of BPNN respectively; b ¹ and b ² are the hidden A vector of bias values for the layer and bias values for the output layer. F ¹ is the hidden layer function vector (that is, the output of the sigmoid function). The same analysis is carried out for sub-BPNN2 and sub-BPNN3.

(2)复合神经网络训练。利用不同吨位的标准砝码(如0.5吨、1吨、3吨、6吨、12吨、24吨、36吨等)加载在汽车衡承载面上的不同位置，系统通过称重传感器、调理电路、模数转换电路和下位机10，采集112组6路称重传感器信号，经式(3)所示的数据预处理方法(Q＝50，M＝5)，获得112组训练样本与测试样本，其中70组用于各子BPNN训练，42组用于各子BPNN测试。下位机10通过串行通信接口将这些样本传送至上位机13，上位机13利用训练软件进行子神经网络离线训练，各子BPNN的隐层神经元个数均为5。网络结束训练后，上位机13将各子BP的参数(如W¹、W²、b¹、b²)下载到下位机10中，为汽车衡在线称重与误差补偿作准备。汽车衡复合神经网络训练方法如图1所示。(2) Composite neural network training. Using standard weights of different tonnages (such as 0.5 tons, 1 ton, 3 tons, 6 tons, 12 tons, 24 tons, 36 tons, etc.) , analog-to-digital conversion circuit and lower computer 10, collect 112 groups of 6 road load cell signals, through the data preprocessing method (Q=50, M=5) shown in formula (3), obtain 112 groups of training samples and test samples , of which 70 groups are used for each sub-BPNN training, and 42 groups are used for each sub-BPNN test. The lower computer 10 transmits these samples to the upper computer 13 through the serial communication interface, and the upper computer 13 uses the training software to perform offline training of the sub-neural network, and the number of hidden layer neurons of each sub-BPNN is 5. After the network training, the upper computer 13 downloads the parameters of each sub-BP (such as W ¹ , W ² , b ¹ , b ² ) to the lower computer 10 to prepare for the on-line weighing and error compensation of the truck scale. The training method of the compound neural network of the truck scale is shown in Figure 1.

(3)汽车衡在线称重与误差补偿。汽车衡在线称重与误差补偿时，系统通过称重传感器、调理电路、模数转换电路和下位机10，采集6路称重传感器输出信号，经式(3)所示的数据预处理方法(Q＝50，M＝5)，获得称重信号向量X，粗略估计被测载荷的重量，并根据国家标准《JJG539-97数字指示秤检定规程》的规定，自适应获得复合神经网络输出权值W；同时将称重信号向量X作为子BP神经网络1、2、3的输入，根据保存在下位机10中的子BPNN1、子BPNN2、子BPNN3的参数(如W¹、W²、b¹、b²)，分别计算出BPNN1、子BPNN2、子BPNN3的输出y₁、y₂、y₃；最后利用式(8)完成汽车衡在线称重和误差补偿，获得被测载荷的称重结果。汽车衡在线称重方法如图2所示。利用汽车衡检定的国家标准《JJG539-97数字指示秤检定规程》，对采用该方法的汽车衡进行现场检定，表3偏载误差检定结果，为表4为线性度误差检定结果。(3) Truck scale online weighing and error compensation. During online weighing and error compensation of the truck scale, the system collects 6 load cell output signals through the load cell, the conditioning circuit, the analog-to-digital conversion circuit and the lower computer 10, and the data preprocessing method shown in formula (3) ( Q=50, M=5), obtain the weighing signal vector X, roughly estimate the weight of the measured load, and according to the provisions of the national standard "JJG539-97 Digital Indicator Scale Verification Regulations", adaptively obtain the output weight of the composite neural network W; while weighing signal vector X as the input of sub-BP neural network 1, 2, 3, according to the parameters (such as W ¹ , W ² , b ¹ ) of sub-BPNN1, sub-BPNN2 and sub-BPNN3 stored in lower computer 10 , b ² ), respectively calculate the output y ₁ , y ₂ , y ₃ of BPNN1, sub-BPNN2, and sub-BPNN3; finally use formula (8) to complete the online weighing and error compensation of the truck scale, and obtain the weighing result of the measured load . The truck scale online weighing method is shown in Figure 2. Using the national standard "JJG539-97 Digital Indicating Scale Verification Regulations" for truck scale verification, the truck scale using this method was verified on-site. Table 3 shows the results of eccentric load error verification, and Table 4 shows the results of linearity error verification.

表3偏载误差检定结果Table 3 eccentric load error verification results

表4线性度误差检定结果Table 4 Linearity Error Verification Results

表3中，1#表示1号称重传感器所在位置的加载区域，其它的有相同的意义。由表3、4可以看出，采用这种方法的汽车衡偏载误差远小于国家标准规定的允许误差，线性度误差同样小于国家标准规定的允许误差，误差补偿效果明显。In Table 3, 1# indicates the loading area where the No. 1 load cell is located, and the others have the same meaning. It can be seen from Tables 3 and 4 that the unbalanced load error of the truck scale using this method is much smaller than the allowable error specified by the national standard, and the linearity error is also smaller than the allowable error specified by the national standard, and the error compensation effect is obvious.

本发明说明书中未作详细描述的内容属于本领域专业技术人员公知的现有技术。The contents not described in detail in the description of the present invention belong to the prior art known to those skilled in the art.

Claims

1. truck scale weighting error automatic compensating method; It is characterized in that: utilize complex neural network to approach the nonlinear function between the truck scale input-output; Thereby accomplishing the truck scale pin-point accuracy weighs and error compensation; This method comprises complex neural network structure, neural network off-line training and three aspects of truck scale on-line weighing, and concrete steps are following:

(1) according to priori, promptly the error character of national standard " JJG539-97 numeral self-indicating scale vertification regulation " regulation is constructed 3 sub neural networks;

(2) utilize the standard test weight of different tonnages; Be carried in the diverse location on the truck scale loading end respectively, system's slave computer obtains N road LOAD CELLS output signal through the separate weighing-up wave Acquisition Circuit in N road; After the data pre-service, obtain weighing-up wave vector X _j(j=1,2 ..., K), and composing training sample (X _j, y _j); Slave computer transfers to host computer through communication interface with training sample, utilizes training software, accomplishes 3 sub neural network off-line trainings, obtains relevant parameters, and is downloaded to slave computer, for the truck scale on-line measurement is prepared;

(3) when truck scale online weighing and error compensation, slave computer obtains N road LOAD CELLS output signal at first through the separate weighing-up wave Acquisition Circuit in N road, after the data pre-service, obtains N road weighing-up wave vector X; With the input of weighing-up wave vector X, calculate the output y that obtains each sub neural network as sub neural network 1,2,3 _i(i=1,2,3); The weight of the tested load of guestimate, and according to the regulation of national standard " JJG539-97 numeral self-indicating scale vertification regulation ", self-adaptation obtains complex neural network output weights W, promptly

w_{1} = \{\begin{matrix} 1 & if & {\tilde{y}}_{i} \leq 500 e \\ 0 & if & 500 e < {\tilde{y}}_{i} \leq 2000 e \\ 0 & if & 2000 e < {\tilde{y}}_{i} \leq M_{\max} \end{matrix}

w_{2} = \{\begin{matrix} 0 & if & {\tilde{y}}_{i} \leq 500 e \\ 1 & if & 500 e < {\tilde{y}}_{i} \leq 2000 e \\ 0 & if & 2000 e < {\tilde{y}}_{i} \leq M_{\max} \end{matrix}

w_{3} = \{\begin{matrix} 0 & if & {\tilde{y}}_{i} \leq 500 e \\ 0 & if & 500 e < {\tilde{y}}_{i} \leq 2000 e \\ 1 & if & 2000 e < {\tilde{y}}_{i} \leq M_{\max} \end{matrix}

\{\begin{matrix} W = (\begin{matrix} w_{1} \\ w_{2} \\ w_{3} \end{matrix}) \\ Σ_{j = 1}^{3} w_{j} = 1 \end{matrix}

In the formula, w ₁, w ₂, w ₃3 components for W; Output y according to each sub neural network _iWith complex neural network output weights, the output weighting fusion with sub neural network obtains final weighing results, and accomplishes the compensation of truck scale weighting error simultaneously, promptly

y = w_{1} y_{i} + w_{2} y_{2} + w_{3} y_{3} = Σ_{j = 1}^{3} w_{j} y_{j} .

2. truck scale error character according to claim 1; Be by national standard " the JJG539-97 numeral self-indicating scale vertification regulation " defined of truck scale calibrating; It is divided into 3 different calibrating sections with the total range of truck scale, and the permissible error of each calibrating section is different.

3. complex neural network model according to claim 1 is made up of sub neural network maker, sub neural network, estimator, weights regulator etc.

4. can be radial basis function neural network (RBFNN), BP neural network (BPNN), perhaps other feedforward neural network according to claim 1 and the described sub neural network of claim 3.

5. identical with the structure of described 3 sub neural networks of claim 4 according to claim 1, claim 3, all be the three-layer network of N input 1 output, wherein N is the number of truck scale LOAD CELLS.

6. the data preprocessing method of weighing according to claim 1 is a kind of digital filtering method, Q LOAD CELLS output voltage of this method continuous sampling, and it is obtained S by series arrangement from small to large _Im, remove forward and backward each M sampled value, to all the other Q-2M S _ImCarry out mean filter, promptly

x_{i} = \frac{1}{Q - 2 M} Σ_{m = M + 1}^{Q - M} S_{im}

7. the weighing-up wave Acquisition Circuit that N according to claim 1 road is separate; Comprise weighing-up wave modulate circuit, analog to digital conversion (A/D conversion) circuit etc., wherein modulate circuit comprises weighing-up wave amplifying circuit, weighing-up wave filtering circuit and impedance matching circuit.

8. slave computer according to claim 1 is high-performance single-chip microcomputer, dsp processor or other embedded system devices; Described host computer is a PC.