Dynamic force calibration method and measurement method based on gaalbp

Abstract

The application relates to the technical field of dynamic force measurement, and specifically provides a dynamic force calibration method and a measurement method based on GAALBP. The GAALBP fusion algorithm is used for dynamic calibration, and the BP network is used for dynamic measurement. The weights and threshold values of the BP network are optimized by using GA, the learning rate of the BP network is adjusted according to the output error of the BP network during training, the training speed and precision are improved, GA and BP are alternately used in the training process, and the BP neural network is prevented from falling into a local optimal solution. Fault-tolerant and non-fault-tolerant measurements of a measurement platform are calibrated separately, and the measurement platform is measured after being calibrated by using a suitable calibration model. The application considers the errors caused by the nonlinearity of the system itself and fault-tolerant measurement, improves the fault-tolerant and non-fault-tolerant measurement precision of the measurement platform, reduces the interdimensional coupling degree, the adaptive learning rate improves the calibration efficiency, and when the sensor of the measurement platform fails, the sensor does not need to be replaced, thereby avoiding unnecessary workload.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dynamic force measurement, and specifically provides a dynamic force calibration method and a measurement method based on GAALBP with fault tolerance. BACKGROUND

[0002] With the continuous exploration of the vast universe, the performance of the Chinese space station telescope in stability, pointing accuracy and other aspects has been greatly improved. In order to explore how the dynamic disturbance forces generated by the vibration sources in the spacecraft affect the performance of the telescope, it is necessary to measure these dynamic forces on the ground as a basis for evaluation. On the one hand, as the size and mass of the spacecraft vibration sources become larger and larger, the ground measurement equipment must have high load capacity and wide installation surface. The increase in the mass and volume of the measurement equipment inevitably introduces nonlinear terms. On the other hand, unexpected situations such as overload impact or signal transmission line breakage may occur, and part of the sensors on the measurement platform may fail during operation. However, the volume and mass of the vibration source installed on the measurement platform are huge, and the disassembly and installation process is difficult and time-consuming. Moreover, even if the faulty sensor is replaced, the measurement platform must be reassembled and dynamically calibrated, which obviously increases the workload.

[0003] Currently, in order to improve the load capacity and installation surface of the measurement equipment, a split-load measurement platform is generally used, such as the one disclosed in Chinese patent CN109990888B. This measurement platform additionally installs a split-load mechanism between the load disc and the base. In order to make the measurement equipment have fault tolerance performance, the measurement equipment is usually provided with redundant measurement output signals, such as the technical solution disclosed in Chinese patent CN110514341B. In order to solve the nonlinear problem in the measurement equipment, Chinese patent CN111272334A discloses a multi-dimensional force sensor calibration decoupling method based on particle swarm optimization BP neural network; and Chinese patent CN113820062B discloses a way of using improved elephant trunk algorithm to optimize BP neural network for temperature nonlinear compensation.

[0004] Although the existing more common split load type measuring equipment has higher load capacity and larger installation surface, the mass and volume of the load platform thereof are also larger, at this time, the system has strong nonlinearity, and there is a large precision error in measurement, and the past dynamic force measurement technology has not proposed a solution; at present, the dynamic force measurement equipment with redundant output performance has a large decrease in precision in fault-tolerant measurement under the traditional measurement mode, and cannot meet the measurement demand; the traditional calibration based on intelligent algorithm is mostly used for static calibration, and when it is applied to dynamic calibration and measurement, the training speed is very slow, and the calibration and measurement period of the platform is greatly increased. The existing technology mostly has good performance in the realization of a single functional demand, and at present, there is no measurement method that can well meet the above strict dynamic measurement requirements.

[0005] Therefore, for the dynamic measurement platform with great load capacity and installation surface, a dynamic force calibration and measurement method considering the nonlinearity term of the system and having fault-tolerant measurement capability is urgently needed. SUMMARY

[0006] The present application provides a dynamic force calibration method and a measurement method based on GAALBP, adopts GAALBP fusion algorithm for dynamic calibration, uses BP network for dynamic measurement, optimizes the weight and threshold value of the BP network (back propagation neural network) by using GA (genetic algorithm), adjusts the learning rate of the BP network according to the output error of the BP network during training, improves the training speed and accuracy, and alternately carries out GA and BP during the training process to avoid the BP neural network from falling into a local optimal solution.

[0007] The dynamic force calibration method based on GAALBP provided by the present application comprises the following steps:

[0008] S1, a plurality of sets of calibration data of the measurement platform are collected ci (ω), F oi (ω)} are divided into a training group and a verification group, and the output signal V ci (ω) of the sensor is collected as an input sample, and the six-dimensional force F oi (ω) obtained through the sensor is taken as an output sample;

[0009] S2, a BP network is established, the weight and threshold value of the BP network are optimized by using GA, the learning rate of the BP network is adjusted according to the output error during training, and the specific process is as follows:

[0010] S21, the weight and threshold value of the BP network are generated;

[0011] S22, input the input sample corresponding to the a-th frequency point into the P BP networks for training, calculate the output error according to the output value of the BP network, and determine whether the BP convergence condition is met:

[0012] If the BP convergence condition is met, go to S24;

[0013] If the BP convergence condition is not met, go to S23;

[0014] S23, adaptively adjust the learning rate of the BP network according to the learning rate adjustment criterion, update the weight and threshold value of the BP network, and go to S22;

[0015] S24, calculate the fitness in GA, and determine whether the GA convergence condition is met:

[0016] If the GA convergence condition is met, the training of the a-th frequency point is completed;

[0017] If the GA convergence condition is not met, go to S25;

[0018] S25, encode the weight and threshold value of the BP network, select the coded data for genetic operation according to the fitness, and decode to go to S22;

[0019] S3, repeat S2 to complete the training of all frequency points, that is, obtain the calibration model.

[0020] Preferably, in S1, the test force is input multiple times to the multiple physical points of the measurement platform, the sensor signal F ci (ω) and the output signal V ci (ω) of the sensor are collected, and the sensor signal F ci (ω) is converted into a six-dimensional force F oi (ω), wherein the collection frequency is f and the collection time is t.

[0021] Preferably, the calibration data {V ci (ω), F oi (ω)} is normalized for pretreatment, which is used to eliminate the influence of variables of different orders of magnitude on network training.

[0022] Preferably, in S21, the value range of the weight of the BP network and the value range of the threshold value are set, and the weight and the threshold value of the BP network are randomly generated.

[0023] Preferably, the learning rate adjustment criterion is as follows:

[0024] When then:

[0025] lr(n+1)=0.7×lr(n);

[0026] w(n+1) = w(n-1) + lr(n+1) x Aw;

[0027] b(n+1) = b(n-1) + lr(n+1) x Ab;

[0028] When then:

[0029] lr(n+1) = lr(n);

[0030] w(n+1) = w(n) + lr(n+1) x Aw;

[0031] b(n+1) = b(n) + lr(n+1) x Ab;

[0032] When then:

[0033] lr(n+1) = 1.05 x lr(n);

[0034] w(n+1) = w(n) + lr(n+1) x Aw;

[0035] b(n+1) = b(n) + lr(n+1) x Ab;

[0036] wherein, error represents an output error, lr represents a learning rate of the BP network, w represents a weight value of the BP network, b represents a threshold value of the BP network, Aw represents a change amount of the weight value, Ab represents a change amount of the threshold value, and n represents a round.

[0037] Preferably, the fitness in the GA is calculated according to the following formula:

[0038]

[0039] wherein, F jk-real (ω a ) represents the kth real input force in the jth verification group at the a th frequency point, F jk-pre (ω a ) represents the kth predicted force in the jth verification group at the a th frequency point; J represents the number of verification groups, and K represents the number of physical points.

[0040] Preferably, in the S25, the genetic operation can adopt one or more of gene selection, crossover and mutation in the GA.

[0041] A dynamic force measurement method based on GAALBP, comprising the following steps:

[0042] The working condition of the sensor is determined by using the coherence function, and the expression of the coherence function is as follows:

[0043]

[0044] wherein G fx (ω) represents the cross spectrum of the sensor input signal and the output signal, G ff (ω) represents the auto spectrum of the sensor output signal, G xx (ω) represents the auto spectrum of the sensor input signal, γ 2 represents the function value, 0≤γ 2 ≤1;

[0045] According to the working condition of the sensor, it is determined whether the measurement platform meets the measurement requirement, if it meets the measurement requirement, the output signal V ci (ω) of the sensor with normal working condition is collected as the input sample, the six-dimensional force F oi (ω) obtained by the sensor with normal working condition is collected as the output sample, the dynamic force calibration method based on GAALBP in any one of claims 1-7 is used to calibrate the measurement platform, the data to be measured is collected, and the dynamic force measurement is performed by using the calibrated measurement platform.

[0046] Preferably, when γ 2 >0.8, the working condition of the sensor is determined as normal;

[0047] When 0.3<γ 2 <0.8, the working condition of the sensor is determined as being disturbed by environmental noise;

[0048] When γ 2 <0.3, the working condition of the sensor is determined as failure.

[0049] Preferably, when the working conditions of all sensors are disturbed by environmental noise, the measurement requirement is not met, and the measurement is stopped.

[0050] When the number of sensors with failure working condition is greater than 50% of the total number of sensors, the measurement requirement is not met, and the measurement is stopped.

[0051] Compared with the prior art, the present application can achieve the following beneficial effects:

[0052] The calibration method of the present application uses GAALBP fusion algorithm for dynamic calibration, considers the errors caused by the nonlinearity of the system itself and fault-tolerant measurement, improves the accuracy of fault-tolerant and non-fault-tolerant measurement of the measurement platform, reduces the interdimensional coupling degree, and the adaptive learning rate improves the efficiency of dynamic calibration.

[0053] The measurement method of the present application describes in detail the fault detection mode and the measurement scheme, when the measurement platform encounters sensor failure, it is not necessary to replace the sensor, unnecessary workload is avoided, and the measurement platform can maintain high measurement accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a structural schematic diagram of a split-load dynamic force measurement platform provided according to an embodiment of the present application;

[0055] Figure 2 is a flowchart of a dynamic force calibration method based on GAALBP provided according to an embodiment of the present application.

[0056] Figure 3 is a flowchart of a dynamic force measurement method based on GAALBP provided according to an embodiment of the present application.

[0057] The reference signs in the drawings include:

[0058] Load table 1, base 2, split load column 3, three-way force sensor 4. DETAILED DESCRIPTION

[0059] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the following description, the same modules are denoted by the same reference signs. In the case of the same reference signs, their names and functions are also the same. Therefore, detailed descriptions thereof will not be repeated.

[0060] In order to make the objectives, technical solutions, and advantages of the present application clearer, further detailed descriptions will be given below in combination with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not constitute a limitation on the present application.

[0061] Figure 1 A structure of a split-load dynamic force measurement platform provided according to an embodiment of the present application is shown.

[0062] As shown in Figure 1 , the split-load dynamic force measurement platform to which the embodiment of the present application is directed includes one load platform 1, four bases 2, one split load column 3, and four three-way force sensors 4. The measurement platform has a total of 12 redundant output signals and can completely measure spatial six-dimensional force. In the conventional technology, dynamic calibration and dynamic measurement are usually based on the least square method, without considering the nonlinear problem of a large vibration source dynamic force measurement platform, and the measurement accuracy and decoupling ability are poor in fault-tolerant measurement. The embodiment of the present application adopts a GAALBP fusion algorithm for dynamic calibration and uses a BP network for dynamic measurement.

[0063] Figure 2 A flow of a dynamic force calibration method based on GAALBP provided according to an embodiment of the present application is shown.

[0064] As shown in Figure 2As shown in the figure, this invention proposes a dynamic force calibration method based on GAALBP. This method is applicable to both fault-tolerant and non-fault-tolerant measurements. The specific process is as follows:

[0065] S1. Input 25 test forces of different magnitudes to 6 physical points on load platform 1, and collect force sensor signals F. ci (ω) and the output signals V from the 12 sensors ci (ω), sensor signal F ci (ω) is the input force of the sensor, which is the sensor signal F. ci (ω) Transformed to the specified reference coordinate system, the six-dimensional force F oi (ω), for example Figure 1 The coordinate system shown. Determine the calibration data {V} ci (ω), F oi (ω)},(i=1,2,…,150), a total of 150 sets of calibration data were obtained. During the data acquisition process, attention should be paid to the six-dimensional force F in the 150 sets of calibration data. oi (ω) needs to be able to cover six-dimensional forces in space.

[0066] The 150 sets of calibration data obtained from the acquisition and measurement platform {V ci (ω), F oi (ω)} is divided into 120 training groups and 30 validation groups, and the calibration data {V} will be used to... ci (ω), F oi V in (ω)} ci (ω) is used as the input sample, F oi (ω) is used as the output sample. Each collection is recorded as a frequency point a. When training, the initial value of a is 1. The frequency point is not the same as the physical point. The frequency point belongs to the data concept, while the physical point is a point that actually exists in space.

[0067] S2. Establish the required BP network and optimize its weights and thresholds using GA. The BP network is a multi-layer feedforward neural network, well-suited for solving nonlinear problems. However, BP networks have some drawbacks, such as slow learning convergence and inability to guarantee convergence to the global minimum. Therefore, GA is used to optimize the network's weights and thresholds, helping the BP network escape local optima. The learning rate of the BP network is adjusted in real-time based on the training error to improve training speed and accuracy. GA and BP are used alternately multiple times to prevent the BP neural network from getting trapped in local optima. The establishment of the BP network and its use for dynamic measurement are relatively mature designs; the specific topology and design of the BP network will not be elaborated here. The specific process of weight and threshold optimization and real-time learning rate adjustment is as follows:

[0068] S21. First, the calibration data {V ci(ω), F oi (V ci (ω), F oi (V

[0069] The value range of the weight of the BP network and the value range of the threshold value are set according to the actual situation, and a set of the weight and the threshold value of the BP network are randomly generated in the value range.

[0070] S22, input the input sample corresponding to the a-th frequency point into the P BP networks for training, the value of P is determined according to the population number in the GA, the output error is calculated according to the output value of the BP network, and whether the BP convergence condition is met is judged:

[0071] If the BP convergence condition is met, learning rate adjustment is not needed in this round, and S24 is entered;

[0072] If the BP convergence condition is not met, learning rate adjustment is needed in this round, and S23 is entered.

[0073] S23, the learning rate of the BP network is adaptively adjusted according to the learning rate adjustment criterion, if the error ratio of this round to the last round is greater than 1.04, the weight and the threshold value of this round are discarded, the values of the last round are used, and the learning rate is reduced, if the error ratio of this round to the last round is less than 1, the learning rate is increased, and except for the above cases, the original learning rate is maintained, and the learning rate adjustment criterion is specifically expressed as follows:

[0074] When The adjustment rule is:

[0075] lr(n+1)=0.7×lr(n);

[0076] w(n+1)=w(n-1)+lr(n+1)×Δw;

[0077] b(n+1)=b(n-1)+lr(n+1)×Δb;

[0078] When The adjustment rule is:

[0079] lr(n+1)=lr(n);

[0080] w(n+1)=w(n)+lr(n+1)×Δw;

[0081] b(n+1)=b(n)+lr(n+1)×Δb;

[0082] When The adjustment rule is:

[0083] lr(n+1) = 1.05 x lr(n) ;

[0084] w(n+1) = w(n) + lr(n+1) x Aw;

[0085] b(n+1) = b(n) + lr(n+1) x Ab;

[0086] Wherein, error represents output error, lr represents learning rate of BP network, w represents weight value of BP network, b represents threshold value of BP network, Aw represents weight value change, Ab represents threshold value change, n represents round, (n-1) represents last round of this round, (n+1) represents next round of this round. After learning rate, weight value and threshold value are updated, return to S22 to rejudge until convergence condition is met.

[0087] S24, calculate fitness in GA according to the following formula:

[0088]

[0089] Wherein, F jk-real (ω a ) represents the kth real input force in the jth verification group at the a th frequency point, F jk-pre (ω a ) represents the kth predicted force in the jth verification group at the a th frequency point; J represents the number of verification groups, that is, J = 30, K represents the number of physical points, that is, K = 6.

[0090] Determine whether the GA convergence condition is met:

[0091] If the GA convergence condition is met, that is, the training of the a th frequency point is completed, and the trained model is obtained.

[0092] If the GA convergence condition is not met, go to S25.

[0093] S25, encode the weight value and threshold value of the BP network, and select the coded data for genetic operation according to the fitness, the genetic operation can adopt one or more of gene selection, crossover and mutation in GA, and decoding is convenient for transmitting the weight value and threshold value to the BP network for optimization, and after decoding, go to S22.

[0094] S3, determine whether the training of all frequency points has been completed, if the condition is met, end the dynamic calibration, otherwise select the data of the a+1 th frequency point and repeat S2 to complete the training of all frequency points, that is, obtain the calibration model.

[0095] Based on the above calibration method, the embodiment of the application provides a dynamic force measurement method based on GAALBP, comprising the following steps:

[0096] Because the sensor 4 of the measurement platform can be partially damaged or the output signal fails, the traditional measurement method needs to replace the sensor 4, and the method is to first determine the working condition of each sensor 4 by using the coherence function, and the expression of the coherence function is as follows:

[0097]

[0098] Wherein, G fx (ω) represents the cross spectrum of the input signal and the output signal of the sensor 4, G ff (ω) represents the self spectrum of the output signal of the sensor 4, G xx (ω) represents the self spectrum of the input signal of the sensor 4, and γ 2 represents the function value, 0≤γ 2 ≤1.

[0099] The working condition of the sensor 4 is determined according to the function value γ 2 , and the specific determination rule is as follows:

[0100] When γ 2 >0.8, the working condition of the sensor 4 is determined as normal, and the output signal is a valid signal;

[0101] When 0.3<γ 2 <0.8, the working condition of the sensor 4 is determined as being disturbed by environmental noise, and the output signal has an error;

[0102] When γ 2 <0.3, the working condition of the sensor 4 is determined as invalid, and the output signal is an invalid signal.

[0103] According to the working condition of the sensor 4, whether the measurement platform meets the measurement requirements is determined, and when one of the following situations occurs, the measurement requirements cannot be met:

[0104] First, when the working conditions of all sensors 4 are disturbed by environmental noise, the measurement requirements are not met, and the measurement is stopped.

[0105] Second, when the number of sensors 4 with invalid working conditions is greater than 50% of the total number of sensors 4, the measurement requirements are not met, and the measurement is stopped, that is, when the number of invalid sensors 4 is greater than two, the measurement platform is determined to be damaged.

[0106] After the working condition determination, when all the sensors 4 are normal, it is non-fault-tolerant measurement; when the sensors 4 are partially invalid, the remaining normal sensors 4 are used for measurement, which is fault-tolerant measurement. If the measurement requirements are met, the output signal V ci (ω) of the sensor 4 with normal working condition is collected as the input sample, and the six-dimensional force Foi (d) Dynamic calibration of the measurement platform using the aforementioned GAALBP-based dynamic force calibration method as the output sample.

[0107] According to the actual working condition, the data to be measured is collected, and the dynamic force measurement is performed using the calibrated measurement platform.

[0108] Figure 3 The flow of the GAALBP-based dynamic force measurement method according to an embodiment of the present application is shown.

[0109] As shown in Figure 3 To achieve the convenience of the measurement process and avoid the need to judge and calibrate before each measurement, the measurement method of the present application can also calibrate all measurable conditions first to obtain multiple calibration models. When measuring and applying, only the appropriate calibration model needs to be selected from the calibration model library according to the condition of the sensor 4, and direct measurement can be performed. This method requires calibration of multiple conditions in the early stage, but can be directly applied in the later measurement. Moreover, the calibration process of multiple conditions using a neural network is not complex, and the specific process is as follows:

[0110] a. Collect sufficient dynamic calibration data. The six-dimensional force F oi (ω) needs to cover the six-dimensional force in space.

[0111] b. Dynamic calibration is performed for fault-tolerant and non-fault-tolerant measurements. When calibrating non-fault-tolerant measurements, 12 output signals from four sensors 4 are used for calibration.

[0112] When calibrating fault-tolerant measurements, there are several different sensor combinations. The four three-dimensional force sensors 4 are numbered as sensor (1), sensor (2), sensor (3), and sensor (4). Then the sensor combinations are as follows:

[0113] (1, 2), (1, 3), (1, 4), (2, 3), (2, 4), (3, 4), (1, 2, 3), (1, 2, 4), (1, 3, 4), (2, 3, 4).

[0114] Dynamic calibration is performed using output signals from the above ten sensor combinations. For example, for sensors with combination (1, 2, 3), 9 output signals from 3 sensors are used to perform dynamic calibration in combination with the known input force of the input sensor.

[0115] A total of 11 calibration models can be obtained for the above fault-tolerant and non-fault-tolerant measurements.

[0116] c. First, use the coherence function to determine the working condition of each sensor 4. The expression of the coherence function is as follows:

[0117]

[0118] wherein, G fx represents the cross spectrum of the input and output signals of sensor 4, G ff represents the auto spectrum of the output signal of sensor 4, G xx represents the auto spectrum of the input signal of sensor 4, G 2 represents the function value, 0≤G 2 ≤1.

[0119] The working condition of sensor 4 is determined according to the function value G 2 , and the specific determination rule is the same as described above, which is not repeated here.

[0120] d. Determine whether the measurement platform meets the measurement requirements according to the working condition of sensor 4. The case of not meeting the measurement requirements is the same as described above, which is not repeated here.

[0121] e. Select the calibration model that fits according to the working condition of all sensors 4. When all sensors are working normally, the calibration model of non-fault-tolerant measurement is preferentially selected through experimental verification.

[0122] f. Collect the data to be measured, and measure using the selected calibration model.

[0123] The dynamic force calibration method and measurement method based on GAALBP of the application are not only suitable for the split load type force platform mentioned in the application, but also suitable for other forms of dynamic force measurement platforms. In addition, the number of redundant outputs of the measurement platform and the number of output signals selected for fault-tolerant measurement can also be adjusted according to the requirements to achieve better measurement results. In addition, a certain sensor can only fail a certain channel, and the calibration and measurement process is essentially the same as described above.

[0124] A large number of experiments were conducted to verify the feasibility and superiority of the method of the application, and the experimental results are as follows:

[0125] The measurement platform as shown in Figure 1 , the dynamic average measurement accuracy of non-fault-tolerant and fault-tolerant based on the traditional least squares method is 12.10% and 26.00% respectively, and the average coupling degree between the dimensions is 8.65% and 16.49% respectively; while using the method of the application, the dynamic average measurement accuracy of non-fault-tolerant and fault-tolerant is 2.53% and 5.50% respectively, and the average coupling degree between the dimensions is 1.27% and 2.10% respectively, thus it can be seen that the method of the application has feasibility and superiority, improves the accuracy of non-fault-tolerant and fault-tolerant measurement, and compensates for the error caused by system nonlinearity and fault-tolerant measurement.

[0126] In addition, the same measurement platform is used to carry out a non-fault-tolerant measurement experiment, the dynamic measurement precision and coupling degree based on the traditional RBF are 6.08% and 4.58%, the calibration time is about 30 minutes, the dynamic measurement precision and coupling degree based on the traditional BP are 5.05% and 2.17%, the calibration time is about 2 hours, the dynamic measurement precision and coupling degree based on the traditional GABP are 4.65% and 1.99%, the calibration time is about 4 days, and the dynamic calibration time of the application is about 1 day, which is due to the adaptive adjustment of the learning rate. As can be seen, the measurement precision of the application is better than other traditional intelligent algorithms, and the calibration time is greatly reduced compared with the GABP algorithm with lower precision, and the calibration and measurement performance of the application is optimal in terms of comprehensive precision, coupling degree and calibration time.

[0127] Although the embodiments of the present application have been shown and described above, it should be understood that the above-described embodiments are exemplary, and should not be construed as limiting the present application. Those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

[0128] The specific embodiments of the application described above do not constitute a limitation on the scope of protection of the application. Any various other corresponding changes and modifications made according to the technical concept of the application shall be included in the scope of protection of the claims of the application.

Claims

1. A dynamic force calibration method based on GAALBP, characterized in that, The method comprises the following steps: S1, collect multiple sets of calibration data of the measurement platform {V ci ( ω ), F oi ( ω )} are divided into training group and validation group, and the output signal V ci ( ω ) of the sensor is collected as the input sample, and the six-dimensional force F oi ( ω ) obtained through the sensor is taken as the output sample; S2, establishing a BP network, optimizing the weight and threshold of the BP network by using GA, adjusting the learning rate of the BP network according to the output error during the training of the BP network, and the specific process is as follows: S21, generating the weight and threshold of the BP network; S22, the first a The input samples corresponding to each frequency point are input into P BP networks for training. The output error is calculated based on the output values ​​of the BP networks, and it is determined whether the BP convergence condition is met. If the BP convergence condition is met, go to S24; If the BP convergence condition is not met, go to S23; S23, adaptively adjusting the learning rate of the BP network according to the learning rate adjustment criterion, updating the weight and threshold of the BP network, and going to S22; The learning rate adjustment criterion is as follows: When then there is: ； ； ； When then there is: ； ； ； When then there is: ； ； ； wherein, error represents an output error, lr represents a learning rate of the BP network, w represents a weight of the BP network, b represents a threshold of the BP network, △ w represents a change amount of the weight, △ b represents a change amount of the threshold, n represents a round; S24, calculating the fitness in GA, and judging whether the GA convergence condition is met: If the GA convergence condition is met, i.e., the training of the a first frequency point is completed; If the GA convergence condition is not met, go to S25; S25, encoding the weight and threshold of the BP network, selecting the coded data for genetic operation according to the fitness, and decoding to go to S22; S3, repeating S2 to complete the training of all frequency points, that is, obtaining the calibration model.

2. The GAALBP-based dynamic force calibration method of claim 1, wherein, S1 inputting test force to multiple physical points of the measurement platform multiple times, collecting sensor signals F ci ( ω ) and output signals V ci ( ω ) of the sensor, and converting sensor signals F ci ( ω ) into six-dimensional force F oi ( ω ), wherein the collection frequency is f and the collection time is t.

3. The GAALBP-based dynamic force calibration method of claim 1, wherein, The calibration data {V ci ( ω ), F oi ( ω )} is normalized for preprocessing to eliminate the influence of variables of different orders of magnitude on network training.

4. The GAALBP-based dynamic force calibration method of claim 1, wherein, In S21, the value range of the weight of the BP network and the value range of the threshold are set, and the weight and threshold of the BP network are randomly generated.

5. The GAALBP-based dynamic force calibration method of claim 2, wherein, The fitness in GA is calculated according to the following formula Fitness : ； wherein, F jk-real ω a represents the first predicted force in the first verification group at the first frequency point, a j k F jk-pre ω a represents the first predicted force in the first verification group at the first frequency point, a j k J represents the number of groups of verification groups, K represents the number of physical points.​​​​​​​​ 6. The GAALBP-based dynamic force calibration method of claim 1, wherein, In S25, the genetic operation can adopt one or more of gene selection, crossover and mutation in GA.

7. A dynamic force measurement method based on GAALBP, characterized in that, The method comprises the following steps: The working condition of the sensor is judged by using the coherence function, and the expression of the coherence function is as follows: ； wherein G fx ω denotes the cross-spectrum of the sensor input signal and the output signal, G ff ω denotes the auto-spectrum of the sensor output signal, G xx ω denotes the auto-spectrum of the sensor input signal, γ 2 denotes the function value, 0≤ γ 2 ≤1;​​​ Determine whether the measurement platform meets the measurement requirements based on the sensor's operating status. If it does, collect the output signal V from the sensor that is operating normally. ci ( ω Using this as input sample, the six-dimensional force F obtained by a sensor in normal working condition is... oi ( ω As the output sample, the measurement platform is dynamically calibrated using the GAALBP-based dynamic force calibration method as described in any one of claims 1-6; the data to be measured is collected, and dynamic force measurement is performed using the calibrated measurement platform.

8. The GAALBP-based dynamic force measurement method of claim 7, wherein, When γ 2 > 0.8, the sensor operating condition is determined to be normal; When 0.3 < γ 2 When the value is less than 0.8, the sensor's operating status is determined to be affected by environmental noise interference. When γ 2 <0.3, the sensor operating condition is determined to be failed.

9. The GAALBP-based dynamic force measurement method of claim 8, wherein, When the working conditions of all sensors are disturbed by environmental noise, the measurement requirement is not met, and the measurement is stopped. When the number of failed sensors is greater than 50% of the total number of sensors, the measurement requirement is not met, and the measurement is stopped.

Images