An Adaptive Genetic Algorithm Neural Network

Abstract

The present invention discloses an adaptive genetic algorithm neural network for compensating the temperature of a MEMS accelerometer; including: S1, constructing a dynamic model of a MEMS accelerometer; S2, deriving a nonlinear relationship between the output error and temperature of the MEMS accelerometer from the dynamic model; S3, constructing a BP neural network model, and deriving the principle of the BP neural network for obtaining a local optimal solution through a series of nonlinear function transformations; S4, introducing an adaptive genetic algorithm when updating the weight of the BP neural network to prevent falling into a local optimum; S5, verifying the effectiveness of the algorithm through actual experiments. The present invention can effectively avoid the process of weight update falling into a local optimum, improve the robustness of the compensation algorithm, and significantly improve the accuracy of sensitive devices.

Description

Technical Field

[0001] The invention relates to the technical field of temperature compensation of MEMS inertial systems, and in particular to a neural network of an adaptive genetic algorithm. Background Art

[0002] MEMS inertial group is widely used in attitude measurement and control system due to its advantages of low cost, small size, low power consumption, high reliability and easy mass production. The current attitude measurement and control system is mainly composed of GPS Beidou satellite navigation system and inertial combined navigation. The significant advantage of MEMS inertial group is that it does not rely on external information, is not subject to external interference, does not emit electromagnetic energy to the outside, and has good concealment.

[0003] At present, the accuracy of inertial measurement units is often constrained by the requirements of small size and low cost. Improving the accuracy of inertial units through manufacturing processes and materials can no longer meet the requirements. Therefore, it is necessary to evaluate the performance of inertial devices. The accuracy of inertial devices can be further improved through algorithm compensation. Summary of the invention

[0004] In view of the above problems, the present invention provides an adaptive genetic algorithm neural network for compensating the temperature of a MEMS meter.

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

[0006] An adaptive genetic algorithm neural network for compensating the temperature of a MEMS accelerometer comprises the following steps:

[0007] S1. Construct the dynamic model of MEMS accelerometer;

[0008] S2, the nonlinear relationship between the output error and temperature of the MEMS accelerometer is derived from the dynamic model;

[0009] S3, construct a BP neural network model, and deduce the principle of BP neural network to obtain local optimal solution through a series of nonlinear function transformations;

[0010] S4. Introduce adaptive genetic algorithm when updating BP neural network weights to prevent falling into local optimum;

[0011] S5. Verify the effectiveness of the algorithm through practical experiments.

[0012] As a preferred embodiment of the above solution, in step S1, the dynamic model of the MEMS accelerometer is:

[0013]

[0014] Where m is the mass of the object, x is the size of the displacement, α is the damping of the entire elastic system, k is the elastic stiffness of the system, and a is the acceleration generated by the inertial object.

[0015] As a preferred embodiment of the above solution, step S2 specifically includes:

[0016] Divide both sides of formula (1.1) by m to solve the kinematic equation of the system:

[0017]

[0018] In the formula, is the natural frequency of the inertial object, is the damping ratio of the system;

[0019] The influence of temperature on the accelerometer lies mainly in the change of elastic coefficient. The accelerometer has the following relationship to the change of elastic coefficient:

[0020]

[0021] Where ω0 is the resonant frequency of the system, k0 is the system stiffness at time T0, and k ET is the temperature coefficient of elastic modulus;

[0022] Based on the above formula, we can know that temperature has a direct impact on the output of the accelerometer. The material of the sensor is more sensitive to temperature changes. The same MEMS inertial group system can measure the relationship between different temperatures and the actual output error. Therefore, it is necessary to use an efficient algorithm with strong nonlinear fitting capabilities to fit the temperature and sensitive data.

[0023] As a preferred embodiment of the above solution, step S3 specifically includes:

[0024] The construction of BP neural network includes setting the maximum number of learning times, the initial network learning rate, and the error transfer function;

[0025] In the process of network training, through the cyclic iteration of network weights and thresholds, the predicted value of the BP neural network is finally approached to the true value.

[0026] During the network training process, the input layer and hidden layer are constructed, including the input layer weights, bias size, and error loss value;

[0027] Construct the error propagation equation, calculate the partial derivative between each neuron in the output layer and the error function, and adjust the size of the weight bias;

[0028] Adjust the weights and error values ​​by reversely derivatizing the global error loss function.

[0029] As a preferred embodiment of the above solution, step S4 specifically includes:

[0030] Starting from the initial total population of M populations, after selection, compilation, and crossover operations, the optimal population is selected by testing accuracy;

[0031] Initialize the population, where M(t) is the population, satisfying M(t)={m1(t),m2(t),...,m Nt (t)},N t is the size of the population at the tth generation, N0 = N, N is a constant;

[0032] Calculate the fitness and life cycle of each population. The life cycle is the number of generations that an individual can survive in the algorithm iteration. The size of the life cycle is determined by the fitness of the individuals in the population.

[0033] The appropriate fitness function is iteratively selected through chromosome crossover mutation, the fitness of individuals is recalculated, and the population is repeatedly updated using crossover mutation;

[0034] Boundary condition check: In order to prevent individual genes from exceeding the solution range after crossover mutation, a boundary condition check mechanism is established. max And M(t+1)<β min , the mechanism satisfies the formula M(t+1)=β min +η(β max -β min ), where β max With β min are the upper and lower boundary values ​​of the population gene.

[0035] As a preferred embodiment of the above scheme, the life cycle of an individual satisfies the following relationship:

[0036]

[0037] In the formula, g avg (t,m) is the average value in the tth generation and the best fitness value in the tth generation, g min (t,m) is the worst fitness value in the tth generation. After being processed by the above formula, livetime(t+1,m) is the life cycle of the next generation of individuals.

[0038] As a preferred embodiment of the above scheme, the crossover of chromosomes is specifically as follows:

[0039] In the current population, the fitness of two adjacent individuals is compared in turn, and the one with the larger fitness is taken as the fitness F, where P cis the crossover rate, which will change adaptively with the evolution of the population. If the fitness value of the selected individual is lower than the current fitness value of the population, a greater crossover probability will be given; if the fitness value of the selected individual population is higher than the current fitness value, a lower crossover probability will be given:

[0040]

[0041] Where P c is the adaptive crossover rate, P c1 ∈(0,1),P c2 ∈(0,1) is a constant, and P c1 >P c2 , F max is the maximum fitness of the population, F is the maximum fitness between adjacent neighbors, and F a is the average fitness of the current population.

[0042] Due to the above structure, the beneficial effects of the present invention are:

[0043] The present invention proposes a BP neural network model by constructing the kinematic equation of the meter and analyzing the nonlinear relationship between the output error and temperature of the MEMS meter. The most notable feature of the BP neural network model is the normalization preprocessing and parallel calculation of complex data, which is suitable for large-scale data sets and complex structural models. Based on the original BP neural network compensation algorithm, a new adaptive genetic algorithm is proposed. This method can effectively avoid the weight update falling into the local optimal process, improve the robustness of the compensation algorithm, and significantly improve the accuracy of sensitive devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2 This is a comparison chart of the prediction curves of the traditional BP neural network and the optimized BP neural network. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, this embodiment provides an adaptive genetic algorithm neural network for compensating the temperature of a MEMS accelerometer, comprising the following steps:

[0049] S1. The dynamic model of the MEMS accelerometer is constructed as follows:

[0050]

[0051] Where m is the mass of the object, x is the size of the displacement, α is the damping of the entire elastic system, k is the elastic stiffness of the system, and a is the acceleration generated by the inertial object.

[0052] S2. The nonlinear relationship between the output error and temperature of the MEMS accelerometer is derived from the dynamic model. Specifically:

[0053] Divide both sides of formula (1.1) by m to solve the kinematic equation of the system:

[0054]

[0055] In the formula, is the natural frequency of the inertial object, is the damping ratio of the system;

[0056] The influence of temperature on the accelerometer lies mainly in the change of elastic coefficient. The accelerometer has the following relationship to the change of elastic coefficient:

[0057]

[0058] Where ω0 is the resonant frequency of the system, k0 is the system stiffness at time T0, and k ET is the temperature coefficient of elastic modulus;

[0059] Based on the above formula, it can be seen that temperature has a direct impact on the output of the accelerometer. The material of the sensor is more sensitive to temperature changes, which will cause zero drift and affect the accuracy of the sensor data output. In the actual temperature compensation algorithm, the peripheral circuit and the traditional analog algorithm can be used to compensate for the temperature. Due to the inaccuracy of the model, the output accuracy of the sensor's sensitive information is limited and requires repeated debugging, so a new temperature compensation technology is needed;

[0060] The same MEMS inertial system can measure the relationship between different temperatures and the actual output error, so it is necessary to use an efficient algorithm with strong nonlinear fitting capabilities to fit the temperature and sensitive data.

[0061] S3. Construct a BP neural network model and deduce the principle of BP neural network to obtain local optimal solution through a series of nonlinear function transformations, specifically:

[0062] During the training process, the network weights and thresholds are iterated cyclically, and the predicted value of the BP neural network is finally close to the true value. The initialization of the BP neural network is to take a random number in the interval (-1,1), set the maximum number of learning times, error transfer function, and calculation accuracy;

[0063] During the neural network training process, the input layer and hidden layer are constructed, including the input layer weight, bias size, and error loss value. In the neural network, the input layer, output layer, and hidden layer satisfy the following relationship:

[0064] x={x1,x2,...,x n} (1.4)

[0065] h i ={h i1 ,h i2 ,...,h im} (1.5)

[0066] h o ={h o1 ,h o2 ,...,h om} (1.6)

[0067] u i = {u i1 ,u i2 ,...,u is} (1.7)

[0068] u o = {u o1 ,u o2 ,...,u os} (1.8)

[0069] d o ={d1,d2,...,d s} (1.9)

[0070] Among them, the input variable is x, and the input and output of the hidden layer are divided into h i 、h o , the input and output of the output layer are u i 、u o , the expected value is d o ;

[0071] The relationship between the input layer and the hidden layer is as follows:

[0072]

[0073] h oh (k) = f(h ih (k))(1.11)

[0074]

[0075] y o (k) = f(y ih (k))(1.13)

[0076] where ω ih is the combination of the weights of the neurons in the previous layer received by the h-th neuron, x i (k) is the data of the input sample, b h is the weight bias, f(h ih (k)) is a nonlinear function, which is used to couple the nonlinear relationship between input and output;

[0077] Construct the error propagation equation, calculate the partial derivative between each neuron in the output layer and the error function, adjust the weight bias, and the global error loss function of the neural network satisfies the following relationship:

[0078]

[0079] Adjust the weights and error values ​​by reversely derivatizing the global error loss function.

[0080] S4. When the traditional BP neural network algorithm reaches the point where the error change is flat during the training of the network model, the weight change in the iteration becomes significantly smaller, and the algorithm model is difficult to learn further, and the algorithm is prone to fall into the local minimum. Therefore, an adaptive genetic algorithm is introduced when the BP neural network weights are updated to prevent falling into the local optimum. The genetic algorithm is an evaluation function of the initial population and individual fitness. The specific implementation method is as follows:

[0081] Starting from the initial total population of M populations, after selection, compilation, and crossover operations, the optimal population is selected by testing accuracy;

[0082] Initialize the population, where M(t) is the population, satisfying N t is the size of the population at the tth generation, N0 = N, N is a constant;

[0083] Calculate the fitness and life cycle of each population. The life cycle is the number of generations that an individual can survive in the algorithm iteration. The size of the life cycle is determined by the fitness of the individuals in the population. The life cycle of an individual satisfies the following relationship:

[0084]

[0085] In the formula, g avg (t,m) is the average value in the tth generation and the best fitness value in the tth generation, g min(t,m) is the worst fitness value in the tth generation. After being processed by the above formula, livetime(t+1,m) is the life cycle of the next generation of individuals;

[0086] The appropriate fitness function is iterated through the crossover mutation of chromosomes to recalculate the fitness of individuals and update the population repeatedly using crossover mutation: the fitness of two adjacent individuals in the current population is compared in turn, and the one with the larger fitness is taken as the fitness F, where P c is the crossover rate, which will change adaptively with the evolution of the population. If the fitness value of the selected individual is lower than the current fitness value of the population, a greater crossover probability will be given; if the fitness of the selected individual population is higher than the existing fitness, a lower crossover probability will be given, which is conducive to protecting excellent individuals from harm:

[0087]

[0088] Where P c is the adaptive crossover rate, P c1 ∈(0,1),P c2 ∈(0,1) is a constant, and P c1 >P c2 , F max is the maximum fitness of the population, F is the maximum fitness between adjacent neighbors, and F a is the average fitness of the current population;

[0089] Boundary condition check: In order to prevent individual genes from exceeding the solution range after crossover mutation, a boundary condition check mechanism is established. max And M(t+1)<β min , the mechanism satisfies the formula M(t+1)=β min +η(β max -β min ), where β max With β min are the upper and lower boundary values ​​of the population gene.

[0090] S5. Verify the effectiveness of the algorithm through actual experiments. Specifically:

[0091] In the actual experiment, 500 MEMS accelerometers were temperature compensated. The temperature started from 1 degree Celsius and continued to rise to 60 degrees Celsius. When the temperature reached 60 degrees, the temperature was kept constant for one hour to complete the heating process. Then the temperature was cooled from 1 degree Celsius to -40 degrees Celsius and the temperature was kept constant for one hour to complete the cooling process. Until the end of the experiment, all temperature compensation experimental data were collected.

[0092] The MEMS output value when the temperature changes in the experiment is used as the training set of the BP neural network and input into the traditional BP neural network and the optimized BP neural network. Figure 2 As shown. The horizontal axis represents the number of iterations, and the vertical axis represents the accuracy of the final prediction. From the curve, it can be seen that the accuracy after optimization is significantly higher than that of the traditional BP neural network. The accuracy of the optimized BP neural network curve is 0.985, and the accuracy of the unoptimized BP neural network is 0.954.

[0093] In order to verify the effectiveness of the proposed algorithm, the MEMS inertial group was compared under multiple methods, and the experimental results are shown in the following table:

[0094]

[0095] Among them, RMES is the mean square error, R_square is the coefficient of determination, RMES represents the difference between the true value and the ideal value, R_square represents the fitness of the fitting model and the sample, the smaller the RMES and the larger the R_square, the better the model fit. It can be seen from the table that the proposed method has the smallest RMES value and the largest R_square value, which has the highest fit with the model and the best effect.

[0096] The present invention proposes a BP neural network model by constructing the kinematic equation of the meter and analyzing the nonlinear relationship between the output error and temperature of the MEMS meter. Compared with traditional temperature compensation algorithms such as the least squares method and the polynomial fitting method, the most notable feature of the BP neural network model is the normalization preprocessing and parallel calculation of complex data, which is suitable for large-scale data sets and complex structural models. On the basis of the original BP neural network compensation algorithm, an adaptive genetic algorithm method is proposed to update the weights. This method updates the weights and thresholds by encoding and decoding the input variables, which can effectively avoid the weights from falling into the local optimum, improve the robustness of the compensation algorithm, and the zero position after compensation is consistent with the scale factor and the design value, and the accuracy of the sensitive device is significantly improved.

[0097] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An adaptive genetic algorithm neural network for compensating the temperature of a MEMS accelerometer, characterized in that: The following steps are involved: S1. Construct the dynamic model of MEMS accelerometer. S2, the nonlinear relationship between the output error and temperature of the MEMS accelerometer is derived from the dynamic model; S3, construct a BP neural network model, and deduce the principle of BP neural network to obtain local optimal solution through a series of nonlinear function transformations; S4. Introduce adaptive genetic algorithm when updating BP neural network weights to prevent falling into local optimum; S5. Verify the effectiveness of the algorithm through practical experiments.

2. The adaptive genetic algorithm neural network according to claim 1, characterized in that: In step S1, the dynamic model of the MEMS accelerometer is: Where m is the mass of the object, x is the size of the displacement, α is the damping of the entire elastic system, k is the elastic stiffness of the system, and a is the acceleration generated by the inertial object.

3. The adaptive genetic algorithm neural network according to claim 1, characterized in that: Step S2 specifically includes: Divide both sides of formula (1.1) by m to solve the kinematic equation of the system: In the formula, is the natural frequency of the inertial object, is the damping ratio of the system; The influence of temperature on the accelerometer lies mainly in the change of elastic coefficient. The accelerometer has the following relationship to the change of elastic coefficient: Where ω0 is the resonant frequency of the system, k0 is the system stiffness at time T0, and k ET is the temperature coefficient of elastic modulus; Based on the above formula, we can know that temperature has a direct impact on the output of the accelerometer. The material of the sensor is more sensitive to temperature changes. The same MEMS inertial group system can measure the relationship between different temperatures and the actual output error. Therefore, it is necessary to use an efficient algorithm with strong nonlinear fitting capabilities to fit the temperature and sensitive data.

4. The adaptive genetic algorithm neural network according to claim 1, characterized in that: Step S3 specifically includes: The construction of BP neural network includes setting the maximum number of learning times, the initial network learning rate, and the error transfer function; In the process of network training, through the cyclic iteration of network weights and thresholds, the predicted value of the BP neural network is finally approached to the true value. During the network training process, the input layer and hidden layer are constructed, including the input layer weights, bias size, and error loss value; Construct the error propagation equation, calculate the partial derivative between each neuron in the output layer and the error function, and adjust the size of the weight bias; Adjust the weights and error values ​​by reversely derivatizing the global error loss function.

5. The adaptive genetic algorithm neural network according to claim 1, characterized in that: Step S4 specifically includes: Starting from the initial total population of M populations, after selection, compilation, and crossover operations, the optimal population is selected by testing accuracy; Initialize the population, where M(t) is the population, satisfying M(t)={m1(t),m2(t),...,m Nt (t)},N t is the size of the population at the tth generation, N0 = N, N is a constant; Calculate the fitness and life cycle of each population. The life cycle is the number of generations that an individual can survive in the algorithm iteration. The size of the life cycle is determined by the fitness of the individuals in the population. The appropriate fitness function is iteratively selected through chromosome crossover mutation, the fitness of individuals is recalculated, and the population is repeatedly updated using crossover mutation; Boundary condition check: In order to prevent individual genes from exceeding the solution range after crossover mutation, a boundary condition check mechanism is established. max And M(t+1)<β min , the mechanism satisfies the formula M(t+1)=β min +η(β max -β min ), where β max With β min are the upper and lower boundary values ​​of the population gene.

6. The adaptive genetic algorithm neural network according to claim 5, characterized in that: The life cycle of an individual satisfies the following relationship: In the formula, g avg (t,m) is the average value in the tth generation and the best fitness value in the tth generation, g min (t,m) is the worst fitness value in the tth generation. After being processed by the above formula, livetime(t+1,m) is the life cycle of the next generation of individuals.

7. The adaptive genetic algorithm neural network according to claim 5, characterized in that: The specific crossover of chromosomes is: In the current population, the fitness of two adjacent individuals is compared in turn, and the one with the larger fitness is taken as the fitness F, where P c is the crossover rate, which will change adaptively with the evolution of the population. If the fitness value of the selected individual is lower than the current fitness value of the population, a greater crossover probability will be given; if the fitness value of the selected individual population is higher than the current fitness value, a lower crossover probability will be given: Where P c is the adaptive crossover rate, P c1 ∈(0,1),P c2 ∈(0,1) is a constant, and P c1 >P c2 , F max is the maximum fitness of the population, F is the maximum fitness between adjacent neighbors, and F a is the average fitness of the current population.

Images