BP神经网络编程应用全解析-如何用C语言实现PID算法

教程大全 2026-03-04 01:38:07 浏览次

#include #include #include #include // PID 控制器结构体typedef struct {double Kp, Ki, Kd;// PID参数double integral;// 积分项double prev_error;// 上一次误差} PIDController;// 神经网络层结构typedef struct {int num_nodes;// 节点数量double *outputs;// 输出值double *errors;// 误差项} NeuralLayer;// BP神经网络结构typedef struct {NeuralLayer input_layer;NeuralLayer hidden_layer;NeuralLayer output_layer;double **w_ih;// 输入层到隐藏层权重double **w_ho;// 隐藏层到输出层权重double learning_rate;// 学习率} BPNeuralNetwork;// 初始化PID控制器void initPID(PIDController *pid, double Kp, double Ki, double Kd) {pid->Kp = Kp;pid->Ki = Ki;pid->Kd = Kd;pid->integral = 0.0;pid->prev_error = 0.0;}// PID控制计算double computePID(PIDController *pid, double setpoint, double actual, double dt) {double error = setpoint - actual;pid->integral += error * dt;double derivative = (error - pid->prev_error) / dt;pid->prev_error = error;return pid->Kp * error + pid->Ki * pid->integral + pid->Kd * derivative;}// 初始化神经网络层void initLayer(NeuralLayer *layer, int num_nodes) {layer->num_nodes = num_nodes;layer->outputs = (double *)malloc(num_nodes * sizeof(double));layer->errors = (double *)malloc(num_nodes * sizeof(double));}// 初始化BP神经网络void initBPNN(BPNeuralNetwork *nn, int input_nodes, int hidden_nodes, int output_nodes, double lr) {// 初始化各层initLayer(&nn->input_layer, input_nodes);initLayer(&nn->hidden_layer, hidden_nodes);initLayer(&nn->output_layer, output_nodes);nn->learning_rate = lr;// 分配权重内存nn->w_ih = (double **)malloc(input_nodes * sizeof(double *));nn->w_ho = (double **)malloc(hidden_nodes * sizeof(double *));// 初始化权重（随机小值）srand(time(NULL));FOR (int i = 0; i < input_nodes; i++) {nn->w_ih[i] = (double *)malloc(hidden_nodes * sizeof(double));for (int j = 0; j < hidden_nodes; j++) {nn->w_ih[i][j] = (rand() % 100) / 100.0 * 0.2 - 0.1; // [-0.1, 0.1]}}for (int i = 0; i < hidden_nodes; i++) {nn->w_ho[i] = (double *)malloc(output_nodes * sizeof(double));for (int j = 0; j < output_nodes; j++) {nn->w_ho[i][j] = (rand() % 100) / 100.0 * 0.2 - 0.1;}}}// Sigmoid激活函数double sigmoid(double x) {return 1.0 / (1.0 + exp(-x));}// Sigmoid导数double sigmoidDerivative(double x) {return x * (1.0 - x);}// 神经网络前向传播void forwardPropagation(BPNeuralNetwork *nn, double *inputs) {// 设置输入层输出for (int i = 0; i < nn->input_layer.num_nodes; i++) {nn->input_layer.outputs[i] = inputs[i];}// 计算隐藏层输出for (int j = 0; j < nn->hidden_layer.num_nodes; j++) {double sum = 0.0;for (int i = 0; i < nn->input_layer.num_nodes; i++) {sum += nn->input_layer.outputs[i] * nn->w_ih[i][j];}nn->hidden_layer.outputs[j] = sigmoid(sum);}// 计算输出层输出for (int k = 0; k < nn->output_layer.num_nodes; k++) {double sum = 0.0;for (int j = 0; j < nn->hidden_layer.num_nodes; j++) {sum += nn->hidden_layer.outputs[j] * nn->w_ho[j][k];}nn->output_layer.outputs[k] = sigmoid(sum);}}// 神经网络反向传播void backwardPropagation(BPNeuralNetwork *nn, double *targets) {// 计算输出层误差for (int k = 0; k < nn->output_layer.num_nodes; k++) {double output = nn->output_layer.outputs[k];nn->output_layer.errors[k] = (targets[k] - output) * sigmoidDerivative(output);}// 计算隐藏层误差for (int j = 0; j < nn->hidden_layer.num_nodes; j++) {double sum = 0.0;for (int k = 0; k < nn->output_layer.num_nodes; k++) {sum += nn->output_layer.errors[k] * nn->w_ho[j][k];}nn->hidden_layer.errors[j] = sum * sigmoidDerivative(nn->hidden_layer.outputs[j]);}// 更新隐藏层到输出层权重for (int j = 0; j < nn->hidden_layer.num_nodes; j++) {for (int k = 0; k < nn->output_layer.num_nodes; k++) {nn->w_ho[j][k] += nn->learning_rate *nn->output_layer.errors[k] *nn->hidden_layer.outputs[j];}}// 更新输入层到隐藏层权重for (int i = 0; i < nn->input_layer.num_nodes; i++) {for (int j = 0; j < nn->hidden_layer.num_nodes; j++) {nn->w_ih[i][j] += nn->learning_rate *nn->hidden_layer.errors[j] *nn->input_layer.outputs[i];}}}// 使用神经网络调整PID参数void adjustPIDWithNN(BPNeuralNetwork *nn, PIDController *pid, double error, double dt) {// 准备输入: 归一化的误差、积分、微分double inputs[3] = {error / 10.0,// 假设误差范围在[-10,10]pid->integral / 100.0,(error - pid->prev_error) / dt / 10.0};// 前向传播获取PID调整量forwardPropagation(nn, inputs);// 输出层对应Kp, Ki, Kd的调整量（范围0.5~2.0倍）double k_adjust[3];for (int i = 0; i < 3; i++) {k_adjust[i] = 0.5 + 1.5 * nn->output_layer.outputs[i]; // 映射到[0.5, 2.0]}// 应用调整pid->Kp *= k_adjust[0];pid->Ki *= k_adjust[1];pid->Kd *= k_adjust[2];}// 训练神经网络（简化示例）void trainNN(BPNeuralNetwork *nn, double actual, double setpoint, double dt) {double error = setpoint - actual;double targets[3] = {1.0, 1.0, 1.0}; // 理想调整量为1（不调整）// 当误差较大时需要调整if (fabs(error) > 0.5) {targets[0] = (error > 0) ? 1.2 : 0.8; // 根据误差方向调整}// 准备输入double inputs[3] = {error / 10.0,0, // 简化示例(error - 0) / dt / 10.0 // 简化示例};forwardPropagation(nn, inputs);backwardPropagation(nn, targets);}// 模拟被控对象（一阶惯性系统）double PLANtmodel(double input, double *state) {*state = 0.9 * (*state) + 0.1 * input;return *state;}int main() {PIDController pid;BPNeuralNetwork nn;double setpoint = 5.0;// 目标值double actual = 0.0;// 实际值double control;// 控制量double plant_state = 0.0;// 被控对象状态const double dt = 0.1;// 控制周期// 初始化PID和神经网络initPID(&pid, 0.5, 0.1, 0.01);initBPNN(&nn, 3, 4, 3, 0.1); // 3输入, 4隐藏节点, 3输出// 控制循环for (int step = 0; step < 200; step++) {// 计算控制量control = computePID(&pid, setpoint, actual, dt);// 应用控制量到被控对象actual = plantModel(control, &plant_state);// 每10步调整一次PID参数if (step % 10 == 0) {double error = setpoint - actual;adjustPIDWithNN(&nn, &pid, error, dt);}// 每5步训练一次神经网络（简化）if (step % 5 == 0) {trainNN(&nn, actual, setpoint, dt);}// 打印结果printf("Step %d: Setpoint=%.2f, Actual=%.4f, Control=%.4fn",step, setpoint, actual, control);}return 0;}