Mining trucks number, mining trucks type, engine speed, oil temperature, transmission gear, reverse state, parking state, braking state, design slope, loading point, unloading point, operation state, mining trucks load, speed, acceleration, angular acceleration, operation duration, movement distance, lifting height, actual slope, vibration intensity, peak factor, kurtosis parameter, weather level, temperature, humidity and average precipitation. The data is collected every second as a set of original data. The original data is merged by feature clustering according to the mining trucks number on the continuous time axis to form one input sequence. 4.3 Network Structure Design The model is composed of multiple layers of LSTM units stacked to enhance the ability to express complex time-series features. Each LSTM layer performs feature abstraction on the input sequence through non-linear mapping and extracts high-level time-series features layer by layer. The model adopts a lightweight architecture of "input layer - 2 LSTM layers - Dropout layer - fully connected layer - output layer" to ensure fitting accuracy while reducing computational complexity. The specific structure is as follows: LSTM layers: The first layer is set with 64 neurons, which is responsible for extracting basic time-series features such as speed, slope and speed. The second layer is set with 32 neurons to further explore multi-factor coupling features such as strong vibration, large slope and sudden acceleration. Both layers use the tanh activation function, taking into account the feature expression ability and computational efficiency. Dropout layer: The dropout ratio is set to 0.2, which randomly shields the connections of some neurons to avoid the model's over-reliance on local samples and improve the generalization ability. Fully connected layer: It is set with 16 neurons and uses the ReLU activation function to integrate the high-dimensional features output by the LSTM layer. Output layer: The linear activation function is used to output the average instantaneous fuel consumption (L/h) corresponding to the per-second time window as the core index of energy efficiency. 4.4 Model Training Strategy In the process of model training, an adaptive optimization algorithm based on gradient descent is used to update the network parameters, and the mean square error is used as the loss function to guide the model to continuously approach the real fuel consumption level. By setting the learning rate and training rounds reasonably, the model can avoid over-fitting while ensuring the convergence efficiency. The design of the above model structure and training strategy focuses on constructing a stable and energy efficiency modeling framework, rather than over-fitting for a specific data set. Optimizer: Adam optimizer, with an initial learning rate of 0.001, which decays by 10% every 10 rounds to balance the convergence speed and later accuracy. Loss function: Mean Square Error (MSE), which accurately describes the deviation between the predicted value and the real fuel consumption.
RkJQdWJsaXNoZXIy MTM0Mzk2