The traditional acquisition of cutting power requires complex cutting power models and rarely considers the impact of tool wear, thus designing a CNC milling machine cutting power prediction model based on Variational Mode Decomposition – Sparrow Search Algorithm – Long Short-Term Memory (VMD-SSA-LSTM), which considers the impact of tool wear and can predict cutting power with high accuracy. Artificial intelligence machine vision technology is used to analyze and process images of tool wear to obtain digital features, and then a prediction model is established: using VMD to decompose operational data, SSA algorithm to optimize LSTM hyperparameters, and inputting the decomposed data into the LSTM network to obtain the cutting power prediction value. Finally, taking face milling as an example, the proposed prediction model is compared with BP neural networks, LSTM neural networks, and traditional models to verify the effectiveness and superiority of the proposed model.

China has a large number of machine tools and consumes a lot of energy, especially CNC machine tools, so accurate and effective energy consumption prediction models for machine tools are crucial for energy efficiency evaluation and optimization. In the modeling of cutting power, most existing studies focus on cutting power models considering process parameters and idle power models, while neglecting the significant impact of tool wear on cutting power and processing quality. Tool wear continuously affects the cutting power of machine tools; however, current research often employs artificial classification methods, without considering the continuous impact of time-varying tool wear on machine tool cutting power. To address these issues, this paper introduces online tool wear extraction technology to obtain real-time dynamic wear values of tools. Additionally, combining the characteristics of cutting power and the advantages and disadvantages of neural networks, this paper establishes a CNC milling machine cutting power prediction model based on VMD-SSA-LSTM considering tool wear.

The acquisition of cutting power is achieved by connecting the CW500 sensor to the machine tool power supply to obtain real-time voltage and current data during the CNC milling process, with a collection frequency of 0.1s. The connection method of the sensor is shown in Figure 2.

Figure 2 Cutting Power Acquisition Platform

The subject of this study is the face milling cutter. Considering the convenience of image acquisition and the complexity of measurement algorithms, the maximum wear width of the rear tool face is chosen as the evaluation index. First, the high-definition images obtained by the industrial camera are pre-processed, then the Canny operator, sub-pixel edge detection, and overlapping edge detection are used to extract the boundaries of tool wear, and finally, the maximum wear width is obtained by using the bottom edge line of the rear tool face as one side of the rectangle and enclosing the minimum rectangle. The main operation process is shown in Figure 3, and the image processing results are shown in Figure 4.

Figure 3 Image Processing Flow

Figure 4 Image Processing Results

The VMD-SSA-LSTM CNC milling machine cutting power prediction model has the following specific prediction steps:

(1) Select the historical operation information of the CNC milling machine as model input.

(2) Use the VMD method to decompose the original cutting power sequence to obtain sub-sequence components.

(3) Data normalization. Due to the large difference between the spindle speed and the back feed amount, the iteration speed decreases. After normalization, the feature values are between 0 and 1, which improves the model’s convergence speed, and dimensionless processing can enhance the model’s accuracy.

(4) First, set the sparrow population size, maximum iteration times, and parameter ranges (number of hidden layer neurons, training times, and initial learning rate) for the search range, then use the minimum root mean square error as the objective function in the optimization algorithm, and finally establish the model combining the sparrow search algorithm with the long short-term neural network (SSA-LSTM).

(5) Each component is input into the SSA-LSTM prediction model to obtain individual prediction models.

(6) Finally, the predicted values of each prediction model are summed correspondingly to obtain the predicted value of the cutting power.

The processing equipment used in the experiment is the Jiasite V-11 CNC milling machine, and the input power signal of the milling machine is collected through the CW500 power measurement device. The experimental material is a 400mm×200mm×50mm plate of 45# steel, and the diameter of the tool holder used in the experiment is a 16mm double-edged tool, with the tool being a face milling cutter with JDL APMT1135 carbide inserts. An industrial camera from Hikvision MV-CS050-10GM is used to capture images of the wear on the rear face of the milling cutter. Considering the limitations of the internal spatial structure of the machine tool, the image acquisition system is deployed on the right side of the spindle, consisting of a camera, lens MVL-MY-018-150-MP, ring light source MV-LRDS-120-45-W, camera support, data transmission line, and light source controller, as shown in Figure 5. The Jiasite CNC milling machine mills the plate as shown in Figure 6. The experimental design includes 25 sets of processing schemes, with each scheme having 40 passes, each pass being 800 mm.

Figure 5 Experimental Deployment Diagram

Figure 6 Face Milling Cutter Path

To verify the effectiveness of the proposed method for extracting the maximum tool wear, this paper uses an industrial microscope for accuracy experiments. A sample of 30 groups is measured for tool wear values with a microscope. To reduce measurement random errors, each group of control group wear measurements is taken three times to obtain an average value denoted as Wa, while the wear amount measured using the image technology method in this paper is denoted as Wb. The measurement deviation ∆ and relative error δ are used as evaluation indicators, and the industrial microscope measurement process is shown in Figure 7. The detection system is compared with the control group data, with a measurement deviation of less than 0.05mm and an average relative error of 4.70%, indicating excellent accuracy and meeting the requirements for milling cutter wear detection.

Figure 7 Industrial Microscope Detection

Select the first 70% of experimental data as training samples for the model, and the remaining 30% as test samples. Using the trained model, predictions are made on the test set, and the specific prediction results are shown in Figure 8.

Figure 8 Results of VMD-SSA-LSTM Prediction

As shown in Figure 8, the predicted values of cutting power are close to the actual values. The prediction results and accuracy of the three models are shown in Figure 9, indicating that the prediction effect of VMD-SSA-LSTM is better than that of the single LSTM and SSA-LSTM models.

Figure 9 Comparison of Prediction Results and Effects of Three Models

Calculations show that the average absolute percentage error of the BP neural network prediction is 13.58%, the average absolute percentage error of the LSTM neural network is 8.95%, the average absolute percentage error of the VMD-LSTM is 5.12%, and the average absolute percentage error of the VMD-SSA-LSTM is 1.53%. By introducing the sparrow search algorithm into the VMD-LSTM model, the optimization of neural network hyperparameters is facilitated. The VMD-SSA-LSTM model has better evaluation indicators than the VMD-LSTM model, and after adding the SSA algorithm, the average absolute percentage error is reduced by 3.59%, with both the average absolute error and root mean square error being lower than those of the single LSTM neural network model and VMD-LSTM model.