A flexible manufacturing process for industrial robots of new energy vehicle parts

Abstract

This invention discloses a flexible machining process for industrial robots used in the production of new energy vehicle parts. It relates to the field of flexible machining technology for industrial robots. The process utilizes historical load curves and virtual simulation to construct an energy consumption-electromagnetic predictor, predicting the power peak time-domain sequence. Based on the prediction results, it instantly switches servo switching strategies and activates a distributed noise suppression unit to suppress electromagnetic interference. Within the noise suppression window, multimodal signals are simultaneously acquired, and thermal-signal quantization is used to generate a thermal-noise coupling coefficient to filter reliable data. A digital twin compensation model receives confirmation data and outputs a small tool position correction and pulse cooling command, achieving coordinated adjustment of heat source, power, and cooling. After machining, residual dimensional errors and energy consumption indicators are summarized, and the predictor weights are updated adaptively to improve the prediction accuracy for the next workpiece. Through the synergistic effect of prediction, noise suppression, quantization, compensation, and adaptation, machining accuracy and energy efficiency are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of flexible manufacturing technology for industrial robots, specifically a flexible manufacturing process for industrial robots used in new energy vehicle parts. Background Technology

[0002] In flexible machining units for new energy vehicle components, industrial robots are typically equipped with high-speed electric spindles to perform complex processes such as milling and drilling on battery trays and high-voltage busbars. During high-speed cutting, the continuous temperature rise of the spindle and workpiece is considered a core factor causing dimensional accuracy drift. To monitor the machining status in real time, the system simultaneously deploys multimodal sensors, including vision, force / torque, and vibration sensors. These highly sensitive electronic devices must operate in parallel with high-power drive components within a confined space. When the toolpath enters the peak load range, the transient power spikes generated by the electric spindle and servo drive increase overall energy consumption and induce broadband electromagnetic noise. Electromagnetic interference is commonly observed in manufacturing environments, interfering with signal acquisition and causing equipment malfunctions. Complex metal structures further exacerbate the coupling between wireless and wired channels. To suppress thermal deformation and signal drift, researchers have begun to introduce digital twin models. By optimizing sensor layout and implementing thermal error compensation through a virtual environment, the virtual and physical machining chains are kept synchronized.

[0003] However, due to the temporal and spatial overlap of high-speed cutting and multi-sensor collaboration, heat sources, power spikes, and electromagnetic fields intertwine, forming a complex coupling phenomenon that amplifies each other. The gradual heating of the spindle and workpiece causes minute displacements in the geometric reference, leading to a loss of reference for the real-time measurements required for tool position compensation. Subsequently, radiated and conducted interference from the drive system and high-frequency cables enters the sensor link, causing jitter in force signals, visual coordinates, and vibration spectra, affecting the controller's error judgment. When cooling flow and energy consumption scheduling fail to keep pace with load changes, thermal distortion and signal drift jointly compress the credibility boundary of the digital twin model, reducing the predictive compensation effect. As a result, hole positions and slots gradually deviate from design requirements, busbar assembly becomes misaligned, rework rates increase, and the cycle time and quality of the entire flexible unit are repeatedly strained, ultimately weakening the high-precision and high-flexibility machining advantages that this invention aims to achieve.

[0004] Therefore, the present invention provides a flexible processing technology for industrial robots used in new energy vehicle parts. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a flexible machining process for industrial robots used in new energy vehicle parts manufacturing. It utilizes historical load curves and virtual simulation to construct an energy consumption-electromagnetic predictor to predict the power peak time-domain sequence. Based on the prediction results, it instantly switches servo switching strategies and activates a distributed noise suppression unit to suppress electromagnetic interference. Within the noise suppression window, it synchronously acquires multimodal signals, generates a thermal-noise coupling coefficient through thermal-signal quantization, and filters reliable data. A digital twin compensation model receives confirmation data and outputs a tool position micro-correction amount and pulse cooling commands, achieving coordinated adjustment of heat source, power, and cooling. After machining, it summarizes residual dimensional errors and energy consumption indicators, and uses feedback to adaptively update the predictor weights, improving the prediction accuracy for the next workpiece. This invention solves the technical problems described in the background section.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: a flexible processing technology for industrial robots of new energy vehicle parts, comprising: constructing an energy consumption-electromagnetic predictor using historical load curves and virtual simulation results, and outputting the power peak time domain sequence of the next tool path segment;

[0009] The servo switch strategy is switched in real time according to the power peak time domain sequence, and the distributed noise suppression unit is activated to perform shielding and harmonic suppression.

[0010] Multimodal signals are acquired synchronously within the noise suppression window. The thermal-noise coupling coefficient is generated by calculating the displacement potential of the coupling between cutting contact thermal energy and structural stiffness and the short-time Fourier entropy of the signal to determine the data availability.

[0011] The digital twin compensation model receives data confirmed by the thermal-noise coupling coefficient, outputs a small tool position correction and a pulse cooling command, and injects them into the next toolpath segment according to a unified timestamp, thereby realizing the linkage adjustment of heat source, power and cooling.

[0012] After processing, the residual dimensional error and energy consumption index are summarized and fed back to the energy consumption-electromagnetic predictor to adaptively update the weights, so as to improve the prediction accuracy of the next workpiece.

[0013] Furthermore, power consumption data of electric spindle, power consumption data of servo drive and power data generated by virtual simulation are collected and preprocessed. The preprocessing steps include removing outliers, filling missing values ​​by linear interpolation and normalization.

[0014] The peak power, peak duration, and peak interval features are extracted from historical load curve data, and the simulated peak power, simulated peak duration, and simulated peak interval features are extracted from virtual simulation data.

[0015] Furthermore, a prediction model is constructed using a long short-term memory network. The input features include the power peak, peak duration, and peak interval of historical load curve data, the simulated power peak, simulated peak duration, simulated peak interval of virtual simulation data, and the current toolpath parameters. The output is the power peak time-domain sequence of the next toolpath segment.

[0016] Furthermore, the model is trained using the training set, the mean square error between the predicted power sequence and the actual sequence is used as the loss function, the parameters are optimized using gradient descent, and the model performance is evaluated using five-fold cross-validation.

[0017] The current processing conditions and preprocessed features are input into the trained model to generate the power spike time-domain sequence of the next toolpath segment, and this sequence is used as the time reference.

[0018] Furthermore, the controller receives and parses the power spike time-domain sequence, extracts the start time, peak amplitude, and duration of each power spike, and decides whether to switch the servo drive to low-power mode within a preset time window before the power spike arrives, based on whether the peak amplitude exceeds a preset threshold.

[0019] Furthermore, a low-power mode is achieved by reducing the feed rate and adjusting the electric spindle speed, and the normal mode is restored after the power peak ends. At the same time, the distributed noise suppression unit is activated within the power peak time window.

[0020] The electromagnetic shielding device is activated and the active filter is enabled to attenuate electromagnetic noise and cancel harmonic currents; the signal-to-noise ratio of the sensor signal is monitored within the noise suppression window, and the gain of the active filter is adjusted when the signal-to-noise ratio is lower than the preset threshold until the signal-to-noise ratio reaches the preset threshold. Finally, the power peak time window and the corresponding noise suppression status are passed to the subsequent steps.

[0021] Furthermore, multimodal signals from vision sensors, force and torque sensors, and vibration sensors are acquired simultaneously; cutting contact heat energy is calculated based on the contact area between the tool and the workpiece, cutting speed, cutting force, and thermal conversion coefficient, and displacement potential is calculated in conjunction with structural stiffness;

[0022] Short-time Fourier transforms are performed on force, torque, and vibration signals, and short-time Fourier entropy is calculated to measure signal purity. Displacement potential and short-time Fourier entropy are input into a fuzzy decision tree model to generate thermal-noise coupling coefficients, and data is filtered based on these coefficients.

[0023] Furthermore, the digital twin compensation model receives available data confirmed by the thermal-noise coupling coefficient and extracts real-time status information during the machining process, including tool position and workpiece temperature.

[0024] Multiphysics coupling simulation is used to predict machining errors, and a small tool position correction is calculated to offset the prediction error. The tool position is then adjusted in real time by the industrial robot control system.

[0025] Based on the workpiece temperature and cutting contact heat energy, the cooling requirements are calculated, and pulse cooling commands are generated to control the coolant flow rate. The tool position micro-correction amount and pulse cooling commands are injected into the next toolpath planning according to a unified timestamp to achieve the linkage adjustment of heat source, power and cooling.

[0026] Furthermore, the actual dimensions of the workpiece are measured using high-precision measuring equipment and compared with the design dimensions to calculate the residual dimensional error. At the same time, the energy consumption of the electric spindle, servo drive, and cooling system during the processing are collected to calculate the total energy consumption.

[0027] The power peak time-domain sequence output by the energy consumption-electromagnetic predictor is compared with the actual acquired power peak time-domain sequence to calculate the power prediction error. The machining error predicted by the digital twin compensation model is compared with the actual measured residual size error to calculate the machining error prediction deviation.

[0028] Furthermore, a machine learning model based on neural networks is adopted to adaptively adjust the weight parameters by optimizing the objective function. The objective function is composed of the integral square of the power prediction error over the processing time period, the modulus square of the processing error prediction deviation, and the weighted combination of the total energy consumption.

[0029] The updated weight parameters are applied to the energy consumption-electromagnetic predictor to generate the power spike time-domain sequence of the next workpiece. Based on this, the control parameters of the servo switching strategy and the distributed noise suppression unit are optimized, as well as the tool position fine-tuning and pulse cooling commands of the digital twin compensation model.

[0030] (III) Beneficial Effects

[0031] This invention provides a flexible manufacturing process for industrial robots used in new energy vehicle parts, which has the following advantages:

[0032] An energy consumption-electromagnetic predictor is constructed using historical load curves and virtual simulation results. It outputs the power spike time-domain sequence of the next tool path segment, providing an accurate time reference for subsequent noise suppression scheduling. This allows the process to predict the timing of power spikes before processing, providing forward-looking guidance for noise suppression measures.

[0033] The controller switches the servo switch strategy and activates the distributed noise suppression unit in real time according to the power peak time domain sequence, achieving targeted shielding and harmonic suppression from the source. Compared with traditional noise suppression methods, this solution can accurately intervene in high-risk segments, avoid redundant filtering, significantly improve the noise suppression effect, reduce unnecessary energy consumption, and enhance the stability and reliability of the processing.

[0034] By calculating the displacement potential of the coupling between cutting contact thermal energy and structural stiffness, as well as the short-time Fourier entropy difference of the signal, a thermal-noise coupling coefficient is generated to determine the availability of the data. Compared with traditional single signal processing, the accuracy of data evaluation can be significantly improved through thermal-signal dual quantization. By combining thermal effects with signal purity, comprehensive data quality assurance is achieved.

[0035] By adjusting the tool position in real time and dynamically controlling the coolant flow, machining errors can be effectively compensated, workpiece temperature can be controlled, machining accuracy and quality can be significantly improved, the coordination and real-time performance of process parameters can be ensured, the adverse effects of thermal distortion and signal drift on machining accuracy in traditional methods can be overcome, and the coordinated control of heat source, power and cooling can be achieved, giving full play to the advantages of digital twin technology in flexible manufacturing.

[0036] By summarizing residual dimensional errors and energy consumption indicators, the weights are adaptively updated to the energy consumption-electromagnetic predictor, thereby improving the prediction accuracy of the next workpiece. This closed-loop feedback mechanism not only optimizes the predictor's performance but also gradually improves machining accuracy and energy efficiency by adjusting weight parameters, ensuring continuous improvement in the cycle time and quality of the flexible machining unit. Compared to traditional static processes, the dynamic optimization capability significantly enhances the process's adaptability and long-term stability. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the flexible manufacturing process of automotive parts industrial robots according to the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1 This invention provides a flexible manufacturing process for industrial robots used in the production of new energy vehicle parts, including:

[0040] Step 1: Collect and preprocess the power consumption data of the electric spindle, the power consumption data of the servo drive, and the power data generated by the virtual simulation. The preprocessing steps include removing outliers, filling in missing values ​​through linear interpolation, and performing normalization.

[0041] Extract peak power, peak duration, and peak interval features from historical load curve data, and simultaneously extract simulated peak power, simulated peak duration, and simulated peak interval features from virtual simulation data;

[0042] A prediction model is constructed using a long short-term memory network. The input features include the power peak, peak duration, and peak interval of historical load curve data, the simulated power peak, simulated peak duration, simulated peak interval of virtual simulation data, and the current toolpath parameters. The output is the power spike time-domain sequence of the next toolpath segment.

[0043] The model is trained using the training set, and the mean square error between the predicted power sequence and the actual sequence is used as the loss function. The parameters are optimized using gradient descent, and the model performance is evaluated using five-fold cross-validation. The current processing conditions and preprocessed features are input into the trained model to generate the power peak time-domain sequence of the next cutting path segment, and this sequence is used as a time reference to pass to subsequent steps.

[0044] Step one includes the following:

[0045] Step 101: Data Collection and Preprocessing

[0046] Power consumption data from the electric spindle and servo drive are collected, measured in watts, with a time resolution set to milliseconds to ensure detailed capture of power variations. Simultaneously, toolpath parameters, including cutting speed, feed rate, and toolpath, are input into virtual simulation software to simulate the machining process of the next toolpath segment, generating corresponding simulated power data. In the preprocessing stage, outlier removal is first performed on historical load curve data and simulated power data. Specifically, data points with power values ​​exceeding physically reasonable ranges, such as power spikes exceeding the equipment's rated value, are identified and removed. Then, missing values ​​are filled using linear interpolation by calculating the average of two adjacent data points. Finally, all data is normalized to the 0-1 range using linear transformations of the maximum and minimum values. Normalization aims to eliminate the influence of unit or magnitude differences in power data, ensuring consistency in subsequent processing.

[0047] Historical load curve data reflects the energy consumption patterns in actual processing, while simulation data provides the basis for predicting the next toolpath segment. Preprocessing ensures the integrity and consistency of the data, improves the accuracy of subsequent feature extraction and model prediction, effectively reduces noise interference, and provides high-quality input data for subsequent analysis.

[0048] Step 102, Feature Extraction

[0049] After completing the data preprocessing, key features are extracted from the historical load curve data, including the power peak, peak duration and peak interval for each toolpath segment.

[0050] Peak power refers to the maximum power value recorded within a toolpath machining cycle, measured in watts; peak duration refers to the duration for which the power value exceeds twice the average power value of that toolpath segment, measured in seconds; peak interval refers to the time difference between two adjacent power peaks, measured in seconds. Similarly, the corresponding simulated power peak, simulated peak duration, and simulated peak interval are extracted from the power data generated by virtual simulation, using the same extraction method as historical data. During feature extraction, the time window is set to the complete machining cycle of each toolpath segment, i.e., the entire time period from when the tool enters the workpiece to when it leaves the workpiece, ensuring that the features comprehensively reflect the power variation characteristics during machining. Peak power, peak duration, and peak interval characterize the key temporal behaviors of energy consumption and electromagnetic interference during machining, directly affecting the accuracy of subsequent predictions.

[0051] Step 103, Model Building

[0052] Based on feature extraction, an energy consumption-electromagnetic predictor is constructed, and a long short-term memory network is used as the core algorithm.

[0053] Long Short-Term Memory (LSTM) networks are artificial neural networks suitable for processing time-series data, capable of capturing the dependencies of power data over time. The model's input features include power peaks, peak durations, and peak intervals extracted from historical load curve data, as well as simulated power peaks, simulated peak durations, and simulated peak intervals extracted from virtual simulation data. It also includes current toolpath parameters, such as depth of cut and feed rate, which are input in numerical form.

[0054] The model output is the time-domain sequence of power spikes in the next cutting path segment, specifically the predicted power value sequence at discrete time points, including the occurrence time and amplitude of each power peak. The temporal characteristics of power data are complex, and traditional static models are difficult to predict accurately. However, long short-term memory networks can effectively fit the dynamic change pattern of power spikes by remembering historical information and current input. In this way, they can combine historical data and simulation data to generate high-precision prediction results, providing a reliable basis for subsequent process optimization.

[0055] Step 104: Model Training and Optimization

[0056] Model training uses historical load curve data and the corresponding actual power sequence of the next cutter path segment as the training set;

[0057] The input features are paired with the actual power sequence, and the system learns using a Long Short-Term Memory (LSTM) network. The training objective is to make the predicted power sequence as close as possible to the actual sequence. To this end, the loss function is defined as the average squared difference between the predicted and actual power sequences. This loss function is calculated by summing the squares of the differences between the predicted and actual values ​​at each time point and then taking the average.

[0058] By repeatedly adjusting the model parameters and gradually reducing the loss function using gradient descent, while simultaneously optimizing the learning rate and the number of hidden layer nodes (the learning rate determines the step size of parameter adjustments, and the number of hidden layer nodes affects the model's complexity), the loss function value was stabilized. To evaluate the model's performance, a five-fold cross-validation method was used: the training set was divided into five equal parts, with four parts used to train the model and the remaining part used to validate the prediction error. The final prediction error was required to be controlled within 5%. Through extensive training on large amounts of data and parameter tuning, the model learned deep-seated patterns in power variation and possessed generalization ability. The trained model exhibited high prediction accuracy and could adapt to the power peak prediction requirements under different processing conditions.

[0059] Step 105: Output of power spike time-domain sequence

[0060] After model training is complete, for the prediction of the next toolpath segment, the current machining conditions and preprocessed features are input into the Long Short-Term Memory (LSTM) network. The current machining conditions include toolpath parameters (such as cutting speed and feed rate) and workpiece material properties (such as hardness), and the preprocessed features include power peaks, peak durations, and peak intervals from historical data and simulation data.

[0061] The model calculates the output power spike time-domain sequence, specifically a series of power values ​​at discrete time points, including the predicted amplitude and occurrence time of each power peak. The output power spike time-domain sequence is represented with time on the horizontal axis and power value on the vertical axis, which can intuitively reflect the power change trend of the next tool path section. The power spikes of the next tool path section directly affect the timing and intensity of electromagnetic interference, and accurately predicting its time-domain distribution is the key to subsequent noise suppression. This sequence provides detailed power change information, which facilitates targeted intervention in subsequent process steps.

[0062] Step 106: Time Base Provision

[0063] The power spike time-domain sequence is used as a time reference and passed to the noise suppression scheduling stage in step two. The peak occurrence time and predicted peak amplitude in the power spike time-domain sequence clearly identify the occurrence time and intensity of the power spike, specifically the time point and power value corresponding to each peak. This information provides the noise suppression unit in step two with a precise timing reference, enabling noise suppression measures to intervene accurately during high-risk periods of power spike occurrence, such as by adjusting processing parameters or activating noise suppression equipment to reduce electromagnetic interference. The timing characteristics of power spikes are highly correlated with the occurrence of electromagnetic interference; understanding their distribution in advance can optimize the noise suppression strategy. Based on this time reference, step two can achieve efficient interference suppression, improving the stability and accuracy of the processing.

[0064] In practice, a complete power spike prediction process is formed, from data collection and preprocessing to time base provision. The combination of historical load curve data and virtual simulation results ensures the comprehensiveness and foresight of the predictions; the application of long short-term memory networks enables accurate modeling of time-series data; and the output of the power spike time-domain sequence provides clear guidance for subsequent process steps. In high-speed cutting, the impact of electromagnetic interference on machining accuracy mainly stems from power spikes. By predicting their time-domain distribution, interference problems can be effectively addressed, significantly improving the stability of flexible machining units, reducing machining errors, and providing technical support for high-precision manufacturing.

[0065] Step 2: The controller receives and parses the power spike time-domain sequence, extracts the start time, peak amplitude, and duration of each power spike, and decides whether to switch the servo drive to low-power mode within a preset time window before the power spike arrives, based on whether the peak amplitude exceeds a preset threshold. Low-power mode is achieved by reducing the feed speed and adjusting the electric spindle speed, and normal mode is restored after the power spike ends. At the same time, the distributed noise suppression unit is activated within the power spike time window, the electromagnetic shielding device is activated, and the active filter is enabled to attenuate electromagnetic noise and cancel harmonic current. The signal-to-noise ratio of the sensor signal is monitored within the noise suppression window. When the signal-to-noise ratio is lower than the preset threshold, the gain of the active filter is adjusted until the signal-to-noise ratio reaches the preset threshold. Finally, the power spike time window and the corresponding noise suppression status are passed to the subsequent steps.

[0066] Step two includes the following:

[0067] Step 201: Reception and parsing of power spike time-domain sequences

[0068] Step 2 receives the power spike time-domain sequence from Step 1. This sequence contains a series of power values ​​at discrete time points, specifically including the occurrence time of each power spike and the predicted peak power amplitude; the occurrence time is recorded in seconds, and the peak power amplitude is expressed in watts.

[0069] When parsing this sequence, the start time, peak amplitude, and duration of each power spike are extracted. The duration is determined by calculating the time difference between the power value decreasing to the average level at adjacent time points; specifically, it is the time interval from the peak time point until the power value recovers to the average power value during system operation. The timing information of the power spikes directly determines the activation timing of the servo drive and noise suppression unit. Accurate extraction of this information ensures the accuracy of subsequent processing measures, thereby providing precise timing basis for the activation of the servo switching strategy and noise suppression unit, thus improving the timeliness and pertinence of intervention measures.

[0070] The controller adjusts the working mode of the servo drive within a preset time window before each power spike arrives, based on the occurrence time and peak amplitude in the power spike time-domain sequence.

[0071] The preset time window is set to 0.05 seconds, a value determined based on the response speed of the machining unit, ensuring the servo drive has sufficient time to complete mode switching. When the predicted power peak amplitude exceeds a preset threshold, the controller switches the servo drive to low-power mode. The preset threshold is set to 80% of the rated power of the electric spindle; for example, if the rated power of the electric spindle is 2000 watts, the threshold is 1600 watts. In low-power mode, the servo drive feed rate is reduced to 70% of its normal value to reduce transient current surges; the electric spindle speed is reduced by 10% (revolutions per minute) to smooth the power variation curve. After the power peak duration ends, the controller restores the servo drive to normal mode to maintain machining efficiency. Mode switching is achieved through pulse width modulation signals, with the controller sending commands at millisecond-level time resolution. By adjusting the servo drive's operating mode, the amplitude of power peaks can be effectively reduced, thereby reducing electromagnetic interference; energy consumption and electromagnetic noise can be reduced, while maintaining the stability of the machining process.

[0072] Step 202: Start the distributed noise suppression unit

[0073] Within the time window of each power spike, a distributed noise suppression unit is activated. This unit includes an electromagnetic shielding device and an active filter. The time window is defined as starting from a preset time window before the power spike occurs and ending at the end of the power spike duration. The electromagnetic shielding device uses a local shielding cover made of a high-permeability material, such as permalloy, covering the area around the electric spindle and servo drive to attenuate broadband electromagnetic noise. The shielding cover is switched on and off by a relay with a response time of less than 10 milliseconds. The active filter detects harmonic currents in the system in real time and injects reverse compensation current to cancel the harmonics. The harmonic current is obtained through high-speed sampling at a sampling rate of 10 kHz, and the compensation current is generated by the inverter to make the total system current close to a sine wave, thereby reducing electromagnetic interference. The noise suppression unit shuts down after the power spike duration ends to reduce unnecessary energy consumption. Electromagnetic shielding and active filtering can suppress the propagation and generation of electromagnetic noise at the source, protecting the sensor from interference, improving the purity of the sensor signal, and ensuring the reliability of subsequent data acquisition.

[0074] Within each power peak time window, the signal-to-noise ratio (SNR) of the sensor signal is monitored to evaluate the noise suppression effect. The SNR is calculated as the logarithm of the ratio of effective signal power to noise power, specifically 10 multiplied by the base-10 logarithm of the signal-to-noise power ratio, in decibels (dB). If the SNR is lower than a preset threshold of 20 dB, the gain of the active filter is adjusted, with a gain range of 1 to 5 in 0.5 increments, until the SNR meets the requirements.

[0075] The verification results, including the signal-to-noise ratio (SNR) sequence within the time window, are passed to step three as a reference for multimodal signal acquisition. Real-time monitoring of the SNR allows for timely detection of insufficient noise suppression, which can be optimized by adjusting the active filter gain. This ensures the quality of the sensor signal and provides high-quality data support for subsequent thermal-noise coupling coefficient calculations.

[0076] The power peak time window and its corresponding noise suppression status, including servo mode switching records and signal-to-noise ratio sequences, are transmitted to step three. This data provides a noise suppression window for the multi-modal signal synchronous acquisition in step three, ensuring signal purity. Recording the noise suppression status provides a reference for subsequent data analysis and processing, helping to accurately assess signal usability. The noise suppression window and status information are crucial for data acquisition and analysis in step three, ensuring data validity and reliability. This enhances the overall coordination of the process flow and improves processing accuracy and quality.

[0077] In practice, step two achieves precise intervention of power spikes and effective suppression of electromagnetic interference. The instantaneous switching of the servo switch strategy and the activation of the distributed noise suppression unit work together to reduce the impact of power spikes and electromagnetic noise. Verification of the noise suppression effect and its connection with subsequent steps ensure high-quality sensor signals and the continuity of the process flow. Addressing the impact of electromagnetic interference on machining accuracy in high-speed cutting, a combination of prediction and intervention controls the generation and propagation of interference at its source. This significantly improves the stability and machining quality of the flexible machining unit, providing technical support for high-precision manufacturing.

[0078] Step 3: Simultaneously acquire multimodal signals from vision sensors, force and torque sensors, and vibration sensors. Set the sampling frequency to 1kHz and control the timestamp alignment error to within 1 millisecond. Calculate the cutting contact heat energy based on the contact area between the tool and the workpiece, cutting speed, cutting force, and thermal conversion coefficient, and calculate the displacement potential in conjunction with the structural stiffness.

[0079] Short-time Fourier transforms are performed on force, torque, and vibration signals with a window width of 0.1 seconds and 50% overlap. The short-time Fourier entropy is calculated to measure signal purity. The displacement potential and short-time Fourier entropy are input into a fuzzy decision tree model to generate thermal-noise coupling coefficients, which range from 0 to 1. Data is filtered based on the thermal-noise coupling coefficients. Data with a thermal-noise coupling coefficient greater than or equal to 0.8 is considered usable and is then passed to subsequent steps.

[0080] Step three includes the following:

[0081] Step 301: Synchronous acquisition of multimodal signals within the noise suppression window

[0082] Within the noise suppression window determined in step two, signals from the vision sensor, force / torque sensor, and vibration sensor are simultaneously acquired. The vision sensor is responsible for acquiring image data, the force / torque sensor is responsible for acquiring force and torque signals, and the vibration sensor is responsible for acquiring acceleration signals.

[0083] All signals are sampled with time as the independent variable, at a uniform frequency of 1000 times per second. Timestamps from all sensors are aligned, with errors controlled within 1 millisecond to ensure data synchronization. The reason for synchronous acquisition of multimodal signals within the noise suppression window is that electromagnetic interference is suppressed at this time, resulting in higher signal quality suitable for subsequent data analysis. This provides high-quality raw data for thermal-signal quantization calculations, thereby improving the reliability of the analysis results.

[0084] Cutting contact heat energy is calculated based on the contact area between the tool and the workpiece, the cutting speed, and the cutting force. The contact area is measured in square meters (m²), the cutting speed in meters per second (m / s), and the cutting force in Newtons (N). Cutting contact heat energy is obtained by multiplying the contact area, cutting speed, cutting force, and thermal conversion coefficient, where the thermal conversion coefficient is a dimensionless parameter representing the proportion of cutting energy converted into heat energy. Combined with structural stiffness (in N / m), displacement potential is calculated. Displacement potential represents the tendency of structural displacement caused by heat energy, measured in meters, and is specifically calculated by dividing the cutting contact heat energy by the structural stiffness. Displacement caused by heat energy affects machining accuracy; quantifying the displacement potential allows for an assessment of the degree of influence of thermal deformation. This provides a key indicator reflecting the thermal effect for the generation of the thermal-noise coupling coefficient.

[0085] Step 302, Thermal-Signal Quantization: Measurement of Signal Purity

[0086] Short-time Fourier transforms were performed on force, torque, and vibration signals with a transform window width of 0.1 seconds and a window overlap ratio of 50%, yielding the time-frequency spectra of each signal. The time-frequency spectrum reflects the energy distribution of the signal at different times and frequencies. The short-time Fourier entropy of each signal was calculated. The entropy value was obtained by weighted summation of the logarithm of the normalized power spectral density, where the normalized power spectral density is the square modulus of the time-frequency spectrum divided by the total energy. A smaller entropy value indicates that the signal is more concentrated at a specific frequency and has higher signal purity; a larger entropy value indicates that the signal frequency distribution is more dispersed and the signal is more susceptible to interference. Electromagnetic interference can cause signal spectrum distortion, and the entropy value can effectively quantify the impact of interference on signal quality. Therefore, it provides a quantitative basis for signal quality in the generation of thermal-noise coupling coefficients.

[0087] The displacement potential and signal entropy are input into the fuzzy decision tree model. The fuzzy decision tree generates thermal-noise coupling coefficients based on predefined fuzzy rules and membership functions. The coefficient values ​​are set to the range of 0 to 1, where 1 indicates that the data is completely usable and 0 indicates that the data is completely unusable.

[0088] Examples of fuzzy rules include: a higher thermal-noise coupling coefficient when both displacement potential and signal entropy are small; and a lower thermal-noise coupling coefficient when either displacement potential or signal entropy is large. Through multi-layer fuzzy judgment, the thermal-noise coupling coefficient at each time point is output. The thermal-noise coupling coefficient can be used to comprehensively consider the combined impact of thermal deformation and electromagnetic interference on data quality, while fuzzy decision trees can effectively handle uncertainties and nonlinear relationships; providing an accurate comprehensive standard for assessing data availability.

[0089] Multimodal signals are screened based on the thermal-noise coupling coefficient: if the thermal-noise coupling coefficient is greater than or equal to 0.8, the data is considered usable; if the thermal-noise coupling coefficient is less than 0.8, the data is considered unusable and needs to be removed or corrected in subsequent steps.

[0090] The timestamps of available data and their corresponding signal values ​​are passed to step four as input to the digital twin compensation model. Only by ensuring that the data input to the digital twin compensation model is of high quality can the reduction in compensation effect due to data errors be avoided, thus improving the accuracy and reliability of the digital twin compensation model.

[0091] In practice, by synchronously acquiring multimodal signals, calculating displacement potential and signal entropy, and generating thermal-noise coupling coefficients, accurate assessment of data availability is achieved. Displacement potential reflects the impact of thermal deformation on machining accuracy, signal entropy quantifies the damage to signal quality caused by electromagnetic interference, and the fuzzy decision tree integrates the coupling effect of both, providing high-quality input data for the digital twin compensation model. Thermal deformation and electromagnetic interference are the main factors affecting machining accuracy. Quantifying their coupling effect can effectively improve data reliability and machining accuracy; it also enhances the stability of the flexible machining unit, improves machining quality, and provides support for high-precision manufacturing.

[0092] Step 4: The digital twin compensation model receives available data confirmed by the thermal-noise coupling coefficient, extracts real-time status information during the machining process, including tool position and workpiece temperature; uses multi-physics coupling simulation to predict machining errors, calculates the tool position micro-correction amount to offset the prediction error, and adjusts the tool position in real time through the industrial robot control system.

[0093] Based on the workpiece temperature and cutting contact heat energy, the cooling requirements are calculated, and pulse cooling commands are generated to control the coolant flow rate. The tool position micro-correction amount and pulse cooling commands are injected into the next toolpath planning according to a unified timestamp to realize the linkage adjustment of heat source, power and cooling. Real-time adjustment data is recorded and transmitted to subsequent steps to support process optimization.

[0094] Step four includes the following:

[0095] Step 401: Receiving and parsing available data

[0096] The available data, confirmed by the thermal-noise coupling coefficient, is received from step three. This data includes timestamps and corresponding multimodal signal values, such as visual sensor images, force signals, torque signals, and vibration signals. This data is parsed to extract real-time status information during the machining process, including tool position and workpiece temperature. Tool position is represented by a three-dimensional coordinate vector in millimeters; workpiece temperature is expressed in degrees Celsius. This data serves as the input parameters for the digital twin compensation model. The reception and parsing of available data ensures high-quality input data to the digital twin compensation model, preventing a decrease in compensation effectiveness due to data errors; it also provides accurate real-time status information for subsequent compensation calculations, improving the reliability of machining accuracy.

[0097] The digital twin compensation model uses multiphysics coupling simulation to integrate thermodynamics, structural mechanics, and electromagnetics to simulate the impact of thermal deformation and electromagnetic interference on machining accuracy during the machining process.

[0098] The model input parameters include real-time tool position, workpiece temperature, cutting force (in Newtons), vibration acceleration (in meters per second squared), and sensor layout and workpiece material properties (such as thermal conductivity and specific heat capacity) in the virtual environment. The model output is the predicted machining error, represented as a three-dimensional error vector in micrometers, reflecting dimensional deviations caused by thermal deformation and electromagnetic interference. The rationale for constructing the digital twin compensation model is that it is necessary to comprehensively consider the influence of multiple physical fields on machining accuracy. Traditional single-field models cannot accurately predict errors; through multi-physics coupled simulation, machining errors can be predicted more accurately, providing a reliable basis for subsequent compensation.

[0099] Step 402: Calculation of tool position micro-correction amount

[0100] Based on the machining error output by the digital twin compensation model, a small tool position correction is calculated to offset the prediction error. The small tool position correction is a displacement vector in three-dimensional space, with units of micrometers, and is calculated by inverting the predicted machining error.

[0101] The tool position adjustment is performed in real time by the industrial robot control system to ensure machining accuracy. The rationale for calculating the tool position adjustment is that adjusting the tool position can directly compensate for machining errors and improve machining accuracy. This allows for real-time dynamic adjustment of the toolpath, reducing machining errors and thus improving the dimensional consistency of the machined parts.

[0102] The cooling requirement is calculated based on the workpiece temperature and cutting contact heat energy (in joules, calculated from cutting force and cutting speed). The cooling requirement is the non-negative portion of the difference between the cutting contact heat energy and the workpiece temperature control threshold (in joules, determined by the workpiece material's heat capacity and allowable temperature rise). A pulse cooling command is generated to control the coolant flow rate (in liters per second): the coolant flow rate is at its maximum when the cooling requirement is greater than 0; otherwise, it is 0. By dynamically controlling the coolant flow rate, the workpiece temperature can be effectively controlled, reducing the impact of thermal deformation on machining accuracy. This enables precise temperature control, reduces errors caused by thermal deformation, and improves machining quality.

[0103] By injecting minute toolpath adjustments and pulse cooling commands into the next toolpath segment with a unified timestamp, the heat source, power, and cooling are linked and adjusted. Adjusting the toolpath reduces cutting forces and cut contact heat, while pulse cooling controls workpiece temperature to maintain thermal dynamic balance. The rationale for unified timestamp injection and linked adjustment is to ensure the coordination of tool position adjustment and cooling measures, avoiding compensation failures due to time asynchrony. This enables real-time collaborative optimization of process parameters, improving machining accuracy and stability.

[0104] The real-time adjustment records during the machining process, including tool position micro-correction amounts, pulse cooling commands, and predicted machining errors, are transmitted to step five as the basis for calculating residual dimensional errors and energy consumption indicators. These data support the adaptive updating of the energy consumption-electromagnetic predictor in step five, improving the machining accuracy of subsequent workpieces. The real-time adjustment records provide feedback information for subsequent process optimization, enabling continuous process improvement. This can enhance the overall coordination of the process flow and improve machining quality and efficiency.

[0105] In use, the digital twin compensation model receives available data and outputs minute tool position corrections and pulse cooling commands to achieve coordinated adjustment of heat source, power, and cooling. The digital twin compensation model, based on multiphysics coupling simulation, predicts machining errors and compensates for them through tool position fine-tuning; the pulse cooling command controls the workpiece temperature according to real-time cooling requirements, reducing the impact of thermal deformation. The injection of a unified timestamp ensures the coordination and real-time nature of the adjustment measures, providing reliable data support for step five.

[0106] Taking into account the impact of thermal deformation and electromagnetic interference on machining accuracy, real-time compensation and control are used to improve machining accuracy. This significantly enhances the machining quality and stability of flexible machining units, providing technical support for high-precision manufacturing.

[0107] Step 5: After processing, use high-precision measuring equipment to measure the actual size of the workpiece and compare it with the design size to calculate the residual dimensional error. At the same time, collect the energy consumption of the electric spindle, servo drive, and cooling system during the processing to calculate the total energy consumption. Compare the power peak time-domain sequence output by the energy consumption-electromagnetic predictor in Step 1 with the actual collected power peak time-domain sequence to calculate the power prediction error. Compare the processing error predicted by the digital twin compensation model in Step 4 with the actual measured residual dimensional error to calculate the processing error prediction deviation.

[0108] A neural network-based machine learning model is employed to adaptively adjust weight parameters by optimizing the objective function. The objective function consists of a weighted combination of the squared integral of the power prediction error over the processing time period, the squared modulus of the processing error prediction deviation, and the total energy consumption. The updated weight parameters are applied to the energy consumption-electromagnetic predictor to generate the power peak time-domain sequence for the next workpiece. Based on this, the servo switching strategy and control parameters of the distributed noise suppression unit in step two are optimized, as well as the tool position fine-tuning and pulse cooling commands of the digital twin compensation model in step four. Feedback data, including residual dimensional error and energy consumption data, are provided through the processing results in step four, forming a closed-loop optimization mechanism.

[0109] Step five includes the following:

[0110] Step 501: Summary of Residual Dimensional Errors and Energy Consumption Indicators

[0111] After machining, high-precision measuring equipment is used to measure the actual dimensions of the workpiece and compare them with the design dimensions to calculate the residual dimensional error. The residual dimensional error refers to the deviation between the actual dimensions of the workpiece and the design dimensions in the X, Y, and Z directions, measured in micrometers.

[0112] Simultaneously, energy consumption indicators are collected during the machining process, including electric spindle energy consumption, servo drive energy consumption, and cooling system energy consumption, in kilowatt-hours. Total energy consumption is calculated by adding the energy consumption of the electric spindle, servo drive, and cooling system. Residual dimensional error and total energy consumption are used as feedback data for subsequent analysis and optimization. Residual dimensional error reflects the actual effect of machining accuracy, while energy consumption indicators reflect the energy efficiency of the machining process. By summarizing these data, accurate feedback information can be provided to the optimization predictor. This ensures that subsequent analysis provides accurate actual data, guaranteeing the targeted and reliable nature of optimization measures.

[0113] Step 502: Comparative Analysis of Feedback Data and Prediction Results

[0114] The power peak time-domain sequence output by the energy consumption-electromagnetic predictor in step one is compared with the actual power peak time-domain sequence acquired during processing to calculate the power prediction error. The power prediction error is calculated by subtracting the predicted power value from the actual power value, in kilowatts. Furthermore, the processing error predicted by the digital twin compensation model in step four is compared with the actual measured residual dimensional error to calculate the processing error prediction deviation. The processing error prediction deviation is calculated by subtracting the predicted processing error from the residual dimensional error, in micrometers. The reason for conducting comparative analysis of feedback data and prediction results is that by comparing actual data with predicted data, the accuracy of the energy consumption-electromagnetic predictor and the digital twin compensation model can be evaluated, errors existing in the prediction process can be identified, and specific improvement directions can be provided for optimizing the predictor.

[0115] The energy consumption-electromagnetic predictor employs a neural network-based machine learning model, whose weight parameters are adaptively adjusted by optimizing the objective function. The objective function consists of the following parts:

[0116] The weighted combination of the integral square of the power prediction error over the processing time period, the modulus square of the processing error prediction deviation, and the total energy consumption is used. The integral square represents the cumulative effect of the power prediction error over the entire processing time period, the modulus square represents the magnitude of the processing error prediction deviation, and the adjustment coefficients in the weighted combination are used to control the relative importance of power prediction accuracy, processing accuracy, and energy consumption optimization. These adjustment coefficients need to be pre-calibrated according to specific process requirements. The weight update method uses gradient descent, with the learning rate set between 0.001 and 0.01. The gradient is calculated using numerical or analytical methods. By optimizing the objective function and adjusting the weight parameters, the prediction accuracy of the predictor and the energy efficiency of the processing process can be improved. This enables continuous optimization of the predictor, gradually improving the accuracy and energy efficiency of the processing technology.

[0117] The updated weighting parameters are applied to the energy consumption-electromagnetic predictor to generate the power spike time-domain sequence for the next workpiece. Based on more accurate power prediction data, the servo switching strategy and control parameters of the distributed noise suppression unit in step two are further optimized, as are the tool position fine-tuning and pulse cooling commands of the digital twin compensation model in step four. By improving prediction accuracy through updated weighting parameters, more reliable process parameters can be provided for subsequent machining processes, enabling continuous improvement of the machining process and thus improving machining accuracy and energy efficiency.

[0118] The feedback data in step five originates from the processing results in step four, including the actual measured residual dimensional errors and the energy consumption data recorded in real time during processing. The energy consumption-electromagnetic predictor built in step one provides initial prediction values, while in step five, a closed-loop optimization mechanism is formed through error comparison and weight updates. By comparing and optimizing the actual processing data with the predicted data, a closed-loop feedback mechanism for the processing technology is established to ensure continuous improvement of the process. This can enhance the coordination of the entire process flow, thereby improving processing quality and efficiency.

[0119] First, residual dimensional errors and energy consumption indicators are summarized and then compared with the prediction results to calculate power prediction error and machining error prediction deviation. Based on these error data, the weight parameters of the energy consumption-electromagnetic predictor are adaptively updated by optimizing the objective function, thereby improving the prediction accuracy of the next workpiece. This process forms a closed-loop feedback mechanism, ensuring that the flexible machining unit for new energy vehicle parts gradually optimizes machining accuracy and energy efficiency while suppressing thermal deformation and electromagnetic interference. Through feedback from actual data and optimization of the predictor, continuous improvement of the machining process can be achieved, enhancing machining accuracy and energy efficiency while strengthening the stability and reliability of the process.

[0120] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0121] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A flexible manufacturing process for industrial robots of new energy vehicle parts, characterized in that: include, An energy consumption-electromagnetic predictor is constructed using historical load curves and virtual simulation results, and outputs the power peak time-domain sequence of the next cutter path segment; The servo switch strategy is switched in real time according to the power peak time domain sequence, and the distributed noise suppression unit is activated to perform shielding and harmonic suppression. Multimodal signals are acquired synchronously within the noise suppression window. The thermal-noise coupling coefficient is generated by calculating the displacement potential of the coupling between cutting contact thermal energy and structural stiffness and the short-time Fourier entropy of the signal to determine the data availability. The digital twin compensation model receives data confirmed by the thermal-noise coupling coefficient, outputs a small tool position correction and a pulse cooling command, and injects them into the next toolpath segment according to a unified timestamp, thereby realizing the linkage adjustment of heat source, power and cooling. After processing, the residual dimensional error and energy consumption index are summarized and fed back to the energy consumption-electromagnetic predictor to adaptively update the weights, so as to improve the prediction accuracy of the next workpiece.

2. The flexible manufacturing process for industrial robots of parts according to claim 1, characterized in that: Collect and preprocess power consumption data of electric spindle, power consumption data of servo drive and power data generated by virtual simulation. The preprocessing steps include removing outliers, filling missing values ​​by linear interpolation and normalization. The peak power, peak duration, and peak interval features are extracted from historical load curve data, and the simulated peak power, simulated peak duration, and simulated peak interval features are extracted from virtual simulation data.

3. The flexible manufacturing process for industrial robots of parts according to claim 2, characterized in that: A prediction model is constructed using a long short-term memory network. The input features include the power peak, peak duration, and peak interval of historical load curve data, the simulated power peak, simulated peak duration, simulated peak interval of virtual simulation data, and the current toolpath parameters. The output is the power peak time-domain sequence of the next toolpath segment.

4. The flexible manufacturing process for industrial robots of parts according to claim 3, characterized in that: The model is trained using the training set, the mean square error between the predicted power sequence and the actual sequence is used as the loss function, the parameters are optimized using gradient descent, and the model performance is evaluated using five-fold cross-validation. The current processing conditions and preprocessed features are input into the trained model to generate the power spike time-domain sequence of the next toolpath segment, and this sequence is used as the time reference.

5. The flexible manufacturing process for industrial robots of parts according to claim 4, characterized in that: The controller receives and parses the power spike time-domain sequence, extracts the start time, peak amplitude and duration of each power spike, and decides whether to switch the servo drive to low power mode within a preset time window before the power spike arrives, based on whether the peak amplitude exceeds a preset threshold.

6. The flexible manufacturing process for industrial robots of parts according to claim 5, characterized in that: A low-power mode is achieved by reducing the feed rate and adjusting the electric spindle speed, and the normal mode is restored after the power peak ends. At the same time, the distributed noise suppression unit is activated within the power peak time window. The electromagnetic shielding device is activated and the active filter is enabled to attenuate electromagnetic noise and cancel harmonic currents; the signal-to-noise ratio of the sensor signal is monitored within the noise suppression window, and the gain of the active filter is adjusted when the signal-to-noise ratio is lower than the preset threshold until the signal-to-noise ratio reaches the preset threshold. Finally, the power peak time window and the corresponding noise suppression status are passed to the subsequent steps.

7. The flexible manufacturing process for industrial robots of parts according to claim 6, characterized in that: Simultaneously acquire multimodal signals from vision sensors, force and torque sensors, and vibration sensors; calculate cutting contact heat energy based on the contact area between the tool and the workpiece, cutting speed, cutting force, and thermal conversion coefficient, and calculate displacement potential in combination with structural stiffness; Short-time Fourier transforms are performed on force, torque, and vibration signals, and short-time Fourier entropy is calculated to measure signal purity. Displacement potential and short-time Fourier entropy are input into a fuzzy decision tree model to generate thermal-noise coupling coefficients, and data is filtered based on these coefficients.

8. The flexible manufacturing process for industrial robots of parts according to claim 7, characterized in that: The digital twin compensation model receives available data confirmed by the thermal-noise coupling coefficient and extracts real-time status information during the machining process, including tool position and workpiece temperature. Multiphysics coupling simulation is used to predict machining errors, and a small tool position correction is calculated to offset the prediction error. The tool position is then adjusted in real time by the industrial robot control system. Based on the workpiece temperature and cutting contact heat energy, the cooling requirements are calculated, and pulse cooling commands are generated to control the coolant flow rate. The tool position micro-correction amount and pulse cooling commands are injected into the next toolpath planning according to a unified timestamp to achieve the linkage adjustment of heat source, power and cooling.

9. The flexible manufacturing process for industrial robots of parts according to claim 8, characterized in that: The actual dimensions of the workpiece are measured using high-precision measuring equipment and compared with the design dimensions to calculate the residual dimensional error. At the same time, the energy consumption of the electric spindle, servo drive, and cooling system during the processing are collected to calculate the total energy consumption. The power peak time-domain sequence output by the energy consumption-electromagnetic predictor is compared with the actual acquired power peak time-domain sequence to calculate the power prediction error. The machining error predicted by the digital twin compensation model is compared with the actual measured residual size error to calculate the machining error prediction deviation.

10. The flexible manufacturing process for industrial robots of parts according to claim 9, characterized in that: A machine learning model based on neural networks is adopted, and the weight parameters are adaptively adjusted by optimizing the objective function. The objective function is composed of the integral square of the power prediction error over the processing time period, the modulus square of the processing error prediction deviation, and the weighted combination of the total energy consumption. The updated weight parameters are applied to the energy consumption-electromagnetic predictor to generate the power spike time-domain sequence of the next workpiece. Based on this, the control parameters of the servo switching strategy and the distributed noise suppression unit are optimized, as well as the tool position fine-tuning and pulse cooling commands of the digital twin compensation model.

Images