Self-adaptive control method for complex curved surface grinding

Abstract

The invention discloses a self-adaptive control method for complex curved surface grinding. The self-adaptive control method comprises the steps that acoustic emission signals in the grinding process are collected and preprocessed; extracting time domain features and frequency domain features from the preprocessed acoustic emission signals; the time domain features and the frequency domain features are combined to conduct threshold value judgment so as to recognize the current grinding state, and if the current grinding state is an abnormal state, a self-adaptive fuzzy PID control algorithm is used for adjusting current grinding parameters; or combining the extracted time domain features and frequency domain features with the force / torque signal and the vibration signal to construct a comprehensive feature vector, inputting the comprehensive feature vector into a machine learning model to identify the current grinding state, and if the current grinding state is an abnormal state, adjusting the current grinding parameters by using a self-adaptive fuzzy PID control algorithm; and the process is repeated until the current grinding state returns to the normal state.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and precision machining technology of robots, specifically relating to an acoustic emission monitoring and parameter adaptive closed-loop control method for complex curved surface robot grinding / belt grinding. Background Technology

[0002] With the rapid development of industries such as aerospace, high-end equipment, new energy vehicles and medical devices, the application of complex curved surface parts is becoming increasingly widespread, such as aero-engine blades, turbine disk tenons, titanium alloy shells, artificial joints, precision mold cavities, etc. These parts usually have the following characteristics: large curvature changes and complex spatial postures; obvious material anisotropy (titanium alloys, nickel-based high-temperature alloys, composite materials, etc.); and extremely high requirements for surface roughness and integrity.

[0003] Traditional manual grinding and polishing methods rely heavily on worker experience, resulting in high labor intensity and poor stability. In contrast, industrial robot belt grinding, with its flexibility and programmability, has gradually become an important means of precision machining of complex curved surfaces. However, existing robotic grinding systems generally generate trajectories through offline programming and set fixed process parameters (belt linear speed, feed rate, normal force, etc.), and make virtually no dynamic adjustments during production. Therefore, to improve the controllability of processing, researchers have begun to introduce online monitoring technologies such as force sensors to monitor normal force, accelerometers to monitor vibration, or infrared / thermal imagers to monitor temperature, and combine them with traditional machine learning algorithms to achieve classification and identification of wear states or working conditions.

[0004] Currently, many solutions extract fewer parameters, such as only extracting a small number of time-domain features like RMS and counts, while ignoring frequency-domain and time-frequency information. However, under highly non-stationary working conditions like grinding complex curved surfaces, a single feature is difficult to accurately reflect phenomena such as grinding mode transitions, belt passivation, and local transient impacts, making it difficult to accurately characterize the grinding state. Furthermore, existing methods for identifying grinding states are mostly used for post-processing analysis or alarm prompts, failing to form a closed loop of "monitoring – identification – parameter adjustment – ​​re-monitoring". Summary of the Invention

[0005] This invention provides an adaptive control method for grinding complex curved surfaces, which solves the technical problems of existing detection methods, such as the need for large amounts of data acquisition, low wear detection speed, poor real-time performance, and excessive size and weight of the overall detection device, resulting in poor portability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An adaptive control method for grinding complex curved surfaces includes the following steps: Step 1: Collect and preprocess the acoustic emission signals during the grinding process; Step 2: Extract time-domain and frequency-domain features from the preprocessed acoustic emission signal; Step 3: Combine time-domain features and frequency-domain features to make a threshold judgment to identify the current grinding state. If the current grinding state is abnormal, use an adaptive fuzzy PID control algorithm to adjust the current grinding parameters. Step 4: Repeat steps 1 to 3 until the current grinding state returns to normal.

[0007] Furthermore, the acoustic emission signal within a certain period after preprocessing is denoted as... The time-domain characteristic energy density is calculated using the following formula. , Where N represents the total number of acoustic emission signals sampled over a period of time, and t represents the sampling time; Then use the formula Acoustic emission signal Perform a Fast Fourier Transform and calculate the centroid of the frequency characteristic spectrum using the following formula. High-frequency energy ratio ; .

[0008] Furthermore, the grinding state is completed using the following formula. Identification, in, For a pre-defined fixed threshold, m = 1, 2, 3, 4.

[0009] Furthermore, the following formula is used to apply the fixed threshold. Make adjustments online. in, This is the online threshold after the i-th fixed threshold is corrected at time t; For offline calibration of fixed thresholds, , respectively corresponding to fixed thresholds ; The average curvature corresponding to the current contact point; is the curvature sensitivity coefficient corresponding to the i-th fixed threshold.

[0010] Furthermore, a bandpass filter is used to filter the acoustic emission signal. Filtering is performed, and then wavelet thresholding is used to refine the filtered acoustic emission signal. Noise reduction is performed to suppress high-frequency noise, and finally, normalization is applied.

[0011] An adaptive control method for grinding complex curved surfaces includes the following steps: Step 1: Collect acoustic emission signals, force / torque signals, and vibration signals during the grinding process, and preprocess the acoustic emission signals; Step 2: Extract time-domain and frequency-domain features from the preprocessed acoustic emission signal; Step 3: Combine the extracted time-domain features, frequency-domain features, force / torque signals, and vibration signals to construct a comprehensive feature vector. Input this vector into the machine learning model to identify the current grinding state. If the current grinding state is abnormal, use an adaptive fuzzy PID control algorithm to adjust the current grinding parameters. Step 4: Repeat steps 1 to 3 until the current grinding state returns to normal.

[0012] Furthermore, the acoustic emission signal within a certain period after preprocessing is denoted as... The time-domain characteristic energy density is calculated using the following formula. Single event peak value P, Where N represents the total number of acoustic emission signals sampled over a period of time, and t represents the sampling time; Then use the formula Acoustic emission signal Perform a Fast Fourier Transform and calculate the centroid of the frequency characteristic spectrum using the following formula. High-frequency energy ratio and the transient spectrum drift rate LTSD of the main frequency, This is used to construct a comprehensive feature vector. ,in, The force information represents the real-time normal force applied to a complex curved surface; It represents the vibration acceleration on a complex curved surface during the grinding process.

[0013] Furthermore, the machine learning model employs a Support Vector Machine (SVM), whose decision function... The formula is as follows: in, Indicates the first The Lagrange multipliers corresponding to each support vector This represents the radial basis function (RBF) kernel function based on adaptive kernel width. Identifier The class labels corresponding to the support vectors are given by b, which represents the bias term of the classification hyperplane, and N represents the total number of support vectors. Indicates the first The comprehensive feature vector corresponding to each support vector is derived from the offline training sample set. This represents the comprehensive feature vector corresponding to the current time t; Based on the comprehensive feature vector Local divergence Calculate the radial basis function (RBF) kernel parameters. , in, Representing the eigenvector The local divergence within the current window is obtained based on variance-based estimation. Indicates the regulating factor. This represents the initial value.

[0014] Furthermore, the adaptive fuzzy PID control algorithm first utilizes fuzzy algorithms to address grinding energy error. and the rate of change of grinding energy error As input, the proportional gain K of the PID controller is adaptively adjusted. p Integral gain K i and differential gain K d Then adjust the grinding parameters. , respectively, represent normal force, feed rate, and belt linear velocity.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Improve the machining quality and consistency of complex curved surfaces By using multi-scale AE features and an improved SVM, the method accurately identifies states such as under-wear, overload, and passivation. It also uses acoustic emission energy as the core feedback quantity to adjust parameters in real time, thereby keeping the grinding energy level at different positions of complex curved surfaces relatively stable. Experimental results show that in belt grinding of complex curved surface workpieces such as aero-engine blades, the method of this invention can control the surface roughness Ra fluctuation within ±5%, which is significantly better than the traditional fixed parameter method.

[0016] 2. Effectively suppresses surface burns and microcrack defects. By identifying overload conditions and early signs of burns in the early stages, and combining energy closed-loop control with coordinated parameter adjustment, the peak temperature rise in the grinding zone can be significantly reduced. Under the same material removal conditions, there are no obvious burn marks or significant microcracks on the workpiece surface, and fatigue life and surface integrity are significantly improved.

[0017] 3. Improve the accuracy and robustness of state recognition The combination of multimodal and multiscale features with adaptive kernel width and class-weighted SVM enables the working condition classification accuracy to reach 90%–95% or more under complex curvature, material property changes and belt abrasion conditions, and the F1 value of abnormal states is significantly higher than that of traditional SVM. 4. Achieve intelligent and adaptive grinding processes This invention organically integrates AE monitoring, status recognition and robot parameter control to construct an intelligent grinding control framework of "self-sensing-self-judgment-self-adjustment", which reduces reliance on operator experience and reduces on-site debugging workload. This method has good versatility and can be extended to robotic belt grinding and polishing scenarios for complex curved surface workpieces of different materials and types. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram illustrating the process of using a machine learning model to identify grinding methods in this invention. Figure 3 This is a comparative diagram of traditional SVM and the ISVM of this invention; Figure 4 This is a flowchart illustrating the adaptive fuzzy PID control algorithm of the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of the present invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the adaptive control method for grinding complex curved surfaces of the present invention. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0020] Specifically as follows: like Figure 1 As shown, this invention provides an adaptive control method for grinding complex curved surfaces. By real-time acquisition of acoustic emission signals, force signals applied during the grinding process, and vibration signals of the grinding surface, the method analyzes their characteristic parameters and establishes a correspondence with the grinding state to identify the grinding state. Then, an adaptive fuzzy PID control algorithm is used to automatically adjust key parameters such as grinding force, belt speed, and feed rate, thereby improving surface grinding quality, processing stability, and processing efficiency. Specifically: Step 1: Signal Acquisition and Synchronization 1. Multi-source signals During the grinding of complex curved surfaces using a robot-driven abrasive belt, multiple signals such as acoustic emission signals, force / torque signals, and vibration signals are stably acquired and aligned on the same time reference. The final output is: Acoustic emission signal: The original waveform of the AE channel with an effective bandwidth of 100–800 kHz and a sampling rate of ≥2 MHz; Force / torque signal: Time-series data of force / torque channel ≥ 1kHz (including at least the normal force applied to the complex curved surface) ); Vibration signal: Time-series data of vibration channel ≥10kHz; The time synchronization error of each channel is ≤ 0.1 ms, and a unified timestamp and metadata are written.

[0021] 2. Equipment and installation (location, orientation, and distance should be explained clearly at once) (1) Acoustic emission channel (AE) Acoustic emission signals are acquired using a broadband piezoelectric acoustic emission sensor with a bandwidth range of [missing information]. The sampling frequency is set to The sensor can be tightly contacted with the workpiece surface through a clamp, and the installation position should be no more than 15 mm away from the contact point of the sanding belt to ensure the rigidity of the acoustic path and the fidelity of the signal.

[0022] (2) Force / Torque Channel (FT) Installed between the robot's end effector and the grinding head fixture, sampling frequency The main monitoring function is the normal force. and tangential force .

[0023] (3) Vibration Channel (ACC) A triaxial accelerometer is mounted on the grinding head housing, with a sampling frequency of... It is used to monitor flutter and structural response.

[0024] All data acquisition systems for the above channels should have synchronous sampling capabilities and be equipped with a shared clock and an external trigger signal interface to ensure consistent time reference.

[0025] Step 2: Signal Preprocessing For the original acoustic emission signal Filtering, denoising, and normalization are performed to eliminate mechanical background noise, sensor drift, and random interference. 1. Bandpass Filtering The radio frequency band of the sound generator is set to , The filtered output signal is: (1) in, This represents the Fourier transform. This step preserves the grinding-related AE components and initially suppresses background noise.

[0026] 2. Adaptive denoising uses wavelet thresholding to suppress high-frequency noise. For the first... Layer wavelet coefficients : (2) in, This represents the standard deviation of the noise level in this layer. This represents the number of sample points.

[0027] This process, while preserving as much AE burst and spike information as possible, suppresses high-frequency random noise, and the reconstructed signal is denoted as... .

[0028] 3. Normalization and baseline correction To standardize the amplitude scale and eliminate slow drift, zero-mean normalization is used. The time window is used as the unit (window length is the same as in step three). (Consistent), calculate the moving average baseline within the window. Then normalize: (3) in, The moving average baseline is used. Signal energy retention rate after processing. This approach ensures comparability of data from different operating conditions and batches at the same amplitude scale while effectively suppressing sensor drift and low-frequency baseline drift. The normalized sound signal... It serves as the input for subsequent feature extraction and grinding condition recognition.

[0029] Step 3: Feature Extraction Normalized sound signal In length Time-domain, frequency-domain, and time-frequency features are extracted within a sliding window. A typical sampling frequency is... .

[0030] 1. Temporal characteristics Used to describe grinding energy and stability, it is one of the sensitive indicators in the grinding process. The following parameters are extracted: : (4) Single event peak value: (5) 2. Frequency Domain Characteristics Fast Fourier Transform Extract the following parameters: Musical focus: (6) Clock speed: (7) High-frequency energy ratio: (8) To characterize the transient variation trend of the acoustic signal AE during complex surface grinding, this invention adds the Local Transient Spectral Drift (LTSD) to the frequency domain characteristics to reflect the rate of change of the dominant frequency over a short period of time. (9) in, This reflects changes in grinding method (downward shift when grinding type changes from cutting to friction). It can characterize the decreasing trend of abrasive belt sharpness; Significant drift occurs under passivation or flutter conditions.

[0031] 3. Feature fusion and real-time output To achieve multi-source collaboration, a comprehensive feature vector is constructed: (10) in, This represents the real-time normal force applied to a complex curved surface; It represents vibration acceleration; LSTD represents the transient drift rate of the dominant frequency.

[0032] Step 4: Grinding Condition Identification 1. Threshold discrimination method (fast identification mode) In the grinding process of complex curved surfaces, the acoustic emission signals under different working conditions show significant differences, as shown below. Therefore, this feature vector can be used as input, and the grinding state category can be output through the threshold discrimination method. This method has the characteristics of low computational load and fast response, and can achieve a millisecond-level response speed, making it suitable for online real-time identification.

[0033] (1) Under-wear state (S1) When a low level of acoustic emission energy is detected, it typically manifests as The value is relatively low, and the proportion of high-frequency energy is also relatively low. (Lower). Physically, this means insufficient effective contact between the abrasive belt and the workpiece, resulting in a low material removal rate and a risk of "insufficient contact / insufficient removal".

[0034] (2) Normal state (S2) When the acoustic emission energy level is in a stable range, it typically exhibits the following characteristics: Stable, with a relatively concentrated distribution of principal energies in the spectrum and a spectral centroid. Within the normal range. This means that the grinding energy input is uniform, the processing is stable, and the surface quality is consistent.

[0035] (3) Overload state (S3) When the acoustic emission energy increases significantly and is accompanied by spectral spread, it typically manifests as High and Increased frequency (higher proportion of energy) results in a wider spectral distribution. Physically, this means that excessive instantaneous load can easily lead to belt blockage, increased friction, and elevated temperature, posing a risk of surface burns or defects.

[0036] (4) Passivation state (S4) When the cutting capability of the abrasive belt decreases, it is usually manifested as a decrease in high-frequency energy. (Lower), spectral center of gravity Downward shift. Physically, this means that the abrasive grains in the belt become dull, the cutting action weakens, and the material removal method may shift from "cutting-dominated" to "friction / plowing-dominated," leading to a deterioration in processing efficiency and surface quality.

[0037] Based on the experimental calibration results, the judgment interval for key features is defined as follows: When the grinding process is in an under-grinding state, the overall energy level of the acoustic emission signal is low, and the judgment condition is as follows: Root mean square value High-frequency energy ratio ; center of gravity .

[0038] This condition indicates insufficient contact between the abrasive belt and the workpiece, resulting in low material removal efficiency.

[0039] When the grinding process is stable and within the normal operating range, its characteristic parameters satisfy: ; ; .

[0040] This state corresponds to an ideal grinding condition with uniform energy input and stable cutting behavior.

[0041] When the grinding load is too high, the acoustic emission signal energy increases significantly. The criterion for determining this is: ; ; .

[0042] This condition is usually associated with belt blockage, increased friction, and localized temperature rise, posing a risk to processing quality.

[0043] When the abrasive grains of the abrasive belt become passivated, the high-frequency components of the acoustic emission signal attenuate. The criterion for this is: ; ; .

[0044] The state determination logic is as follows: (11) Where m = 1, 2, 3, 4, These correspond to the grinding states of "under-grinding, normal, overload, and passivation," respectively.

[0045] In the rapid discrimination mode, the present invention first obtains a set of fixed thresholds calibrated offline based on a large amount of grinding experimental data under under-grinding, normal, overload, and passivation conditions. This serves as a global benchmark for initial online status identification. However, in complex surface grinding, the curvature differences at different locations (convex / concave / edge) lead to significant variations in the effective contact area, causing natural fluctuations in acoustic emission energy even under normal grinding conditions. Directly using the aforementioned fixed threshold can easily result in misjudgments due to geometric factors. Therefore, this invention introduces curvature feedback from the robot in real time. , For the robot at the current moment The local curvature index at the contact point between the abrasive belt / grinding head and the workpiece is used to characterize the influence of complex surface geometry on the effective contact area and acoustic emission energy.

[0046] In one implementation, the curvature Take the average curvature of the workpiece surface at the contact point. ,Right now ,in, The two principal curvatures at this point are calculated as follows: "Contact point" refers to the point at which contact occurs. The point on the workpiece surface where the abrasive belt / grinding head makes effective contact with the workpiece is denoted as . in, The nominal CAD / NURBS parameter surface of the workpiece or the actual surface obtained by measurement.

[0047] At this point At this point, a tangential plane exists on the surface of the workpiece. With unit normal direction .make , Let the two principal directions be mutually orthogonal within the tangent plane (i.e., the characteristic directions corresponding to the shape operator / second fundamental form), then the corresponding two principal curvatures are defined as follows: in, (Or sorted according to convention), the two are points. The maximum / minimum normal curvature of the surface.

[0048] The curvature index is the average curvature: The offline fixed threshold is dynamically corrected online. Considering that grinding condition identification involves the RMS value, frequency domain characteristics, and vibration characteristics of the acoustic emission signal, this invention sets a threshold set consisting of five benchmark thresholds. The formulas for calculating the curvature-related online dynamic thresholds, corresponding to the lower RMS limit, upper RMS limit, high-low frequency energy ratio threshold, and spectral centroid threshold, are as follows: (12) in, This is the online threshold after the i-th fixed threshold is corrected at time t; A fixed threshold for offline calibration; The average curvature corresponding to the current contact point; This is the curvature sensitivity coefficient corresponding to the i-th fixed threshold (for example, for RMS features, the larger the curvature, the smaller the contact area, so the coefficient is positive to increase the threshold; for vibration features, changes in curvature may cause changes in stiffness, so the coefficient needs to be adjusted according to the calibration).

[0049] 2. Pattern classification based on machine learning Feature extraction is performed on the zero-mean normalized signal. The obtained time-domain and frequency-domain features are used for subsequent grinding state identification. The feature update cycle is 10 ms, and the calculation delay is no more than 5 ms, which meets the requirements of real-time control. This invention uses LTSD features as a sensitive indicator to judge transient fluctuations in grinding (such as sudden changes in local stress, instantaneous blockage of the abrasive belt, etc.), which complements steady-state features such as RMS features and η features. This helps to identify instantaneous changes in grinding state caused by curvature and stress changes in local areas of complex curved surfaces, thereby improving the accuracy of identification.

[0050] like Figure 2 As shown, based on the above multimodal features, a state recognition model for the grinding process is constructed to realize real-time discrimination of four typical working conditions in the grinding process of complex curved surface robots: "under-grinding, normal, overload, and passivation". Among them, under-grinding: the main frequency gradually increases with the improvement of contact → LTSD>0; passivation: high frequency components decay → LTSD<0; overload / burn precursor: the main frequency jumps significantly → |LTSD| increases sharply.

[0051] To improve recognition accuracy and robustness, supervised learning algorithms are used for model training in the offline stage.

[0052] Input signal: the above-mentioned comprehensive feature vector ; Output label: Grinding condition category These correspond to the grinding states of "under-grinding, normal, overload, and passivation," respectively.

[0053] (1) Classifier selection To improve adaptability to the highly non-stationary scenario of grinding complex curved surfaces, such as Figure 3 As shown, this invention introduces an adaptive kernel width adjustment mechanism and a class weighting strategy on the basis of traditional SVM to construct an "Improved-SVM (ISVM)".

[0054] A. Adaptive kernel width adjustment mechanism Based on the comprehensive feature vector Local divergence For RBF kernel parameters Perform adaptive adjustments: (13) in, Indicates the regulating factor. Representation and synthesis of eigenvectors (t) An initial kernel width parameter, matching the dimensions and numerical ranges of each dimension, is used to characterize the basic similarity scale of the feature space. Its value is determined through offline sample statistics or cross-validation. Represents the comprehensive feature vector The local divergence within the current window can be obtained based on variance-based estimation, calculated as follows: Step 1: Determine the data window by taking the feature vectors corresponding to the current time t and M samples from its preceding time (M is the window size, for example, M=10 or M=20): Step 2: Calculate the mean vector, which is the center vector within the window. Calculate the average of these M vectors. : Step 3: Calculate the local divergence D(t), which is the variance estimate. The variance is the average squared Euclidean distance between each eigenvector and the center vector within the calculation window. This is D(t). In this way, for under-wear / passivation state features that are sparsely distributed, the kernel width will automatically increase, making it easier to separate; for sudden and drastic changes such as overload, the kernel width will automatically decrease, making it more sensitive.

[0055] B. The decision function of ISVM is in the form of: The decision function of the Improved Support Vector Machine (ISVM) is defined as follows: The meanings of each symbol are defined as follows: Current moment The comprehensive feature vector is used to characterize the instantaneous state of the complex surface grinding process, and is composed of acoustic emission features, force features and vibration features; : No. The feature vectors corresponding to each support vector are derived from the offline training sample set; The number of support vectors; The first one obtained through training Lagrange multipliers corresponding to each support vector; : No. Each support vector corresponds to a category label, and its value corresponds to the grinding state category. : Bias term of the classification hyperplane; : Symbolic function, used to output the classification result of grinding state; The radial basis function (RBF) kernel function based on adaptive kernel width has the following form: Compared with the traditional SVM, this improved ISVM has improved the adaptiveness of classification boundaries, the recognition rate of abnormal states, and the robustness to interference from complex curvatures, making it more suitable for signal pattern classification in complex surface grinding.

[0056] (2) Training and verification Using the experimental dataset for five-fold cross-validation, the average classification accuracy can reach over 95%, and the F1-score is ≥ 0.93.

[0057] (3) Model lightweighting and online updates To adapt to real-time robot control, a sliding window is used to update model parameters: (15) in, To achieve adaptive learning.

[0058] This method can effectively eliminate natural energy fluctuations caused by curvature changes and improve the stability of actual working condition determination.

[0059] Step 5: Key Parameter Adaptive Fuzzy PID Control Algorithm like Figure 4 As shown, when the system identifies different grinding states, this invention employs an adaptive fuzzy PID control algorithm to coordinately adjust key process parameters during the grinding process based on the physical characteristics of the corresponding state, in order to achieve stable processing and anomaly suppression, as detailed below: When the condition is identified as under-grinding, it indicates insufficient contact between the abrasive belt and the workpiece, resulting in a low material removal rate. The system then increases the normal force applied to the workpiece surface. Or reduce the feed rate If necessary, increase the linear speed of the sanding belt appropriately. This enhances the effective cutting action per unit time and increases the grinding energy input.

[0060] When the system is identified as being in a normal grinding state, it maintains the current normal force. Feed rate and the linear velocity of the sanding belt The process remains unchanged, allowing the grinding process to continue operating under stable energy input conditions to ensure consistent surface quality.

[0061] When an overload condition is identified, it indicates that the grinding energy input is too high, posing a risk of temperature rise and surface burn. The system reduces the normal force. Reduce feed rate Simultaneously reduce the linear speed of the sanding belt. This reduces instantaneous cutting load and heat accumulation, thereby suppressing the further development of overload and thermal damage.

[0062] When the system identifies the belt as being in a passivated state, it indicates that the belt's cutting ability has decreased but has not yet completely failed. The system prioritizes increasing the belt's linear speed. Or slightly increase normal force The system performs short-term adaptive compensation to delay the impact of passivation on processing stability. When the grinding state is still determined to be passivated after compensation, or the state recognition confidence is continuously lower than the preset threshold, the system triggers a sanding belt replacement or automatic dressing command to prevent continued processing from causing surface quality deterioration.

[0063] 1. Control objectives and function outputs After completing the grinding condition identification, the system needs to automatically adjust the grinding parameters based on the acoustic emission feedback and identification results to maintain a constant energy input and surface quality for complex surfaces under different curvature and material conditions. The main control objectives include: maintaining the target grinding energy. Stable; limiting grinding temperature rise Not exceeding the threshold; improving surface roughness consistency To achieve coordinated optimization of energy, force, and velocity, the final output includes three adaptive adjustment parameters: , respectively, represent normal force, feed rate, and belt linear velocity.

[0064] The control system is based on acoustic emission energy feedback signals. A closed-loop regulation circuit is constructed, and its core logic is as follows, in order to achieve a dynamic closed loop of energy error, control output, and parameter correction.

[0065] 2. Error Definition and Energy Calculation Let the RMS of the acoustic emission signal be... The system uses empirical proportional coefficients. Calculate the current grinding energy: (16) in: This represents the ratio of sound energy to grinding energy (calibrated and determined, taken as 0.8–1.2 J / V²). This indicates the effective energy intensity of the acoustic emission signal.

[0066] Energy error is defined as: (17) Its energy error change rate is defined as: (18) This error signal serves as the input to the adaptive controller.

[0067] 3. Implementing Adaptive PID Control Using Fuzzy Algorithms (1) Basic control law (19) in: These represent the proportional, integral, and derivative gains, respectively. u(t) represents the comprehensive control quantity output by the fuzzy adaptive PID controller at time t, used to characterize the adjustment requirement of the current grinding process relative to the target grinding energy. (2) Fuzzy adjustment mechanism Grinding energy error and error change rate As input variables for the fuzzy controller, they are divided into seven sets of linguistic variables: {NB, NM, NS, ZO, PS, PM, PB}. NB represents negative large, NM represents negative medium, NS represents negative small, ZO represents zero, PS represents positive small, PM represents positive medium, and PB represents positive large. A 7×7 fuzzy rule table is designed based on this, and is obtained through offline simulation and experimental data calibration.

[0068] A fuzzy algorithm is used to adaptively adjust the proportional gain Kp, integral gain Ki, and derivative gain Kd. The adjustment formula is as follows: (20) in, These represent the proportional control gain, integral control gain, and derivative control gain, respectively. These represent the initial set values ​​of the proportional, integral, and derivative control gains, respectively. These represent the results obtained by the fuzzy controller based on the error. and error change rate The calculated gain correction amount; This indicates the error between the current grinding energy and the target grinding energy; Indicates the error Rate of change over time.

[0069] when big, When the value is large, increase the proportional gain. Suppress overshoot; when Small, Hours, increase integral gain This improves steady-state accuracy.

[0070] 4. Parameter mapping and real-time correction Based on control output Three key parameters are corrected through mapping relationships. : (twenty one) in: This represents the gain factor, with typical values ​​of 0.3, 0.1, and 0.05. This indicates the output value of the previous control cycle.

[0071] In summary, by using fuzzy algorithms to adjust the various parameters of the PID controller to optimize its response characteristics, and then using parameter mapping relationships to convert the control output into adjustment amounts for the actual grinding process parameters, the two processes work together at different control levels to achieve adaptive control of complex surface grinding processes, which helps to improve the robustness and accuracy of the control.

[0072] The following example, using a robot-driven sander belt to grind aero-engine blades, illustrates the adaptive control process in detail: 1. Process Objects and Equipment Configuration Workpiece: A turbine blade for a certain type of aero-engine, made of a nickel-based superalloy, with significant curvature variation, affecting surface roughness. High requirements for surface integrity; Equipment: Six-axis industrial robot + belt grinding end effector, equipped with a belt grinding unit driven by a speed adjustable motor; Sensor placement: Install broadband AE sensors at the blade root or on the fixture; install a six-dimensional force / torque sensor at the robot end effector; install a three-dimensional accelerometer on the grinding head housing; Data acquisition: AE sampling rate 2 MHz, FT sampling rate 2 kHz, ACC sampling rate 20 kHz, unified clock and trigger control are achieved through a multi-channel synchronous acquisition card.

[0073] 2. Offline calibration and model training Grinding tests were conducted on different areas of the blade under typical grinding parameter combinations (different normal forces, feed rates, belt speeds, and belt wear degrees), and AE / FT / ACC data were collected simultaneously. Data is divided into 10 ms time windows, multi-scale features are extracted, and combined with surface roughness, temperature rise and surface defect detection results, the working conditions of each window are labeled with under-wear, normal, overload, passivation and other labels. An improved SVM classifier is trained using the above-mentioned labeled feature set, and hyperparameters such as adaptive kernel width and class weights are determined through cross-validation. A prediction model between process parameters and energy AE is established to obtain the energy prediction residual threshold.

[0074] 3. Online processing and adaptive control process Following a pre-planned NURBS trajectory, the robot drives the belt grinding head to sweep and grind along the complex curved surface of the blade. During the processing, AE / FT / ACC signals are acquired in real time and preprocessed such as bandpass filtering, wavelet denoising, and sliding normalization. Multi-scale features are calculated with a period of 10 ms, input into an improved SVM and combined with prediction residual criteria to identify the current working condition in real time; based on the state identification results and energy error, the fuzzy adaptive PID controller outputs parameter adjustment amounts to automatically correct the normal force, feed speed and belt speed. If the system identifies a situation where the sanding belt is severely passivated or has an excessively high risk of overload, it will issue a prompt to replace or repair the sanding belt.

[0075] 4. Effect Verification Compared with traditional fixed-parameter grinding processes, the method of this invention can significantly reduce surface roughness fluctuations at different locations on the blade and lower the incidence of burn and microcrack defects; under typical process conditions, the workpiece surface roughness... Fluctuations are reduced by about 40%–60%, surface burn defects are basically eliminated, and the consistency of processing between blades is significantly improved.

[0076] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples. Various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims.

[0077] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications or variations that can be made by those skilled in the art without creative effort within the scope of the appended claims are still within the scope of protection of this patent.

Claims

1. An adaptive control method for grinding complex curved surfaces, characterized in that... Includes the following steps: Step 1: Collect and preprocess the acoustic emission signals during the grinding process; Step 2: Extract time-domain and frequency-domain features from the preprocessed acoustic emission signal; Step 3: Combine time-domain features and frequency-domain features to make a threshold judgment to identify the current grinding state. If the current grinding state is abnormal, use an adaptive fuzzy PID control algorithm to adjust the current grinding parameters. Step 4: Repeat steps 1 to 3 until the current grinding state returns to normal.

2. The adaptive control method for grinding complex curved surfaces according to claim 1, characterized in that: Let the acoustic emission signal within a certain period after preprocessing be denoted as . The time-domain characteristic energy density is calculated using the following formula. , Where N represents the total number of acoustic emission signals sampled over a period of time, and t represents the sampling time; Then use the formula Acoustic emission signal Perform a Fast Fourier Transform and calculate the centroid of the frequency characteristic spectrum using the following formula. High-frequency energy ratio ; 。 3. The adaptive control method for grinding complex curved surfaces according to claim 2, characterized in that: The grinding state is achieved using the following formula. Identification, in, For a pre-defined fixed threshold, m = 1, 2, 3, 4.

4. The adaptive control method for grinding complex curved surfaces according to claim 3, characterized in that: Use the following formula to apply a fixed threshold Make adjustments online. in, This is the online threshold after the i-th fixed threshold is corrected at time t; For offline calibration of fixed thresholds, , respectively corresponding to fixed thresholds ; The average curvature corresponding to the current contact point; is the curvature sensitivity coefficient corresponding to the i-th fixed threshold.

5. The adaptive control method for grinding complex curved surfaces according to claim 2, characterized in that: Using a bandpass filter to filter acoustic emission signals Filtering is performed, and then wavelet thresholding is used to refine the filtered acoustic emission signal. Noise reduction is performed to suppress high-frequency noise, and finally, normalization is applied.

6. An adaptive control method for grinding complex curved surfaces, characterized in that... Includes the following steps: Step 1: Collect acoustic emission signals, force / torque signals, and vibration signals during the grinding process, and preprocess the acoustic emission signals; Step 2: Extract time-domain and frequency-domain features from the preprocessed acoustic emission signal; Step 3: Combine the extracted time-domain features, frequency-domain features, force / torque signals, and vibration signals to construct a comprehensive feature vector. Input this vector into the machine learning model to identify the current grinding state. If the current grinding state is abnormal, use an adaptive fuzzy PID control algorithm to adjust the current grinding parameters. Step 4: Repeat steps 1 to 3 until the current grinding state returns to normal.

7. The adaptive control method for grinding complex curved surfaces according to claim 6, characterized in that: Let the acoustic emission signal within a certain period after preprocessing be denoted as . The time-domain characteristic energy density is calculated using the following formula. Single event peak value P, Where N represents the total number of acoustic emission signals sampled over a period of time, and t represents the sampling time; Then use the formula Acoustic emission signal Perform a Fast Fourier Transform and calculate the centroid of the frequency characteristic spectrum using the following formula. High-frequency energy ratio and the transient spectrum drift rate LTSD of the main frequency, This is used to construct a comprehensive feature vector. ,in, The force information represents the real-time normal force applied to a complex curved surface; It represents the vibration acceleration on a complex curved surface during the grinding process.

8. The adaptive control method for grinding complex curved surfaces according to claim 7, characterized in that: The machine learning model uses a support vector machine (SVM) with a decision function. The formula is as follows: in, Indicates the first The Lagrange multipliers corresponding to each support vector This represents the radial basis function (RBF) kernel function based on adaptive kernel width. Identifier The class labels corresponding to the support vectors are given by b, which represents the bias term of the classification hyperplane, and N represents the total number of support vectors. Indicates the first The comprehensive feature vector corresponding to each support vector is derived from the offline training sample set. This represents the comprehensive feature vector corresponding to the current time t; Based on the comprehensive feature vector Local divergence Calculate the radial basis function (RBF) kernel parameters. , in, Representing the eigenvector The local divergence within the current window is obtained based on variance-based estimation. Indicates the regulating factor. This represents the initial value.

9. The adaptive control method for grinding complex curved surfaces according to claim 1 or 6, characterized in that: The adaptive fuzzy PID control algorithm first uses fuzzy algorithms to measure grinding energy error. and the rate of change of grinding energy error As input, the proportional gain K of the PID controller is adaptively adjusted. p Integral gain K i and differential gain K d Then adjust the grinding parameters. , respectively, represent normal force, feed rate, and belt linear velocity.

Images