An on-line deformation detection method for high-performance titanium alloy part machining process

Abstract

The application discloses a kind of deformation on-line detection methods in high-performance titanium alloy part machining process, it is related to titanium alloy part machining technical field.The method steps include: titanium alloy blank is annealed and is treated with slow cooling, and low-stress blank is obtained;The chatter signal is input into the random forest model based on intrinsic mode decomposition, and the chatter state is output;The rough machining state is judged based on the chatter state, so that the rough machining part is obtained;The rough machining part is stationary and three-dimensional scanning, and three-dimensional point cloud sequence is obtained;The point cloud baseline data is calculated by iterative closest point method and the difference with tempering point cloud data, and the tempering deformation deviation is obtained, the finishing cutting parameters are input into BP neural network model, and the finishing deformation compensation amount is output to obtain total compensation amount;Based on total compensation amount, the part is finished, and the deformation value of the part is obtained in real time by laser displacement sensor, and the deformation value is greater than preset threshold, then early warning is carried out.

Description

Technical Field

[0001] This invention relates to the field of titanium alloy parts processing technology, specifically to an online deformation detection method during the processing of high-performance titanium alloy parts. Background Technology

[0002] Titanium alloys, due to their high specific strength, excellent corrosion resistance, and good high-temperature performance, are widely used in high-end manufacturing fields such as aerospace, energy equipment, and medical devices. With their high strength, excellent corrosion resistance, and high specific strength, titanium alloys have become a core material in aerospace (such as engine turbine blades and airframe load-bearing structures), medical devices (such as artificial joints and orthopedic implants), and high-end equipment manufacturing. The dimensional accuracy and morphological stability of titanium alloy parts directly determine the service performance and safety reliability of the end products.

[0003] In traditional methods, measuring tools such as calipers and micrometers are used to perform static offline measurements on the key dimensions of titanium alloy parts. If the measured data is greater than a preset threshold, the part is judged to be deformed; if it is less than the preset threshold, the part is judged to be normal.

[0004] However, existing technologies lack real-time detection methods for deformation of titanium alloy parts during the machining stage, making it impossible to monitor and suppress deformation during the cutting process in real time. This results in workpiece surface deformation and out-of-tolerance geometric accuracy, ultimately leading to a waste of titanium alloy materials and time. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an online deformation detection method for high-performance titanium alloy parts during machining, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online deformation detection method for high-performance titanium alloy parts during machining, comprising the following steps: Step S1: Place the titanium alloy blank in an electric furnace for annealing and slow cooling to obtain a low-stress titanium alloy blank. Step S2: Perform rough machining on the titanium alloy low-stress blank, and simultaneously collect rough machining temperature data, rough machining displacement data, and chatter signal data; input the chatter signal data into a random forest model based on intrinsic mode decomposition, and output the chatter state value; Step S3: Determine the state of the rough machining process by using chatter state value, rough machining temperature data, and rough machining displacement data. Adjust the process parameters in the rough machining process according to the state determination results and continue machining to finally obtain the rough-machined titanium alloy part. Step S4: Let the rough-machined titanium alloy part stand still until a preset period is reached to obtain a stable titanium alloy part. During the preset period, the rough-machined titanium alloy part is scanned in three dimensions by a laser scanner to obtain a three-dimensional point cloud data sequence. The adjacent deformation of the three-dimensional point cloud sequence is calculated by the iterative nearest point method. Based on the adjacent deformation, the trend stability analysis of the three-dimensional point cloud data sequence is performed to obtain the titanium alloy point cloud baseline data. Step S5: Temper the titanium alloy stable part to obtain a tempered titanium alloy part and perform a three-dimensional scan on the tempered titanium alloy part to obtain tempered point cloud data; compare the tempered point cloud data with the titanium alloy point cloud baseline data to calculate the tempering deformation deviation. Step S6: Before finishing the tempered titanium alloy parts, obtain the finishing cutting parameters, input the finishing cutting parameters into the BP neural network model, and output the finishing deformation compensation amount; by combining the tempering deformation deviation and the finishing deformation compensation amount, calculate the total finishing compensation amount, perform finishing processing on the tempered titanium alloy parts based on the total finishing compensation amount, and use a laser displacement sensor to obtain the deformation value of the tempered titanium alloy parts in real time during finishing. If the deformation value is greater than a preset threshold, an early warning is issued, thereby realizing the deformation detection of titanium alloy parts.

[0007] Preferably, the step of placing the titanium alloy blank in an electric furnace for annealing and slow cooling to obtain a low-stress titanium alloy blank includes the following specific steps: The titanium alloy billet is smoothly placed into the center of the working area of ​​the box furnace, ensuring that it is not in contact with the heating elements and that there is a uniform gap around it to ensure the circulation of hot air. Then, the furnace door is closed tightly, and the furnace temperature is raised to 750°C at a rate not exceeding 150°C / hour, and the holding time is started. This temperature is maintained for 2 hours to allow the internal structure of the titanium alloy billet to fully recrystallize and grow grains, thereby eliminating the original residual stress caused by previous forging or rolling. After the holding time is completed, the heating power is cut off, and the titanium alloy billet is allowed to cool naturally and slowly in the sealed furnace chamber, with the cooling rate controlled below 50°C / hour, so as to avoid the generation of new thermal stress due to rapid cooling. When the furnace temperature drops below 200°C, the furnace door can be opened to remove the titanium alloy billet, thereby obtaining a low-stress titanium alloy billet.

[0008] Preferably, the step of inputting the flutter signal data into a random forest model based on intrinsic mode decomposition and outputting the flutter state includes the following steps: The intrinsic mode decomposition formula is:

[0009] in, For flutter signal data, Let j be the j-th order eigenmode function. For low-frequency residuals, t is the time index, j is the order index, and n is the total order of the intrinsic mode functions, n=3; By merging the first three essential mode functions, the reconstructed effective signal is obtained:

[0010] in, To reconstruct a valid signal, For the first-order eigenmode function, For the second-order eigenmode function, It is the 3rd order eigenmode function; The power spectral energy S is obtained by performing a Fourier transform on the reconstructed effective signal. ):

[0011] in, for Power spectral energy at that location To reconstruct a valid signal, To reconstruct the i-th frequency component of the valid signal; Based on the power spectral energy, calculate the power spectral entropy:

[0012] in, For power spectral entropy, Let m be the probability density of the i-th frequency component, m be the total number of frequency components in the Fast Fourier Transform, and i be the index of the frequency component. Calculate the reconstructed valid signal The signal kurtosis peak is used to form a flutter feature vector by combining the power spectral entropy E and the signal kurtosis peak in each sample group; The flutter feature vectors collected in real time are input into the random forest model, and the flutter state is output:

[0013] in, The classification results are from the random forest model, where c is the class label and Q is the number of decision trees. Let q be the classification result of the q-th decision tree, where q is the index of the decision tree. This represents the classification result of the q-th decision tree. When equal to category c, The value is 1 if it is not 0 otherwise. For optimization operations, it means finding the category label c that maximizes the total number of votes in the voting mechanism of the random forest model.

[0014] Preferably, the step of determining the state of the roughing process using chatter status, roughing temperature data, and roughing displacement data includes the following steps: The comprehensive risk index R is calculated using chatter state values, rough machining temperature data, and rough machining displacement data.

[0015] Where R is the comprehensive risk index. Temperature weighting, WD represents the temperature threshold, where WD is the roughing temperature. Here, WY represents the displacement weight, and WY represents the roughing displacement. Displacement threshold For flutter state weights, This represents the risk value for flutter conditions. + + =1.

[0016] Preferably, the step of resting the rough-machined titanium alloy part includes the following steps: The rough-machined blank enters the stage of natural stress release and data monitoring: First, the rough-machined workpiece is removed from the machine tool and transferred to a stable environment with constant temperature and no vibration for static placement for more than 12 hours, so that the internal residual stress can be fully and naturally released.

[0017] Preferably, the calculation of adjacent deformations of the 3D point cloud sequence using the iterative nearest point method includes the following specific steps: Will and Registration is performed using the iterative nearest-point method to obtain the transformation matrix. Thus Precise alignment In the coordinate system, the registered point cloud is obtained. :

[0018] Calculate the point cloud after registration Each point in the middle and its position Calculate the Euclidean distance between the nearest neighbors and the average value:

[0019] in, Let be the adjacent deformation of the k-th region at time t+1 and time t. Let h be the number of point clouds in the k-th region, and h be the index of the point cloud points. For the registered point cloud The h-th point cloud point and the point cloud The Euclidean distance between its nearest neighbor and the nearest neighbor.

[0020] Preferably, the step of performing trend stability analysis on the three-dimensional point cloud data sequence based on adjacent deformations to obtain titanium alloy point cloud baseline data includes the following specific steps: The maximum value of adjacent deformations in all regions of the rough-machined titanium alloy part is taken as the overall deformation evaluation index for that time period. If the deformation evaluation index calculated at the last time interval is less than or equal to the preset stability threshold, the three-dimensional shape of the part is determined to have reached a stable state, and the three-dimensional point cloud data at the last moment is used as the titanium alloy point cloud baseline data. .

[0021] Preferably, the step of tempering the titanium alloy stable part to obtain a tempered titanium alloy part and performing a three-dimensional scan on the tempered titanium alloy part to obtain tempered point cloud data includes the following specific steps: The titanium alloy stable parts were placed in a tempering furnace and subjected to a standardized heat treatment process of 600°C for 2 hours, followed by slow cooling to room temperature at a rate of less than 50°C / hour. After tempering, the titanium alloy stable parts were immediately transferred to a constant temperature measurement chamber. After the temperature in the constant temperature measurement chamber was kept uniform and stable, the tempered titanium alloy parts were obtained. The tempered titanium alloy parts were then subjected to three-dimensional scanning to obtain tempering point cloud data.

[0022] Preferably, the step of comparing the tempered point cloud data with the titanium alloy point cloud baseline data to calculate the tempering deformation deviation includes the following specific steps: Calculate the dimensional deviation between the tempered point cloud data and the titanium alloy point cloud baseline data:

[0023] in, The size deviation of the k-th region. This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. This refers to the size of the k-th region in the registered tempered point cloud data.

[0024] Preferably, the step of calculating the total compensation amount for finishing by combining the tempering deformation deviation and the finishing deformation compensation amount includes the following specific steps: The finishing cutting parameters are input into the BP neural network model, and the finishing deformation compensation amount is output:

[0025] in, To compensate for deformation during finishing, Input function for neural network model, For finishing cutting parameters, including cutting speed, feed per tooth, axial depth of cut, radial depth of cut and number of tool teeth; By combining the tempering deformation deviation and the finishing deformation compensation amount, the total finishing compensation amount is calculated as follows:

[0026] in, This represents the total compensation amount for the k-th region. The size deviation of the k-th region. This represents the theoretical processing allowance for the k-th region. >0, = - , This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. For the original design dimensions of the titanium alloy machined part in the k-th region, when In this case, the total compensation for finishing is not calculated. = - , This is the amount of compensation for deformation during finishing.

[0027] This invention provides an online deformation detection method for high-performance titanium alloy parts during machining, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The flutter signal data is input into the random forest model based on intrinsic mode decomposition, and the flutter state value is output, which realizes the leap from passive detection to active early warning in the processing process. The EMD method can adaptively process nonlinear and non-stationary vibration signals in titanium alloy processing and effectively extract the weak features in the flutter incubation stage; combined with the powerful classification ability of the random forest model, it can accurately identify its critical state before the flutter occurs violently (incubation stage), thus winning valuable time for active intervention.

[0028] (2) The adjacent deformation of the three-dimensional point cloud sequence is calculated by the iterative nearest point method. The iterative nearest point method can efficiently and accurately register the three-dimensional point cloud obtained by scanning at different time points and calculate the sub-millimeter level of small deformation. This allows for the accurate capture of the slow and subtle deformation trend of the part caused by the release of internal stress during the static process. It can objectively determine when the part enters a stable equilibrium state from the active stress release stage (i.e., the deformation evaluation index ≤ stability threshold), thereby scientifically determining the required static time and avoiding the problem of insufficient time or excessive waiting caused by estimation based on experience. This provides a stable and reliable geometric benchmark for the subsequent tempering process.

[0029] (3) The finishing cutting parameters are input into the BP neural network model, and the finishing deformation compensation amount is output. This realizes the intelligent closed-loop control of the machining error prediction and compensation. By learning historical machining data, the BP neural network can establish a complex nonlinear mapping relationship between cutting parameters and deformation amount, accurately predict the elastic deformation amount that will be generated during finishing under specific parameters, and pre-assign its offset to the machining allowance parameter in the CAM system when generating the CNC toolpath. This allows the toolpath to actively give up the expected deformation space. After machining is completed and deformation is recovered, the part size can fall exactly within the tolerance range. This transforms post-correction into pre-avoidance, greatly improving the first-pass yield of machining accuracy. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the online deformation detection method for high-performance titanium alloy parts during the processing proposed in this invention. Figure 2 This invention provides a step hierarchy diagram of rough-machined titanium alloy parts obtained from an online deformation detection method during the machining process of high-performance titanium alloy parts. Figure 3 This is a step hierarchy diagram of obtaining tempering deformation deviation in an online deformation detection method during the machining process of high-performance titanium alloy parts proposed in this invention. Detailed Implementation

[0032] 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.

[0033] Please see Figures 1-3 The present invention provides a technical solution: an online deformation detection method for high-performance titanium alloy parts during the machining process.

[0034] Step S1: Place the titanium alloy blank in an electric furnace for annealing and slow cooling to obtain a low-stress titanium alloy blank.

[0035] The titanium alloy billet is carefully placed into the center of the working area of ​​the box furnace, ensuring it is not in contact with the heating elements and that there is a uniform gap around it to allow for hot air circulation. The furnace door is then tightly closed, and the furnace temperature is uniformly raised to 750°C at a rate not exceeding 150°C / hour, and a holding time is initiated. This temperature is maintained for 2 hours to allow the internal structure of the titanium alloy billet to fully recrystallize and grow grains, thereby eliminating the original residual stress generated during previous forging or rolling. After the holding time, the heating power is cut off, allowing the billet to cool naturally and slowly within the sealed furnace chamber, controlling the cooling rate to be below 50°C / hour to avoid generating new thermal stress due to rapid cooling. When the furnace temperature drops below 200°C, the furnace door can be opened to remove the titanium alloy billet. At this point, the titanium alloy billet has undergone stress pretreatment and stabilization, its internal structure is uniform, and the stress level is significantly reduced, providing a stable material basis for subsequent rough machining.

[0036] It should be noted that the combined process of heat preservation and slow cooling in the furnace for titanium alloy blanks aims to thoroughly reduce stress and stabilize the microstructure. During the previous forging and rolling process, the internal lattice of the titanium alloy blank generates and stores significant original residual stress due to the huge plastic deformation. High-temperature annealing causes recrystallization inside the blank, effectively eliminating the original residual stress accumulated during forging and rolling. Strict slow cooling control avoids the introduction of new thermal stress due to sudden temperature changes, thereby significantly improving the dimensional stability of the material from the source. This provides a low-stress, high-stability material foundation for all subsequent processing stages, fundamentally reducing the risk of deformation in subsequent cutting processes.

[0037] Step S2: Perform rough machining on the titanium alloy low-stress blank, and simultaneously collect rough machining temperature data, rough machining displacement data, and chatter signal data; input the chatter signal data into a random forest model based on intrinsic mode decomposition, and output the chatter state value.

[0038] First, the low-stress titanium alloy blank is clamped on a CNC machine tool. A small-diameter carbide end mill is selected and conservative cutting parameters are set to control the single cutting load. Then, high-precision laser displacement sensors and infrared temperature sensors are installed near the machine tool spindle and at key points in the machining area, and communication and debugging between the sensing system and the CNC system are completed. The PCB accelerometer (e.g., PCB352C22) is attached to a high-deformation area of ​​the low-stress titanium alloy blank (e.g., thin-walled sections of the part, complex cavity sidewalls). It should be avoided in the workpiece clamping area (≥30mm from the edge of the fixture) and the direct cutting action area (≥50mm from the milling cutter path) to prevent clamping stiffness from interfering with the signal or chips from damaging the sensor. Before attachment, the oxide layer on the part surface is wiped with alcohol. High-temperature resistant adhesive is used for attachment, and the adhesive layer thickness is controlled to be ≤0.1mm. It is cured for 15 minutes to ensure a rigid connection between the sensor and the part and to avoid vibration signal attenuation. At the same time, the sensor, displacement sensor and infrared temperature sensor are set to a unified trigger signal, and the sampling frequency is set to 1000Hz to meet the high-frequency signal acquisition requirements of chatter and adapt to the high-speed scenarios of high-performance part processing. Signal acquisition is performed during roughing. One signal acquisition is triggered after each feed increment (e.g., 5mm feed increment). Each acquisition lasts for 3 seconds (corresponding to 3000 sample points, 1000Hz×3s) to ensure coverage of vibration characteristics within a single cutting cycle. Pulse noise during the sampling process is removed using the signal mean ± 3 times the standard deviation as a threshold to prevent invalid data from entering the subsequent model. After acquisition, the acceleration signal, displacement data, and temperature data are stored aligned with the timestamp.

[0039] Experimental data collection for roughing process parameters and chatter conditions: Based on roughing process parameters (spindle speed 1500-3000 r / min, axial depth of cut 1-5 mm, feed rate 0.02-0.1 mm / z), sample collection experiments were designed for three types of chatter conditions: normal cutting, chatter initiation, and chatter occurrence, ensuring that the sample data fully matched the machining scenario. For the normal cutting samples (e.g., spindle speed 2000 r / min, axial depth of cut 2 mm, feed rate 0.05 mm / z), the conditions of displacement ≤ 0.02 mm and temperature ≤ 400℃ during machining were met. 50 sets of parameters were collected for each set, for a total of 150 sets, labeled "1". For chatter initiation samples, the parameters needed to be adjusted to "critical". For samples exhibiting chatter, the parameters are adjusted to the "supercritical state" (e.g., spindle speed 2000 r / min, axial depth of cut 3.5 mm, feed rate 0.08 mm / z). During machining, the displacement is controlled within 0.02-0.04 mm, the temperature is controlled within 400-450℃, and the vibration amplitude is observed to begin to fluctuate. 30 sets of parameters are collected for each set, for a total of 90 sets, labeled "2". For samples exhibiting chatter, the parameters are adjusted to the "supercritical state" (e.g., spindle speed 2000 r / min, axial depth of cut 4.5 mm, feed rate 0.1 mm / z). During machining, the displacement must be greater than 0.04 mm, the temperature greater than 450℃, and a sudden increase in vibration amplitude must be observed. 30 sets of parameters are collected for each set, for a total of 90 sets, labeled "3".

[0040] Intrinsic mode decomposition (IMF) was performed on all the chatter signal data corresponding to the collected roughing process parameters. The vibration signal from titanium alloy milling contains a large amount of cutting noise (nonlinear and non-stationary). IMF can be adaptively decomposed into intrinsic mode functions (IMFs). The first three IMFs contain more than 90% of the chatter information. After removing the residual trend term, the effective signal can be obtained. Therefore, the IMF formula is:

[0041] in, For flutter signal data, Let j be the j-th eigenmode function. For low-frequency residuals, t is the time index, j is the order index, and n is the total order of the intrinsic mode functions, n=3.

[0042] It should be noted that the EMD decomposition steps are as follows: Find all the maxima / minima of s(t), fit the upper envelope U(t) and lower envelope L(t) of the signal using cubic spline interpolation; calculate the mean of the envelopes. : ; Calculate the difference between the original vibration signal and the mean of the envelope: Check if the IMF conditions are met (number of zero crossings ≈ number of extreme points, envelope symmetry); if not, use For the new signal, repeat the above steps until the first-order IMF is obtained. ;make Repeat the above steps to obtain , Stop at (The residual term has no fluctuation).

[0043] Extract flutter-sensitive features from the 3rd order IMF reconstructed signal: power spectral entropy reflects the frequency distribution dispersion (frequency concentration during flutter, entropy value drops sharply), and kurtosis reflects the impact characteristics (impact is enhanced during flutter, kurtosis increases sharply).

[0044] By merging the first three IMFs, a reconstructed effective signal is obtained:

[0045] in, To reconstruct a valid signal, For the first-order eigenmode function, For the second-order intrinsic mode function, It is the third-order intrinsic mode function.

[0046] The power spectral energy S is obtained by performing a Fourier transform on the reconstructed effective signal. ):

[0047] in, for Power spectral energy at that location To reconstruct a valid signal, To reconstruct the i-th frequency component of the valid signal.

[0048] Based on the power spectral energy, calculate the frequency component probability density and power spectral entropy:

[0049] in, Let m be the probability density of the i-th frequency component, m be the total number of frequency components in the Fast Fourier Transform, and i be the index of the frequency component.

[0050]

[0051] in, For power spectral entropy, Let m be the probability density of the i-th frequency component, m be the total number of frequency components in the Fast Fourier Transform, and i be the index of the frequency component.

[0052] Calculate the reconstructed valid signal The signal kurtosis peak (calculated by reconstructing the mean and standard deviation of the effective signal) is used to form a flutter feature vector by combining the power spectral entropy E and the signal kurtosis peak in each sample group.

[0053] It should be noted that the kurtosis and power spectral entropy are used to form a feature vector for flutter state identification because they can complement each other to capture the nonlinear and non-stationary characteristics of titanium alloy milling flutter signals from two non-redundant key dimensions in the time and frequency domains. Kurtosis, as a time-domain indicator, quantifies the impact characteristics of a signal. During normal cutting, the vibration signal in titanium alloy milling is stable. After entering the chatter incubation stage, the vibration amplitude begins to fluctuate, leading to increased impact and a slow increase in kurtosis. When chatter occurs, the violent vibration generates a significant pulse impact, causing a sharp increase in kurtosis. This accurately characterizes the time-domain impact intensity changes caused by chatter. Power spectral entropy, as a frequency-domain indicator, reflects the dispersion of the signal frequency distribution. During normal cutting, energy is concentrated at the tooth passage frequency and low harmonics, resulting in a concentrated frequency distribution with low uncertainty and a high power spectral entropy value (e.g., above 0.35). During the chatter incubation stage, energy diffuses to multiple frequencies, increasing distribution uncertainty and decreasing entropy. When chatter occurs, energy further transfers to higher harmonics, causing the frequency distribution to become more concentrated but with a frequency band shift, and the entropy value continues to decrease to below 0.2. This effectively captures the frequency domain energy reconstruction pattern caused by chatter. Combining these two indicators can comprehensively cover the essential signal characteristics of different stages of chatter, avoiding the problem of missing detection due to local signal changes caused by a single indicator (only time domain or only frequency domain), ensuring that the feature vector can stably distinguish the three chatter states.

[0054] The flutter feature vectors and corresponding flutter labeled samples in each sample group are divided into training and test sets in a 7:3 ratio. The training set is then input into a random forest model for training. During training, the core parameters of the model are configured according to the dimensions of the flutter feature vectors (two-dimensional features: E and peak) and the sample size (e.g., 330 samples): a parameter grid is constructed with the number of decision trees (100, 150, 200) and the maximum depth (5, 6, 7, 8). Cross-validation is performed on each parameter combination on the training set, and the average accuracy of cross-validation is used as the evaluation metric. Finally, the parameter combination with the highest average accuracy (e.g., decision tree = 150, maximum depth = 7) is selected as the final hyperparameters of the model to maximize the model's generalization ability and avoid overfitting. The Gini coefficient is used as the splitting criterion (to improve classification efficiency). During training, a subset (approximately 2 / 3 of the total training set) is randomly selected from the training set using Bootstrap sampling (sampling with replacement) to train a single decision tree; the samples not selected naturally form the validation set for that tree. Subsequently, at each node of the decision tree, due to the low feature dimensionality (only two features, E and peak), the algorithm no longer performs random feature selection, but directly uses both features to find the optimal split point (based on minimizing the Gini coefficient). This process is repeated until a predetermined number of decision trees (e.g., 150) are generated, at which point training stops.

[0055] It should be noted that after model training, the model performance needs to be verified in real time using a test set, focusing on the accuracy of identifying the three types of flutter states (normal, gestation, and occurrence). The overall accuracy should be ≥95%, with the identification rate of the gestation state being no less than 93% (to avoid the risk of processing deformation due to missed detection). If the accuracy is not up to standard, it can be optimized by increasing the number of decision trees (e.g., increasing to 200), adjusting the maximum depth (e.g., adjusting to 7), or supplementing boundary samples (e.g., adding flutter critical state samples). When the model's accuracy on the test set is stable above 95% for three consecutive iterations, and the confusion matrix shows that the misclassification rate of the three types of samples is ≤5%, training should be stopped and the model parameters (including tree structure, split threshold, voting weights, etc.) saved. Finally, a trained random forest model that can be directly used for flutter state identification is obtained.

[0056] The flutter signal data collected in real time is input into a random forest model based on intrinsic mode decomposition, and the flutter state is output:

[0057] in, The classification results are from the random forest model, where c is the class label and Q is the number of decision trees. Let q be the classification result of the q-th decision tree, where q is the index of the decision tree. This represents the classification result of the q-th decision tree. When equal to category c, The value is 1 if it is not 0 otherwise. For optimization operations, it means finding the category label c that maximizes the total number of votes in the voting mechanism of the random forest model.

[0058] If the classification result of the random forest model is 1, the flutter state is normal cutting; if the classification result of the random forest model is 2, the flutter state is flutter initiation; and if the classification result of the random forest model is 3, the flutter state is flutter occurrence.

[0059] Step S3: Determine the state of the rough machining process using chatter status value, rough machining temperature data, and rough machining displacement data. Adjust the process parameters in the rough machining process according to the state determination results and continue machining to finally obtain the rough-machined titanium alloy part.

[0060] Define flutter state risk value When flutter state value = 1 =0; when the flutter state value = 2 =0.5; when the flutter state value = 3, =1.0.

[0061] Using chatter state values, roughing temperature data, and roughing displacement data, a comprehensive risk index R is calculated.

[0062] Where R is the comprehensive risk index. Temperature weighting, WD represents the temperature threshold, where WD is the roughing temperature. Here, WY represents the displacement weight, and WY represents the roughing displacement. Displacement threshold For flutter state weights, This represents the risk value for flutter conditions. + + =1.

[0063] It's important to clarify the temperature threshold and displacement threshold. The temperature threshold represents the upper limit of the critical temperature for the stability of the machine tool-workpiece-tool system under specific titanium alloy materials (such as TC4) and a given roughing process. When the real-time temperature WD exceeds this threshold, it indicates a surge in overheating risk, which may lead to rapid tool wear, surface burning of the part, or changes in the material's microstructure, resulting in uncontrollable deformation. For example, for TC4 material, the temperature threshold can be set to 400°C. The displacement threshold represents the upper limit of the critical displacement for workpiece vibration or tool deformation under a given clamping scheme and cutting parameters. When the real-time displacement WY exceeds this threshold, it directly indicates excessive cutting force or insufficient rigidity of the machining system, chatter is about to occur or has already occurred, the machining state is extremely unstable, and continued machining is very likely to lead to geometric deviations in the part. For example, for TC4 material, the displacement threshold can be set to 0.02 mm.

[0064] It should be noted that the weights for temperature, displacement, and flutter state are as follows: temperature weight can be 0.5, displacement weight can be 0.3, and flutter state weight can be 0.2. Temperature has the highest weight (0.5) because it is the root cause of material softening and thermal deformation, and can serve as a leading indicator of process deterioration, enabling early warning. Displacement has the second highest weight (0.3), as it is a direct quantitative result of deformation and is directly related to part quality, but it has a certain lag. Flutter state has the lowest weight (0.2) because although it is a strong signal of process instability, it is a derived parameter and relies on model inference, so its stability is slightly lower than the direct measurement value of physical sensors.

[0065] The system is judged and parameters are adjusted based on the risk index R: Low risk (R<0.3): The system is judged to be stable. Continue machining with the current roughing parameters (such as spindle speed and feed rate) while continuously collecting three types of data to ensure that the system is in a stable state.

[0066] Medium risk (0.3≤R<0.7): The system is determined to be in a critical state, triggering the early warning mechanism. First, reduce the feed rate by 10%-15% (to reduce cutting load), and simultaneously activate the high-pressure oil mist cooling system to enhance heat dissipation (keeping the temperature below 450℃). After adjustment, continuously monitor for 15 seconds (approximately 3 spindle rotation cycles). If the R value drops to the low-risk range, resume machining; if the R value does not improve or increases, further reduce the axial depth of cut by 15%-20%.

[0067] High risk (R≥0.7): System instability is determined, and roughing should be stopped immediately. After stopping the machine, check the tool wear (if the flank wear is >0.2mm, the tool needs to be replaced), and readjust the core parameters: spindle speed ±300 r / min (avoid the chatter sensitive range), and reduce the axial depth of cut by 25%. After adjustment, verify the system vibration status through a no-load test (ensure the chatter status returns to 1), then restart roughing and increase the data acquisition frequency to 2 seconds / time, tracking the R value change in real time.

[0068] Step S4: Let the rough-machined titanium alloy part stand still until a preset period is reached to obtain a stable titanium alloy part. During the preset period, the rough-machined titanium alloy part is 3D scanned by a laser scanner to obtain a 3D point cloud data sequence. The adjacent deformation of the 3D point cloud sequence is calculated by the iterative nearest point method. Based on the adjacent deformation, the trend stability analysis of the 3D point cloud data sequence is performed to obtain the titanium alloy point cloud baseline data.

[0069] The rough-machined titanium alloy part is left to stand still, and a three-dimensional scan of the rough-machined titanium alloy part is performed using a laser scanner to obtain a three-dimensional point cloud data sequence; based on the deformation trend of the feature region in the three-dimensional point cloud data sequence, a stability analysis is performed to obtain the titanium alloy point cloud baseline data and the stable titanium alloy part.

[0070] After rough machining, the rough-machined titanium alloy parts are transferred to a constant-temperature, vibration-free stable environment for at least 12 hours to allow for the full and natural release of internal residual stress. To efficiently monitor deformation trends, 3D point cloud data is collected using a laser scanner only at three time points: the beginning of the resting period, the intermediate point (e.g., 6 hours), and the end of the resting period. All scans are performed in a unified global coordinate system, eliminating the need for rigid body registration between point clouds. Before scanning, several key feature regions must be predefined based on the part's CAD model. (k=1,2,...,K), such as thin-walled, ribbed, cavity sidewalls and other easily deformable parts.

[0071] For any two consecutive scans of the point cloud (e.g.) and ( ), and the regions corresponding to the same key feature need to be extracted from them respectively. The point cloud subset, denoted as and ; then, As a source point cloud to be moved, As a fixed target point cloud, fine registration is performed using the iterative nearest point method: this algorithm iteratively calculates the correspondence between nearest points and minimizes the mean square error between corresponding points, ultimately solving for an optimal rigid transformation matrix. Apply this matrix to the source point cloud Transform to target point cloud The optimal alignment position yields the registered point cloud. .

[0072] Will and Registration is performed using the iterative nearest-point method to obtain the transformation matrix. Thus Precise alignment In the coordinate system, the registered point cloud is obtained. :

[0073] Calculate the point cloud after registration Each point in the middle and its position Calculate the Euclidean distance between the nearest neighbors and the average value:

[0074] in, Let be the adjacent deformation of the k-th region at time t+1 and time t. Let h be the number of point clouds in the k-th region, and h be the index of the point cloud points. For the registered point cloud The h-th point cloud point and the point cloud The Euclidean distance between its nearest neighbor and the nearest neighbor.

[0075] Subsequently, the maximum value of adjacent deformations in all regions is taken as the overall deformation evaluation index for that time period. If the deformation evaluation index calculated at the last time interval is less than or equal to the preset stability threshold, the three-dimensional shape of the part is determined to have reached a stable state. At this point, the three-dimensional point cloud data at the last moment is used as the titanium alloy point cloud baseline data. This part is a titanium alloy stabilizing component. If the deformation evaluation index calculated at the last time interval is greater than the preset stability threshold, the resting time needs to be extended, and the scanning and calculation process described above needs to be repeated until the stability conditions are met.

[0076] Step S5: Temper the titanium alloy stable part to obtain a tempered titanium alloy part, and perform a three-dimensional scan on the tempered titanium alloy part to obtain tempered point cloud data; compare the tempered point cloud data with the titanium alloy point cloud baseline data to calculate the tempering deformation deviation.

[0077] The titanium alloy stable parts were placed in a tempering furnace and subjected to a standardized heat treatment process: a temperature of 600°C for 2 hours, followed by slow cooling to room temperature at a rate of less than 50°C / hour. Immediately after tempering, the workpieces were transferred to a constant temperature measurement chamber. After the temperature in the chamber stabilized uniformly (reaching the standard measurement temperature of (23±0.5)°C), stable tempered titanium alloy parts were obtained. Three-dimensional scanning of the tempered titanium alloy parts yielded tempering point cloud data. By using the Iterative Closest Point (ICP) method, Registration In the coordinate system, we obtain ':(1) Nearest point search: for Each point in, (1) Find the corresponding point with the closest Euclidean distance in the middle and initially establish the point pair association; (2) Transformation matrix calculation: Based on the current matched point pair, calculate an optimal rigid transformation matrix (containing rotation matrix and translation vector) through singular value decomposition (SVD) or least squares method to minimize the mean square error between the current matched point pairs; (3) Transformation application: Apply the calculated rotation matrix and translation vector to the entire Dot clouds, making them point towards (4) Iterative optimization: Repeat steps (1) to (3) to continuously find the nearest point and optimize the transformation matrix until the iteration termination condition is met (usually set to the change in mean square error being less than a preset threshold, such as 1×). (Or reaching the maximum number of iterations, such as 100). The algorithm ultimately outputs an accurate transformation matrix, and... Transform into Spatial alignment This provides a baseline for subsequent dimensional deviation calculations. Extracting... and 'Dimension value of the kth region' and Calculate the dimensional deviation between the tempered point cloud data and the titanium alloy point cloud baseline data:

[0078] in, The size deviation of the k-th region. This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. This refers to the size of the k-th region in the registered tempered point cloud data.

[0079] It should be noted that the dimension value of the k-th region is obtained through two methods: geometric fitting and direct measurement. The geometric fitting method utilizes high-precision algorithms such as the least squares method to automatically fit the point cloud data into a standard geometric entity. For example, for hole-like features, a cylinder fitting function is used to extract the diameter as the dimension value; for wall thickness features, two parallel planes are fitted and their perpendicular distance is calculated as the dimension value. The direct measurement method relies on interactive tools in the point cloud software, allowing manual or automatic selection of point sets within the region to directly calculate statistical values ​​such as average distance or extreme values. These operations are all automated using the powerful point cloud analysis module built into the PMLAB DIC-3D system or MATLAB's point cloud processing toolbox (such as the pcfitcylinder and pcfitplane functions), ensuring the efficiency and accuracy of dimension extraction and providing a reliable data foundation for subsequent deviation calculations.

[0080] Step S6: Before finishing the tempered titanium alloy parts, obtain the finishing cutting parameters, input the finishing cutting parameters into the BP neural network model, and output the finishing deformation compensation amount; by combining the tempering deformation deviation and the finishing deformation compensation amount, calculate the total finishing compensation amount, perform finishing processing on the tempered titanium alloy parts based on the total finishing compensation amount, and use a laser displacement sensor to obtain the deformation value of the tempered titanium alloy parts in real time during finishing. If the deformation value is greater than a preset threshold, an early warning is issued, thereby realizing the deformation detection of titanium alloy parts.

[0081] The tempered titanium alloy parts are precision machined, and the precision machining process parameters (such as spindle speed, feed rate, and depth of cut) are used as input features and input into a pre-trained BP neural network model to output the elastic deformation amount.

[0082] The BP neural network model consists of an input layer, hidden layers, and an output layer. The input layer has 5 nodes, corresponding to 5 finishing cutting parameters: cutting speed, feed per tooth, axial depth of cut, radial depth of cut, and number of tool teeth. These parameters are the main factors affecting milling force and workpiece deformation. The hidden layer's node count is determined using a self-generated node method. This method initializes a network with only one hidden layer node, gradually increases the number of nodes during training, and monitors the change in validation set error. The increase stops when the error decrease rate becomes no longer significant (i.e., the error begins to increase), and the optimal number of nodes is determined. This adaptive method determined the hidden layer to have 8 nodes. The hidden layer neurons use a hyperbolic tangent sigmoid transfer function to capture complex nonlinear features. The output layer has 1 node, and the output is the predicted elastic deformation of the workpiece. The output layer uses a pure linear function (Purelin) as its transfer function.

[0083] It should be noted that the BP neural network model was trained using supervised learning with historical machining data. Orthogonal experimental design was employed to design titanium alloy cutting experiments to explore the intrinsic influence of cutting parameters on machining deformation and to provide a reliable training and validation dataset for the subsequent BP neural network model. The experiments identified five key influencing factors: cutting speed (30, 60, 90 m / min), feed per tooth (0.03, 0.06, 0.09 mm / z), axial depth of cut (0.5, 1.0, 1.5 mm), radial depth of cut (0.3, 0.6, 0.9 mm), and the number of tool teeth (2, 3, 4). Eighteen sets of experiments were arranged based on the standard L18(3^7) orthogonal array. The data acquisition system consists of a Kistler 9265B triaxial piezoelectric force gauge (sampling frequency 1000Hz) and a Keyence LK-H series laser displacement sensor. The former is installed at the bottom of the workpiece to synchronously collect triaxial dynamic cutting forces, while the latter is aligned with the middle of the sidewall of the thin-walled part to monitor the elastic deformation of the part caused by the cutting force in real time, and the force-deformation data is accurately synchronized through timestamps. Each experiment generates a complete sample containing an input parameter vector (cutting speed, feed rate, depth of cut, cutting width, number of tool teeth) and output deformation. The final 18-sample dataset is divided using the hold-out method. To maximize the use of limited data and ensure the model's generalization ability, 15 samples are selected as the training set, and the remaining 3 samples are used as independent test sets for the final evaluation of the model's prediction accuracy.

[0084] The learning rate of the BP neural network model was set to 0.05; the maximum number of training iterations was 10,000; and the training objective was to achieve an error of less than 1×10⁻⁶. During training, based on the training set, the model calculates predicted values ​​through forward propagation and updates network weights and biases through backpropagation of the error to minimize the mean squared error between the predicted values ​​and the actual deformation. The model is considered converged when the training error reaches a preset target or when the error no longer decreases significantly in consecutive iterations. After training, all parameters, including the network structure, weights, and biases, are saved to form a deployable prediction model.

[0085] The finishing cutting parameters are input into the BP neural network model, and the finishing deformation compensation amount is output:

[0086] in, To compensate for deformation during finishing, Input function for neural network model, These are the cutting parameters for finishing.

[0087] Obtain the point cloud of the tempered parts that have been determined to be qualified from step S5. ', and stable titanium alloy point cloud baseline data obtained from step S4. And the original design point cloud data of titanium alloy machined parts. .

[0088] For the k-th key feature region, its total finishing compensation consists of two parts:

[0089] in, This represents the total compensation amount for the k-th region. The size deviation of the k-th region. This represents the theoretical processing allowance for the k-th region. >0, = - , This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. For the original design dimensions of the titanium alloy machined part in the k-th region, when In this case, the total compensation for finishing is not calculated. This is the amount of compensation for deformation during finishing.

[0090] It should be noted that the theoretical processing margin for the k-th region... The measured baseline point cloud was analyzed using the Iterative Closest Point (ICP) algorithm. Point cloud design High-precision registration is performed, and the optimal spatial transformation matrix is ​​solved to align the design point cloud to the coordinate system of the measured point cloud, achieving spatial pose unification between the digital model and the physical entity. After registration, for the k-th specific region (such as a hole, wall, or rib), geometric features are extracted from the two aligned point clouds respectively. Fitting algorithms such as least squares (e.g., cylinder fitting to obtain diameter or plane fitting to obtain thickness) are used to reconstruct ideal geometric elements (such as cylinders or planes) from the discrete point cloud, and their key dimensional parameters and measured values ​​are output respectively. Design value Finally, the machining allowance is obtained by subtracting the two extracted scalar dimension values. This value quantifies the material allowance of the current part relative to the design target. A positive number indicates that material needs to be removed, while a zero or negative number indicates that the part has exceeded the tolerance and should be scrapped without compensation.

[0091] The calculated total compensation amount is then applied to the CNC programming to eliminate tempering deformation. Specifically, in the CAM (Computer-Aided Manufacturing) software, the "machining allowance" parameter for the finishing operation is located, and... The value is assigned to this parameter. The value and direction of the value are used to offset the finishing toolpath at equal intervals: if A positive value indicates excess material in the part, and the machining allowance is reduced accordingly (e.g., if the original allowance is 0.1mm and the compensation is 0.05mm, then the new allowance is set to 0.05mm), causing the toolpath to shift inwards towards the material. A negative value indicates material loss, usually indicating that the part is scrapped or requires special treatment such as welding repair to ensure machining feasibility. In actual operation, simply specify the final machined surface in the CAM system and directly set the machining allowance parameter to... The software can automatically generate optimized toolpaths. Subsequently, the generated G-code is executed to complete the finishing process, and a final inspection is performed using equipment such as a coordinate measuring machine (CMM) to verify that all critical dimensions fall within the design tolerance zone, thus forming a closed-loop quality control from measurement to compensation.

[0092] During the finishing process, a high-precision laser displacement sensor is used to monitor the actual relative position changes between the tool path trajectory and the critical weak points of the workpiece in real time. First, the theoretical tool path after applying the total compensation is used as the zero deviation baseline. The sensor measures the instantaneous deviation between the workpiece surface and the theoretical path in real time to capture abnormal deformation caused by unpredictable factors such as tool wear, residual stress release, or chatter. This instantaneous deviation value is compared with a preset strict online tolerance threshold. If the threshold is exceeded, an early warning is immediately triggered and the process parameters can be automatically adjusted to achieve real-time closed-loop control and prevent irreversible out-of-tolerance deformation.

[0093] This invention proposes an online deformation detection method for high-performance titanium alloy parts during machining. Through a series of interconnected steps, including annealing pretreatment, real-time monitoring of rough machining, static and non-contact scanning detection, tempering treatment and online verification, and closed-loop control of finishing, a process-detection-verification closed-loop system covering the entire process is constructed.

[0094] The flutter signal data is input into a random forest model based on intrinsic mode decomposition (EMD) to output flutter state values, thus realizing a leap from passive detection to proactive early warning in the processing. The EMD method can adaptively process nonlinear and non-stationary vibration signals in titanium alloy processing, effectively extracting weak features in the flutter incubation stage; combined with the powerful classification ability of the random forest model, it can accurately identify the critical state before flutter occurs violently (incubation stage), gaining valuable time for proactive intervention.

[0095] The calculation of adjacent deformations in a 3D point cloud sequence using the iterative nearest point method is crucial for accurately quantifying micro-deformation. This method efficiently and precisely registers 3D point clouds acquired at different time points, calculating sub-millimeter-level micro-deformations. This allows for the accurate capture of the slow, subtle deformation trends of parts during resting due to internal stress release. It enables the objective determination of when a part transitions from an active stress release phase to a stable equilibrium state (i.e., deformation evaluation index ≤ stability threshold), thus scientifically determining the required resting time. This avoids the problems of insufficient time or excessive waiting caused by empirical estimation, providing a stable and reliable geometric benchmark for subsequent tempering processes.

[0096] By inputting the finishing cutting parameters into a BP neural network model and outputting the finishing deformation compensation amount, intelligent closed-loop control of machining error prediction and compensation is achieved. The BP neural network, through learning historical machining data, can establish a complex nonlinear mapping relationship between cutting parameters and deformation, accurately predicting the elastic deformation that will occur during finishing under specific parameters. When generating the CNC toolpath, it pre-assigns the offset to the machining allowance parameters in the CAM system, allowing the toolpath to proactively yield the expected deformation space. After machining is completed and deformation is recovered, the part dimensions will fall precisely within the tolerance range, transforming post-processing correction into pre-processing avoidance, greatly improving the first-pass yield of machining accuracy.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online deformation detection during the machining process of high-performance titanium alloy parts, characterized in that: Includes the following steps: Step S1: Place the titanium alloy blank in an electric furnace for annealing and slow cooling to obtain a low-stress titanium alloy blank. Step S2: Roughly machine the titanium alloy low-stress blank, and simultaneously collect rough machining temperature data, rough machining displacement data, and chatter signal data; input the chatter signal data into a random forest model based on intrinsic mode decomposition, and output the chatter state value; Step S3: Determine the state of the roughing process by using chatter state value, roughing temperature data, and roughing displacement data. Adjust the process parameters in the roughing process according to the state determination results and continue processing to finally obtain the rough-machined titanium alloy part. Step S4: Let the rough-machined titanium alloy part stand still until a preset period is reached to obtain a stable titanium alloy part. During the preset period, the rough-machined titanium alloy part is scanned in three dimensions by a laser scanner to obtain a three-dimensional point cloud data sequence. The adjacent deformation of the three-dimensional point cloud sequence is calculated by the iterative nearest point method. Based on the adjacent deformation, the trend stability analysis of the three-dimensional point cloud data sequence is performed to obtain the titanium alloy point cloud baseline data. Step S5: Temper the titanium alloy stable part to obtain a tempered titanium alloy part and perform a three-dimensional scan on the tempered titanium alloy part to obtain tempered point cloud data; compare the tempered point cloud data with the titanium alloy point cloud baseline data to calculate the tempering deformation deviation. Step S6: Before finishing the tempered titanium alloy parts, obtain the finishing cutting parameters, input the finishing cutting parameters into the BP neural network model, and output the finishing deformation compensation amount; by combining the tempering deformation deviation and the finishing deformation compensation amount, calculate the total finishing compensation amount, perform finishing processing on the tempered titanium alloy parts based on the total finishing compensation amount, and use a laser displacement sensor to obtain the deformation value of the tempered titanium alloy parts in real time during finishing. If the deformation value is greater than a preset threshold, an early warning is issued, thereby realizing the deformation detection of titanium alloy parts.

2. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 1, characterized in that: The process of placing the titanium alloy blank in an electric furnace for annealing and slow cooling to obtain a low-stress titanium alloy blank includes the following specific steps: The titanium alloy billet is smoothly placed into the center of the working area of ​​the box furnace, ensuring that it is not in contact with the heating elements and that there is a uniform gap around it to ensure the circulation of hot air. Then, the furnace door is closed tightly, and the furnace temperature is raised to 750°C at a rate not exceeding 150°C / hour, and the holding time is started. This temperature is maintained for 2 hours to allow the internal structure of the titanium alloy billet to fully recrystallize and grow grains, thereby eliminating the original residual stress caused by previous forging or rolling. After the holding time is completed, the heating power is cut off, and the titanium alloy billet is allowed to cool naturally and slowly in the sealed furnace chamber, with the cooling rate controlled below 50°C / hour, so as to avoid the generation of new thermal stress due to rapid cooling. When the furnace temperature drops below 200°C, the furnace door can be opened to remove the titanium alloy billet, thereby obtaining a low-stress titanium alloy billet.

3. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 2, characterized in that: The process of inputting the flutter signal data into a random forest model based on intrinsic mode decomposition and outputting the flutter state includes the following steps: The intrinsic mode decomposition formula is: ； in, For flutter signal data, Let j be the j-th eigenmode function. For low-frequency residuals, t is the time index, j is the order index, and n is the total order of the intrinsic mode functions, n=3; By merging the first three essential mode functions, the reconstructed effective signal is obtained: ； in, To reconstruct a valid signal, For the first-order eigenmode function, For the second-order eigenmode function, It is the 3rd order eigenmode function; The power spectral energy S is obtained by performing a Fourier transform on the reconstructed effective signal. ): ； in, for Power spectral energy at that location To reconstruct a valid signal, To reconstruct the i-th frequency component of the valid signal; Based on the power spectral energy, calculate the power spectral entropy: ； in, For power spectral entropy, Let m be the probability density of the i-th frequency component, m be the total number of frequency components in the Fast Fourier Transform, and i be the index of the frequency component. Calculate the reconstructed valid signal The signal kurtosis peak is used to form a flutter feature vector by combining the power spectral entropy E and the signal kurtosis peak in each sample group; The flutter feature vectors collected in real time are input into the random forest model, and the flutter state is output: ； in, The classification results are from the random forest model, where c is the class label and Q is the number of decision trees. Let q be the classification result of the q-th decision tree, where q is the index of the decision tree. This represents the classification result of the q-th decision tree. When equal to category c, The value is 1 if it is not 0 otherwise. For optimization operations, it means finding the category label c that maximizes the total number of votes in the voting mechanism of the random forest model.

4. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 3, characterized in that: The process of determining the state of the roughing process using chatter status, roughing temperature data, and roughing displacement data includes the following steps: The comprehensive risk index R is calculated using chatter state values, roughing temperature data, and roughing displacement data. ； Where R is the comprehensive risk index. Temperature weighting, WD represents the temperature threshold, where WD is the roughing temperature. Here, WY represents the displacement weight, and WY represents the roughing displacement. Displacement threshold For flutter state weights, This represents the risk value for flutter conditions. + + =1.

5. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 4, characterized in that: The process of resting the roughly machined titanium alloy part includes the following steps: The rough-machined blank enters the stage of natural stress release and data monitoring: First, the rough-machined workpiece is removed from the machine tool and transferred to a stable environment with constant temperature and no vibration for static placement for more than 12 hours, so that the internal residual stress can be fully and naturally released.

6. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 5, characterized in that: The calculation of adjacent deformations of a 3D point cloud sequence using the iterative nearest point method includes the following specific steps: Will and Registration is performed using the iterative nearest-point method to obtain the transformation matrix. Thus Precise alignment In the coordinate system, the registered point cloud is obtained. : ； Calculate the point cloud after registration Each point in the middle and its position Calculate the Euclidean distance between the nearest neighbors and the average value: ； in, Let be the adjacent deformation of the k-th region at time t+1 and time t. Let h be the number of point clouds in the k-th region, and h be the index of the point cloud points. For the registered point cloud The h-th point cloud point and the point cloud The Euclidean distance between its nearest neighbor and the nearest neighbor.

7. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 6, characterized in that: The process of performing trend stability analysis on the three-dimensional point cloud data sequence based on adjacent deformations to obtain titanium alloy point cloud baseline data includes the following specific steps: The maximum value of adjacent deformations in all regions of the rough-machined titanium alloy part is taken as the overall deformation evaluation index for that time period. If the deformation evaluation index calculated at the last time interval is less than or equal to the preset stability threshold, the three-dimensional shape of the part is determined to have reached a stable state, and the three-dimensional point cloud data at the last moment is used as the titanium alloy point cloud baseline data. .

8. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 7, characterized in that: The process of tempering the titanium alloy stable part to obtain a tempered titanium alloy part and then performing a three-dimensional scan on the tempered titanium alloy part to obtain tempered point cloud data includes the following specific steps: The titanium alloy stable parts were placed in a tempering furnace and subjected to a standardized heat treatment process of 600°C for 2 hours, followed by slow cooling to room temperature at a rate of less than 50°C / hour. After tempering, the titanium alloy stable parts were immediately transferred to a constant temperature measurement chamber. After the temperature in the constant temperature measurement chamber was kept uniform and stable, the tempered titanium alloy parts were obtained. The tempered titanium alloy parts were then subjected to three-dimensional scanning to obtain tempering point cloud data.

9. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 8, characterized in that: The step of comparing the tempered point cloud data with the titanium alloy point cloud baseline data to calculate the tempering deformation deviation includes the following specific steps: Calculate the dimensional deviation between the tempered point cloud data and the titanium alloy point cloud baseline data: ； in, The size deviation of the k-th region. This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. This refers to the size of the k-th region in the registered tempered point cloud data.

10. The online deformation detection method for high-performance titanium alloy parts during machining according to claim 9, characterized in that: The calculation of the total compensation for finishing by combining the tempering deformation deviation and the finishing deformation compensation includes the following specific steps: The finishing cutting parameters are input into the BP neural network model, and the finishing deformation compensation amount is output: ； in, To compensate for deformation during finishing, Input function for neural network model, For finishing cutting parameters, including cutting speed, feed per tooth, axial depth of cut, radial depth of cut and number of tool teeth; By combining the tempering deformation deviation and the finishing deformation compensation amount, the total finishing compensation amount is calculated as follows: ； in, This represents the total compensation amount for the k-th region. The size deviation of the k-th region. This represents the theoretical processing allowance for the k-th region. >0, = - , This represents the dimension value of the k-th region in the titanium alloy point cloud baseline data. For the original design dimensions of the titanium alloy machined part in the k-th region, when In this case, the total compensation for finishing is not calculated. This is the amount of compensation for deformation during finishing.

Images