Adaptively adjusted spline shaft broaching process method

Abstract

The application discloses a self-adaptive adjusting spline shaft broaching process method and relates to the field of data processing.The spline shaft is clamped and configured with a broach set, and multiple source data are synchronously collected in the broaching process, time-frequency domain, temperature gradient, force fluctuation and appearance characteristics are extracted, a characteristic vector is constructed to input a process evaluation model, and adjustment instructions are output in real time to optimize the broaching speed and broaching force parameters.After processing, white light interference detection is carried out, tooth profile precision and roughness data are obtained, and the whole process monitoring data, adjustment records and quality results are stored in association, which is used for continuously optimizing the evaluation model weight.The application has the advantages that the real-time parameter self-adaptive adjustment is realized through the multi-source sensing collection of processing data and model fusion evaluation, the processing abnormality is effectively avoided, the tooth surface processing precision and quality stability are ensured, the model iteration is optimized, and the processing efficiency and economy are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to an adaptive adjustment method for broaching spline shafts. Background Technology

[0002] As a key mechanical structure for transmitting torque, splines are subject to limitations such as low machining efficiency and difficulty in ensuring consistent precision when machined using traditional methods like milling and gear shaping. With the rapid development of industries such as automobiles and construction machinery, the demand for high-volume, high-strength spline shafts has surged, making broaching a standout process due to its unique advantages.

[0003] Current broaching methods generally suffer from core disadvantages such as low intelligence, poor processing adaptability, and insufficient quality control. Most methods use fixed process parameters, failing to dynamically adjust based on spline shaft tooth profile parameters, material properties, and real-time processing conditions. They lack synchronous monitoring and data acquisition from multi-source sensors, making it difficult to detect abnormal vibrations, excessive temperatures, and excessive force fluctuations during broaching. This easily leads to chatter, tooth surface defects, and excessive tool wear, resulting in unstable tooth profile accuracy, surface roughness, and a high scrap rate. Furthermore, existing methods lack integrated evaluation and closed-loop optimization mechanisms for process conditions. Post-processing data is not effectively correlated and stored, preventing targeted optimization of process parameters. Adapting to spline shafts of different specifications and materials requires repeated adjustments, resulting in low efficiency. Over-reliance on human experience leads to high costs for manual intervention, making it difficult to balance processing accuracy, efficiency, and economy, and ultimately failing to meet the demands of high-precision, large-scale spline shaft processing. Summary of the Invention

[0004] To improve existing methods, an adaptive adjustment method for spline shaft broaching is provided. This method achieves real-time adaptive adjustment of parameters by acquiring machining data from multiple sources and evaluating the model fusion. This effectively avoids machining anomalies, ensures the machining accuracy and quality stability of the tooth surface, and improves machining efficiency and economy.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An adaptive adjustment method for broaching splined shafts includes:

[0007] The spline shaft workpiece to be processed is clamped on the broaching machine table and tightened. A combination broaching set is selected according to the tooth profile parameters and material properties of the spline shaft. The initial process parameters of broaching speed, broaching force and coolant flow rate are set.

[0008] Embedded vibration and temperature sensors are placed at the broach teeth. When broaching is started, real-time data is collected synchronously during the continuous feed of the broach.

[0009] The collected vibration signals are transformed in the time and frequency domain to extract the energy amplitude and resonance peak shift characteristics of specific frequency bands. Gradient analysis and hot spot region identification are performed on the temperature field data. The broaching force fluctuation data are smoothed and decomposed into trends. Texture analysis and contour deviation calculation are performed on the tooth surface morphology image to construct the feature vector of the current process state.

[0010] The acquired feature vectors are input into the pre-trained process evaluation model. Based on the fusion evaluation results of vibration energy level, temperature gradient, force fluctuation amplitude and tooth surface quality index, the model outputs adjustment instructions for adjusting at least one of broaching speed, broaching force and coolant flow rate. After the adjustment is executed, the model collects and evaluates data for the next monitoring cycle until all tooth grooves of the spline shaft are broached.

[0011] After machining, white light interferometry is performed on the spline tooth surface to obtain data on tooth profile error, surface roughness, and tooth direction accuracy. Multi-source monitoring data, adjustment event records, and quality data from the entire machining process are linked and stored, and the parameter weights of the process evaluation model are optimized.

[0012] Preferably, the synchronous acquisition of real-time data specifically includes:

[0013] Vibration spectrum signals of the area where the broach and workpiece mesh are acquired using a vibration sensor;

[0014] Temperature field distribution data of the tool-chip contact area and the workpiece substrate are collected using a temperature sensor.

[0015] Real-time broaching force and torque fluctuation data are collected through the broaching machine spindle servo system;

[0016] The machine vision unit continuously acquires images of the machined tooth surface morphology along the broaching direction.

[0017] Preferably, the step of performing time-frequency domain transformation on the acquired vibration signal to extract the energy amplitude and resonance peak shift characteristics of a specific frequency band specifically includes:

[0018] An adaptive filter is used to filter out the vibration of the broaching machine foundation and environmental noise from the continuous time-domain signal collected by the vibration sensor. The filtered signal is then divided into segments at equal intervals, with each segment corresponding to a complete broaching tooth groove machining cycle.

[0019] A continuous wavelet transform is performed on each signal segment, and Morlet wavelets are used as basis functions. A two-dimensional time spectrum is obtained by scaling and time shifting.

[0020] Based on the wavelet time spectrum, the frequency domain is divided into the knife-tooth cutting-out characteristic frequency band, the material shear-dominated frequency band, and the flutter potential frequency band;

[0021] By integrating the squared amplitude of the wavelet coefficients within each sub-band along the time axis on the time spectrum, the energy value of each band within each time window is calculated.

[0022] The maximum value, mean value, and fluctuation variance of energy in each frequency band within each processing cycle are statistically analyzed and used as characteristic parameters of the vibration energy distribution state.

[0023] On the power spectral density curve of each time window, identify the resonance peak caused by the inherent characteristics of the process system and record its center frequency value;

[0024] The trajectory of the center frequency of the same resonance peak as the broach feed position changes during the complete machining cycle is tracked, and the maximum offset of the center frequency of the resonance peak relative to its initial position and the average rate of frequency drift are calculated as characteristic quantities of stiffness change or abnormal load.

[0025] Preferably, the step of performing continuous wavelet transform on each signal segment, using Morlet wavelets as basis functions, and obtaining a two-dimensional time-spectrum graph through scaling and time shifting specifically includes:

[0026] The complex Morlet wavelet is selected as the mother wavelet, and the corresponding scale parameter range is calculated based on the sampling frequency and the target analysis frequency range.

[0027] Multiple scale values ​​are generated by discretizing the scale in a logarithmic or linear manner within the scale range, and a corresponding Morlet wavelet kernel function is generated for each scale:

[0028] For each signal segment, convolution calculation is performed in the time domain or frequency domain. In the time domain, the wavelet kernel function of each scale is convolved with the signal segment to calculate the coefficients. In the frequency domain, the signal segment and the wavelet kernel function are subjected to FFT respectively, and then multiplied in the frequency domain and then subjected to IFFT.

[0029] The coefficients at all scales are arranged in scale order to form a two-dimensional complex matrix. The wavelet scale map is obtained by calculating the modulus matrix of the coefficients.

[0030] The scale coordinates are converted into actual frequency coordinates to obtain the final two-dimensional time-frequency spectrum.

[0031] Preferably, the gradient analysis and hotspot region identification of the temperature field data specifically includes:

[0032] The discrete temperature values ​​collected by the temperature sensor are used to reconstruct a complete and continuous two-dimensional temperature field distribution map through spatial interpolation algorithm.

[0033] The reconstructed temperature field is mapped onto a dynamic mesh, with mesh cells aligned with the tool-workpiece contact area. Each mesh cell corresponds to a temperature node with a defined physical location.

[0034] For gridded temperature field data, the radial and axial temperature gradient components of each node are calculated in the spatial domain to generate the corresponding temperature gradient vector field. By setting a gradient magnitude threshold, regions with gradient magnitudes higher than the average level are filtered out.

[0035] Potential hotspot areas are identified based on both absolute temperature threshold and local temperature difference. By tracking the dynamic evolution trajectory of hotspots, they are classified into instantaneous high-temperature zones or persistent heat accumulation zones.

[0036] Preferably, the smoothing and trend decomposition of the broaching force fluctuation data specifically includes:

[0037] Based on the original broaching force data, instantaneous spike outliers caused by signal interference are removed, and data compensation is performed on the removed points using linear interpolation.

[0038] A sliding weighted average filter is used to perform initial smoothing of the data, with the weighting coefficients decreasing symmetrically along the center of the window.

[0039] The smoothed data were analyzed using a local weighted regression method to separate the trend components that characterize the slow changes in tool wear and material hardness.

[0040] Subtract the extracted trend term from the original smoothed data to obtain the fluctuation term component. Perform a fast Fourier transform on the fluctuation term data to extract its main frequency component and corresponding amplitude. Compare the main frequency component with the passing frequency of the blade or its harmonics and output the comparison result.

[0041] Calculate the standard deviation, kurtosis, and skewness of the fluctuation term data to quantify the intensity and distribution characteristics of the fluctuation, and extract the slope of the trend term, the amplitude of the dominant frequency of the fluctuation, and the characteristic parameters of the intensity of random fluctuations.

[0042] Preferably, the texture analysis and contour deviation calculation of the tooth surface morphology image specifically includes:

[0043] The side surface contour of a single spline tooth is identified by an edge detection algorithm, and the tooth surface area image to be analyzed is located and cropped.

[0044] For the segmented tooth surface region image, calculate its gray-level co-occurrence matrix, extract texture statistics, and calculate the gradient histogram of the image in multiple directions;

[0045] The extracted texture feature vectors are input into the trained classifier to identify and label specific defect regions in the image, and output the distribution density and area ratio of various defects.

[0046] The actual edge contour of the tooth surface region is spatially aligned with the theoretical ideal contour generated based on the spline design parameters. Along the normal direction of the theoretical contour, the distance deviation between the actual contour points and the theoretical contour points is densely sampled and calculated.

[0047] The maximum positive deviation, maximum negative deviation, and standard deviation of all sampling points are statistically analyzed and used as quantitative indicators to evaluate tooth profile error and contour accuracy.

[0048] Preferably, the step of inputting the extracted texture feature vector into the trained classifier to identify and label specific defect regions in the image, and outputting the distribution density and area ratio of various defects specifically includes:

[0049] The texture feature vector is input into a pre-trained fully convolutional network, which decodes and fuses the input features through fully connected layers, upsamples and reconstructs a multi-channel classification heatmap corresponding to the spatial size of the original input image;

[0050] Each channel of the classification heatmap corresponds to a preset defect category. By comparing the values ​​between channels pixel by pixel, the defect category of each pixel is determined, and a defect segmentation map is generated.

[0051] Morphological operations are performed on the marked defective pixels to connect adjacent defective pixels of the same type, forming a continuous defective region;

[0052] The number of independent connected regions contained in each defect category is counted and divided by the total area of ​​the effective tooth surface analysis area to obtain the number of defects per unit area.

[0053] The total number of pixels covered by each defect category is counted and divided by the total number of pixels in the effective tooth surface analysis area to calculate the defect area ratio.

[0054] Preferably, the step of inputting the acquired feature vector into a pre-trained process evaluation model, and evaluating the results based on the fusion of vibration energy level, temperature gradient, force fluctuation amplitude, and tooth surface quality indicators, specifically includes:

[0055] The acquired feature vectors are input into a pre-trained deep neural network model, which performs nonlinear transformations and feature fusion through multiple hidden layers to learn complex correlation patterns between different features.

[0056] At the end of the network, the deep neural network model outputs a multi-dimensional state vector, where each dimension of the multi-dimensional state vector corresponds to a quantified index of the process state after fusion evaluation.

[0057] The state vector output by the model is compared with the preset process thresholds at each level, and a composite judgment is made based on the fused state.

[0058] Based on the specific process state category determined, the corresponding combination of adjustment instructions is mapped from the preset strategy library.

[0059] Compared with the prior art, the advantages of the present invention are:

[0060] This method achieves intelligent closed-loop control of the broaching process, effectively overcoming the limitations of traditional fixed-parameter broaching. It synchronously collects vibration, temperature, broaching force, and tooth surface morphology data throughout the broaching process using multi-source sensors. Combined with professional data processing technology, it accurately extracts process state characteristics and relies on a pre-trained evaluation model to achieve multi-index fusion evaluation and real-time parameter adaptive adjustment. This allows for timely avoidance of problems such as chatter, overheating, and abnormal force fluctuations, ensuring the machining accuracy and quality stability of the tooth surface. After machining, white light interferometry detection and full-process data association storage enable accurate verification of finished product quality and optimize model parameter weights, improving the adaptability to subsequent machining. This method balances versatility and accuracy, adapting to spline shaft machining with different tooth profile parameters and material properties. It significantly reduces tool wear and scrap rate, while minimizing manual intervention and improving machining efficiency, achieving a synergistic improvement in machining quality, efficiency, and economy. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the adaptive adjustment spline shaft broaching process proposed in this invention. Detailed Implementation

[0062] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0063] See Figure 1 As shown, the adaptive adjustment spline shaft broaching process includes:

[0064] S1: Mount the spline shaft workpiece to be processed onto the broaching machine table and clamp it. Select a combination broaching tool set according to the tooth profile parameters and material properties of the spline shaft, and set the initial process parameters of broaching speed, broaching force and coolant flow rate.

[0065] S2: An embedded vibration sensor and temperature sensor are arranged at the broach teeth position to start the broaching process and collect real-time data synchronously during the continuous feed of the broach.

[0066] S3: Perform time-frequency domain transformation on the collected vibration signal, extract the energy amplitude and resonance peak shift characteristics of a specific frequency band, perform gradient analysis and hot spot region identification on the temperature field data, perform smoothing and trend decomposition on the broaching force fluctuation data, perform texture analysis and contour deviation calculation on the tooth surface morphology image, and construct the feature vector of the current process state.

[0067] S4: Input the acquired feature vector into the pre-trained process evaluation model. Based on the fusion evaluation results of vibration energy level, temperature gradient, force fluctuation amplitude and tooth surface quality index, output the adjustment command for adjusting at least one of broaching speed, broaching force and coolant flow rate. After the adjustment is executed, data collection and evaluation are carried out in the next monitoring cycle until all tooth grooves of the spline shaft are broached.

[0068] S5: After machining, perform white light interference detection on the spline tooth surface to obtain data on tooth profile error, surface roughness and tooth direction accuracy. Link and store multi-source monitoring data, adjustment event records and quality data from the entire machining process, and optimize the parameter weights of the process evaluation model.

[0069] In S1, broaches are selected from the standard broach module library based on the spline shaft's tooth profile parameters and material properties. The roughing module selects broaches with large chip space, a large rake angle, and a large tooth rise for efficient removal of most of the excess material. The semi-finishing and finishing modules select broaches with decreasing tooth rise, a small rake angle, and finishing edges or specific tooth profile corrections to gradually approach the final dimensions and tooth profile accuracy, achieving good surface quality. The calibration module selects broaches with zero tooth rise, an appropriate clearance angle and cutting edge width, and high-precision ground teeth for final dimension calibration, accuracy stabilization, and surface smoothing.

[0070] Broaching speed is set based on workpiece material, tool material, and expected machining quality, referring to empirical formulas and incorporating database recommendations. A commonly used speed estimation formula is:

[0071]

[0072] in: For broaching speed, The basic cutting speed constant depends on the tool-workpiece material combination; This is a correction factor for the workpiece material, which is related to the material's hardness and strength. This is the correction factor for the tool coating; This is the correction factor for tool wear condition.

[0073] The estimated total broaching force serves as a benchmark for verifying the load-bearing capacity of the machine tool and fixture, as well as for subsequent closed-loop control. The total broaching force can be approximately estimated as follows:

[0074]

[0075] in, For the total cutting force, It is the cutting force per unit area, which is related to the workpiece material and tool geometry parameters, and can be obtained by referring to tables or through experiments. The cutting cross-sectional area of ​​a single cutting tooth is equal to the tooth rise multiplied by the cutting width. This refers to the effective number of teeth simultaneously participating in the cutting process. In the control system, this number is calculated... As a rated force value, the upper and lower threshold values ​​for force monitoring are set accordingly.

[0076] In S3, an adaptive filter based on the least mean square algorithm is used, and its weight update formula is as follows:

[0077]

[0078] in: Let n be the filter weight vector at time n. The convergence step size factor, Let x(n) be the error between the desired signal and the filter output, and let x(n) be the reference input vector.

[0079] This filter adaptively subtracts components strongly correlated with the broaching machine foundation vibration, as well as fixed environmental noise, from the original vibration signal. Based on the broach feed position fed back in real time by the CNC system, the continuous vibration signal is cut into a series of equal-length signal segments. The start and end points of each segment strictly correspond to the broaching stroke of a complete spline tooth groove.

[0080] For each signal segment, a continuous wavelet transform is performed, expanding it from a one-dimensional time-domain signal into a two-dimensional time-frequency joint distribution image. The Morlet wavelet is chosen as the analysis basis function. The Morlet wavelet is a waveform modulated by a complex exponential function and a Gaussian window, exhibiting good localization properties in both the time and frequency domains. By changing the wavelet's scaling parameter 'a' (control frequency) and translation parameter 'b' (control time), the inner product of the signal and each member of the wavelet family is calculated. The core calculation formula is:

[0081]

[0082] in, These are wavelet coefficients, which are complex numbers and contain amplitude and phase information at that time scale point; It is the mother wavelet. Indicates its complex conjugate; It is a scaling factor, inversely proportional to frequency; It is a translation factor, representing the time position.

[0083] Calculate the squared modulus of the wavelet coefficients at each point (a, b) and use it as the energy density. Convert scale a to the actual frequency through the center frequency relationship, and finally obtain a two-dimensional time-frequency spectrum with time as the horizontal axis, frequency as the vertical axis, and color intensity representing energy strength.

[0084] Based on the dynamic characteristics of the broaching process, the frequency axis is divided into three sub-bands with clear physical meaning on the generated time-frequency spectrum:

[0085] Cutting tooth entry and exit characteristic frequency band: High-frequency region, mainly containing the transient impact energy generated by the sudden contact and separation between each cutting tooth rake angle and the workpiece material. Its frequency range usually covers the higher harmonics of the cutting tooth passing frequency.

[0086] The dominant frequency band for material shearing is the mid-frequency region. This mainly includes the quasi-continuous vibration energy excited by the continuous friction between the tool's rake face and the chip, the flank face and the machined surface, and the shear slip deformation of the material. Its frequency range typically revolves around the dominant frequency of the cutting teeth's passage frequency and its second harmonic.

[0087] Chatter potential frequency band: low-frequency region. It mainly includes the low-order modal frequencies of weak links in the machine tool-tool-workpiece process system. Abnormal growth in this frequency band is a major sign that chatter is about to occur or has already occurred.

[0088] For each defined characteristic frequency band, at each time slice, the wavelet coefficient energies corresponding to all frequency points within that frequency band are summed to obtain the instantaneous energy value of that frequency band at that moment. By iterating through all time points, three curves describing the energy variation of each frequency band over time are obtained. For the energy time series curves within a complete processing cycle, the maximum value, mean, and variance are calculated.

[0089] For the same signal segment, the Welch average periodogram method is used to calculate its power spectral density. On the PSD curve, local peaks significantly higher than the background noise are identified; the frequencies corresponding to these peaks are the resonance peaks of the process system. The identified resonance peaks are matched and tracked across multiple consecutive time windows throughout the entire processing. Utilizing the frequency stability, amplitude, and bandwidth information of the resonance peaks, nearest neighbor data correlation ensures that resonance peaks of the same physical mode are tracked in different time windows.

[0090] For each successfully tracked resonance peak, analyze the trajectory of its center frequency throughout the observation period. Find the maximum absolute deviation of the center frequency of that resonance peak from its initial value. The calculation formula can be simplified to:

[0091]

[0092] in, For the maximum absolute deviation, It is the center frequency of the resonance peak at time t. This is the reference frequency. This value directly reflects the maximum change in the equivalent stiffness of the system during machining, which may be caused by tool wear, thermal deformation, or sudden load changes.

[0093] Calculate the average slope of the resonance peak frequency change trajectory, which is the average frequency change per unit time. This reflects the rate of change of system characteristics; rapid drift may indicate the rapid development of abnormal conditions.

[0094] It receives real-time temperature readings from multiple embedded temperature sensors. Each data point contains its precise measurement and known three-dimensional spatial coordinates through preset calibration. Since the sensors are discrete points, spatial interpolation is performed using Kriging interpolation. (Point to be estimated) temperature We obtain the result by weighted summation of the known points:

[0095]

[0096] Among them, weight This is determined by solving a system of equations based on the variogram and spatial structure, which minimizes the variance of the estimation error and satisfies the unbiasedness condition ∑. = 1. This generates a continuous temperature distribution map covering the tool-workpiece contact area and surrounding region.

[0097] The reconstructed temperature field is mapped onto a dynamic, non-uniform computational grid. The grid is not uniformly divided, but rather refined in areas with high heat flux density or high risk of thermal damage, such as near the cutting edge, chip flute exit, and machined surfaces. Each grid node becomes a temperature data point with clearly defined physical coordinates.

[0098] On the generated structured mesh, the temperature gradient at each node is calculated. For nodes within the mesh, the central difference method is used to calculate their gradient components in the radial (x-direction) and axial (y-direction) directions:

[0099] Radial gradient:

[0100] Axial gradient:

[0101] The temperature gradient vector at that node is Its gradient magnitude is A larger gradient magnitude indicates a more drastic spatial temperature change at that point, typically corresponding to a heat source or heat dissipation boundary. High gradient regions are identified by calculating the average and standard deviation of the gradient magnitude across the entire temperature field. A threshold is set, and all mesh nodes that meet this threshold are marked as high temperature gradient regions.

[0102] Based on the identified high-gradient regions, a more stringent dual condition is applied to precisely locate hotspots: The absolute temperature threshold criterion sets a global safety threshold. This threshold is determined based on the physical properties of the workpiece material or process experience. Any node whose temperature exceeds the global safety threshold is immediately marked as a region requiring special attention. The local relative temperature difference criterion avoids misjudgments due to overall temperature increases and simultaneously identifies local anomalies. A circular neighborhood of radius R is taken centered on the candidate node, and the average temperature of all non-directly cutting nodes within this neighborhood is calculated. If the temperature of a candidate node is significantly higher than its surrounding environment, it is considered to have localized abnormal heat concentration. Continuous mesh regions that simultaneously satisfy both of the above conditions are merged, marked as a potential hotspot region, and its contour, area, and highest temperature within the region are recorded.

[0103] If a hot spot has a short lifespan, and its temperature change rate and area expansion rate rapidly turn from positive to negative after its appearance, it is identified as an instantaneous high-heat zone. This type usually corresponds to an instantaneous cutting heat source for a single cutting tooth, which is a normal process phenomenon with low risk. If a hot spot has a long lifespan, and its temperature change rate and area expansion rate remain positive or have a significantly higher mean than the baseline within the observation window, it is identified as a persistent heat accumulation zone. This indicates that the heat generation rate in this region is consistently higher than the heat dissipation rate, which is a key risk source leading to tool thermal wear, workpiece thermal damage, or dimensional thermal deformation, and adaptive cooling adjustment must be triggered.

[0104] In trend term extraction based on local weighted regression, the trend term reflects the slow drift of the broaching force caused by factors such as progressive tool wear and gradual changes in workpiece material hardness. Local weighted regression is a nonparametric fitting method that uses a low-order polynomial to perform weighted least squares fitting within a local neighborhood of each data point. The value of the fitted polynomial at that point is the estimated value of the trend term.

[0105] For local fitting of a target point t0 in a time series, a neighborhood window is selected that includes h data points before and after it. A weight is assigned to each point t within the window; the weighting function is typically a cubic function.

[0106]

[0107] in This is the distance from the farthest point within the window to t0. Then, using these weights, a straight line or parabola is fitted using weighted least squares. The value of the fitted curve at t0 is the estimate of the trend term.

[0108] Repeat the above local fitting process for each point in the sequence, and connect the trend estimates of all points to form a complete trend term sequence. This sequence smoothly captures the macroscopic trend of the tensile force.

[0109] Subtracting the extracted trend term from the smoothed force data yields a pure fluctuation term sequence. This sequence eliminates slow-changing trends and concentrates force fluctuation information caused by irregularities in chip formation, dynamic interaction between the tool and workpiece, and potential chatter.

[0110] Perform a Fast Fourier Transform on the wave term sequence to calculate its power spectral density. Identify significant peaks in the power spectrum and record their frequencies and corresponding amplitudes. Crucially, determine whether these dominant frequencies are equal to or close to the theoretically calculated passing frequency of the cutting teeth. = (Number of broach teeth × Feed rate) / (Workpiece circumference), or an integer multiple thereof. A match indicates that the fluctuation mainly originates from the periodicity of cutting; an independent peak that does not match may suggest regenerative chatter or other abnormal vibrations.

[0111] The classifier-based defect region identification and quantization employs a sliding window method based on a deep learning-based semantic segmentation model. The model outputs a probability map of the same size as the input image, where each pixel contains the probability of belonging to a particular defect class. By setting a probability threshold and performing connected component analysis, a binarized defect region mask is obtained. For each defect class, the number of all independent connected regions in its mask is counted and divided by the effective area of ​​the tooth surface region to obtain the distribution density. For each defect class, the total area occupied by all pixels in its mask is calculated and divided by the total area of ​​the tooth surface region to obtain the area percentage.

[0112] Based on the design parameters of the spline shaft, generate a discrete point set of its theoretical tooth profile on the end face. By solving the rigid body transformation, the actual tooth profile point set extracted from the image is spatially aligned with the theoretical point set to minimize the distance between corresponding points in the two sets. The optimal transformation is found using the iterative nearest-point algorithm. The core objective is to minimize the following objective function:

[0113]

[0114] in, Let be a rotation matrix. It is a translation vector. , These are the actual and theoretical points corresponding to the registration.

[0115] After registration, for each sampling point on the theoretical profile, find the nearest point on the actual profile along its normal direction, calculate the signed distance deviation, and based on all the calculated deviation values, calculate the maximum positive deviation, the maximum negative deviation, the total profile error, and the standard deviation of the deviation.

[0116] In S5, the pre-trained process evaluation model is a deep feedforward neural network whose input layer receives the aforementioned temporal feature tensor. Subsequently, the data passes through multiple fully connected hidden layers, each performing a nonlinear transformation. The core transformation formula can be expressed as:

[0117]

[0118] in, It is the activation value of the (l-1)th layer. and These are the weight matrix and bias vector of the l-th layer. It is an activation function. Through this series of nonlinear transformations, the model automatically learns and deeply fuses complex, high-order correlation patterns between features from different sensors.

[0119] The last layer of the network typically does not use a non-linear activation function, and its number of neurons corresponds to the dimension of the process state to be evaluated. This layer outputs a multi-dimensional state vector. Each dimension is a continuous scalar value, representing a quantitative evaluation score for a specific process aspect after the model integrates all input features. For example:

[0120] Dynamic stability index is a comprehensive measure of vibration energy and the amplitude of the dominant frequency of force fluctuations. The higher the value, the greater the risk of instability such as flutter.

[0121] The heat load risk index is a comprehensive measure of temperature gradient, hot spot persistence, and area. A higher value indicates a higher risk of overheating.

[0122] Surface quality degradation index, which combines the density of tooth surface texture defects and profile deviation. The higher the value, the worse the surface quality.

[0123] The progressive degradation trend index is the slope of the comprehensive force trend term and the resonant peak drift rate. The higher the value, the faster the tool wear or system performance declines.

[0124] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0125] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive adjustment method for broaching splined shafts, characterized in that, include: The spline shaft workpiece to be processed is clamped on the broaching machine table and tightened. A combination broaching set is selected according to the tooth profile parameters and material properties of the spline shaft. The initial process parameters of broaching speed, broaching force and coolant flow rate are set. Embedded vibration and temperature sensors are placed at the broach teeth. When broaching is started, real-time data is collected synchronously during the continuous feed of the broach. The collected vibration signals are transformed in the time and frequency domain to extract the energy amplitude and resonance peak shift characteristics of specific frequency bands. Gradient analysis and hot spot region identification are performed on the temperature field data. The broaching force fluctuation data are smoothed and decomposed into trends. Texture analysis and contour deviation calculation are performed on the tooth surface morphology image to construct the feature vector of the current process state. The acquired feature vectors are input into the pre-trained process evaluation model. Based on the fusion evaluation results of vibration energy level, temperature gradient, force fluctuation amplitude and tooth surface quality index, the model outputs adjustment instructions for adjusting at least one of broaching speed, broaching force and coolant flow rate. After the adjustment is executed, the model collects and evaluates data for the next monitoring cycle until all tooth grooves of the spline shaft are broached. After machining, white light interferometry is performed on the spline tooth surface to obtain data on tooth profile error, surface roughness, and tooth direction accuracy. Multi-source monitoring data, adjustment event records, and quality data from the entire machining process are linked and stored, and the parameter weights of the process evaluation model are optimized.

2. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The synchronous acquisition of real-time data specifically includes: Vibration spectrum signals of the area where the broach and workpiece mesh are acquired using a vibration sensor; Temperature field distribution data of the tool-chip contact area and the workpiece substrate are collected using a temperature sensor. Real-time broaching force and torque fluctuation data are collected through the broaching machine spindle servo system; The machine vision unit continuously acquires images of the machined tooth surface morphology along the broaching direction.

3. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The process of performing time-frequency domain transformation on the acquired vibration signal to extract energy amplitude and resonance peak shift characteristics in a specific frequency band specifically includes: An adaptive filter is used to filter out the vibration of the broaching machine foundation and environmental noise from the continuous time-domain signal collected by the vibration sensor. The filtered signal is then divided into segments at equal intervals, with each segment corresponding to a complete broaching tooth groove machining cycle. A continuous wavelet transform is performed on each signal segment, and Morlet wavelets are used as basis functions. A two-dimensional time spectrum is obtained by scaling and time shifting. Based on the wavelet time spectrum, the frequency domain is divided into the knife-tooth cutting-out characteristic frequency band, the material shear-dominated frequency band, and the flutter potential frequency band; By integrating the squared amplitude of the wavelet coefficients within each sub-band along the time axis on the time spectrum, the energy value of each band within each time window is calculated. The maximum value, mean value, and fluctuation variance of energy in each frequency band within each processing cycle are statistically analyzed and used as characteristic parameters of the vibration energy distribution state. On the power spectral density curve of each time window, identify the resonance peak caused by the inherent characteristics of the process system and record its center frequency value; The trajectory of the center frequency of the same resonance peak as the broach feed position changes during the complete machining cycle is tracked, and the maximum offset of the center frequency of the resonance peak relative to its initial position and the average rate of frequency drift are calculated as characteristic quantities of stiffness change or abnormal load.

4. The adaptive adjustment spline shaft broaching process method according to claim 3, characterized in that, The step of performing continuous wavelet transform on each signal segment, using Morlet wavelets as basis functions, and obtaining a two-dimensional time-spectrum graph through scaling and time shifting specifically includes: The complex Morlet wavelet is selected as the mother wavelet, and the corresponding scale parameter range is calculated based on the sampling frequency and the target analysis frequency range. Multiple scale values ​​are generated by discretizing the scale in a logarithmic or linear manner within the scale range, and a corresponding Morlet wavelet kernel function is generated for each scale: For each signal segment, convolution calculation is performed in the time domain or frequency domain. In the time domain, the wavelet kernel function of each scale is convolved with the signal segment to calculate the coefficients. In the frequency domain, the signal segment and the wavelet kernel function are subjected to FFT respectively, and then multiplied in the frequency domain and then subjected to IFFT. The coefficients at all scales are arranged in scale order to form a two-dimensional complex matrix. The wavelet scale map is obtained by calculating the modulus matrix of the coefficients. The scale coordinates are converted into actual frequency coordinates to obtain the final two-dimensional time-frequency spectrum.

5. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The gradient analysis and hotspot region identification of the temperature field data specifically includes: The discrete temperature values ​​collected by the temperature sensor are used to reconstruct a complete and continuous two-dimensional temperature field distribution map through spatial interpolation algorithm. The reconstructed temperature field is mapped onto a dynamic mesh, with mesh cells aligned with the tool-workpiece contact area. Each mesh cell corresponds to a temperature node with a defined physical location. For gridded temperature field data, the radial and axial temperature gradient components of each node are calculated in the spatial domain to generate the corresponding temperature gradient vector field. By setting a gradient magnitude threshold, regions with gradient magnitudes higher than the average level are filtered out. Potential hotspot areas are identified based on both absolute temperature threshold and local temperature difference. By tracking the dynamic evolution trajectory of hotspots, they are classified into instantaneous high-temperature zones or persistent heat accumulation zones.

6. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The smoothing and trend decomposition of the broaching force fluctuation data specifically includes: Based on the original broaching force data, instantaneous spike outliers caused by signal interference are removed, and data compensation is performed on the removed points using linear interpolation. A sliding weighted average filter is used to perform initial smoothing of the data, with the weighting coefficients decreasing symmetrically along the center of the window. The smoothed data were analyzed using a local weighted regression method to separate the trend components that characterize the slow changes in tool wear and material hardness. Subtract the extracted trend term from the original smoothed data to obtain the fluctuation term component. Perform a fast Fourier transform on the fluctuation term data to extract its main frequency component and corresponding amplitude. Compare the main frequency component with the passing frequency of the blade or its harmonics and output the comparison result. Calculate the standard deviation, kurtosis, and skewness of the fluctuation term data to quantify the intensity and distribution characteristics of the fluctuation, and extract the slope of the trend term, the amplitude of the dominant frequency of the fluctuation, and the characteristic parameters of the intensity of random fluctuations.

7. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The specific steps of performing texture analysis and contour deviation calculation on the tooth surface morphology image include: The side surface contour of a single spline tooth is identified by an edge detection algorithm, and the tooth surface area image to be analyzed is located and cropped. For the segmented tooth surface region image, calculate its gray-level co-occurrence matrix, extract texture statistics, and calculate the gradient histogram of the image in multiple directions; The extracted texture feature vectors are input into the trained classifier to identify and label specific defect regions in the image, and output the distribution density and area ratio of various defects. The actual edge contour of the tooth surface region is spatially aligned with the theoretical ideal contour generated based on the spline design parameters. Along the normal direction of the theoretical contour, the distance deviation between the actual contour points and the theoretical contour points is densely sampled and calculated. The maximum positive deviation, maximum negative deviation, and standard deviation of all sampling points are statistically analyzed and used as quantitative indicators to evaluate tooth profile error and contour accuracy.

8. The adaptive adjustment spline shaft broaching process method according to claim 7, characterized in that, The process of inputting the extracted texture feature vectors into a trained classifier to identify and label specific defect regions in the image, and outputting the distribution density and area ratio of various defects, specifically includes: The texture feature vector is input into a pre-trained fully convolutional network, which decodes and fuses the input features through fully connected layers, upsamples and reconstructs a multi-channel classification heatmap corresponding to the spatial size of the original input image; Each channel of the classification heatmap corresponds to a preset defect category. By comparing the values ​​between channels pixel by pixel, the defect category of each pixel is determined, and a defect segmentation map is generated. Morphological operations are performed on the marked defective pixels to connect adjacent defective pixels of the same type, forming a continuous defective region; The number of independent connected regions contained in each defect category is counted and divided by the total area of ​​the effective tooth surface analysis area to obtain the number of defects per unit area. The total number of pixels covered by each defect category is counted and divided by the total number of pixels in the effective tooth surface analysis area to calculate the defect area ratio.

9. The adaptive adjustment spline shaft broaching process method according to claim 1, characterized in that, The process of inputting the acquired feature vectors into a pre-trained process evaluation model, and evaluating the results based on the fusion of vibration energy level, temperature gradient, force fluctuation amplitude, and tooth surface quality indicators, specifically includes: The acquired feature vectors are input into a pre-trained deep neural network model, which performs nonlinear transformations and feature fusion through multiple hidden layers to learn complex correlation patterns between different features. At the end of the network, the deep neural network model outputs a multi-dimensional state vector, where each dimension of the multi-dimensional state vector corresponds to a quantified index of the process state after fusion evaluation. The state vector output by the model is compared with the preset process thresholds at each level, and a composite judgment is made based on the fused state. The composite judgment includes: on the generated time spectrum, the frequency axis is divided into high-frequency, mid-frequency and low-frequency regions according to the dynamic characteristics of the broaching process; on the PSD curve, the resonance peak that is significantly higher than the background noise is found, the trajectory of its center frequency change during the entire observation period is analyzed, and the absolute maximum deviation and the average slope of the frequency change trajectory are calculated; by calculating the average value and standard deviation of the gradient amplitude of the entire temperature field, a threshold is set, and all grid nodes that meet the threshold are marked as high temperature gradient regions. In the high temperature gradient region, if the lifetime of a hot spot is very short, and its temperature change rate and area expansion rate quickly turn from positive to negative after it appears, it is determined to be an instantaneous high-temperature zone; if the lifetime of a hot spot is long, and its temperature change rate and area expansion rate are continuously positive or the average value is significantly higher than the baseline within the observation window, it is determined to be a persistent heat accumulation zone. Based on the specific process state category determined, the corresponding combination of adjustment instructions is mapped from the preset strategy library.

Images