In my research, I treat the gear shaping process as a coupled mechanical, metrological, and data-driven problem. The central objective is to improve the shaping accuracy of cylindrical gears, especially when the starting workpiece is a forged gear blank. A forged gear blank often carries geometric inheritance from the forging operation, and this inheritance interacts with fixture errors, tool motion errors, and thermal drift during shaping. My work therefore does not begin with the assumption that a new machine tool must be built. Instead, I begin with the existing four-axis CNC gear shaper and ask how its motion axes can be adjusted in a non-uniform meshing manner so that the final tooth flank accuracy is improved. I focus on the total cumulative pitch deviation and the radial runout tolerance of the gear ring, because these two indices strongly influence transmission accuracy, motion smoothness, and noise. In my method, a forged gear blank is measured after machining, its tooth flank data are evaluated by a full-coordinate method, the next forged gear blank error is predicted by a data-augmented long short-term memory network, and the predicted error is decoupled onto the tool radial feed axis and the tool rotation axis. The resulting non-uniform meshing strategy changes the center distance and the relative meshing position between the cutter and the workpiece. I verify the approach through simulation, machining trials, and digital rolling inspection.
The motivation is practical. High-precision gears are required in aerospace, automotive, machine tools, and precision instruments. Gear shaping remains one of the most important processes for cylindrical gears, especially internal gears, sector gears, eccentric gears, and gears with short cutter relief grooves. Conventional shaping can usually reach an accuracy level between GB 6 and GB 7, whereas many modern transmission systems require GB 5 or better. For some geometries, no ideal gear grinding machine is available, so shaping is the only viable process. If I can improve the accuracy of a forged gear blank after shaping without replacing the machine, I can reduce cost and expand the process window. My method uses the existing motion axes, the existing control system, and the existing measuring equipment. The key is to convert measured and predicted errors into small, coordinated axis corrections.
A forged gear blank is not a passive object. Its initial eccentricity, residual forging stress, hardness variation, and clamping deformation can all influence the final tooth profile. In my experiments, I treat the forged gear blank as the physical carrier of the machine error signature. When a batch of forged gear blanks is machined on the same shaper, the resulting gear errors contain repeatable components and random components. The repeatable components can be predicted and compensated. The random components can be reduced by robust filtering and data augmentation. I therefore use the forged gear blank as the unit of analysis, not just the finished gear.
I built a physical model of the shaping system by simplifying the four-axis CNC gear shaper into three essential motion axes: the tool radial feed axis, the tool rotation axis, and the workpiece rotation axis. The main drive axis mainly affects the tooth lead direction, while the three selected axes dominate the total cumulative pitch deviation and the radial runout tolerance. The physical model is a “machine–cutter–workpiece” system. In this model, the cutter and the forged gear blank rotate in opposite directions, and the radial feed axis changes the center distance. During rough shaping, the cutter feeds radially while the generating motion continues. During finish shaping, the radial feed stops and the cutter and the forged gear blank perform a pure generating motion. This distinction is important because the error compensation strategy must respect the roughing and finishing stages.
I define the tool coordinate system and the workpiece coordinate system as follows. The tool coordinate system is fixed to the cutter, and the workpiece coordinate system is fixed to the forged gear blank. At the initial position, the axes are parallel. As shaping proceeds, the tool and workpiece rotate in opposite directions, and the center distance decreases until the final cutting depth is reached. The transmission ratio remains nominally constant in conventional shaping. In my non-uniform meshing strategy, the transmission ratio is intentionally varied in a controlled way, and the center distance is also varied according to the predicted error. This is why I call the method “non-uniform meshing.” It is not random variation; it is an error-driven variation.
Table 1 lists the main symbols I use throughout the derivation.
| Symbol | Meaning | Unit |
|---|---|---|
| \(F_p\) | Total cumulative pitch deviation | mm |
| \(F_r\) | Radial runout tolerance of gear ring | mm |
| \(f_f\) | Tooth profile deviation | mm |
| \(\delta x\) | Tool radial feed axis motion error | mm |
| \(\delta \theta_c\) | Tool rotation axis angular error | degree |
| \(\delta \theta_w\) | Workpiece rotation axis angular error | degree |
| \(\alpha\) | Pressure angle of the gear | degree |
| \(\alpha_t\) | Pressure angle of the cutter | degree |
| \(m\) | Module of the gear | mm |
| \(m_t\) | Module of the cutter | mm |
| \(z\) | Number of gear teeth | — |
| \(z_t\) | Number of cutter teeth | — |
| \(r_b\) | Base radius of the gear | mm |
| \(\theta\) | Angular position on the base circle | rad or degree |
| \(\omega_c\) | Angular speed of the cutter | degree/s |
| \(\omega_w\) | Angular speed of the workpiece | degree/s |
| \(h\) | Center distance between cutter and workpiece | mm |
| \(N\) | Number of samples | — |
| \(\hat{y}_i\) | Predicted error value | mm |
| \(y_i\) | Measured error value | mm |
I model the tooth profile deviation, the total cumulative pitch deviation, and the radial runout tolerance as linear combinations of the motion errors projected onto the meshing line. The projection direction differs for each error index. For tooth profile deviation, the dominant projection is along the meshing line. I write the model as
$$ f_f = a_x \delta x \sin\alpha + a_c \frac{\pi m z \cos\alpha}{360}\delta\theta_c + a_w \frac{\pi m_t z_t \cos\alpha_t}{360}\delta\theta_w $$
where \(a_x\), \(a_c\), and \(a_w\) are sign coefficients determined by the cutting direction, the tool geometry, and the workpiece geometry. These coefficients are not arbitrary; they follow from whether the error component increases or decreases the effective material removal along the meshing line. I determine them from the kinematic configuration of the shaping process and from the measured error direction.
For the total cumulative pitch deviation, the projection is along the tangent to the pitch circle. The corresponding model is
$$ F_p = b_x \delta x \tan\alpha + b_c \frac{\pi m z \cos\alpha}{360}\delta\theta_c + b_w \frac{\pi m_t z_t \cos\alpha_t}{360}\delta\theta_w $$
where \(b_x\), \(b_c\), and \(b_w\) are again sign coefficients. The radial feed error enters through \(\tan\alpha\), because a radial displacement changes the effective chordal position on the pitch circle. The angular errors of the cutter and the workpiece enter through their arc-length contributions on the pitch circle.
For the radial runout tolerance of the gear ring, the projection is along the radial direction of the gear. I write
$$ F_r = c_x \delta x + c_c \frac{\pi m z \sin\alpha}{360}\delta\theta_c + c_w \frac{\pi m_t z_t \sin\alpha_t}{360}\delta\theta_w $$
where \(c_x\), \(c_c\), and \(c_w\) are sign coefficients. The radial feed error enters directly, while the angular errors enter through their radial components. These three equations form the basis of my decoupling method. If I can predict \(F_p\) and \(F_r\) for the next forged gear blank, I can solve for the required adjustments to \(\delta x\), \(\delta \theta_c\), and \(\delta \theta_w\). In practice, I fix the workpiece rotation axis as the reference and adjust the tool radial feed axis and the tool rotation axis. This is a practical choice because the tool axes are easier to modify on the CNC shaper without disturbing the workholding of the forged gear blank.
Table 2 summarizes the axis roles and the dominant error effects.
| Axis | Symbol | Primary effect | Compensated index |
|---|---|---|---|
| Tool radial feed axis | X | Center distance variation | \(F_p\), \(F_r\), \(f_f\) |
| Tool rotation axis | C2 | Relative meshing angle variation | \(F_p\), \(F_r\), \(f_f\) |
| Workpiece rotation axis | C1 | Generating motion phase | \(F_p\), \(F_r\), \(f_f\) |
| Main drive axis | Z | Reciprocating cutting motion | Lead-related deviations |
I now describe the full-coordinate gear error evaluation method. Traditional evaluation uses a small number of points near the pitch circle. This can be fast, but it is sensitive to local surface condition, probe triggering, temperature, vibration, and residual chips. For a forged gear blank, the surface may have forging scale, hardness variation, or residual stress, so a single-point method can produce a biased result. I therefore sample the entire tooth flank with a coordinate measuring machine, filter the data, fit the involute, and locate the best-fit involute by region traversal. The output is a per-tooth error map, not just a single number.
The sampling procedure is as follows. I place the forged gear blank on the CMM table and ensure that no chips, dust, or burrs interfere with the fixture. I calibrate the probe, set the coordinate system, and define a measurement path that covers the left and right flanks of every tooth. I keep the room temperature within a narrow band. I use a scanning probe or a touch-trigger probe depending on the surface condition. For each tooth, I record a dense set of points \((x,y)\). The raw data are then converted to polar form:
$$ \rho^2 = x^2 + y^2 $$
$$ x = \rho \cos\theta,\qquad y = \rho \sin\theta $$
I compare the radial coordinate \(\rho\) with the base circle radius, the root circle radius, and the tip circle radius. If the gear has \(z \le 41\), the base circle is usually larger than the root circle, so I keep points between the base circle and the tip circle. If \(z \ge 42\), the root circle is usually larger than the base circle, so I keep points between the root circle and the tip circle. This step removes the tip chamfer, the root fillet, and the transition zone. The result is a set of valid involute points for each flank.
Table 3 gives the filtering rules.
| Condition | Lower bound | Upper bound | Retained region |
|---|---|---|---|
| \(z \le 41\) | Base circle radius \(r_b\) | Tip circle radius \(r_a\) | Active involute above base circle |
| \(z \ge 42\) | Root circle radius \(r_f\) | Tip circle radius \(r_a\) | Active involute above root circle |
| All cases | Remove tip chamfer | Remove root fillet | Left and right flanks |
After filtering, I remove outliers. I use a distance-to-line criterion combined with the Pauta criterion. For a point \((x_i,y_i)\), I compute its distance to the line connecting its two neighbors:
$$ d_i = \frac{|A x_i + B y_i + C|}{\sqrt{A^2 + B^2}} $$
where the line is obtained from \((x_{i-1},y_{i-1})\) and \((x_{i+1},y_{i+1})\). I then compute the mean \(\bar{d}\) and the standard deviation \(\sigma\). If
$$ |d_i – \bar{d}| > 3\sigma $$
I mark the point as an outlier and remove it. This step is important for a forged gear blank because local forging defects can create isolated spikes in the measured data. I repeat the procedure for all flanks. In my measurements, I found and removed several dozen outliers from a full gear data set.
For curve fitting, I select four points near the pitch circle: two below and two above. I fit a cubic B-spline segment. The parametric form is
$$ x(t)=A_0+A_1 t+A_2 t^2+A_3 t^3 $$
$$ y(t)=B_0+B_1 t+B_2 t^2+B_3 t^3,\qquad 0\le t\le 1 $$
The coefficients are expressed in terms of the four control points:
$$ A_0=\frac{1}{6}(x_0+4x_1+x_2) $$
$$ A_1=-\frac{1}{2}(x_0-x_2) $$
$$ A_2=\frac{1}{2}(x_0-2x_1+x_2) $$
$$ A_3=-\frac{1}{6}(x_0-3x_1+3x_2-x_3) $$
and similarly for \(B_0,\dots,B_3\) using \(y_0,\dots,y_3\). I then intersect the fitted curve with the pitch circle:
$$ x^2+y^2=\left(\frac{mz}{2}\right)^2 $$
The intersection point \(P\) is used as an anchor for the involute. The involute is generated from the base circle to the tip circle. In polar form, the involute is
$$ r_k=\frac{r_b}{\cos\alpha_k} $$
$$ \theta_k=\tan\alpha_k-\alpha_k $$
and in Cartesian form,
$$ x=r_b(\cos\theta+\theta\sin\theta) $$
$$ y=r_b(\sin\theta-\theta\cos\theta) $$
I rotate the involute so that it passes through \(P\). The rotation is
$$ \begin{bmatrix} x\\ y \end{bmatrix} = \begin{bmatrix} \cos\phi & -\sin\phi\\ \sin\phi & \cos\phi \end{bmatrix} \begin{bmatrix} x’\\ y’ \end{bmatrix} $$
where \(\phi\) is the angle of \(P\). Because the actual base radius and the actual starting angle may deviate from the nominal values, I define an error window \(\Delta r_b\) and \(\Delta \phi\). Inside this window, I generate a family of involutes. I then compute the total distance from the measured points to each involute using the offset curve formula:
$$ x_1 = x_i – \Delta d \frac{\partial y/\partial t}{\sqrt{(\partial x/\partial t)^2+(\partial y/\partial t)^2}} $$
$$ y_1 = y_i – \Delta d \frac{\partial x/\partial t}{\sqrt{(\partial x/\partial t)^2+(\partial y/\partial t)^2}} $$
For each candidate involute, I sum the distances over all valid points on that flank:
$$ D_k=\sum_{i=1}^{N} |\Delta d_i| $$
The best involute is the one that minimizes \(D_k\):
$$ D_{\min}=\min\{D_1,D_2,D_3,\dots\} $$
I repeat this for every left and right flank. The result is a set of base radii and intersection angles. From these, I calculate the total cumulative pitch deviation and the radial runout tolerance. The total cumulative pitch deviation is the maximum absolute difference between the actual arc length and the nominal arc length over any number of pitches. The radial runout tolerance is the difference between the best-fit base radius and the theoretical base radius, projected in the radial direction. This method gives a per-tooth error map, which is more informative than a single-point result.
Table 4 compares the full-coordinate method with the traditional method in terms of data usage and output.
| Item | Traditional single-point method | Full-coordinate method |
|---|---|---|
| Measurement points | Few points near pitch circle | Dense points over active involute |
| Sensitivity to local defects | High | Low after filtering |
| Output | One value per index | Per-tooth and per-flank map |
| Use for compensation | Limited | Directly usable |
| Robustness for forged gear blank | Moderate | High |
After evaluating a batch of gears, I store the results in a database. I use the total cumulative pitch deviation of the left flank, the total cumulative pitch deviation of the right flank, and the radial runout tolerance as three time series. Each new forged gear blank adds a new row. The data are not independent, because they come from the same machine, the same cutter, the same fixture, and similar forged gear blanks. I therefore use a sequence model. I selected a long short-term memory network because it can retain long-range dependencies and because it is less prone to vanishing gradients than a plain recurrent neural network.
To improve generalization, I apply data augmentation before the LSTM. The data set from a small batch of forged gear blanks is limited. I use B-spline interpolation to generate additional samples between measured error points. The interpolation is smooth and preserves the local shape of the error curve. For four original points, I generate a dense set of intermediate points. The augmented data are normalized and split into training and test sets. The training set is used to update the LSTM weights, and the test set is used to evaluate prediction accuracy.
The LSTM cell is governed by the following equations:
$$ f_t=\sigma(W_f[h_{t-1},x_t]+b_f) $$
$$ i_t=\sigma(W_i[h_{t-1},x_t]+b_i) $$
$$ \tilde{C}_t=\tanh(W_C[h_{t-1},x_t]+b_C) $$
$$ C_t=f_t C_{t-1}+i_t \tilde{C}_t $$
$$ o_t=\sigma(W_o[h_{t-1},x_t]+b_o) $$
$$ h_t=o_t \tanh(C_t) $$
Here \(f_t\) is the forget gate, \(i_t\) is the input gate, \(o_t\) is the output gate, \(C_t\) is the cell state, and \(h_t\) is the hidden state. The forget gate decides which old information to discard. The input gate decides which new information to store. The output gate decides what to expose to the next layer. This architecture is well suited to the error sequence of a forged gear blank because the error at tooth \(k\) is correlated with the errors at neighboring teeth and with the previous forged gear blank.
I compared three variants: a single-layer LSTM, a two-layer LSTM, and a bidirectional LSTM. All three used the same augmented data. The evaluation metrics are root mean square error and mean absolute error:
$$ RMSE=\sqrt{\frac{1}{N}\sum_{i=1}^{N}(\hat{y}_i-y_i)^2} $$
$$ MAE=\frac{1}{N}\sum_{i=1}^{N}|\hat{y}_i-y_i| $$
Table 5 gives the model architecture I used for the selected single-layer LSTM.
| Parameter | Value |
|---|---|
| Input dimension | 4 |
| Time step | Error data length per batch |
| Batch size | 5 |
| LSTM output neurons | 300 |
| First fully connected layer | 1200 neurons |
| Second fully connected layer | 1 neuron |
| Activation | Sigmoid and tanh |
| Optimizer | Adam |
| Learning rate | 0.005 |
| Iterations | 300 |
| Training/test split | 80/20 |
Table 6 compares the prediction accuracy of the three models for the left-flank total cumulative pitch deviation, the right-flank total cumulative pitch deviation, and the radial runout tolerance.
| Error index | Model | RMSE | MAE |
|---|---|---|---|
| Left-flank \(F_p\) | Single-layer LSTM | 0.0546 | 0.0624 |
| Left-flank \(F_p\) | Two-layer LSTM | 0.0983 | 0.0872 |
| Left-flank \(F_p\) | Bidirectional LSTM | 0.1479 | 0.1512 |
| Right-flank \(F_p\) | Single-layer LSTM | 0.0589 | 0.0648 |
| Right-flank \(F_p\) | Two-layer LSTM | 0.0920 | 0.0854 |
| Right-flank \(F_p\) | Bidirectional LSTM | 0.1498 | 0.1473 |
| \(F_r\) | Single-layer LSTM | 0.0554 | 0.0619 |
| \(F_r\) | Two-layer LSTM | 0.0975 | 0.0887 |
| \(F_r\) | Bidirectional LSTM | 0.1509 | 0.1603 |
The single-layer LSTM produced the lowest errors. The bidirectional model sometimes produced abrupt changes, which are undesirable for compensation. The two-layer model was more sensitive to overfitting on the small data set. I therefore selected the single-layer data-augmented LSTM for predicting the error of the next forged gear blank. The prediction output is a set of per-tooth error values for the left flank, the right flank, and the radial runout. These values are then used as the target for decoupling.
The decoupling step converts the predicted gear errors into motion commands. I start with the total cumulative pitch deviation. Its radial component on the tool radial feed axis is
$$ \delta x_{F_p}=\frac{F_p}{\tan\alpha} $$
Its angular component on the tool rotation axis is
$$ \delta\theta_{c,F_p}=\frac{360F_p}{mz\cos\alpha} $$
Its angular component on the workpiece rotation axis is
$$ \delta\theta_{w,F_p}=\frac{360F_p}{m_t z_t\cos\alpha_t} $$
Because I keep the workpiece rotation axis as the reference, I combine the cutter and workpiece contributions into an equivalent tool rotation correction:
$$ \omega_c^{\mathrm{corr}}=\frac{360F_p}{mz\cos\alpha}+\frac{360F_p}{m_t z_t\cos\alpha_t} $$
For the radial runout tolerance, the radial feed correction is
$$ \delta x_{F_r}=F_r $$
The angular correction on the tool rotation axis is
$$ \delta\theta_{c,F_r}=\frac{360F_r}{mz\sin\alpha} $$
The angular correction on the workpiece rotation axis is
$$ \delta\theta_{w,F_r}=\frac{360F_r}{m_t z_t\sin\alpha_t} $$
The combined equivalent tool rotation correction is
$$ \omega_c^{\mathrm{corr}}=\frac{360F_r}{mz\sin\alpha}+\frac{360F_r}{m_t z_t\sin\alpha_t} $$
In implementation, I apply the radial correction to the tool radial feed axis and the angular correction to the tool rotation axis. If the predicted deviation is positive, I reduce the corresponding axis command by the absolute correction. If the predicted deviation is negative, I increase the command. If the predicted deviation is zero, I leave the command unchanged. This produces a non-uniform meshing path: the center distance is no longer constant, and the relative meshing angle is no longer constant. The variations are small, but they are synchronized with the tooth index and with the predicted error map.
Table 7 gives an example of decoupling for a forged gear blank. The original center distance is 10 mm, and the original tool rotation increment is 0.5°.
| Tooth index | Predicted \(F_p\) / mm | Predicted \(F_r\) / mm | Radial correction / mm | Tool rotation correction / degree |
|---|---|---|---|---|
| 1 | -0.0095 | -0.0291 | -0.0198 | -0.0409 |
| 2 | -0.0148 | -0.0213 | -0.0339 | -0.0423 |
| 3 | -0.0134 | -0.0382 | -0.0142 | -0.0399 |
| 4 | -0.0124 | -0.0413 | -0.0162 | -0.0396 |
| 5 | -0.0080 | -0.0393 | -0.0149 | -0.0402 |
I then implement the corrected path in the CNC controller. The path is no longer a standard constant-ratio generating motion. It is a modulated generating motion. I use the controller’s curve table function to define the center distance and the relative rotation as functions of the tooth index. The cutter still reciprocates along the main drive axis, but the radial feed and the tool rotation are adjusted for each tooth. This is the practical meaning of non-uniform meshing in my method.
Before machining a forged gear blank, I verify the corrected path in a virtual environment. I build a simplified machine model with the tool radial feed axis, the tool rotation axis, and the workpiece rotation axis. I import the cutter model and the forged gear blank model. I select a compatible CNC controller and load the corrected G-code. I run an interference check and a material removal simulation. If the simulation shows no collision and the tool path is correct, I proceed to the real machine. This step is important because the corrected path is non-standard, and a small programming error could cause interference. Table 8 lists the simulation checks.
| Check | Purpose | Result required |
|---|---|---|
| Machine model | Confirm axis definitions | X, C1, C2 aligned |
| Cutter model | Confirm tool geometry | Module, teeth, pressure angle correct |
| Forged gear blank model | Confirm stock size | No undercut before shaping |
| Controller | Confirm G-code compatibility | No syntax error |
| Interference | Avoid collision | No collision |
| Material removal | Check tooth completeness | All teeth cut |
The simulation results showed that the non-uniform meshing path was feasible. The corrected G-code ran without collision, and the forged gear blank was cut to the expected shape. I then moved to physical machining trials. I used a four-axis CNC gear shaper and a batch of forged gear blanks. I first machined a reference group without compensation. I measured the resulting gears with the full-coordinate method. I stored the error data in the database. I then used the data-augmented LSTM to predict the error of the next forged gear blank. I decoupled the prediction and generated a corrected path. I machined the next forged gear blank with the corrected path. I measured the result again.
Table 9 gives the gear and cutter parameters used in the experiment.
| Parameter | Gear | Cutter |
|---|---|---|
| Module | 6 mm | 6 mm |
| Number of teeth | 34 | 21 |
| Pressure angle | 20° | 20° |
| Workpiece type | Forged gear blank | Shaping cutter |
| Machine | Four-axis CNC gear shaper | — |
Table 10 compares the measured errors before and after compensation. The values are representative maximum deviations.
| Index | Before compensation / mm | After compensation / mm | Improvement / mm |
|---|---|---|---|
| Left-flank \(F_p\) | 0.032 | 0.020 | 0.012 |
| Right-flank \(F_p\) | 0.036 | 0.025 | 0.011 |
| Radial runout \(F_r\) | 0.035 | 0.017 | 0.018 |
| GB 6 limit | 0.045 | 0.045 | — |
| GB 5 limit | 0.028 | 0.028 | — |
The results show that the compensated process brought the forged gear blank into the GB 5 range for the selected indices. The left-flank total cumulative pitch deviation decreased by about 10 to 12 micrometers, and the radial runout tolerance decreased by about 15 to 18 micrometers. The right-flank deviation also decreased. These improvements are significant for a process that uses the same machine, the same cutter, and the same fixture. The only change is the non-uniform meshing path derived from the predicted error of the forged gear blank.
After machining, I performed a digital rolling inspection. The purpose was to evaluate the transmission behavior of the gears made before and after compensation. I built a standard gear model and two test gear models: one from the uncompensated forged gear blank and one from the compensated forged gear blank. I imported the models into a multibody dynamics environment. I defined contact, stiffness, damping, friction, and initial conditions. I applied a constant rotational speed to the standard gear and measured the center distance variation and the angular speed variation of the test gear. A smoother transmission corresponds to smaller variations.
Table 11 gives the digital rolling inspection results.
| Quantity | Before compensation | After compensation |
|---|---|---|
| Center distance variation | 0.15 mm | 0.10 mm |
| Angular speed variation | 200 deg/s | 120 deg/s |
| Transmission smoothness | Moderate | High |
| Motion transfer accuracy | Lower | Higher |
The compensated gear showed smaller center distance variation and smaller angular speed variation. This indicates improved transmission stability and improved motion transfer accuracy. The digital rolling inspection therefore supports the conclusion that the non-uniform meshing method improves the effective accuracy of the forged gear blank after shaping.
I also compared the experimental results with the relevant accuracy grades. Table 12 summarizes the comparison.
| Accuracy index | Before compensation | After compensation | GB 5 requirement | Meets GB 5 |
|---|---|---|---|---|
| Left-flank \(F_p\) | 0.032 mm | 0.020 mm | 0.028 mm | Yes |
| Right-flank \(F_p\) | 0.036 mm | 0.025 mm | 0.028 mm | Yes |
| Radial runout \(F_r\) | 0.035 mm | 0.017 mm | 0.018 mm | Yes |
The mechanism behind the improvement can be explained as follows. The predicted error map identifies which teeth are likely to have larger deviations. The decoupling step converts those deviations into a small radial correction and a small angular correction. The radial correction changes the effective cutting depth along the tooth height. The angular correction changes the relative meshing phase along the tooth profile. When these corrections are applied tooth by tooth, the actual cutting edge removes material from the high-error regions and leaves material in the low-error regions. The result is a more uniform error distribution. The forged gear blank therefore becomes a gear with smaller total cumulative pitch deviation and smaller radial runout.
My method has several practical advantages. First, it does not require a new machine tool. It uses the existing CNC shaper and the existing measuring equipment. Second, it is predictive. The data-augmented LSTM predicts the error of the next forged gear blank before machining, so the correction can be applied in the current cycle rather than after the fact. Third, it is per-tooth. The compensation is not a single global offset; it varies with the tooth index and with the predicted error map. Fourth, it is verifiable. The corrected path can be checked in simulation before cutting a forged gear blank. Fifth, it is extensible. As more forged gear blanks are machined, the database grows, and the prediction model can be retrained.
There are also limitations. The accuracy of the method depends on the quality of the error measurement. If the CMM data contain large noise or if the fixture is unstable, the prediction will be less reliable. The method also assumes that the dominant repeatable errors come from the tool radial feed axis and the tool rotation axis. If thermal drift or spindle vibration dominates, the prediction may need additional inputs. In my current work, I focus on the geometric and kinematic error components. I do not yet include a full thermal model or a full vibration model. These are directions for future work.
Another limitation is the size of the data set. A batch of forged gear blanks provides limited samples. Data augmentation improves the generalization of the LSTM, but it cannot replace real data. As more forged gear blanks are machined, the model can be updated with new measurements. I expect the prediction accuracy to improve as the database grows. In addition, the sampling density of the CMM measurement affects the error evaluation. A higher sampling density can capture more detail, but it also increases measurement time. I selected a sampling density that balances accuracy and productivity.
For future work, I plan to extend the method in several ways. I will measure multiple axial layers of the tooth flank so that the lead direction can be evaluated together with the profile and pitch. I will add temperature sensors and vibration sensors to the shaper so that the prediction model can include thermal and dynamic features. I will increase the number of forged gear blanks in the database so that the LSTM can learn more stable patterns. I will also investigate a real-time implementation in which the error prediction and decoupling are performed inside the CNC controller. This would make the method suitable for production environments.
I also plan to design an on-machine measurement device for the gear shaping process. The device would measure the forged gear blank at the same station, without removing it from the fixture. This would reduce re-clamping error and provide faster feedback. If the on-machine measurement is integrated with the prediction model, the shaper could adjust the non-uniform meshing path in real time. This would move the method from predictive compensation to closed-loop compensation. A fully closed-loop system would include measurement, prediction, decoupling, machining, and verification in one cycle.
In summary, I have developed a non-uniform meshing method for improving gear shaping accuracy. I established the physical and process model of a four-axis CNC gear shaper. I derived the mapping from axis motion errors to tooth profile deviation, total cumulative pitch deviation, and radial runout tolerance. I designed a full-coordinate gear error evaluation method based on CMM data, B-spline fitting, and region traversal. I built a data-augmented LSTM model to predict the error of the next forged gear blank. I decoupled the predicted error onto the tool radial feed axis and the tool rotation axis. I designed a non-uniform meshing path with varying center distance and varying relative meshing angle. I verified the path in simulation. I machined forged gear blanks before and after compensation and measured the results. The compensated forged gear blank reached GB 5 accuracy for the selected indices. The digital rolling inspection showed smaller center distance variation and smaller angular speed variation. The method is therefore feasible, practical, and extensible. It improves the accuracy of a forged gear blank after shaping without requiring a new machine tool, and it provides a foundation for future closed-loop precision shaping systems.