LNN-Transform-based worm grinding wheel gear grinding machine main shaft vibration prediction method
By using the LNN-Transformer model to perform high-dimensional sequence transformation and global feature capture of the process parameters of the worm gear grinding machine, the problem of spindle vibration caused by improper selection of worm gear grinding process parameters was solved, and accurate spindle vibration prediction and machining quality control were achieved.
Patent Information
- Application Number
- CN202511664148.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-02-03
AI Technical Summary
In the existing technology, the selection of process parameters for worm wheel grinding lacks scientific basis, which makes it impossible to effectively predict spindle vibration, affecting machining quality and machine tool service life, and the reliance on experience leads to uncertainty.
A method based on LNN-Transformer is adopted, which uses a liquid neural network to perform high-dimensional sequence transformation and dynamic feature processing of process parameters, and combines a Transformer encoder for global feature capture to establish a relationship model between grinding parameters and spindle vibration, thereby achieving accurate prediction.
It improves the accuracy and robustness of spindle vibration prediction, enhances the ability to identify complex vibration modes, ensures machining quality, and extends machine tool service life.
Smart Images
Figure CN121457032A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of spindle vibration prediction technology in machining, specifically to a method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer. Background Technology
[0002] In recent years, the new energy vehicle industry has risen rapidly. Compared with traditional fuel vehicles, the gear speed in new energy vehicles is increased by about 500%, reaching up to 30,000 rpm. This significant increase in speed leads to harsher operating conditions for the gears, placing higher demands on gear machining quality. Worm wheel grinding, with its ultra-high precision, adaptability to hard materials, and surface integrity control, has become the mainstream machining process for high-end gears (such as those used in new energy vehicles and aerospace gears). In worm wheel grinding, optimization of the grinding process (grinding parameters, cooling and lubrication, multi-process combinations, etc.) and selection of grinding tools (wear resistance of the grinding wheel, self-sharpening properties of the abrasive grains) all affect the vibration of the grinding wheel spindle. This vibration can cause uneven stress on the bearings, reduce the service life of various machine tool components, and affect the production quality of the machined parts.
[0003] Inappropriate selection of grinding process parameters can lead to tooth surface burns or microcracks, accumulation of grinding heat, and thus affect the quality of the machined gears. It can also accelerate the wear of the spindle bearings and thermal deformation of the guide rails of the gear grinding machine, resulting in severe spindle vibration and reducing the service life of the gear grinding machine. At the same time, since process parameters are mostly determined based on experience, they are subjective and uncertain, making it impossible to reduce spindle vibration while ensuring processing efficiency.
[0004] Therefore, it is necessary to establish a relationship model between process parameters and spindle vibration, so as to predict the corresponding spindle vibration based on process parameters, and at the same time optimize the process parameters to obtain the best process parameter input. How to provide a method that can predict the corresponding spindle vibration based on process parameters has become an urgent technical problem to be solved. Summary of the Invention
[0005] This application aims to address the shortcomings of existing technologies by proposing a method for predicting the spindle vibration of a worm gear grinding machine based on LNN-Transformer. The method utilizes LNN to extract and process the implicit dynamic characteristics of the input parameters, establishes the relationship between the grinding parameters of the gear grinding machine and the processing vibration signal, and uses the Transformer module to predict the spindle vibration corresponding to the input grinding parameters.
[0006] To achieve the objectives of this application, this application provides a method for predicting the vibration of the spindle of a worm gear grinding machine based on LNN-Transformer, the method comprising: Input the process parameters for the worm gear grinding machine; The input process parameters are converted into a high-dimensional parameter set using the input parameter high-dimensional sequence conversion module, and dynamic feature processing is performed on the high-dimensional parameter set to obtain a feature set containing hidden dynamic characteristics. The vibration prediction module is used to capture global features of the feature group and output the predicted value of the spindle vibration based on the global features. The high-dimensional sequence conversion module for input parameters employs a liquid neural network (LNN), and the vibration prediction module utilizes a Transformer encoder to capture global features of the feature group.
[0007] Furthermore, the vibration prediction module includes a cascaded Transformer encoder and a regression network, wherein the regression network includes a layer normalization unit, at least one fully connected layer, and an inverse normalization processing unit connected in sequence.
[0008] Based on a Transformer-based global feature capture architecture and regression network, this design fully leverages the spatial and temporal correlations within feature groups through techniques such as multi-head attention, residual connections, layer normalization, and denormalization. This design enables the model to process multiple feature subspaces in parallel, enhancing its ability to identify complex vibration modes while maintaining computational efficiency. Ultimately, it outputs accurate vibration predictions through a complete encoder structure.
[0009] Furthermore, the conversion of the input process parameters into a high-dimensional parameter set includes using a linear layer to upscale the three-dimensional process parameters to a high-dimensional parameter set. This technique of using a linear layer to upscale the three-dimensional process parameters to a high-dimensional parameter set enhances the model's ability to represent the features of low-dimensional process parameters. This upscaling operation can map the complex nonlinear relationships implicit in the original parameters to a high-dimensional space, providing a richer information foundation for subsequent dynamic feature extraction, thereby improving the accuracy and robustness of vibration prediction.
[0010] Furthermore, the dynamic feature processing of the high-dimensional parameter set to obtain a feature set containing hidden dynamic characteristics includes: The gating structure is initialized and the gating signal is calculated based on the high-dimensional parameter set. The high-dimensional parameter set is then processed using the initialized gating structure through linear transformation, ordinary differential equation description and discretization. The implicit dynamic characteristics are extracted from the processed parameter set to obtain a feature set containing the hidden dynamic characteristics.
[0011] The dynamic feature processing mechanism based on gating structures and ordinary differential equations effectively captures the temporal dependencies in the process parameter sequence. Through the synergistic effect of gating structures and discretization processing, it can adaptively learn and retain key information from historical states, while flexibly incorporating new input features, enabling the model to have a stronger ability to model the complex dynamic characteristics of the spindle vibration process.
[0012] Furthermore, the calculation formula for the gating signal is as follows: ; in, i, f, o These are the input gate, forget gate, and output gate, respectively. i t , f t , o t These are the input gating signal, the forgetting gating signal, and the output gating signal, respectively. W i、 W f、 W o These are the input gate weight matrix, the forget gate weight matrix, and the output gate weight matrix, respectively. b i、 b f、 b o These represent the input gate bias, forget gate bias, and output gate bias, respectively, where σ is the activation function. x t The input information is t, where t is the time.
[0013] The specifically defined gating signal calculation formula enables fine-grained control of the input information flow. The synergistic effect of the input gate, forget gate, and output gate can dynamically adjust the information transmission path, ensuring that the model can retain important historical features and respond promptly to new operating conditions during long-term sequence processing, thereby avoiding gradient vanishing or exploding problems.
[0014] Furthermore, the calculation formula for the ordinary differential equation is: ; Where W is the hidden state weight matrix, U is the cyclic weight matrix, b is the dynamic system bias, and τ is the liquid time constant. h Indicates the hidden state at the current time step. h t-1 This represents the hidden state of the previous time step.
[0015] By introducing an ordinary differential equation describing the liquid time constant, the model is endowed with the ability to adapt to dynamic characteristics at different time scales. This technique enables the system to automatically adjust its dynamic response speed according to the characteristics of the vibration signal, maintaining stable predictive performance in the face of abrupt changes or gradual processes.
[0016] Furthermore, the Euler method is used for discretization, and the calculation formula is as follows: ; Where Δt is the time step and tanh() is the activation function. h new This represents the hidden state generated at time t-1. Discretization using the Euler method transforms the continuous dynamic system into a computable form, maintaining the advantages of liquid neural networks in continuous-time modeling while ensuring engineering feasibility. This numerical computation method improves computational efficiency while ensuring accuracy, making the model suitable for real-time industrial prediction scenarios.
[0017] Furthermore, the extraction of implicit dynamic characteristics from the processed parameter set includes: The processed parameter set is gated to calculate the hidden state at the current time step and extract its implicit dynamic characteristics. The calculation formula is as follows: ; in, h t To extract the implicit dynamic characteristics, ⊙ represents element-wise multiplication. f t ⊙ h t-1 To preserve historical states for the forgetting gate control, i t ⊙ h new The input gate controls the injection of new states.
[0018] By employing gating and hidden state computation mechanisms, effective extraction of dynamic characteristics is achieved. This technique combines historical state preservation with balanced control of new state injection, enabling the model to capture subtle transient features and long-term evolution patterns in vibration signals, significantly improving prediction sensitivity.
[0019] Furthermore, the formula for extracting the global features is as follows: ; in, These are the first-layer weight matrix and the second-layer weight matrix, respectively. b 1, b 2 represents the first-layer bias term and the second-layer bias term, respectively; max(0,·) is the LeakyReLU activation function.
[0020] A two-layer fully connected structure with LeakyReLU activation function is used for global feature extraction, which ensures nonlinear modeling capability while avoiding neuron failure. This design enhances the model's efficiency in representing high-dimensional features, ensures that key vibration feature information is preserved during dimensionality reduction, and ultimately achieves accurate vibration prediction output.
[0021] The beneficial effects of this application are: This application utilizes a liquid neural network (LNN) to transform the input process parameters into a high-dimensional parameter set and performs dynamic feature processing, effectively extracting feature sets containing hidden dynamic characteristics. This processing method can fully exploit the complex nonlinear relationships implicit in the process parameters, especially significantly enhancing the ability to capture time-series dynamic features. Subsequently, a Transformer structure including an encoder is used to globally capture the feature set, and its powerful attention mechanism comprehensively analyzes the intrinsic correlations between features, ultimately outputting accurate spindle vibration prediction values.
[0022] First, a liquid neural network is used for dynamic feature processing, enabling the model to adaptively learn the patterns of process parameter changes over time, thus enhancing the system's adaptability to non-stationary operating conditions. Second, by combining the global feature capture mechanism of the Transformer, the limitations of traditional methods in long-sequence dependency modeling are overcome, allowing for the simultaneous consideration of local details and global trends. Finally, the entire method achieves an organic combination of feature extraction and prediction through modular design, improving prediction accuracy and robustness while maintaining model interpretability, providing reliable technical support for spindle condition monitoring and preventive maintenance in practical industrial applications.
[0023] This method integrates LNN and Transformer encoder structures to extract implicit dynamic characteristics from input static parameters and transform the input static process parameters into high-dimensional feature sequences. At the same time, it combines Transformer-Encoder to capture global features between features, making the extracted parameter features more complete and the predicted values more reliable. Attached Figure Description
[0024] Figure 1 The overall framework diagram of the vibration prediction method for the spindle of a worm gear grinding machine based on LNN-Transformer is shown in the present invention. Figure 2 The overall framework diagram of the vibration prediction method for the spindle of a worm gear grinding machine based on LNN-Transformer is shown in the present invention. Figure 3 A comparison of prediction results for a worm gear grinding machine spindle vibration prediction method based on LNN-Transformer; Figure 4 A performance index comparison chart of the vibration prediction method for the spindle of a worm gear grinding machine based on LNN-Transformer. Detailed Implementation
[0025] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0026] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.
[0027] The present invention provides a method for predicting the vibration of a worm gear grinding machine spindle based on LNN-Transformer. The execution subject includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in the embodiments of this application: a server, a terminal, or a computer. In other words, a method for predicting the vibration of a worm gear grinding machine spindle based on LNN-Transformer can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.
[0028] Reference Figure 1 and Figure 2 In one embodiment, the present invention provides a method for predicting the vibration of a worm gear grinding machine spindle based on LNN-Transformer, the method comprising: Input the process parameters for the worm gear grinding machine, including the grinding wheel linear velocity. v s、 Axial feed rate v w and radial feed fr。 In some embodiments, this method uses a CNC worm gear grinding machine as the experimental object. Based on the principle of controlled variables, a three-factor, multi-level orthogonal experimental design is employed. Signals are collected across the three factors according to the experimental design. The level of each process parameter is set as follows: grinding wheel linear velocity... v s 40m / s, 50m / s, 60m / s; Axial feed speedv w 50 mm / min, 100 mm / min, 150 mm / min, 200 mm / min; radial feed rate fr The sample sizes were 0.04 mm, 0.1 mm, 0.16 mm, and 0.22 mm. A total of 48 (3×4×4) different parameter combinations were obtained. The sampling frequency was set to 10 kHz. The root mean square (RMS) value was selected as the vibration characteristic value for vibration analysis. After data acquisition, radial basis function interpolation was used to expand the difference samples of the acquired data, and the value range of each process parameter was divided into 15 intervals, ultimately yielding 3375 (15×15×15) sets of sample data.
[0029] The input process parameters are converted into a high-dimensional parameter set by the input parameter high-dimensional sequence conversion module, and dynamic feature processing is performed on the high-dimensional parameter set to obtain a feature set containing hidden dynamic characteristics. The process of converting the input process parameters into a high-dimensional parameter set includes using a linear layer to upgrade the three-dimensional process parameters to a high-dimensional parameter set.
[0030] Three gating structures are initialized and defined based on a high-dimensional parameter set, and gating signals are calculated. Using these initialized gating structures, the high-dimensional parameter set is processed for implicit dynamic features through linear transformation, ordinary differential equation description, and discretization. Implicit dynamic characteristics are extracted from the processed parameter set to obtain a feature set containing these hidden dynamic characteristics. A high-dimensional mapping is performed on each gating signal value through linear transformation; the continuous-time changes of the hidden state are described using ordinary differential equations; and the ordinary differential equations are discretized using the Euler method.
[0031] In some embodiments, after the initialization phase, three gating signal values are first calculated, and the calculation formula for the gating signals is as follows: ; in, i, f, o These are the input gate, forget gate, and output gate, respectively. i t The input gating signal is used to control the proportion of new information flowing into the model. f t The forgetting gate signal is used to determine which information to discard from historical states. o t The output gating signal is used to control which information is output from the current state to subsequent processing stages. W i ,, W f , W oThese are the input gate weight matrix, the forget gate weight matrix, and the output gate weight matrix, respectively. Their dimensions depend on the input information. x t The feature dimensions and gating signal design requirements are determined to process the input information. x t Perform a linear transformation to extract relevant features. b i , b f , b o These are the input gate bias, forget gate bias, and output gate bias, used to adjust the result after the linear transformation and optimize the range of values for the gate signal. σ is the activation function, which uses the Sigmoid function to map the result after the linear transformation to the [0,1] interval, thereby achieving selective control of information. x t The input information is specifically the pre-processed process parameter data related to the spindle vibration of the worm gear grinding machine.
[0032] As shown in the gating signal calculation formula above, process parameters serve as input information. x t Part of the process, after being processed with the weight matrix and bias, is used to obtain the gating signal through the activation function σ. The gating signal (input gate, forget gate, output gate) controls the injection of new states, the retention of historical states, and the output of the current state, respectively. The process parameters affect the gating signal, thereby affecting the model's updating and processing of the hidden states.
[0033] In some embodiments, a high-dimensional mapping is performed on each gated signal value through a linear transformation, and dynamic system evolution is carried out. The continuous-time changes of the hidden state are described by an ordinary differential equation (ODE), the calculation formula of which is: ; Where W is the hidden state weight matrix, b is the dynamic system bias, and τ is the liquid time constant. In the formula, W is the hidden state weight matrix, U is the cyclic weight matrix, b is the dynamic system bias, and τ is the liquid time constant (controlling the dynamic response speed). h Indicates the hidden state at the current time step. h t-1 This represents the hidden state of the previous time step.
[0034] In some embodiments, the Euler method is used for discretization, and the calculation formula is as follows: ; Where Δt is the time step and tanh() is the activation function. h newThis represents the hidden state generated at time t-1.
[0035] In some embodiments, extracting implicit dynamic characteristics from the processed parameter set includes: The processed parameter set is gated to calculate the hidden state at the current time step and extract its implicit dynamic characteristics. The calculation formula is as follows: ; in, h t To extract the implicit dynamic characteristics, ⊙ represents element-wise multiplication. f t ⊙ h t-1 To preserve historical states for the forgetting gate control, i t ⊙ h new The input gate controls the injection of new states. Based on the implicit dynamic characteristics, the final high-dimensional feature sequence output is generated. y t . h t By integrating historical dynamic information with new input states through a gating mechanism and undergoing dynamic system evolution and optimization, it possesses hidden dynamic characteristics and represents a high-dimensional, serialized core feature representation. This generates the final high-dimensional feature sequence output. y t "Actually, it is" h t As serialization descriptions of the output, both essentially share the same feature. y t Right now h t The output sequence is used for subsequent global feature capture tasks.
[0036] After the above processing, a high-dimensional feature sequence is finally generated. The high-dimensional feature sequence is the result obtained after high-dimensional sequence transformation of the input process parameters and related signals. It integrates the influence of process parameters on the entire processing process and can reflect the complex relationship between process parameters and hidden states. It can be used for subsequent applications such as process quality prediction and intelligent control. For example, in process manufacturing workshops, this high-dimensional feature sequence can be used to accurately predict process quality and provide support for intelligent decision-making and optimization control of the production line.
[0037] The vibration prediction module performs global feature capture on the feature group and outputs the predicted value of the spindle vibration based on the global features. The vibration prediction module includes a cascaded Transformer encoder and a regression network, wherein the regression network includes a layer normalization unit, at least one fully connected layer, and an inverse normalization processing unit connected in sequence.
[0038] The process of globally capturing features from the feature set and outputting predicted values for spindle vibration includes: The Transformer encoder is used to output feature sets containing implicit dynamic characteristics, especially high-dimensional feature sequences. y t A linear transformation is performed, and then three weight matrices are used to generate queries, keys, and values. Attention sparsity is calculated by randomly selecting a portion of the keys for each query through uniform sampling. Q, K, and V are then divided into h sub-matrices according to the number of attention heads to obtain the dimension of each head.
[0039] After the multi-head subspace is segmented, scaled dot product attention is performed, and attention weights are calculated independently for each head. The hidden state at each position in the sequence of the feature group is mapped independently, which enhances the model's ability to capture complex features. A regression network is used to element-wise superimpose the input tensor and the sub-layer processed output through residual connections. Layer normalization is then used to standardize the feature vector numerical distribution, further optimizing the stability during gradient propagation. Finally, the output is dimensionality-reduced through a fully connected layer and inversely normalized to obtain the final vibration feature value prediction. Specifically, dimensionality reduction and inverse normalization of the output yield the final RMS prediction value. The Transformer treats the high-dimensional representation of each LNN output as a sequence of length 1, utilizing its powerful self-attention mechanism to enhance feature representation capabilities.
[0040] The global feature extraction formula is as follows: ; in, These are the first-layer weight matrix and the second-layer weight matrix, respectively. b 1, b 2 represents the first-layer bias term and the second-layer bias term, respectively; max(0,·) is the LeakyReLU activation function.
[0041] The high-dimensional sequence conversion module for input parameters is implemented using a liquid neural network (LNN), and the vibration prediction module uses a Transformer encoder to capture global features of the feature group.
[0042] This method consists of a high-dimensional sequence conversion module for input parameters and a vibration prediction module. In the sequence conversion module, feature extraction and processing of the input grinding parameters are performed using a liquid neural network (LNN). Since the input data for this model is a combination of grinding process parameters (…),… v s , vw , fr The data (i.e., grinding wheel linear velocity, grinding wheel axial feed rate, and grinding wheel radial feed) are three-dimensional data, and the data received by Transformer is mostly high-dimensional data. Adding a linear layer will upgrade the combination of three-dimensional process parameters to a high-dimensional parameter group, and LNN will be used to extract and process the implicit dynamic features of the high-dimensional parameter group.
[0043] First, after standardization preprocessing, a gating mechanism (input gate, forget gate, output gate) is used to dynamically filter key information and suppress noise, simulating the physical transmission logic of parameters to vibration response. Then, the input is Eulerian discretized and evolved through the continuous-time dynamic system (differential equation) of LNN, and a high-dimensional hidden state is generated using the LeakyReLU activation function to capture the nonlinear sensitivity and potential dynamic trajectory of process parameter combinations. Finally, the old and new states are fused by gating weighting and the output feature vector rich in dynamic semantics is used for global relationship modeling in the downstream Transformer.
[0044] The LNN first performs initialization, defining and initializing three gating structures (input gate, forget gate, and output gate) and calculating three gating signals. (These three gating structures (input gate, forget gate, and output gate) are key components of the LNN (Liquid Neural Network) used to process combinations of grinding process parameters (wheel linear velocity, wheel axial feed rate, and wheel radial feed). The LNN initializes these three gating structures and calculates the gating signals, and then, through subsequent linear transformations, ordinary differential equation descriptions, and discretization, achieves the extraction and processing of implicit dynamic features of the three-dimensional grinding process parameter combinations.) After calculating the gated signal, it is mapped to a high dimension through linear transformation, and the changes in its hidden state are described by ordinary differential equations (ODE). Finally, the ODE is discretized using the Euler method. Finally, the gated signal is combined with the dynamic system state to calculate the hidden state at the current time step and extract its implicit dynamic characteristics.
[0045] In the spindle vibration prediction module, an Only-Encoder module is used to receive a feature set containing hidden dynamic characteristics, and basic static feature extraction and processing are performed on this feature set. The input sequence is processed through three weight matrices to generate queries, keys, and values. Attention is achieved by randomly selecting a portion of the keys for each query through uniform sampling. After segmenting the multi-head subspace, scaled dot product attention is performed, and attention weights are calculated independently for each head.
[0046] This method integrates LNN and Transformer encoder structures to extract implicit dynamic characteristics from input static parameters and transform the input static process parameters into high-dimensional feature sequences. At the same time, it combines Transformer-Encoder to capture global features between features, making the extracted parameter features more complete and the predicted values more reliable.
[0047] Reference Figure 3 , Figure 3 Figures (a)-(e) show the prediction results and absolute errors of BP, SVR, ELM, CART, and our proposed model, respectively. As can be seen from the figures, the model used in our method shows a high degree of consistency between the predicted and actual RMS curves of the spindle vibration, and is closer to the true curve trend compared to the other four models. Furthermore, regarding the absolute error of our model, except for a few errors above 0.1, most errors are between 0 and 0.05, with the minimum value close to 0. Among the other four models, SVR has the worst overall prediction result, while CART has the best overall prediction result, but some predictions have large errors, with the highest value approaching 0.3, which is significantly different from our proposed model. This demonstrates the accuracy and effectiveness of our proposed model in predicting the RMS of spindle vibration.
[0048] Reference Figure 4 The maximum value of each model's performance metric was used as a benchmark for normalization, resulting in a performance metric comparison chart. The chart shows that the performance of this model significantly outperforms the other four models. Specifically, the R² values for BP, SVR, ELM, and CART are 95.09%, 97.35%, 96.37%, and 95.94%, respectively, while the R² value of this model is 99.16%, the highest and closest to 1 among all models, indicating that this model has the best fit to the data. Furthermore, the MAE value of this model is 0.0511, and the RMSE value is 0.0741, more than half that of the other models. Additionally, the MAPE value of this model is 0.05%, significantly lower than the other models, indicating that this model has the highest prediction accuracy and validating its feasibility.
[0049] In the description of this specification, the references to terms such as "an embodiment," "some embodiments," "example," "specific example," "a implementation," "a preferred implementation," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0050] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.
Claims
1. A method for predicting the vibration of the spindle of a worm gear grinding machine based on LNN-Transformer, characterized in that, The method includes: Input the process parameters for the worm gear grinding machine; The input process parameters are converted into a high-dimensional parameter set using the input parameter high-dimensional sequence conversion module, and dynamic feature processing is performed on the high-dimensional parameter set to obtain a feature set containing hidden dynamic characteristics. The vibration prediction module is used to capture global features of the feature group and output the predicted value of the spindle vibration based on the global features. The high-dimensional sequence conversion module for input parameters employs a liquid neural network (LNN), and the vibration prediction module utilizes a Transformer encoder to capture global features of the feature group.
2. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer as described in claim 1, characterized in that, The vibration prediction module includes a cascaded Transformer encoder and a regression network, wherein the regression network includes a layer normalization unit, at least one fully connected layer, and an inverse normalization processing unit connected in sequence.
3. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer as described in claim 1, characterized in that, The process of converting the input process parameters into a high-dimensional parameter set includes using a linear layer to upgrade the three-dimensional process parameters to a high-dimensional parameter set.
4. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer as described in claim 3, characterized in that, The dynamic feature processing of the high-dimensional parameter set to obtain a feature set containing hidden dynamic characteristics includes: The gating structure is initialized and the gating signal is calculated based on the high-dimensional parameter set. The high-dimensional parameter set is then processed using the initialized gating structure through linear transformation, ordinary differential equation description and discretization. The implicit dynamic characteristics are extracted from the processed parameter set to obtain a feature set containing the hidden dynamic characteristics.
5. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer according to claim 4, characterized in that, The formula for calculating the gating signal is as follows: ； in, i, f, o These are the input gate, forget gate, and output gate, respectively. i t , f t , o t These are the input gating signal, the forgetting gating signal, and the output gating signal, respectively. W i、 W f、 W o These are the input gate weight matrix, the forget gate weight matrix, and the output gate weight matrix, respectively. b i、 b f、 b o These represent the input gate bias, forget gate bias, and output gate bias, respectively, where σ is the activation function. x t The input information is t, where t is the time.
6. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer as described in claim 4, characterized in that, The calculation formula for the ordinary differential equation is: ； Where W is the hidden state weight matrix, U is the cyclic weight matrix, b is the dynamic system bias, and τ is the liquid time constant. h Indicates the hidden state at the current time step. h t-1 This represents the hidden state of the previous time step.
7. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer according to claim 4, characterized in that, The Euler method is used for discretization, and the calculation formula is as follows: ； Where Δt is the time step and tanh() is the activation function. h new This represents the hidden state generated at time t-1.
8. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer as described in claim 4, characterized in that, The extraction of latent dynamic characteristics from the processed parameter set includes: The processed parameter set is gated to calculate the hidden state at the current time step and extract its implicit dynamic characteristics. The calculation formula is as follows: ； in, h t To extract the implicit dynamic characteristics, ⊙ represents element-wise multiplication. f t ⊙ h t-1 To preserve historical states for the forgetting gate control, i t ⊙ h new The input gate controls the injection of new states.
9. The method for predicting spindle vibration of a worm gear grinding machine based on LNN-Transformer according to claim 1, characterized in that, The formula for extracting the global features is as follows: ； in, These are the first-layer weight matrix and the second-layer weight matrix, respectively. b 1, b 2 represents the first-layer bias term and the second-layer bias term, respectively; max(0,·) is the LeakyReLU activation function.