Multi-objective broaching process parameter optimization method based on improved NSGA-Ⅲ algorithm

Abstract

The present application relates to a multi-objective broaching process parameter optimization method based on improved NSGA-III algorithm, belongs to the broaching process optimization technical field, solves the problems of inaccurate dynamic signal processing, high feature redundancy, optimization model prone to gradient anomaly and overfitting, slow convergence and uneven solution set distribution of the existing optimization method. First, the dynamic process parameter signals in the broaching process are collected by a sensor array, including cutting force, vibration acceleration, real-time feed speed and cutting depth, and the WPT-HHT joint algorithm is used to extract features to form an initial feature set, then the random forest model is used to select key features and eliminate redundant features, then the improved DNN is constructed to model the fitness function, and finally the multi-objective optimization model is solved based on the improved NSGA-III algorithm to obtain the Pareto optimal broaching process parameter combination. The present application significantly improves the machining quality, tool life and machining efficiency of the broaching process.

Description

Technical Field

[0001] This invention relates to the field of broaching process optimization technology, and in particular to a multi-objective broaching process parameter optimization method based on an improved NSGA-Ⅲ algorithm. Background Technology

[0002] In broaching, machining quality, tool wear, and machining efficiency are closely related to dynamic process parameters such as feed rate, depth of cut, cutting force, and vibration. While existing process optimization methods can achieve basic parameter adjustments, they still suffer from drawbacks in practical applications, including inaccurate dynamic signal processing, unreasonable feature selection, and poor model adaptability. These shortcomings make it difficult to meet the high-precision, high-efficiency, and high-reliability machining requirements of modern broaching processes, and are key factors restricting the upgrading of broaching technology. Currently, widely used process optimization methods rely heavily on traditional signal processing and shallow optimization models. Although they can achieve basic parameter optimization, in actual broaching processes, current methods still suffer from poor signal processing performance, insufficient optimization accuracy, and low efficiency.

[0003] The main problems with existing broaching process optimization methods include:

[0004] 1. When using existing process optimization methods to optimize broaching dynamic parameters, even if the processing scenario and dynamic parameter requirements are similar, most methods only extract simple statistical features such as signal mean and maximum value, and cannot capture time-varying characteristics such as instantaneous frequency and amplitude changes. In addition, the original signal contains a lot of noise, resulting in incomplete feature representation and large deviation in optimization decision.

[0005] 2. Even though some current optimization methods attempt to extract multi-dimensional signal features, directly using all extracted features will significantly increase the complexity of the optimization model, greatly increase the risk of overfitting, double the computational cost, prolong the time consumption, and significantly reduce the generalization ability.

[0006] 3. When optimizing broaching process parameters, there may be a need for coupling dynamic and static parameters and multi-objective collaborative optimization. However, existing optimization models mostly use shallow neural networks, which are prone to getting trapped in local optima. Traditional deep neural networks (DNNs) have gradient vanishing and gradient exploding problems. Furthermore, single static parameter optimization ignores the coupling relationship between dynamic and static parameters, making it difficult to meet the multi-objective collaborative optimization needs such as surface roughness, dimensional accuracy, and tool life.

[0007] 4. When facing high-dimensional target optimization scenarios, the traditional NSGA-II algorithm has poor uniformity of solution distribution and slow convergence speed, and cannot efficiently obtain the Pareto optimal solution set, resulting in low optimization efficiency and difficulty in adapting to the high-efficiency processing requirements of broaching process.

[0008] To address the aforementioned issues, there is an urgent need for an intelligent optimization method that combines accurate dynamic signal processing, efficient feature selection, stable and reliable models, and excellent optimization results. Summary of the Invention

[0009] The purpose of this invention is to address the problems existing in current broaching process optimization methods by providing a multi-objective broaching process parameter optimization method based on an improved NSGA-Ⅲ algorithm. This method is an intelligent optimization approach. On the one hand, it solves the problems of incomplete feature representation and high redundancy through precise dynamic process parameter signal processing and efficient feature selection, thereby improving the accuracy of optimization decisions. On the other hand, through a stable and reliable multi-objective optimization model and an efficient optimization algorithm, it takes into account the needs of multi-objective collaborative optimization, improves optimization efficiency and effectiveness, solves the technical bottleneck of broaching process parameter optimization, and meets the high-precision and high-efficiency machining requirements.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A multi-objective broaching process parameter optimization method based on an improved NSGA-Ⅲ algorithm includes the following steps:

[0012] Step S1: Collect dynamic process parameter signals during the broaching process through a sensor array, including cutting force, vibration acceleration, real-time feed rate, and depth of cut;

[0013] Step S2: Use the WPT-HHT joint algorithm to extract features from the dynamic process parameter signal, capture the time-varying characteristics of the signal, and form an initial feature set;

[0014] Step S3: Construct a random forest model, sort and filter the initial feature set according to feature importance, remove redundant features, and retain key features;

[0015] Step S4: Construct an improved deep neural network, which is a two-layer network structure, and introduces the Leaky ReLU activation function, gradient clipping and L2 regularization mechanism to establish a fitness function with surface roughness, dimensional accuracy and tool wear as optimization objectives;

[0016] Step S5: Using spindle speed, feed per tooth, and depth of cut as optimization variables, and cutting power, broaching force, and machining efficiency as constraints, construct a multi-objective optimization model. Using the fitness function as the evaluation criterion, solve the multi-objective optimization model using the improved NSGA-Ⅲ algorithm to obtain the Pareto optimal broaching process parameter combination.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) The present invention adopts the WPT-HHT joint algorithm, which realizes the accurate representation of dynamic features through adaptive decomposition layer and energy threshold calculation, effectively improves the signal-to-noise ratio, significantly improves the correlation between features and optimization targets after random forest screening, and greatly reduces feature redundancy.

[0019] (2) The improved deep neural network in this invention adopts the Leaky ReLU activation function, gradient clipping and L2 regularization mechanism, which ensures the stability and reliability of the model, effectively avoids gradient vanishing and gradient explosion, significantly improves prediction accuracy and reduces the risk of model overfitting.

[0020] (3) The present invention applies the improved NSGA-Ⅲ algorithm, combined with weight configuration and uniform reference point design, to achieve good optimization effect, making the Pareto solution distribution more uniform and the convergence speed faster. The surface quality and dimensional accuracy are significantly improved after optimization, while effectively extending tool life and improving processing efficiency. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the multi-objective broaching process parameter optimization method in an embodiment of the present invention;

[0022] Figure 2 This is a flowchart illustrating feature extraction using the WPT-HHT joint algorithm in an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the improved two-layer network structure of DNN in an embodiment of the present invention;

[0024] Figure 4 This is a flowchart of a multi-objective optimization process based on the improved NSGA-Ⅲ algorithm in an embodiment of the present invention. Detailed Implementation

[0025] To more clearly illustrate the technical problems, technical solutions, and advantages of the present invention, a detailed description is provided below in conjunction with the accompanying drawings and specific embodiments.

[0026] This invention addresses the problems encountered in broaching processes, such as numerous dynamic process parameters, high signal noise, high feature redundancy, gradient vanishing and overfitting issues in optimization models, slow convergence of traditional multi-objective optimization algorithms, and uneven distribution of Pareto solutions. These problems lead to unstable machining quality, short tool life, low machining efficiency, and difficulty in guaranteeing dimensional accuracy. The invention provides a multi-objective broaching process parameter optimization method based on an improved NSGA-Ⅲ algorithm. This method achieves accurate, stable, efficient, and intelligent optimization of broaching process parameters through steps including collecting dynamic process parameter signals during broaching, feature extraction using the WPT-HHT joint algorithm, feature screening using random forest, improved DNN fitness modeling, and improved NSGA-Ⅲ algorithm multi-objective optimization.

[0027] Specifically, firstly, dynamic process parameter signals during the broaching process are acquired using a sensor array. Then, the acquired dynamic process parameter signals are preprocessed for noise reduction and feature extraction using the WPT-HHT joint algorithm to capture the time-varying characteristics of the signals, such as instantaneous frequency and amplitude, forming an initial feature set. Next, the initial feature set is sorted and filtered by feature importance using the random forest algorithm to remove redundant information and retain key features. Subsequently, an improved deep neural network fitness function model is constructed to integrate static process parameters and dynamic features to establish an optimization target prediction relationship. Finally, the multi-objective optimization model is solved using the improved NSGA-Ⅲ algorithm to obtain the Pareto optimal broaching process parameter combination, achieving multi-objective collaborative optimization of surface roughness, dimensional accuracy, and tool life, significantly improving the quality and efficiency of broaching.

[0028] like Figure 1 As shown, the multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm proposed in this embodiment mainly includes the following steps:

[0029] Step S1: Dynamic process parameter signal acquisition: Dynamic process parameter signals during the broaching process are acquired through a sensor array, including cutting force, vibration acceleration, real-time feed rate, and depth of cut.

[0030] Step S2, WPT-HHT joint algorithm feature extraction: The WPT-HHT joint algorithm is used to extract features from the dynamic process parameter signal, capture the time-varying characteristics of the signal, and form an initial feature set;

[0031] Step S3: Ranking and filtering of random forest features: Construct a random forest model, rank and filter the features in the initial feature set according to their importance, remove redundant features, and retain key features;

[0032] Step S4: Improve DNN fitness function modeling: Construct an improved deep neural network, which is a two-layer network structure, and introduce Leaky ReLU activation function, gradient clipping and L2 regularization mechanism to establish a fitness function with surface roughness, dimensional accuracy and tool wear as optimization objectives.

[0033] Step S5: Multi-objective optimization based on the improved NSGA-Ⅲ algorithm: Using spindle speed, feed per tooth, and depth of cut as optimization variables, and cutting power, broaching force, and machining efficiency as constraints, a multi-objective optimization model is constructed. The fitness function is used as the evaluation basis, and the improved NSGA-Ⅲ algorithm is used to solve the multi-objective optimization model to obtain the Pareto optimal broaching process parameter combination.

[0034] Among them, the dynamic process parameter signal acquisition and WPT-HHT joint algorithm feature extraction form the basis of the entire optimization method. In the process of acquiring dynamic process parameter signals in step S1, dynamic signals such as cutting force, vibration acceleration, real-time feed rate, and cutting depth during the broaching process are acquired through a sensor array. The sensor array used as the acquisition device includes a piezoelectric cutting force sensor, an ICP vibration acceleration sensor, and a grating displacement sensor, which can ensure high data synchronization accuracy.

[0035] The feature extraction process using the WPT-HHT joint algorithm in step 2 is as follows: Figure 2 As shown, the dynamic feature extraction based on the WPT-HHT joint algorithm mainly consists of three core processes: signal denoising preprocessing, EMD decomposition, and Hilbert transform. In the signal preprocessing stage, the Wavelet Packet Transform (WPT) algorithm is used for adaptive denoising. The number of decomposition layers is adaptively determined according to the signal-to-noise ratio, the db4 wavelet basis function is selected, and the effective frequency band is screened and the signal is reconstructed by the energy threshold method. Subsequently, the Hilbert-Huang Transform (HHT) algorithm is used to perform Empirical Mode Decomposition (EMD) on the denoised signal. The noise IMF components with low correlation coefficients are removed, and four types of time-frequency features, namely instantaneous frequency, instantaneous amplitude, amplitude envelope, and frequency bandwidth, are extracted. Statistical features such as mean and variance are added to form an initial feature set, which can comprehensively characterize the dynamic change characteristics of the signal.

[0036] The key parameters and operation procedures of the WPT-HHT joint algorithm are as follows:

[0037] (1) Signal noise reduction preprocessing:

[0038] The WPT algorithm is used to denoise the acquired raw dynamic process parameter signals. The specific operation is as follows: First, the number of signal decomposition layers is adaptively determined according to the signal-to-noise ratio of the raw signal. Second, a suitable wavelet is selected as the basis function, and the denoising threshold is calculated by the energy threshold method. Then, the effective frequency band is selected based on the threshold. Finally, the effective frequency band signal is reconstructed by wavelet packet inverse transform to obtain a clean dynamic process parameter signal, which lays a reliable foundation for subsequent dynamic feature extraction.

[0039] (2) EMD decomposition:

[0040] Empirical mode decomposition (EMD) is performed on the purified dynamic process parameter signal after noise reduction. The purified signal is decomposed into several intrinsic mode functions (IMFs), i.e. IMF components. The correlation coefficient between each IMF component and the original dynamic process parameter signal is calculated. Noisy IMF components with low correlation coefficients are removed, and effective IMF components are retained for subsequent Hilbert transform and feature extraction.

[0041] (3) Hilbert Transform:

[0042] Perform Hilbert transform on the retained effective IMF components to extract four types of time-domain-frequency domain features: instantaneous frequency, instantaneous amplitude, amplitude envelope, and frequency bandwidth. This can comprehensively capture the time-varying characteristics of dynamic signals and provide sufficient and reliable data support for feature screening in the subsequent step 3.

[0043] In step 3, the feature importance ranking and selection driven by the random forest algorithm is used to solve the feature redundancy problem. First, a corresponding random forest model is constructed, and the Gini coefficient importance score of each feature is calculated using surface roughness and dimensional accuracy as evaluation indicators. Then, by setting appropriate feature selection thresholds, several key features are selected to significantly reduce feature redundancy. At the same time, a dynamic matching mechanism is established to flexibly adjust the selection threshold according to sensor type, sampling frequency, and optimization target priority, so that the extracted features are adapted to the actual application scenario, improving the model's computational efficiency while ensuring feature expressive power.

[0044] The specific process of using the random forest model to rank and filter the initial feature set based on feature importance in step 3 is as follows;

[0045] First, a random forest model is constructed, and reasonable model parameters are set. Surface roughness and dimensional accuracy are used as evaluation indicators. The importance score of each feature in the initial feature set is calculated by the Gini coefficient.

[0046] Secondly, a feature selection threshold is set, and an appropriate number of key features are selected based on this threshold to construct a high-dimensional feature vector, effectively eliminating redundant features, significantly reducing feature redundancy, and improving the computational efficiency of subsequent models.

[0047] Finally, based on the data acquisition conditions (sensor type, sampling frequency) and the optimization target priority, the set feature selection threshold is dynamically adjusted to achieve dynamic feature matching. In the dynamic matching mechanism, the feature matching rules corresponding to different optimization target priorities are as follows:

[0048] When the optimization objective priority is surface roughness priority: the sensor configuration adopts a combination of cutting force sensor and vibration sensor, sets high frequency sampling specifications for the surface roughness characterization requirements, matches a high feature screening threshold, and filters out core key features to focus on sensitive information related to surface quality;

[0049] When the optimization objective priority is dimensional accuracy: the sensor configuration uses a combination of cutting force sensor and grating displacement sensor, the sampling frequency is matched according to the dimensional accuracy detection requirements, the corresponding feature filtering threshold is set, and a moderate number of key features are selected to take into account the correlation characteristics between position accuracy and cutting state.

[0050] When the optimization objective priority is tool life: the sensor configuration uses a combination of vibration sensors and cutting force sensors, sets high-frequency sampling specifications around the monitoring needs of tool wear, adopts a lower feature screening threshold, and retains more detailed features related to tool wear in order to achieve accurate prediction of tool life.

[0051] In step S4, fitness function modeling based on the improved DNN two-layer network structure is crucial for ensuring optimization accuracy. The first DNN layer takes static process parameters and key features as input to extract deep features of dynamic process parameters. The second DNN layer integrates the static process parameters with the deep dynamic features extracted by the first DNN layer, outputting fitness values ​​corresponding to surface roughness, dimensional accuracy, and tool wear. The improved DNN employs the Leaky ReLU activation function to alleviate gradient vanishing and avoids gradient explosion through gradient clipping. Combined with the Adam optimizer and L2 regularization, it significantly improves model training stability and prediction accuracy, providing a high-precision fitness evaluation basis for subsequent multi-objective optimization.

[0052] Improved two-layer network structure of DNN, such as Figure 3 As shown. This improved DNN adopts a two-layer network architecture design, introducing the Leaky ReLU activation function, gradient clipping, and L2 regularization mechanism, which can effectively overcome the shortcomings of traditional DNN models such as gradient vanishing, gradient exploding, and overfitting. Its network architecture and parameter configuration are as follows:

[0053] (1) Configuration of core network parameters.

[0054] Activation function: The Leaky ReLU activation function is selected. By setting a reasonable leakage coefficient, a nonlinear mapping relationship is constructed to achieve effective transmission of gradient information and solve the gradient vanishing problem.

[0055] Optimizer: The Adam optimizer is used, with an adaptive learning rate strategy and momentum parameters configured to achieve efficient iterative updates of network weights and ensure the convergence of model training.

[0056] Regularization: Introducing L2 regularization mechanism in the loss function A weight penalty term is added to limit the size of the weights to suppress overfitting. For network weights, The regularization coefficient is . The mean squared error is used to quantify the deviation between the model's predicted values ​​and the actual measured values, and it is the basic error term of the loss function.

[0057] Gradient control strategy: Set a gradient clipping threshold, and normalize gradient vectors that exceed the threshold to avoid gradient explosion and ensure the stability of the model training process. The formula for normalizing the gradient vector is:

[0058]

[0059] in, This is the normalized gradient vector after gradient clipping. The original gradient vector obtained through backpropagation. This represents the magnitude of the original gradient vector.

[0060] (2) The first DNN network layer is used for dynamic feature extraction.

[0061] Input layer: The input dimension is 3+m, including three types of static process parameters: spindle speed, feed per tooth, and depth of cut, which serve as the basic input features of the model. At the same time, the m-dimensional key features (high-dimensional feature vectors) obtained in step 3 are synchronously input into this layer to achieve the initial fusion of static process parameters and key features and the in-depth extraction of features. These static process parameters are the core components of the population in the subsequent multi-objective optimization in step 5.

[0062] Hidden layers: A three-layer hidden layer structure is set up, with the number of neurons in each layer in a progressive-decreasing distribution. Each layer is equipped with the Leaky ReLU activation function to realize the nonlinear mapping of static process parameters and key features to fused deep dynamic features.

[0063] Output layer: The output dimension is 8~12 dimensions. The output is based on the key features selected in step 3. After fusing three types of static process parameters, the output is a fused deep dynamic process parameter key feature set obtained by deep feature extraction and nonlinear mapping of the network.

[0064] The training process for the first DNN network layer is as follows: First, the network weights and thresholds are initialized and configured. The output results are calculated through forward propagation, and the prediction accuracy is evaluated based on the mean squared error (MSE). If the MSE meets the preset accuracy requirements, the training is terminated. Otherwise, the gradient is solved through backpropagation, and the weights and thresholds are updated after gradient clipping. The training is iterated until the error meets the set standard.

[0065] (3) The second DNN network layer is used for fitness function modeling.

[0066] Input layer: The input dimension is 11~15 dimensions. It integrates the key feature set of the fusion deep dynamic process parameters output by the first DNN network layer with three types of static process parameters: spindle speed, feed per tooth, and depth of cut. This constructs a multi-dimensional feature input space with deep coupling of dynamic and static parameters.

[0067] Hidden Layers: A three-layer hidden layer structure is set up, with the number of neurons in each layer adapted to the feature dimension scale. Each layer adopts the Leaky ReLU activation function to deeply explore the complex nonlinear relationship between fusion features (i.e., key features of fusion-type deep dynamic process parameters) and optimization objectives.

[0068] Output layer: The output dimension is 3-dimensional, corresponding to three optimization objectives: surface roughness (Ra), dimensional accuracy (IT), and tool wear (VB). This output is the fitness value corresponding to the combination of process parameters, which serves as the core evaluation basis of the NSGA-Ⅲ algorithm in step 5, thereby completing the fitness function modeling.

[0069] The training process for the second DNN network layer is as follows: follow the training logic of the first DNN layer, iterate the training until the mean squared error (MSE) reaches the preset accuracy threshold, and generate a fitness function model that meets the requirements of engineering applications.

[0070] The improved DNN in this step is developed and implemented based on the Python-TensorFlow framework. It constructs a complete code system that includes a dynamic feature extraction network and a fitness function model, covering the entire process of model building, parameter training, and result output, ensuring the stability of the model training process, the repeatability of the results, and the feasibility of engineering implementation.

[0071] In step S5, a multi-objective optimization model is first established using spindle speed, feed per tooth, and depth of cut as optimization variables, incorporating constraints on cutting power, broaching force, and machining efficiency. Then, through population initialization, simulated binary crossover and polynomial mutation, non-dominated sorting, uniform reference point association, and elite retention strategies, the Pareto optimal solution set is efficiently obtained. Compared with traditional optimization algorithms, the uniformity of the solution set distribution and convergence speed are significantly improved, ultimately achieving comprehensive optimization of machining quality, tool life, and machining efficiency.

[0072] The multi-objective optimization process based on the improved NSGA-III algorithm is as follows: Figure 4As shown. The multi-objective optimization based on the improved NSGA-Ⅲ algorithm focuses on achieving multi-objective collaborative optimization of broaching process parameters through a series of ordered steps, including genetic operations, fitness calculation, non-dominated sorting, and reference point association. Ultimately, it obtains the Pareto optimal solution set that meets engineering requirements. The specific optimization process and parameter configuration are as follows:

[0073] First, construct a multi-objective optimization model.

[0074] The multi-objective optimization model constructed in this step uses the core parameters of the broaching process as optimization variables, the machining performance index as the optimization objective, and the actual machining conditions as constraints. The specific settings are as follows:

[0075] Optimization variables: Spindle speed, feed per tooth, and depth of cut were selected as the core optimization variables, and each variable was limited to a reasonable range suitable for broaching conditions;

[0076] Optimization objective: Using a weighted multi-objective optimization approach, the objective function is:

[0077]

[0078] in, The comprehensive objective vector for multi-objective optimization of broaching process is composed of weighted individual objectives of surface roughness, dimensional accuracy, and tool wear. Represents the target vector Perform a cooperative minimization solution; The spindle speed for broaching; This is the feed per tooth of the broach; This refers to the depth of cut. This is a surface roughness prediction function, representing the surface roughness value under given process parameters; This is a dimensional accuracy prediction function, characterizing the dimensional accuracy level under given process parameters; This is a tool wear prediction function, representing the tool wear under given process parameters; These are the weighting coefficients for surface roughness, dimensional accuracy, and tool wear, respectively, and they satisfy... This allows for priority control of different optimization objectives.

[0079] Constraints: Based on the load-bearing capacity of the machining equipment, broaching process requirements, and production efficiency needs, three types of constraints are set: the first is the cutting power constraint, and the cutting power... ≤Preset threshold (where The three main constraints are: 1) cutting torque to avoid equipment overload; 2) broaching force constraint, ensuring the broaching force is less than or equal to a preset threshold to prevent broach breakage and workpiece deformation; and 3) machining efficiency constraint, ensuring machining efficiency... ≥Preset threshold (where (Number of teeth for broaching), to ensure mass production needs.

[0080] To address the priority requirements of different processing scenarios, different combinations of weighting coefficients are configured, as follows:

[0081] Surface roughness priority scenario: Configure weighting coefficients Prioritize ensuring surface processing quality;

[0082] Dimensional accuracy priority scenario: Configure weighting coefficients Prioritize ensuring the accuracy of geometric dimensions;

[0083] Tool life priority scenarios: Configure weighting coefficients Prioritize reducing tool wear and extending tool life.

[0084] This invention enhances the generalization ability of the technology by setting a dynamic matching mechanism for feature screening thresholds and a weight coefficient configuration scheme for different processing scenarios. It can adapt to different acquisition conditions and optimization objectives and has high optimization accuracy within the range of commonly used process parameters.

[0085] Finally, the improved NSGA-III algorithm is used to solve the constructed multi-objective optimization model. After initializing the algorithm population size, number of iterations, iteration termination condition, initial parameter solutions, etc., the improved NSGA-III algorithm is executed. The optimization process of the improved NSGA-III algorithm follows the logic of "population initialization - genetic operations - fitness calculation - non-dominated sorting - reference point and ideal point construction - reference point association and elite retention - iteration termination" step by step, and the specific steps are as follows:

[0086] (1) Population initialization: The Latin hypercube sampling method is used to generate a parent population of a preset size. Each individual in the population represents a set of broaching process parameters including spindle speed, feed per tooth, and depth of cut, corresponding to an independent process optimization scheme. Each individual in the population strictly meets the constraint range of the optimization variables to ensure the uniform distribution and diversity of the population individuals.

[0087] (2) Genetic operations: The genetic operation strategy combining simulated binary crossover (SBX) and polynomial mutation is adopted. Crossover operation is completed by setting the crossover probability and distribution index, and mutation operation is completed by setting the mutation probability and distribution index to generate offspring population; the parent population and offspring population are merged to form a mixed population.

[0088] (3) Fitness calculation: The improved DNN constructed in step 4 is called to calculate the fitness value of each individual in the mixed population. Specifically, each individual in the mixed population represents a set of broaching process parameters including spindle speed, feed per tooth, and depth of cut. This set of parameters is used as the 3D basic input features of the first layer of the improved DNN network. Combined with the key features obtained in step 3, deep feature fusion and extraction are completed. The fitness value is calculated through the two-layer network operation of the improved DNN model. Finally, the three optimization target parameters of surface roughness, dimensional accuracy, and tool wear corresponding to each individual in the mixed population are obtained.

[0089] (4) Non-dominated ranking: Non-dominated ranking is performed based on the fitness value of each individual, the dominance level of each individual is calculated, and individuals with the highest dominance level in the non-dominated layer are selected and retained in the archive set to achieve the initial screening of high-quality individuals.

[0090] (5) Construction of reference points and ideal points: In the three-dimensional target space, a preset number of reference points are generated uniformly using the Das-Dennis method. A corresponding reference vector is formed from the origin to each reference point. This reference vector is used to characterize the optimization direction that is uniformly distributed in the target space. The Pareto solution set is guided to converge towards the uniform and optimal direction through the reference vector. At the same time, the minimum value of each optimization target dimension (i.e., surface roughness, dimensional accuracy, tool wear) in the current population is calculated. That is, the minimum value of the corresponding target value of all individuals is taken to construct the ideal point set and translate it to the origin to provide a benchmark for the distance calculation between individuals and reference points.

[0091] (6) Reference vector association and elite retention: Calculate the vertical distance from each individual in the archive set to each reference vector, and associate each individual with the nearest reference vector; for each reference vector, select the nearest individual to retain until the optimal individual of the preset size is selected as the next generation parent population;

[0092] (7) Iteration termination: Repeat the above genetic operations, fitness calculation, non-dominated sorting, reference point and ideal point construction, reference vector association and elite retention steps until the number of iterations reaches the preset threshold, terminate the iteration process, output the final Pareto optimal solution set, and obtain the corresponding Pareto optimal broaching process parameter combination, thereby providing theoretical support for the selection of actual broaching process parameters.

[0093] The multi-objective broaching process parameter optimization method proposed in this invention, based on the improved NSGA-Ⅲ algorithm, firstly collects dynamic process parameter signals during the broaching process through a sensor array. After WPT noise reduction preprocessing, the dynamic features are extracted using the WPT-HHT joint algorithm. Then, key features are screened and redundant features are removed using random forest. Next, an improved DNN two-layer network model is constructed to model the fitness function. Finally, multi-objective optimization is achieved based on the improved NSGA-Ⅲ algorithm to obtain the Pareto optimal broaching process parameter combination. This method effectively solves the problems of poor signal processing effect, feature redundancy, model instability, and low optimization efficiency in traditional process optimization methods, and significantly improves the machining quality, tool life, and machining efficiency of the broaching process.

[0094] The invention will be further explained and illustrated below with specific examples.

[0095] When verifying the invention through examples, the Python+TensorFlow development tool can be used, combined with a piezoelectric cutting force sensor, an ICP vibration acceleration sensor, and a grating displacement sensor to build a data acquisition platform, and the multi-objective broaching process parameter optimization method can be verified through examples.

[0096] Based on the established multi-sensor data acquisition platform, the verification process sequentially executes the following steps: First, the sensor array is deployed, the installation positions of each sensor are accurately determined and debugged, and the sampling frequency is set to meet the requirements of broaching signal acquisition, acquiring various dynamic signals during the broaching process; Second, the acquired raw dynamic process parameter signals are preprocessed using WPT noise reduction, and time-frequency features are extracted using the WPT-HHT joint algorithm, and statistical features (mean, variance) are supplemented to construct an initial feature set; Next, a random forest model is constructed, key features are selected based on feature importance scores, and redundant features are eliminated; Then, an improved DNN two-layer network is built, the model is trained, and the model prediction accuracy is ensured to meet the preset requirements; Finally, the trained improved DNN model is embedded into the improved NSGA-Ⅲ algorithm framework, optimization variables, optimization objectives, and constraints are reasonably set, and the Pareto optimal solution set is obtained after iterative optimization.

[0097] The following analysis uses broaching with different optimization target priorities as an example, selecting three scenarios: surface roughness priority, dimensional accuracy priority, and tool life priority, to verify the optimization effect of the present invention.

[0098] First, a complete data acquisition system was built, and the installation and debugging of various types of sensors were completed to ensure that the data synchronization accuracy met the test requirements. The dynamic signals of broaching under different combinations of process parameters were collected, a test sample set was constructed, and the training set and test set were divided according to a preset ratio to provide data support for subsequent modeling and verification.

[0099] Second, strictly follow the optimization method steps of this invention to complete signal preprocessing, dynamic feature extraction, random forest feature screening, improved DNN model construction and training, and improved NSGA-Ⅲ multi-objective optimization in sequence. For the three different optimization objective priorities, configure the corresponding weight coefficients and feature screening thresholds respectively to ensure that the optimization process is adapted to the actual processing requirements.

[0100] Third, the optimal combination of process parameters obtained in each scenario is applied to actual broaching, and the core performance indicators such as surface roughness, dimensional accuracy, tool wear and processing efficiency after processing are recorded. The processing effect is compared and analyzed with that of traditional process parameters before optimization, so as to intuitively verify the optimization effect of the method of the present invention.

[0101] Fourth, based on actual mass production needs, the optimal combination of process parameters corresponding to the optimization target priority is selected and applied to batch broaching to verify the stability of the optimization effect; at the same time, experiments are carried out within the range of commonly used broaching process parameters to verify the generalization ability of the method of this invention and ensure that the method can be adapted to different processing conditions.

[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A multi-objective broaching process parameter optimization method based on an improved NSGA-Ⅲ algorithm, characterized in that, Includes the following steps: Step S1: Collect dynamic process parameter signals during the broaching process through a sensor array, including cutting force, vibration acceleration, real-time feed rate, and depth of cut; Step S2: Use the WPT-HHT joint algorithm to extract features from the dynamic process parameter signal, capture the time-varying characteristics of the signal, and form an initial feature set; Step S3: Construct a random forest model, sort and filter the initial feature set according to feature importance, remove redundant features, and retain key features; Step S4: Construct an improved deep neural network, which is a two-layer network structure, and introduces the Leaky ReLU activation function, gradient clipping and L2 regularization mechanism to establish a fitness function with surface roughness, dimensional accuracy and tool wear as optimization objectives; Step S5: Using spindle speed, feed per tooth, and depth of cut as optimization variables, and cutting power, broaching force, and machining efficiency as constraints, construct a multi-objective optimization model. Using the fitness function as the evaluation criterion, solve the multi-objective optimization model using the improved NSGA-Ⅲ algorithm to obtain the Pareto optimal broaching process parameter combination.

2. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1, characterized in that, The WPT-HHT joint algorithm specifically includes: Signal denoising preprocessing: The dynamic process parameter signal is denoised using a wavelet packet transform algorithm; Empirical Mode Decomposition: Perform empirical mode decomposition on the denoised signal to obtain several IMF components. Calculate the correlation coefficient between each IMF component and the original denoised signal, and remove noisy IMF components with correlation coefficients lower than a preset threshold, retaining the effective IMF components. Hilbert Transform: Perform Hilbert transform on the effective IMF components to extract four types of time-frequency features: instantaneous frequency, instantaneous amplitude, amplitude envelope, and frequency bandwidth. Combine these features with statistical characteristics to form an initial feature set.

3. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1 or 2, characterized in that, In step S3, the random forest model uses surface roughness and dimensional accuracy as evaluation indicators, calculates the Gini coefficient of each feature as an importance score, and selects key features according to the set feature selection threshold; at the same time, a dynamic matching mechanism is established to adjust the feature selection threshold according to sensor type, sampling frequency and optimization target priority.

4. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 3, characterized in that, The dynamic matching mechanism includes: When the optimization target priority is surface roughness, a combination of cutting force sensor and vibration sensor is used, a high frequency sampling specification is set, and a higher feature screening threshold is matched. When the optimization objective priority is dimensional accuracy, a combination of cutting force sensor and grating displacement sensor is used, and the sampling frequency and corresponding feature screening threshold are matched according to the dimensional accuracy detection requirements. When the optimization objective priority is tool life, a combination of vibration sensor and cutting force sensor is used, a high-frequency sampling specification is set, and a lower feature screening threshold is matched.

5. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1 or 2, characterized in that, The structure of the improved deep neural network specifically includes a first DNN network layer and a second DNN network layer; The first DNN network layer includes: Input layer: The input dimension is 3+m, including three types of static process parameters: spindle speed, feed per tooth, and depth of cut. These serve as the basic input features of the model, and the key features are also input synchronously. Hidden layers: A three-layer hidden layer structure is set up, with the number of neurons in each layer in a progressive-decreasing distribution. Each layer is equipped with the LeakyReLU activation function to realize the nonlinear mapping of static process parameters and key features to fused deep dynamic features. Output layer: The output dimension is 8~12 dimensions. The output is a fusion-type deep dynamic process parameter key feature set obtained after deep feature extraction and nonlinear mapping of the network. The second DNN network layer includes: Input layer: The input dimension is 11~15 dimensions. It integrates the static process parameters with the key feature set of the fused deep dynamic process parameters output by the first DNN network layer to construct a multi-dimensional feature input space. Hidden layers: A three-layer hidden layer structure is set up, with the number of neurons in each layer adapted to the feature dimension scale. Each layer adopts the Leaky ReLU activation function to deeply explore the complex nonlinear relationship between fused features and optimization objectives. Output layer: The output dimension is 3-dimensional, corresponding to three optimization objectives: surface roughness, dimensional accuracy, and tool wear. The output is the fitness value corresponding to the combination of process parameters.

6. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 5, characterized in that, The gradient clipping is achieved by setting a gradient clipping threshold and normalizing gradient vectors that exceed the threshold; the L2 regularization mechanism is achieved by adding a weight penalty term to the loss function.

7. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1 or 2, characterized in that, The objective function of the multi-objective optimization model is: in, To optimize the multi-objective integrated objective vector for broaching processes; The spindle speed for broaching; This is the feed per tooth of the broach; This refers to the depth of cut. This is a surface roughness prediction function; This is a dimensional accuracy prediction function; This is a function for predicting tool wear. These are the weighting coefficients for surface roughness, dimensional accuracy, and tool wear, respectively, and they satisfy... .

8. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 7, characterized in that, Based on the priority requirements of the processing scenario, different combinations of weight coefficients are configured, as follows: Surface roughness priority scenario: Configure weighting coefficients ; Dimensional accuracy priority scenario: Configure weighting coefficients ; Tool life priority scenarios: Configure weighting coefficients .

9. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1 or 2, characterized in that, The process of solving the multi-objective optimization model using the improved NSGA-Ⅲ algorithm includes the following steps: (1) Population initialization: The parent population was generated using the Latin hypercube sampling method; (2) Genetic operation: The genetic operation strategy of combining simulated binary crossover and polynomial mutation is adopted. Crossover operation is completed by setting the crossover probability and distribution index, and mutation operation is completed by setting the mutation probability and distribution index to generate offspring population, which is then merged with the parent population to form a mixed population. (3) Fitness calculation: The improved deep neural network is invoked to calculate the fitness value of each individual in the mixed population; (4) Non-dominated ranking: Based on the fitness value of each individual, the dominance level of each individual is calculated by non-dominated ranking, and individuals with the highest dominance level in the non-dominated layer are selected and retained in the archive set; (5) Construction of reference points and ideal points: In the three-dimensional target space, reference points are uniformly generated using the Das-Dennis method. At the same time, the individual minimum values ​​of each optimization target dimension are calculated, and the set of ideal points is constructed and translated to the origin of the coordinate system. (6) Reference vector association and elite retention: Calculate the vertical distance from each individual in the archive set to each reference vector, associate the individual with the nearest reference vector, and select the closest individual for each reference vector to retain, forming the next generation parent population; (7) Iteration termination: Repeat steps (2) to (6) above until the number of iterations reaches the preset threshold, terminate the iteration process, output the final Pareto optimal solution set, and obtain the corresponding Pareto optimal broaching process parameter combination.

10. The multi-objective broaching process parameter optimization method based on the improved NSGA-Ⅲ algorithm according to claim 1 or 2, characterized in that, The sensor array includes a piezoelectric cutting force sensor, an ICP vibration acceleration sensor, and a grating displacement sensor.

Images