A control method based on real-time broaching force feedback adjustment and a broaching machine system

Abstract

The application discloses a control method and broaching machine system based on real-time broaching force feedback adjustment, relates to the technical field of broaching machining, and comprises the following steps: constructing a dynamic state vector and inputting a pre-trained mixed multi-neural network prediction model to output a broaching force prediction sequence in a future preset time window; a prediction force change rate vector is generated; if the prediction force change rate vector is greater than a fracture early warning threshold, it is determined that the broaching tool is about to be fractured, the power supply of a main drive motor is cut off, and hydraulic clamping braking is triggered; if the prediction force change rate vector is smaller than a smooth running threshold, a speed optimization space is constructed based on the current feeding speed; a multi-objective fitness function containing an energy consumption item and a stability item is constructed; an optimal feeding speed node is output; a target lower broaching speed instruction is generated; a speed adjustment decision sequence is generated; a pulse signal with S-shaped acceleration and deceleration characteristics is generated; and the pulse signal is used to drive the main drive motor of the broaching machine; and the application improves the broaching machining quality and reduces the safety risk.

Description

Technical Field

[0001] This invention relates to the field of broaching technology, specifically to a control method and broaching machine system based on real-time broaching force feedback adjustment. Background Technology

[0002] Broaching, as a high-precision and high-efficiency metal cutting process, is widely used in high-end equipment fields such as aerospace, precision machinery, and automobile manufacturing. Its machining quality directly determines the assembly accuracy and service reliability of the workpiece. During broaching, the cutting interaction between the broach and the workpiece generates complex broaching forces. The dynamic change of these forces is closely related to factors such as broach wear, workpiece material inhomogeneity, and unreasonable cutting parameters. These forces are the core factors affecting the stability of broaching, tool life, and workpiece machining quality. Existing broaching machine control technologies mostly employ fixed feed rates or simple real-time force feedback adjustment modes, which can only passively respond to sudden changes in the current broaching force and cannot predict future broaching force trends, exhibiting significant limitations. On one hand, traditional control methods lack the fusion and utilization of multi-source sensor information, relying solely on a single broaching force signal for adjustment. This makes them susceptible to sensor noise and environmental interference, resulting in low force detection accuracy, lag in adjustment, and difficulty in coping with dynamic fluctuations in cutting load during broaching, often leading to problems such as broach chatter and workpiece surface quality defects. On the other hand, for extreme conditions where the broach is about to break, traditional methods can only trigger a shutdown after a significant drop in force, failing to provide early warning and easily causing broach damage, workpiece scrap, or even equipment safety accidents. Furthermore, existing broaching speed adjustments mostly rely on empirical settings, failing to incorporate intelligent optimization based on broaching force variation patterns, resulting in excessive energy consumption or low processing efficiency during machining, making it difficult to balance machining stability and energy economy. Therefore, there is an urgent need for a broaching force feedback regulation and control method and system that can integrate multimodal sensor data, realize early prediction of broaching force, intelligently optimize feed speed, and have fracture early warning function, so as to solve the technical pain points of lag and low prediction accuracy in the existing technology. Summary of the Invention

[0003] To solve the above technical problems, a control method and broaching machine system based on real-time broaching force feedback adjustment are provided. This technical solution solves the above problems.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A control method based on real-time broaching force feedback adjustment includes: The spindle motor current signal, bed vibration acceleration signal, and ambient temperature field distribution during the broaching process are collected as multimodal sensing data; a dynamic state vector is constructed and input into a pre-trained hybrid multi-neural network prediction model, which outputs a broaching force prediction sequence within a preset time window; the hybrid multi-neural network prediction model integrates convolutional neural networks, long short-term memory networks, and residual networks; Calculate the force difference between adjacent time steps in the broaching force prediction sequence to generate a vector of predicted force change rate; If the predicted force change rate vector is greater than the fracture warning threshold, it is determined that the cutter is about to break, the power supply to the main drive motor is cut off, and the hydraulic clamping brake is triggered. If the predicted force change rate vector is less than the steady-state operation threshold, a speed optimization space is constructed based on the current feed rate. Based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence, a multi-objective fitness function including energy consumption and stability terms is constructed. A multi-objective fitness function is used to perform particle swarm optimization iteration on the velocity optimization space, and the optimal feed velocity node is output. Extract the speed scalar corresponding to the optimal feed rate node, combine it with the broaching force prediction sequence fluctuation trend to perform feedforward compensation, and generate the target pull-down speed command; The target pull-down speed command is combined according to the time step to generate a speed adjustment decision sequence; the speed adjustment decision sequence is parsed to obtain the servo driver frequency setpoint, and a pulse signal with S-shaped acceleration and deceleration characteristics is generated; The main drive motor of the broaching machine is driven by a pulse signal to achieve adaptive constant force control during the broaching process.

[0005] Furthermore, a dynamic state vector is constructed and input into a pre-trained hybrid multi-neural network prediction model, including: Variational mode decomposition is performed on the spindle motor current signal to extract the eigenmode function components containing cutting force characteristics; A short-time Fourier transform is performed on the bed vibration acceleration signal to extract the frequency domain energy spectrum features; The intrinsic mode function components, frequency domain energy spectrum characteristics, and ambient temperature field distribution are mapped into a three-dimensional tensor. Convolutional neural networks are used to extract the spatial coupling features of three-dimensional tensors to eliminate sensor noise interference. Spatial coupling features are input into a long short-term memory network to capture the time-dependent evolution of the broaching force; A residual network module is connected at the end of the network. By skipping connections, gradient vanishing is avoided, and the tensile force prediction sequence is output.

[0006] Furthermore, a multi-objective fitness function containing energy consumption and stability terms is constructed, including: Extract the average force component and peak force component from the broaching force prediction sequence; The ratio of the average force component to the peak force component is calculated to obtain the force stability index; The average force component, the force stability index, and the cutting specific energy coefficient are weighted and summed using preset weighting coefficients to generate a multi-objective fitness function; among them, the energy consumption term is positively correlated with the average force component, and the stability term is negatively correlated with the force stability index.

[0007] Furthermore, particle swarm optimization iterations are performed on the velocity optimization space using a multi-objective fitness function, including: Initialize the particle swarm's position and velocity; each particle represents a candidate feed velocity node in the velocity optimization space. Calculate the local fitness value of each particle and update the individual's best historical position and the global best historical position; An adaptive inertia weight is introduced, which is dynamically adjusted based on the convergence degree of the global historical best position; When the force stability index is lower than the preset vibration critical value, the learning factor of the velocity component is increased to guide the particles to search in the low-velocity region to suppress flutter; The global historical best position at the end of the output iteration is used as the optimal feed rate node.

[0008] Furthermore, by combining the fluctuation trend of the broaching force prediction sequence with feedforward compensation, a target pull-down speed command is generated, including: Calculate the slope of the broaching force prediction sequence within a future preset time window; If the slope is greater than zero, it is determined that the cutting load is about to increase, and the feed rate is reduced in advance according to the slope ratio; When the slope is less than zero, it is determined that the cutting load is about to decrease, and the feed rate is increased in advance according to the slope ratio; The feedforward compensation is superimposed on the speed scalar corresponding to the optimal feed speed node, and after low-pass filtering, the target pull-down speed command is generated.

[0009] Furthermore, multimodal sensing data is collected during the broaching process, including: The raw signal of the three-phase current of the main shaft is acquired at a high-frequency sampling rate using a Hall current sensor; Vibration signals at the spindle box of the broaching machine were collected using a triaxial piezoelectric accelerometer. The temperature distribution matrix of the broach extension end and the workpiece clamping area was acquired using an infrared thermal imager array. The Clarke and Park transformations are performed on the original three-phase current signals to decouple and obtain the torque current component, which is used as the spindle motor current signal.

[0010] Furthermore, after determining that the broach is about to break, cutting off the power supply to the main drive motor and triggering the hydraulic clamping brake, the following steps are also included: Record multimodal sensing data and broaching force prediction sequence fragments prior to fracture; The recorded data fragments are marked as fault samples and stored in the local database; When the number of fault samples accumulates to a preset number, the online fine-tuning training of the hybrid multi-neural network prediction model is triggered, and the model weights are updated to adapt to the cutting characteristics after tool wear.

[0011] Furthermore, the servo drive frequency setpoint is obtained by analyzing the speed adjustment decision sequence, and a pulse signal with S-shaped acceleration / deceleration characteristics is generated, including: Calculate the acceleration requirement based on the difference in target pull-down speed commands between adjacent time steps in the speed adjustment decision sequence; If the acceleration requirement exceeds the preset mechanical impact threshold, an intermediate transition velocity node is inserted. Based on the initial frequency, target frequency, and inserted intermediate transition velocity nodes, a seven-segment S-curve planning algorithm is used to generate the pulse signal frequency change trajectory; The pulse signal frequency change trajectory is converted into a pulse width modulation signal with varying duty cycle, which is then input to the servo controller of the broaching machine's main drive motor.

[0012] Furthermore, after outputting the optimal feed rate node, it also includes: Substitute the optimal feed rate node into the broaching process dynamics model to calculate the elastic deformation at the broach overhang end; The upper boundary of the shrinkage rate optimization space when the elastic deformation exceeds the workpiece tolerance; Repeat the particle swarm optimization iteration until the elastic deformation meets the workpiece tolerance requirements.

[0013] A broaching machine system based on real-time broaching force feedback adjustment includes: The multimodal sensing module collects spindle motor current signals, bed vibration acceleration signals, and ambient temperature field distribution during the broaching process. The broaching force prediction module constructs a dynamic state vector and inputs it into a pre-trained hybrid multi-neural network prediction model, outputting a broaching force prediction sequence within a preset future time window; The safety monitoring module calculates the difference in force values ​​between adjacent time steps in the broaching force prediction sequence and generates a predicted force change rate vector. If the predicted force change rate vector is greater than the fracture warning threshold, the power supply to the main drive motor is cut off and the hydraulic clamping brake is triggered. The optimization space construction module constructs a velocity optimization space based on the current feed rate if the predicted force change rate vector is less than the steady-state operation threshold. The fitness function construction module constructs a multi-objective fitness function based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence. The velocity optimization module uses a multi-objective fitness function to perform particle swarm optimization iterations on the velocity optimization space and outputs the optimal feed velocity node. The feedforward compensation module extracts the speed scalar corresponding to the optimal feed rate node, performs feedforward compensation based on the fluctuation trend of the broaching force prediction sequence, and generates the target pull-down speed command; The decision sequence generation module generates a speed adjustment decision sequence by combining the target pull-down speed command according to the time step. The signal modulation module analyzes the speed adjustment decision sequence to obtain the servo driver frequency setpoint and generates a pulse signal with S-shaped acceleration and deceleration characteristics. The motor drive module uses pulse signals to drive the main drive motor of the broaching machine, thereby achieving adaptive constant force control during the broaching process.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention collects multimodal sensing data on spindle motor current, bed vibration acceleration, and ambient temperature field. Through a hybrid multi-neural network prediction model integrating convolutional neural networks, long short-term memory networks, and residual networks, it accurately outputs a predicted broaching force sequence within a preset time window. This allows for early detection of broaching force trends, avoiding the drawbacks of traditional methods that only passively respond to current force changes. It effectively suppresses broach chatter and reduces workpiece surface quality defects. 2. By calculating the difference in force values ​​between adjacent time steps in the broaching force prediction sequence, a predicted force change rate vector is generated. When this vector exceeds the fracture warning threshold, it can be determined in advance that the broach is about to break, and the power supply to the main drive motor can be cut off in time to trigger the hydraulic clamping brake. Compared with the traditional method of stopping the machine after a sudden drop in force, the braking response time is significantly shortened, avoiding broach damage, workpiece scrapping, and equipment failure, improving the safety of the machining process, and reducing production costs. 3. When the broaching force is stable, a speed optimization space is constructed. Combining the mean and variance of the broaching force prediction sequence and the workpiece cutting energy coefficient, a multi-objective fitness function containing energy consumption and stability terms is constructed. The optimal feed rate node is output through particle swarm optimization iteration, and feedforward compensation is performed in combination with the broaching force fluctuation trend to achieve adaptive adjustment of the feed rate. Under the premise of ensuring machining stability, the machining efficiency is maximized and the machining energy consumption is reduced, achieving the dual goals of energy saving and high efficiency. 4. This invention integrates multi-source sensor data and extracts effective features through methods such as variational mode decomposition and short-time Fourier transform. Combined with the spatial feature extraction and time-dependent capture capabilities of a hybrid multi-neural network model, it effectively eliminates sensor noise interference, avoids gradient vanishing, and significantly improves the accuracy of broaching force prediction. At the same time, by driving the motor with pulse signals of S-shaped acceleration and deceleration characteristics, it achieves smooth speed regulation, reduces mechanical impact, and extends the service life of equipment and cutting tools. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the steps of the present invention. Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0016] The following description is intended to disclose the invention so that those skilled in the art can implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 As shown, a control method based on real-time broaching force feedback adjustment includes: Step S101: Collect spindle motor current signal, bed vibration acceleration signal, and ambient temperature field distribution during the broaching process as multimodal sensing data; construct a dynamic state vector and input it into a pre-trained hybrid multi-neural network prediction model, outputting a broaching force prediction sequence within a preset future time window; the hybrid multi-neural network prediction model integrates convolutional neural network, long short-term memory network, and residual network; Step S102: Calculate the force value difference between adjacent time steps in the broaching force prediction sequence to generate the predicted force change rate vector; Step S103: If the predicted force change rate vector is greater than the fracture warning threshold, it is determined that the cutter is about to break, the power supply to the main drive motor is cut off and the hydraulic clamping brake is triggered; Step S104: If the predicted force change rate vector is less than the steady-state operation threshold, construct a speed optimization space based on the current feed rate; Step S105: Based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence, construct a multi-objective fitness function that includes energy consumption and stability terms; Step S106: Perform particle swarm optimization iteration on the velocity optimization space using a multi-objective fitness function, and output the optimal feed velocity node; Step S107: Extract the speed scalar corresponding to the optimal feed speed node, combine it with the broaching force prediction sequence fluctuation trend to perform feedforward compensation, and generate the target pull-down speed command; Step S108: Combine the target pull-down speed command according to the time step to generate a speed adjustment decision sequence; parse the speed adjustment decision sequence to obtain the servo driver frequency setpoint, and generate a pulse signal with S-shaped acceleration and deceleration characteristics; Step S109: Use pulse signals to drive the main drive motor of the broaching machine to complete the adaptive constant force control of the broaching process.

[0018] The above scheme uses multimodal sensor data to compensate for the limitations of single signal detection, and combines a hybrid multi-neural network prediction model to achieve early prediction of broaching force, solving the problem of adjustment lag in traditional control methods; it achieves early warning of broach breakage by predicting the force change rate vector, reducing safety risks; it achieves intelligent optimization of feed rate based on multi-objective fitness function and particle swarm optimization, taking into account both processing efficiency and energy economy; and it improves the smoothness of speed regulation and reduces mechanical impact through feedforward compensation and S-shaped acceleration and deceleration control, ultimately achieving adaptive constant force control in the broaching process, improving processing quality and equipment reliability.

[0019] In some embodiments, step S101, acquiring multimodal sensing data during the broaching process, includes: The raw signal of the three-phase current of the main shaft is acquired at a high-frequency sampling rate using a Hall current sensor; Vibration signals at the spindle box of the broaching machine were collected using a triaxial piezoelectric accelerometer. The temperature distribution matrix of the broach extension end and the workpiece clamping area was acquired using an infrared thermal imager array. The Clarke and Park transforms are performed on the original three-phase current signals to decouple and obtain the torque current component, which is used as the spindle motor current signal. Specifically, a closed-loop Hall current sensor with an accuracy class of 0.1 and a measurement range of 0-50A is selected and installed on the power supply line of the broaching machine spindle motor. The sampling frequency is set to 200-500Hz to ensure real-time capture of the dynamic changes in the spindle motor current. This current signal is positively correlated with the broaching force and can indirectly reflect the magnitude of the cutting load. A triaxial piezoelectric accelerometer with a sensitivity of 100mV / g and a measurement range of ±50g is selected and fixedly installed on the broaching machine spindle box near the broach. It is used to collect the vibration signal of the machine bed during broaching. The amplitude and frequency changes of the vibration signal can reflect abnormal working conditions such as broach wear and cutting chatter. The infrared thermal imager array consists of 4-8 infrared thermal imagers, evenly arranged around the broaching area. The acquisition range covers the broach extension end, the workpiece clamping area, and the cutting contact area. The acquisition frequency is 50-100Hz, generating a temperature distribution matrix with a resolution of 640×480 to reflect the influence of ambient temperature and cutting heat on the broaching process. The acquired raw three-phase spindle current signals are decoupled using Clarke and Park transforms: First, the Clarke transform converts the three-phase currents (Ia, Ib, Ic) into current components (Iα, Iβ) in a two-phase stationary coordinate system, eliminating the coupling relationship of the three-phase currents; then, the Park transform converts the current components (Iα, Iβ) in the two-phase stationary coordinate system into current components (Id, Iq) in a two-phase rotating coordinate system, where Id is the excitation current component and Iq is the torque current component; the torque current component Iq is extracted as the spindle motor current signal, which directly reflects the change in spindle output torque and is linearly correlated with the broaching force, effectively eliminating the interference of the excitation current and improving signal correlation.

[0020] In some embodiments, step S101, constructing a dynamic state vector and inputting it into a pre-trained hybrid multi-neural network prediction model, includes: Variational mode decomposition is performed on the spindle motor current signal to extract the eigenmode function components containing cutting force characteristics; A short-time Fourier transform is performed on the bed vibration acceleration signal to extract the frequency domain energy spectrum features; The intrinsic mode function components, frequency domain energy spectrum characteristics, and ambient temperature field distribution are mapped into a three-dimensional tensor. Convolutional neural networks are used to extract the spatial coupling features of three-dimensional tensors to eliminate sensor noise interference. Spatial coupling features are input into a long short-term memory network to capture the time-dependent evolution of the broaching force; A residual network module is connected at the end of the network. Skip connections are used to avoid gradient vanishing, and the predicted tensile force sequence is output. Specifically, variational mode decomposition (VMD) is performed on the spindle motor current signal. The number of decomposition modes is set to 3-5, the penalty factor is 2000, and the convergence accuracy is 10 to the power of -7. By iteratively updating the modal components and center frequency, the current signal is decomposed into multiple intrinsic mode functions (IMF) components. Modal components containing high-frequency noise are removed, and 2-3 IMF components related to the cutting force change are retained. These components can effectively reflect the dynamic change characteristics of the broaching force. A short-time Fourier transform (STFT) is performed on the bed vibration acceleration signal. The sliding window length is set to 256, the overlap rate to 50%, and the sampling frequency is consistent with the vibration signal acquisition frequency. The time-domain vibration signal is converted into a frequency-domain signal, and the frequency-domain energy spectrum features in the 10-1000Hz frequency band are extracted. The energy proportion of each frequency band is calculated, and a frequency-domain energy spectrum feature vector with a dimension of 1 row and 100 columns is generated. This feature can reflect the frequency distribution of the vibration signal and indirectly characterize the broach wear and cutting chatter state. The extracted intrinsic mode function components (1 row N columns, where N is the number of sampling points), frequency domain energy spectrum feature vector (1 row 100 columns), and environmental temperature field distribution matrix (640 rows 480 columns) are dimension-unified. The temperature distribution matrix is ​​then reduced to 1 row 100 columns through linear interpolation. Subsequently, the three are concatenated by channel and mapped to a three-dimensional tensor with a dimension of 3 rows 100 columns. This tensor contains the spatial coupling information of multiple source features of current, vibration, and temperature during the broaching process. The structure of the hybrid multi-neural network prediction model is as follows: The Convolutional Neural Network (CNN) part contains two convolutional layers and one pooling layer. The convolutional layers use 3×3 convolutional kernels with 32 and 64 kernels respectively, and ReLU is used as the activation function. The pooling layer uses 2×2 max pooling to extract spatial coupling features in the three-dimensional tensor and eliminate sensor noise interference. The Long Short-Term Memory (LSTM) network part contains one hidden layer with 128 neurons and a dropout coefficient of 0.3 to capture the time-dependent evolution of the drawing force and adapt to the temporal characteristics of the drawing process. The Residual Network (ResNet) module contains two residual blocks. Each residual block consists of two convolutional layers and one skip connection. The skip connection directly transmits the input features to the output, avoiding the gradient vanishing problem in the training process of deep networks and improving the model's prediction accuracy. Model pre-training process: Collect multimodal sensing data and corresponding actual broaching force data of different workpiece materials (such as aluminum alloy and titanium alloy) and different broach wear states to construct a training dataset with a training set to test set ratio of 8:2; use the Adam optimizer with a learning rate of 0.001, 100 iterations, and mean squared error as the loss function; update the model weights through backpropagation until the model's prediction error on the test set is less than 5%, thus completing pre-training; input the constructed three-dimensional tensor into the pre-trained hybrid multi-neural network prediction model to output a broaching force prediction sequence within a preset time window. The preset time window is set to 0.5-1s, and the time step is 0.01s, i.e., outputting 50-100 broaching force prediction values ​​to form a broaching force prediction sequence.

[0021] In some embodiments, step S102 involves calculating the difference in force values ​​between adjacent time steps of the broaching force prediction sequence to generate a predicted force change rate vector, specifically including: Let the broaching force prediction sequence be F, which contains the broaching force prediction values ​​from the first time step to the nth time step, and be denoted as F1, F2 and so on up to F. n , where F i This represents the predicted tensile force at the i-th time step, where i ranges from 1 to n, and n is the total length of the prediction sequence. Calculate the difference between the predicted values ​​of the cutting force at two adjacent time steps, that is, the predicted value at the (i+1)th time step minus the predicted value at the ith time step, to obtain the force difference ΔF between each adjacent time step. i , where the value of i ranges from 1 to n-1; Difference of all calculated force values ​​ΔF i Arranged sequentially according to the time steps, a predictive force change rate vector ΔF is formed. This vector contains the difference from the first difference ΔF1 to the (n-1)th difference ΔF. n-1 It can intuitively reflect the changing trend of the broaching force. A positive value indicates that the broaching force is increasing, a negative value indicates that the broaching force is decreasing, and the larger the absolute value, the more drastic the change in the broaching force.

[0022] In some embodiments, in step S103, if the predicted force change rate vector is greater than the fracture warning threshold, it is determined that the cutter is about to break, the power supply to the main drive motor is cut off, and the hydraulic clamping brake is triggered, including: Preset fracture warning threshold ΔF th1 This threshold is pre-set based on the broach material, diameter, and broaching parameters, and its value ranges from -500N per 0.01 seconds to -300N per 0.01 seconds (the negative sign indicates a sudden drop in broaching force). Examine each element ΔFi in the predictor force change rate vector ΔF one by one. If any ΔF exists... i Less than the preset fracture warning threshold ΔF th1 If the broaching force suddenly drops beyond the warning threshold, it is determined that the broach is about to break. Immediately output a control signal to cut off the power supply circuit of the broaching machine's main drive motor, and at the same time trigger the hydraulic clamping and braking system to control the hydraulic cylinder to push the brake caliper to clamp the broach spindle, thereby achieving an instantaneous stop of the broaching machine's downward movement, with a braking response time of no more than 0.05s.

[0023] After determining that the broach is about to break, cutting off the power supply to the main drive motor and triggering the hydraulic clamping brake, the process also includes: Record multimodal sensor data (spindle motor current signal, bed vibration acceleration signal, ambient temperature field distribution) and broaching force prediction sequence fragments within 10 seconds before fracture occurs; The recorded data segments are marked as fault samples, labeled with the fault type "broker about to break," and stored in a local database. The database uses SQLite, which supports the storage, querying, and retrieval of samples. When the number of fault samples accumulates to a preset number (e.g., 50), the online fine-tuning training of the hybrid multi-neural network prediction model is triggered. The fine-tuning training adopts the mini-batch gradient descent method, with a batch size of 10, a learning rate of 0.0001, and 20 iterations. The model weights are updated to adapt to the cutting characteristics after tool wear, thereby further improving the prediction accuracy.

[0024] In some embodiments, in step S104, if the predicted force change rate vector is less than the steady-state operation threshold, a velocity optimization space is constructed based on the current feed rate, specifically including: Preset stable operation threshold ΔF th2 This threshold is set in advance according to the requirements of the broaching process, and its value ranges from -50N per 0.01 seconds to 50N per 0.01 seconds. If all elements ΔF in the predicted force change rate vector ΔF... i The absolute values ​​are all less than the stable operation threshold ΔF th2 If the broaching process is stable, the cutting load fluctuation is small. The current actual feed speed v0 of the broaching machine is obtained. This speed is fed back in real time by the servo driver and its value ranges from 0.1 to 1 m / min. Based on the current feed rate v0, the speed adjustment range is set to 80% to 120% of the current speed. That is, the lower limit of the speed optimization space is v0 multiplied by 0.8, and the upper limit is v0 multiplied by 1.2, so as to avoid the cutting instability caused by excessive speed adjustment. The velocity optimization space is divided into several candidate feed velocity nodes with an interval of 0.01 m / min between nodes, forming a discrete velocity optimization space, which provides candidate solutions for subsequent particle swarm optimization iterations.

[0025] In some embodiments, in step S105, based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence, a multi-objective fitness function including an energy consumption term and a stability term is constructed, including: Extract the broaching force prediction sequence F (including F1, F2 up to F) n The average force component F in ) ave and peak force component F peak The average force component is calculated as follows: sum the predicted values ​​of all time steps in the broaching force prediction sequence, and then divide by the total length n of the prediction sequence; the peak force component is the maximum value of all predicted values ​​in the broaching force prediction sequence. The ratio of the average force component to the peak force component is calculated to obtain the force stability index S. The value of S ranges from 0 to 1. The closer S is to 1, the more stable the broaching force and the better the cutting stability. Obtain the workpiece cutting energy coefficient K. This coefficient is preset according to the workpiece material and broach type, and its value ranges from 1 to 5 J / mm3. The larger the cutting energy coefficient, the more energy is consumed per unit volume of material during the cutting process, and the higher the energy consumption. The average force component F is calculated using preset weighting coefficients. aveThe force stability index S and the cutting specific energy coefficient K are weighted and summed to generate a multi-objective fitness function; among them, the energy consumption term is positively correlated with the average force component, and the stability term is negatively correlated with the force stability index; The multi-objective fitness function is constructed as follows: the fitness function value f(v) is equal to the first weight coefficient ω1 multiplied by the average force component F of the corresponding candidate feed velocity node v. ave (v), plus the second weighting coefficient ω2 multiplied by (1 minus the force stability index S(v) of the corresponding candidate feed speed node v), plus the third weighting coefficient ω3 multiplied by the workpiece cutting specific energy coefficient K; where v is the candidate feed speed node, ω1, ω2, ω3 are weighting coefficients, and the sum of the three weighting coefficients is 1. They can be set in advance according to the processing requirements. Usually, the values ​​are ω1=0.4 (energy consumption weight), ω2=0.4 (stability weight), and ω3=0.2 (material weight). The smaller the fitness function value, the better the candidate feed speed node is, which can take into account both energy consumption economy and cutting stability.

[0026] In some embodiments, step S106 involves performing particle swarm optimization iterations on the velocity optimization space using a multi-objective fitness function to output the optimal feed velocity node, including: Initialize the position and velocity of the particle swarm. Each particle corresponds to a candidate feed velocity node in the velocity optimization space. The particle swarm size is set to 30-50. The initial position of the particles is randomly generated within the velocity optimization space (between 80% and 120% of the current velocity v0). The initial velocity is set to 0-0.05 m / min. Calculate the fitness value corresponding to each particle, that is, based on the candidate feed velocity node v represented by the particle, substitute it into the multi-objective fitness function mentioned above to calculate the fitness value corresponding to the node; Update the individual historical best position and global historical best position for each particle: the individual historical best position is the position with the minimum fitness value for that particle in the current iteration; the global historical best position is the position with the minimum fitness value for all particles in the current iteration. An adaptive inertia weight w is introduced, and its magnitude is dynamically adjusted based on the convergence degree of the global historical best position. The inertia weight is calculated as follows: the inertia weight w is equal to the maximum inertia weight w. max Subtract (maximum inertia weight w) max Subtract the minimum inertia weight w min Multiply by (the ratio of the current iteration number t to the maximum iteration number T); where the maximum inertia weight w max Set to 0.9, minimum inertia weight w minThe value is set to 0.4, t is the current iteration number, and T is the maximum iteration number (set to 50~100 rounds). In the early stage of iteration, the inertia weight w is larger to enhance the global search capability, and in the later stage of iteration, the inertia weight w is smaller to enhance the local search capability. Preset vibration critical value S th The value ranges from 0.7 to 0.8. If the force stability index S is less than the critical vibration value S... th This indicates a risk of chattering during the cutting process. In this case, the learning factors c1 and c2 of the velocity components (where c1 is the individual learning factor and c2 is the global learning factor) are increased. Typically, c1 and c2 are adjusted from 2.0 to 2.5 to guide the particles to search in the low-speed region in order to suppress cutting chattering. Update the particle's position and velocity. The position update method is: the particle position x in the next iteration. i+1 Equal to the current particle position x i Add the particle velocity v of the next iteration i+1 The velocity update method is: the particle velocity v in the next iteration. i+1 It equals the current inertial weight w multiplied by the current particle velocity v i Add the individual learning factor c1 multiplied by a random number r1 between 0 and 1, and then multiply by (the particle's individual historical best position p). i Subtract the current particle position x i ), plus the global learning factor c2 multiplied by a random number r2 between 0 and 1, and then multiplied by (the global historical best position g minus the current particle position x). i ); where r1 and r2 are random numbers between 0 and 1, p i Let be the individual historical best position of particle i, and g be the global historical best position; Determine whether the iteration has reached the preset maximum number of iterations T, or whether the change in fitness value of the global historical best position is less than the preset convergence threshold (e.g., 10 to the power of -4). If either of the above two conditions is met, stop the iteration and output the global historical best position at the time of iteration termination as the optimal feed rate node. After outputting the optimal feed rate node, it also includes: Substituting the optimal feed rate node into the broaching process dynamics model, the elastic deformation at the broach overhang is calculated. The calculation method for the elastic deformation at the broach overhang is: the elastic deformation δ equals (the average force component F of the broaching force prediction sequence). ave Multiply by the cube of the broach overhang length L, then divide by (3 multiplied by the elastic modulus E of the broach material, then multiplied by the moment of inertia I of the broach section); where δ is the elastic deformation at the broach overhang end, and F is the average force component F of the broaching force prediction sequence. ave L is the broach overhang length, E is the elastic modulus of the broach material, and I is the moment of inertia of the broach section; Preset workpiece tolerance allowable value δ th The value is set according to the workpiece machining accuracy requirements, and the range is 0.001~0.01mm; If the calculated elastic deformation δ is greater than the preset workpiece tolerance value δ th The upper limit of the speed optimization space is narrowed by adjusting it to 1.05 times the current optimal feed rate node, thus reducing the speed optimization range. Repeat the particle swarm optimization iteration steps described above until the calculated elastic deformation δ is less than or equal to the preset workpiece tolerance value δ. th This ensures that the elastic deformation of the broach meets the workpiece tolerance requirements and improves machining accuracy.

[0027] In some embodiments, in step S107, the speed scalar corresponding to the optimal feed rate node is extracted, and feedforward compensation is performed in conjunction with the broaching force prediction sequence fluctuation trend to generate a target pull-down speed command, including: Extract the velocity scalar v corresponding to the optimal feed rate node opt This scalar is the global optimal position output by the particle swarm optimization iteration, which is also the optimal feed rate value; The slope k of the broaching force prediction sequence within a future preset time window is calculated. The slope k is obtained through a linear regression fitting method. The specific calculation method is as follows: Let the broaching force prediction sequence F include F1, F2, up to F... n The corresponding time series t includes t1, t2 up to t n The slope k equals (total predicted sequence length n multiplied by all time steps t). i Corresponding to the predicted value of the broaching force F i The sum of the products, minus all time steps t i The sum of all predicted tensile forces F i (The product of the sums), then divided by (the total predicted sequence length n multiplied by all time steps t) i The sum of squares, minus all time steps t i The slope k reflects the changing trend of the broaching force; k > 0 indicates that the broaching force is about to increase, and k < 0 indicates that the broaching force is about to decrease. If the slope k is greater than 0, it is determined that the cutting load is about to increase. The feed rate is reduced in advance according to the slope ratio. The calculation method of the feedforward compensation amount Δv1 is: the negative slope k multiplied by the compensation coefficient k1, and then multiplied by the optimal feed rate scalar v. opt Where k1 is the compensation coefficient, with a value ranging from 0.1 to 0.3, and the negative sign indicates a reduction in feed rate; If the slope k is less than 0, it is determined that the cutting load is about to decrease. The feed rate is increased in advance according to the slope ratio. The calculation method for the feedforward compensation Δv2 is: the absolute value of the slope k multiplied by the compensation coefficient k1, and then multiplied by the optimal feed rate scalar v. opt A plus sign indicates an increase in feed rate; The calculated feedforward compensation (Δv1 or Δv2) is superimposed on the velocity scalar v corresponding to the optimal feed rate node. opt The initial velocity command v' is obtained, meaning the initial velocity command v' is equal to the optimal feed rate scalar v. opt Add the feedforward compensation amount Δv (Δv is Δv1 or Δv2); The initial speed command v' is low-pass filtered using a Butterworth low-pass filter with a cutoff frequency of 10Hz to eliminate high-frequency noise in the speed command and prevent excessive speed fluctuations. After filtering, the target pull-down speed command v' is generated. tar9et .

[0028] In some embodiments, in step S108, a speed adjustment decision sequence is generated by combining the target pull-down speed command according to the time step; the speed adjustment decision sequence is parsed to obtain the servo driver frequency setpoint, and a pulse signal with S-shaped acceleration / deceleration characteristics is generated, including: Set the time step to 0.01s, and in chronological order of the time steps, assign the target pull-down speed command v to each time step. tar9et By combining the results, a velocity adjustment decision sequence V is generated, which contains the target velocity v from the first time step. t1 The target velocity v at the second time step t2 Up to the target velocity v at the m-th time step tm Where m is the number of time steps, determined based on the broaching process duration; The speed adjustment decision sequence V is analyzed, and the servo driver frequency setpoint f corresponding to each time step is obtained based on the mapping relationship between the speed of the main drive motor of the broaching machine and the feed speed. i The mapping relationship is calculated as follows: servo drive frequency setpoint f i Equals (target pull-down speed v at the current time step) ti Multiply by the gearbox transmission ratio i), then divide by (π multiplied by the diameter of the broach feed screw d); where i is the gearbox transmission ratio and d is the diameter of the broach feed screw; Based on the difference Δv' between the target pull-down speed commands of two adjacent time steps in the speed adjustment decision sequence (i.e., the target speed v at the (i+1)th time step) ti+1 Subtract the target velocity v at the i-th time step ti), calculate the acceleration demand 'a', which is equal to the velocity difference Δv' divided by the time step Δt (the time step Δt is 0.01s); Preset mechanical impact threshold a th The value ranges from 0.5 to 1 m / s². If the calculated acceleration requirement 'a' is greater than the mechanical impact threshold 'a', th This indicates that the speed change is too rapid and is prone to mechanical shock. In this case, an intermediate transition speed node is inserted. The number of transition speed nodes is determined by the difference between the acceleration requirement and the mechanical shock threshold. The larger the difference, the more transition nodes are needed to ensure a smooth acceleration transition. Based on the initial frequency f0 (the current operating frequency of the servo drive) and the target frequency f t The pulse signal frequency change trajectory is generated using a seven-segment S-curve planning algorithm based on the given frequency value of the current time step and the inserted intermediate transition velocity nodes. The seven S-curves include acceleration, constant speed, and deceleration segments, with smooth acceleration changes in each stage to avoid sudden velocity changes. The pulse signal frequency change trajectory is converted into a pulse width modulation signal (PWM signal) with a duty cycle change. The frequency of the PWM signal is set to 10kHz, and the duty cycle range is 0~100%. The duty cycle is linearly related to the frequency setpoint. The PWM signal is input to the servo controller of the main drive motor of the broaching machine for drive motor speed regulation.

[0029] In some embodiments, in step S109, the main drive motor of the broaching machine is driven by a pulse signal to complete the adaptive constant force control of the broaching process, specifically including: The main drive motor of the broaching machine is a servo motor. The model is determined according to the power requirements of the broaching machine, and usually an AC servo motor of 11kW~15kW is selected. The servo controller is electrically connected to the pulse signal output terminal to receive pulse width modulation signals. The servo controller adjusts the speed of the servo motor according to the change in the duty cycle of the pulse width modulation signal, thereby adjusting the pull-down speed of the broaching machine and achieving precise execution of the target pull-down speed command; During the broaching process, multimodal sensor data is collected in real time. Steps S101 to S109 are repeated to continuously update the broaching force prediction sequence and dynamically adjust the target pull-down speed command to form a closed-loop control, ensuring that the broaching force is always maintained within a reasonable range, thus completing the adaptive constant force control of the broaching process. Reference Figure 2 The second embodiment of the present invention provides a broaching machine system based on real-time broaching force feedback adjustment, comprising: The multimodal sensing module 201 is used to: acquire spindle motor current signals, bed vibration acceleration signals, and ambient temperature field distribution during the broaching process; Specifically, the multimodal sensing module 201 includes a Hall current sensor, a triaxial piezoelectric accelerometer, an infrared thermal imager array, and a signal conditioning unit. The Hall current sensor acquires the raw three-phase current signal of the main spindle, the triaxial piezoelectric accelerometer acquires the vibration acceleration signal of the bed, the infrared thermal imager array acquires the ambient temperature field distribution, and the signal conditioning unit filters, amplifies, and performs analog-to-digital conversion on the acquired raw signals to convert the analog signals into digital signals, which are then transmitted to the control core module. The broaching force prediction module 202 is used to: construct a dynamic state vector and input it into a pre-trained hybrid multi-neural network prediction model, and output a broaching force prediction sequence within a preset future time window; Specifically, the broaching force prediction module 202 includes a feature extraction unit and a model prediction unit. The feature extraction unit performs variational mode decomposition and short-time Fourier transform on the signal transmitted by the multimodal sensing module, extracts intrinsic mode function components and frequency domain energy spectrum features, and constructs a dynamic state vector. The model prediction unit has a built-in pre-trained hybrid multi-neural network prediction model, receives the dynamic state vector, outputs a broaching force prediction sequence, and also has an online model fine-tuning function, which can update the model weights according to fault samples. Safety monitoring module 203 is used to: calculate the difference in force values ​​between adjacent time steps of the broaching force prediction sequence and generate a predicted force change rate vector; if the predicted force change rate vector is greater than the fracture warning threshold, cut off the power supply to the main drive motor and trigger the hydraulic clamping brake; Specifically, the safety monitoring module 203 includes a differential calculation unit, a threshold judgment unit, and a braking control unit. The differential calculation unit calculates the difference in force values ​​between adjacent time steps of the broaching force prediction sequence to generate a predicted force change rate vector. The threshold judgment unit compares the predicted force change rate vector with the fracture warning threshold and the stable operation threshold, and outputs the judgment result. Based on the judgment result, when the broach is about to break, the braking control unit outputs a control signal to cut off the power supply to the main drive motor, triggers the hydraulic clamping brake, and records the fault sample. The optimization space construction module 204 is used to: construct a velocity optimization space based on the current feed rate if the predicted force change rate vector is less than the steady-state operation threshold; Specifically, the optimization space construction module 204 includes a speed acquisition unit and a space construction unit. The speed acquisition unit obtains the current feed speed from the servo driver, and the space construction unit sets the speed adjustment range, divides candidate feed speed nodes, and constructs the speed optimization space based on the current feed speed. The fitness function construction module 205 is used to: construct a multi-objective fitness function based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence; Specifically, the fitness function construction module 205 includes a feature extraction unit and a function construction unit. The feature extraction unit extracts the mean and variance of the broaching force prediction sequence to obtain the workpiece cutting energy coefficient. The function construction unit performs a weighted summation of the above features according to preset weight coefficients to construct a multi-objective fitness function that includes energy consumption and stability terms. The velocity optimization module 206 is used to: perform particle swarm optimization iteration on the velocity optimization space using a multi-objective fitness function, and output the optimal feed velocity node; Specifically, the speed optimization module 206 includes a particle swarm initialization unit, an iterative calculation unit, and an optimal node output unit. The particle swarm initialization unit initializes the position and velocity of the particle swarm; the iterative calculation unit calculates the fitness value of each particle, updates the individual and global historical best positions, dynamically adjusts the inertia weight and learning factor, and executes particle swarm optimization iteration; the optimal node output unit outputs the global historical best position at the end of the iteration as the optimal feed speed node, and at the same time determines whether the elastic deformation of the broach meets the workpiece tolerance requirements, and shrinks the optimization space and iterates again if necessary. The feedforward compensation module 207 is used to: extract the speed scalar corresponding to the optimal feed speed node, combine it with the broaching force prediction sequence fluctuation trend to perform feedforward compensation, and generate the target pull-down speed command; Specifically, the feedforward compensation module 207 includes a speed extraction unit, a trend analysis unit, a compensation calculation unit, and a filtering unit. The speed extraction unit extracts the speed scalar corresponding to the optimal feed speed node; the trend analysis unit calculates the slope of the broaching force prediction sequence and analyzes the trend of broaching force changes; the compensation calculation unit calculates the feedforward compensation amount according to the slope ratio and superimposes it onto the optimal speed scalar; the filtering unit performs low-pass filtering on the preliminary speed command to generate the target pull-down speed command. Decision sequence generation module 208 is used to: generate a speed adjustment decision sequence by combining target pull-down speed commands according to time steps; Specifically, the decision sequence generation module 208 includes a time step setting unit and a sequence combination unit. The time step setting unit sets the time step length, and the sequence combination unit combines the target pull-down speed commands for each time step in the order of the time step length to generate a speed adjustment decision sequence, which is then transmitted to the signal modulation module. The signal modulation module 209 is used to: analyze the speed adjustment decision sequence to obtain the servo driver frequency setpoint, and generate a pulse signal with S-shaped acceleration and deceleration characteristics; Specifically, the signal modulation module 209 includes a frequency analysis unit, an acceleration judgment unit, a curve planning unit, and a pulse generation unit. The frequency analysis unit analyzes the speed adjustment decision sequence to obtain the servo driver frequency setpoint; the acceleration judgment unit calculates the acceleration requirements of adjacent time steps and determines whether an intermediate transition speed node needs to be inserted; the curve planning unit uses a seven-segment S-curve planning algorithm to generate a frequency change trajectory; and the pulse generation unit converts the frequency change trajectory into a pulse width modulation signal and outputs it to the motor drive module. The motor drive module 210 is used to: drive the main drive motor of the broaching machine using pulse signals to complete the adaptive constant force control of the broaching process; Specifically, the motor drive module 210 includes a servo controller and a main drive motor. The servo controller receives the pulse width modulation signal output by the signal modulation module, adjusts the speed of the main drive motor, and then adjusts the pull-down speed of the broaching machine to achieve adaptive constant force control. At the same time, the servo controller provides real-time feedback on the current feed speed to provide data support for the construction of the optimization space.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A control method based on real-time broaching force feedback adjustment, characterized in that... ,include: The spindle motor current signal, bed vibration acceleration signal, and ambient temperature field distribution during the broaching process are collected as multimodal sensing data; a dynamic state vector is constructed and input into a pre-trained hybrid multi-neural network prediction model, which outputs a broaching force prediction sequence within a preset time window; the hybrid multi-neural network prediction model integrates convolutional neural networks, long short-term memory networks, and residual networks; Calculate the force difference between adjacent time steps in the broaching force prediction sequence to generate a vector of predicted force change rate; If the predicted force change rate vector is greater than the fracture warning threshold, it is determined that the cutter is about to break, the power supply to the main drive motor is cut off, and the hydraulic clamping brake is triggered. If the predicted force change rate vector is less than the steady-state operation threshold, a speed optimization space is constructed based on the current feed rate. Based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence, a multi-objective fitness function including energy consumption and stability terms is constructed. A multi-objective fitness function is used to perform particle swarm optimization iteration on the velocity optimization space, and the optimal feed velocity node is output. Extract the speed scalar corresponding to the optimal feed rate node, combine it with the broaching force prediction sequence fluctuation trend to perform feedforward compensation, and generate the target pull-down speed command; The target pull-down speed command is combined according to the time step to generate a speed adjustment decision sequence; the speed adjustment decision sequence is parsed to obtain the servo driver frequency setpoint, and a pulse signal with S-shaped acceleration and deceleration characteristics is generated; The main drive motor of the broaching machine is driven by a pulse signal to achieve adaptive constant force control during the broaching process.

2. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... Constructing a dynamic state vector and inputting it into a pre-trained hybrid multi-neural network prediction model, including: Variational mode decomposition is performed on the spindle motor current signal to extract the eigenmode function components containing cutting force characteristics; A short-time Fourier transform is performed on the bed vibration acceleration signal to extract the frequency domain energy spectrum features; The intrinsic mode function components, frequency domain energy spectrum characteristics, and ambient temperature field distribution are mapped into a three-dimensional tensor. Convolutional neural networks are used to extract the spatial coupling features of three-dimensional tensors to eliminate sensor noise interference. Spatial coupling features are input into a long short-term memory network to capture the time-dependent evolution of the broaching force; A residual network module is connected at the end of the network. By skipping connections, gradient vanishing is avoided, and the tensile force prediction sequence is output.

3. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... Construct a multi-objective fitness function that includes energy consumption and stability terms, including: Extract the average force component and peak force component from the broaching force prediction sequence; The ratio of the average force component to the peak force component is calculated to obtain the force stability index; The average force component, the force stability index, and the cutting specific energy coefficient are weighted and summed using preset weighting coefficients to generate a multi-objective fitness function; among them, the energy consumption term is positively correlated with the average force component, and the stability term is negatively correlated with the force stability index.

4. The control method based on real-time broaching force feedback adjustment according to claim 3, characterized in that... The particle swarm optimization iterative process is performed on the velocity optimization space using a multi-objective fitness function, including: Initialize the particle swarm's position and velocity; each particle represents a candidate feed velocity node in the velocity optimization space. Calculate the local fitness value of each particle and update the individual's best historical position and the global best historical position; An adaptive inertia weight is introduced, which is dynamically adjusted based on the convergence degree of the global historical best position; When the force stability index is lower than the preset vibration critical value, the learning factor of the velocity component is increased to guide the particles to search in the low-velocity region to suppress flutter; The global historical best position at the end of the output iteration is used as the optimal feed rate node.

5. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... Combined with the broaching force prediction sequence fluctuation trend, feedforward compensation is performed to generate the target pull-down speed command, including: Calculate the slope of the broaching force prediction sequence within a future preset time window; If the slope is greater than zero, it is determined that the cutting load is about to increase, and the feed rate is reduced in advance according to the slope ratio; When the slope is less than zero, it is determined that the cutting load is about to decrease, and the feed rate is increased in advance according to the slope ratio; The feedforward compensation is superimposed on the speed scalar corresponding to the optimal feed speed node, and after low-pass filtering, the target pull-down speed command is generated.

6. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... Collect multimodal sensing data during the broaching process, including: The raw signal of the three-phase current of the main shaft is acquired at a high-frequency sampling rate using a Hall current sensor; Vibration signals at the spindle box of the broaching machine were collected using a triaxial piezoelectric accelerometer. The temperature distribution matrix of the broach extension end and the workpiece clamping area was acquired using an infrared thermal imager array. The Clarke and Park transformations are performed on the original three-phase current signals to decouple and obtain the torque current component, which is used as the spindle motor current signal.

7. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... After determining that the baffle is about to break, cutting off the power supply to the main drive motor and triggering the hydraulic clamping brake, the process also includes: Record multimodal sensing data and broaching force prediction sequence fragments prior to fracture; The recorded data fragments are marked as fault samples and stored in the local database; When the number of fault samples accumulates to a preset number, the online fine-tuning training of the hybrid multi-neural network prediction model is triggered, and the model weights are updated to adapt to the cutting characteristics after tool wear.

8. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... The servo drive frequency setpoint is obtained by analyzing the speed adjustment decision sequence, and a pulse signal with S-shaped acceleration / deceleration characteristics is generated, including: Calculate the acceleration requirement based on the difference in target pull-down speed commands between adjacent time steps in the speed adjustment decision sequence; If the acceleration requirement exceeds the preset mechanical impact threshold, an intermediate transition velocity node is inserted. Based on the initial frequency, target frequency, and inserted intermediate transition velocity nodes, a seven-segment S-curve planning algorithm is used to generate the pulse signal frequency change trajectory; The pulse signal frequency change trajectory is converted into a pulse width modulation signal with varying duty cycle, which is then input to the servo controller of the main drive motor of the broaching machine.

9. The control method based on real-time broaching force feedback adjustment according to claim 1, characterized in that... After outputting the optimal feed rate node, it also includes: Substitute the optimal feed rate node into the broaching process dynamics model to calculate the elastic deformation at the broach overhang end; The upper boundary of the shrinkage rate optimization space when the elastic deformation exceeds the workpiece tolerance; Repeat the particle swarm optimization iteration until the elastic deformation meets the workpiece tolerance requirements.

10. A broaching machine system based on real-time broaching force feedback adjustment, characterized in that... A control method based on real-time broaching force feedback adjustment for implementing any one of claims 1-9, comprising: The multimodal sensing module collects spindle motor current signals, bed vibration acceleration signals, and ambient temperature field distribution during the broaching process. The broaching force prediction module constructs a dynamic state vector and inputs it into a pre-trained hybrid multi-neural network prediction model, outputting a broaching force prediction sequence within a preset future time window; The safety monitoring module calculates the difference in force values ​​between adjacent time steps in the broaching force prediction sequence and generates a predicted force change rate vector. If the predicted force change rate vector is greater than the fracture warning threshold, the power supply to the main drive motor is cut off and the hydraulic clamping brake is triggered. The optimization space construction module constructs a velocity optimization space based on the current feed rate if the predicted force change rate vector is less than the steady-state operation threshold. The fitness function construction module constructs a multi-objective fitness function based on the mean, variance, and workpiece cutting energy coefficient of the broaching force prediction sequence. The velocity optimization module uses a multi-objective fitness function to perform particle swarm optimization iterations on the velocity optimization space and outputs the optimal feed velocity node. The feedforward compensation module extracts the speed scalar corresponding to the optimal feed rate node, performs feedforward compensation based on the fluctuation trend of the broaching force prediction sequence, and generates the target pull-down speed command; The decision sequence generation module generates a speed adjustment decision sequence by combining the target pull-down speed command according to the time step. The signal modulation module analyzes the speed adjustment decision sequence to obtain the servo driver frequency setpoint and generates a pulse signal with S-shaped acceleration and deceleration characteristics. The motor drive module uses pulse signals to drive the main drive motor of the broaching machine, thereby achieving adaptive constant force control during the broaching process.

Images