Pressure parameter optimization method in micro-ball polishing process based on multi-source data fusion

Abstract

The present invention provides a method for optimizing pressure parameters in a microsphere grinding and polishing process based on multi-source data fusion, which specifically includes the following steps: step 1, first collecting vibration signals and non-vibration signals in the microsphere grinding and polishing process through multiple sensors; then processing the collected vibration signals to extract main features; then, integrating the extracted vibration signal features with the non-vibration signals to ensure consistency of data dimensions; step 2, inputting the integrated data into a DAE-MMR prediction model to predict the optimal pressure value under the current processing state; the prediction model is constructed based on an autoencoder and a BP neural network, and is obtained through training of historical data of multi-source data collected by multiple sensors; step 3, inputting the predicted pressure value into a control system, and the control system adjusts the pressure in the grinding and polishing process to achieve accurate optimization of the pressure parameters; thereby improving the processing accuracy and efficiency of the microspheres.

Description

Technical Field

[0001] The present invention belongs to the field of micro-ball grinding and polishing, and in particular relates to a method for optimizing pressure parameters in a micro-ball grinding and polishing process based on multi-source data fusion. Background Art

[0002] Diamond-coated microspheres play a vital role in ignition engineering. They are not only related to the research progress of laser inertial confinement fusion (ICF), but are also important components in high-end optical instruments, high-quality bearings, and high-precision measuring instruments, and they hold a very important position in basic components.

[0003] Precisely controlling pressure parameters during microsphere grinding and polishing is crucial for ensuring product quality. Traditional pressure control methods rely on mechanical systems and static measurements, making real-time adjustments difficult. In recent years, with the advancement of vibration signal processing technology, pressure parameter control methods based on vibration signals have become a research hotspot. This method analyzes the characteristic parameters of the equipment's vibration signal and establishes a corresponding mathematical model to precisely control pressure parameters. Its real-time, accurate, and intelligent nature makes it particularly suitable for high-precision manufacturing applications such as optical components, semiconductors, and aerospace. Furthermore, it can be integrated with automated production lines to improve production efficiency and product quality, providing new insights and broad application prospects for intelligent manufacturing and automated control. Therefore, if these pressure parameter control methods can be applied to microsphere grinding and polishing, developing a method for optimizing pressure parameters during the microsphere grinding and polishing process would be of great significance for improving microsphere processing quality and enabling mass production of high-end microspheres. Summary of the Invention

[0004] In response to the above-mentioned problems existing in the existing microsphere processing technology, the present invention provides a method for optimizing pressure parameters in the microsphere grinding and polishing process based on multi-source data fusion; this optimization method collects multiple data sources such as vibration signals and non-vibration signals in the microsphere grinding and polishing process, and inputs them into the DAE-MMR prediction model to predict the optimal pressure value under the current processing state, and then the control system adjusts the pressure to achieve precise optimization of the pressure parameters in the microsphere grinding and polishing process, thereby improving the microsphere processing accuracy and efficiency; wherein, the DAE-MMR prediction model is constructed based on the autoencoder and the BP neural network, combining the feature fusion capability of the autoencoder and the linear regression learning capability of the BP neural network, and can effectively fuse multi-source data with high prediction accuracy.

[0005] The technical solution of this application is:

[0006] The pressure parameter optimization method in the micro-ball grinding and polishing process based on multi-source data fusion specifically includes the following steps: Step 1, first collect the vibration signal and non-vibration signal in the micro-ball grinding and polishing process through multiple sensors; then process the collected vibration signal and extract the main features to distinguish the differences in vibration signals in different processing stages; then, integrate the extracted vibration signal features with the non-vibration signal to ensure the consistency of data dimensions; Step 2, input the integrated data into the DAE-MMR prediction model to predict the optimal pressure value under the current processing state; the prediction model is constructed based on the autoencoder and BP neural network, and is trained with historical data of multi-source data collected by multiple sensors; the specific training process is as follows: a. The vibration signals and non-vibration signals collected by multiple sensors are divided into rough polishing stage and fine polishing stage according to the time period, and further divided into training set and test set; b. Use the training set data to train the autoencoder, extract the encoding part of the autoencoder, and use the BP neural network for regression prediction training until the loss function converges or reaches the predetermined number of training times; c. Input the data in the test set into the trained prediction model, and output the prediction result, and then according to the formula Calculate the mean absolute error MAE to evaluate the performance of the prediction model; where y i is the true pressure value, y′ i To predict the pressure value, N represents the number of samples; if the MAE is not greater than the set threshold, the training is completed, otherwise, repeat step b; in step three, the predicted pressure value is input into the control system, and the control system adjusts the pressure during the grinding and polishing process to achieve accurate optimization of the pressure parameters.

[0007] Compared with the existing technology, the pressure parameter optimization method in the micro-ball grinding and polishing process based on multi-source data fusion of the present application collects multiple data sources such as vibration signals and non-vibration signals (including temperature, speed, torque, etc.) in the micro-ball grinding and polishing process, and inputs them into the DAE-MMR prediction model to predict the optimal pressure value under the current processing state. The control system then adjusts the pressure to achieve precise optimization of the pressure parameters in the micro-ball grinding and polishing process, thereby improving the micro-ball processing accuracy and processing efficiency. Among them, the DAE-MMR prediction model is constructed based on the autoencoder and the BP neural network, combining the feature fusion capability of the autoencoder and the linear regression learning capability of the BP neural network. It can effectively fuse multi-source data and has high prediction accuracy. After the prediction model has been used for a period of time, the vibration signal, non-vibration signal and predicted pressure parameters in the time period can be input into the prediction model, and the prediction model can be trained again to obtain a new prediction model, thereby achieving regular updating and optimization of the prediction model, continuously improving the accuracy of the prediction model, and making it more adaptable to changes in the processing process.

[0008] As an optimization, in the aforementioned pressure parameter optimization method during micro-ball polishing based on multi-source data fusion, the autoencoder includes multiple neural network hidden layers, and the calculation of each layer is as follows: h = σ(Wx + b); wherein x is the input data, W is the encoding weight matrix, b is the encoding bias matrix, σ is the activation function, and h is the hidden layer output matrix; the decoding formula of the autoencoder is: x′ = σ'(W′x + b′), wherein x' is the reconstructed output, W' is the decoding weight matrix, b' is the decoding bias matrix, and σ' is the activation function; in the training step b, the specific process of the autoencoder training is as follows: first, initialize the weight matrices W and W' and the bias matrices b and b' of the encoder and decoder; then, input data x to calculate the hidden layer h, and then calculate the reconstructed output x' through the decoding formula; finally, use the mean square error as the loss function: Train an autoencoder until the error between the input data x and the reconstructed output x' is minimized. Autoencoders can be used for data fusion and dimensionality reduction, which can optimize the training of predictive models and improve their predictive performance and stability.

[0009] Furthermore, in the training step b, the specific process of regression prediction training is as follows: first, initialize the weight matrix and bias vector of the BP neural network (to prevent all neurons from learning the same features and to prevent gradient explosion from causing unstable training); then calculate the output of the hidden layer and the output layer; taking the first layer as an example, the output of the hidden layer is: z 1 =W 1 x+b 1 , a 1 =f(z 1 ), where x is the input feature vector, W 1 is the weight matrix of the first hidden layer, f is the activation function, a 1 is the activation output of the first hidden layer, b 1 is the bias vector of the first hidden layer; the output of the output layer is: y'=W o a L +b o , where W o is the weight matrix of the output layer, b o is the bias vector of the output layer, a L is the activation output of the last hidden layer, and y′ is the predicted pressure value. Finally, the error gradient is calculated through backpropagation to adjust the weights of the BP neural network to optimize the predictive ability of the prediction model. The present invention uses a BP neural network for feature extraction, effectively extracting features from the autoencoder and mapping them to pressure parameters, achieving accurate regression prediction. Furthermore, the present invention uses a random initialization method to break symmetry and avoid initial values ​​that are too large or too small.

[0010] Furthermore, the specific adjustment process of the weight of the BP neural network is as follows: First, the error of the output layer and the hidden layer is calculated; the error of the output layer is: δ o =y′-y, where y is the actual pressure value; taking the lth layer as an example, the error of the hidden layer is: δ l =(W l+1 ) T δ (l+1) ⊙f′(z l ), where f′ is the derivative of the activation function and ⊙ represents element-by-element multiplication; then, the partial derivatives are calculated based on the gradient of the error function with respect to the weights of each layer in the BP neural network: Finally, according to the update formula: Update the weights and biases, where η is the learning rate. The learning rate η is usually selected to a suitable value to ensure that the neural network converges quickly and does not oscillate.

[0011] Furthermore, during the regression prediction training process, GA is used to tune the hyperparameters of the BP neural network. GA (genetic algorithm) automatically adjusts hyperparameters and can perform global searches in complex, high-dimensional hyperparameter spaces, avoiding local optimality. This improves the efficiency and robustness of the prediction model, resulting in better performance in practical applications.

[0012] As an optimization, in the aforementioned method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion, in step one, the main frequency domain features of the vibration signal data are extracted by fast Fourier transform, specifically: first use a low-pass filter to obtain the signal in the required frequency domain range, and then reduce spectrum leakage by signal truncation and period extension to extract more accurate frequency domain features, and on this basis, use a window function to reduce the impact of boundary effects on frequency domain feature extraction. The present invention adopts the fast Fourier transform algorithm, the core function of which is to efficiently convert time domain signals to frequency domain, which is convenient for frequency analysis, filtering and signal processing; it significantly improves the efficiency of processing large-scale data by reducing computational complexity.

[0013] Furthermore, in step 1, a Butterworth low-pass filter is used to remove high-frequency noise and low-frequency drift in the signal, retaining the main frequency components of the signal. The transfer function of the Butterworth low-pass filter is: Where n is the order, f c is the cutoff frequency, ω C =2πf c is the angular frequency of the cutoff frequency; the truncated signal is expressed as: Where x[n] is a given discrete signal, n0 is the sample point index where truncation starts, and n1 is the sample point index where truncation ends. The periodic extension signal is expressed as: x extended [n]=x[mmod N], where n mod N is the modulo operation, which is the signal length x. cut Periodic repetition on [n]; the mathematical expression of the window function is: Where ω[n] is the value of the window function, n is the sample index point of the window function, N is the number of samples of the window function, α and β are used to balance the main lobe width and side lobe attenuation; the calculation formula of the fast Fourier transform is: X[k] = FFT(x extended [n]), where X[k] is the kth frequency component of the frequency domain signal.

[0014] As an optimization, in the aforementioned method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion, a deterministic polishing device is used to grind and polish the micro-balls; the deterministic polishing device includes a workbench and a polishing tank provided on the workbench; a polishing disc is provided in the polishing tank, and a polishing head is provided above the polishing tank; the polishing disc is driven to rotate by a servo motor; the polishing head is connected to a DD motor and driven to rotate by the DD motor; the DD motor is connected to a cylinder and driven to move up and down by the cylinder; a pressure sensor 5 is provided between the DD motor and the cylinder; during the grinding and polishing process, the micro-balls to be polished are mounted on the polishing disc, and polishing liquid is added to the polishing tank so that the micro-balls are completely immersed in the polishing liquid; the polishing head is then driven downward by the cylinder until the polishing head presses on the micro-balls with a certain pressure; then, the servo motor is started to drive the polishing disc to rotate, and the DD motor is started to drive the polishing head to rotate at the same time, so that the micro-balls are ground and polished. The micro-ball grinding and polishing experiment using the polishing device with the above structure is easy to implement and easy to obtain various polishing process parameters. In order to ensure the validity and accuracy of the vibration signal, in this application, the collected vibration signal includes the vibration signal of the polishing plate and the vibration signal of the polishing head.

[0015] Since the polishing disc and the polishing head need to rotate, in this application, a signal slip ring is set on the polishing disc and the polishing head, and an acceleration sensor is set on the signal slip ring, so that the acceleration sensor is connected to an external signal collector through the signal slip ring, and is used to divide the collected vibration signal into three directions of X, Y, and Z axes and transmit it to the signal collector. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the optimization method of the present application;

[0017] Figure 2 Schematic diagram of the structure of the polishing device in the embodiment of the present application;

[0018] The markings in the accompanying drawings are: 1-polishing disc; 2-polishing groove; 3-polishing head; 4-DD motor; 5-pressure sensor; 6-cylinder; 7-drive shaft; 8-workbench. DETAILED DESCRIPTION

[0019] The present application is further described below with reference to the accompanying drawings and embodiments, but they are not intended to limit the present application.

[0020] In the process of micro-ball grinding and polishing, traditional pressure control methods rely on mechanical systems and static measurements, making it difficult to achieve real-time adjustment. The present invention provides a method for optimizing pressure parameters in the process of micro-ball grinding and polishing based on multi-source data fusion. This method collects multiple data sources such as vibration signals and non-vibration signals (including temperature, rotation speed, torque, etc.) in the process of micro-ball grinding and polishing, and inputs them into a DAE-MMR prediction model to predict the optimal pressure value under the current processing state. The control system then adjusts the pressure to achieve precise optimization of pressure parameters in the process of micro-ball grinding and polishing, thereby improving the micro-ball processing accuracy and efficiency. The DAE-MMR prediction model is constructed based on an autoencoder and a BP neural network, and is trained with historical data of multi-source data collected by multiple sensors.

[0021] Example:

[0022] In this embodiment, a deterministic polishing device is used to grind and polish the microspheres. Figure 2 The deterministic polishing device includes a workbench 8 and a polishing tank 2 mounted on the workbench 8. A polishing disc 1 is positioned within the polishing tank 2, and a polishing head 3 is positioned above the polishing tank 2. The polishing disc 1 is driven for rotation by a servo motor. The polishing head 3 is connected to a DD motor 4 for rotation. The DD motor 4 is connected to a drive shaft 7 of a pneumatic cylinder 6 and is driven up and down by the cylinder 6. A pressure sensor 5 is positioned between the DD motor 4 and the cylinder 6. During the grinding and polishing process, the microspheres to be polished are mounted on the polishing disc 1, and polishing fluid is filled into the polishing tank 2, completely submerging the microspheres. The polishing head 3 is then driven downward by the pneumatic cylinder 6 until it presses against the microspheres with a certain pressure. The servo motor is then activated to rotate the polishing disc 1, while the DD motor 4 is simultaneously activated to rotate the polishing head 3, thereby grinding and polishing the microspheres. During the polishing process, polishing fluid is added every 20 minutes to cool the microspheres and remove any waste.

[0023] See also Figure 2 The pressure parameter optimization method in the micro-ball polishing process based on multi-source data fusion specifically includes the following steps.

[0024] In step 1, multiple sensors are used to collect vibration signals and non-vibration signals during the micro-ball grinding and polishing process. The non-vibration signals include temperature, rotation speed, and torque (specifically, the rotation speed of the polishing disc 1 and the polishing head 3, and the torque of the motor output shaft). The collected vibration signals are then processed to extract the main features to distinguish the differences in vibration signals at different processing stages. The extracted vibration signal features are then integrated with the non-vibration signals to ensure consistency in data dimensions.

[0025] To ensure the validity and accuracy of the vibration signal, this embodiment separately collects vibration signals from the polishing plate 1 and polishing head 3. Signal slip rings are installed on the polishing plate 1 and polishing head 3, and a triaxial accelerometer is mounted on the signal slip rings. This allows for connection to an external signal acquisition device via the signal slip rings. The collected vibration signal is then transmitted to the signal acquisition device in the X, Y, and Z directions for data processing. To collect temperature parameters, a PT100 platinum thermal resistor is embedded in the polishing plate 1 to measure the temperature of the grinding interface during micro-ball polishing.

[0026] In this embodiment, the main frequency domain features of the vibration signal data are extracted by fast Fourier transform (that is, the main frequency domain features in the X, Y, and Z axis vibration signals are extracted, and the vibration signals of each axis are Fourier transformed according to the time period to obtain the frequency amplitude distribution of each time period, and then the frequency features of each time period are extracted); first, a low-pass filter is used to obtain the signal in the required frequency domain range, and then the signal is truncate and the cycle is extended to reduce the spectrum leakage to extract more accurate frequency domain features, and on this basis, a window function is used to reduce the impact of boundary effects on the frequency domain feature extraction. The details are as follows:

[0027] First, a Butterworth low-pass filter is used to remove high-frequency noise and low-frequency drift in the signal and retain the main frequency components of the signal. The transfer function of the Butterworth low-pass filter is: Where n is the order, f c is the cutoff frequency, ω C =2πf c is the angular frequency of the cutoff frequency.

[0028] Next, in order to further process the filtered vibration signal, reduce spectrum leakage before Fourier transform, and improve the accuracy of frequency domain features, this embodiment adopts the signal truncation and cycle extension method. For a given discrete signal x[n], the data of sample points n0 to n1 are truncate. The truncated signal is expressed as: Among them, n0 is the sample point index where truncation starts, and n1 is the sample point index where truncation ends. Then, given the truncated signal x cut [n], its period extension signal is expressed as: x extended[n]=x[n modN], where n mod N is the modulo operation, which is the signal length x cut [n] is a periodic repetition. In order to further reduce the spectrum leakage in the Fourier transform, the present embodiment uses a Hamming window to process the phase signal, making the signal edge smoother, thereby reducing spectrum leakage; the mathematical expression of the Hamming window is: Where ω[n] is the value of the window function, n is the sample index point of the window function, N is the number of samples of the window function (the total length of the window function), and α and β are used to balance the main lobe width and side lobe attenuation, which can be modified according to actual needs.

[0029] Finally, in order to further process the signal and extract frequency features, the fast Fourier transform (FFT) algorithm is used in this embodiment to efficiently convert the time domain signal to the frequency domain for frequency analysis, filtering, and signal processing; FFT can significantly improve the efficiency of large-scale data processing by reducing computational complexity.

[0030] The signal data x obtained after the above period is extended extended [n] is used as the input signal, and the frequency domain representation is: X[k] = FFT(x extended [n]); FFT calculation is based on the discrete Fourier transform (DFT) formula: Among them, X[k] is the kth frequency component of the frequency domain signal, x extended [n] is the time domain signal after window function and period extension, is a complex exponential function representing the sine and cosine components in the frequency domain.

[0031] Step 2: Input the integrated data into the DAE-MMR prediction model to predict the optimal pressure value under the current processing state;

[0032] The prediction model is constructed based on an autoencoder and a BP neural network and is trained using historical data from multiple sources collected by multiple sensors. The specific training process is as follows: a. The vibration signals and non-vibration signals collected by multiple sensors are divided into a rough polishing stage and a fine polishing stage according to time periods, and further divided into a training set and a test set;

[0033] b. Using the training set data to train the autoencoder, extracting the encoding part of the autoencoder, and then performing regression prediction training through the BP neural network; the autoencoder includes multiple neural network hidden layers, and the calculation of each layer is as follows: h = σ(Wx + b); where x is the input data, W is the encoding weight matrix, b is the encoding bias matrix, σ is the activation function, and h is the hidden layer output matrix; the decoding formula of the autoencoder is: x′ = σ'(W′x + b′), where x' is the reconstructed output, W' is the decoding weight matrix, b' is the decoding bias matrix, and σ' is the activation function; specifically, the following steps are included:

[0034] S1, first perform autoencoder training: First, initialize the weight matrices W and W' and the bias matrices b and b' of the encoder and decoder; then, input data x to calculate the hidden layer h, and then calculate the reconstructed output x' through the decoding formula; finally, use the mean square error as the loss function: The autoencoder is trained until the error between the input data x and the reconstructed output x' is minimized.

[0035] S2, then perform regression prediction training: In order to prevent all neurons from learning the same features and prevent gradient explosion from causing unstable training, it is necessary to first initialize the weight matrix and bias vector of each layer of the BP neural network through the initialization function (in this embodiment, the random initialization method is used to break the symmetry and avoid the initial value being too large or too small); then, calculate the output of the hidden layer and the output layer; taking the first layer as an example, the output of the hidden layer is: z 1 =W 1 x+b 1 , a 1 =f(z 1 ), where x is the input feature vector, W 1 is the weight matrix of the first hidden layer, f is the activation function, a 1 is the activation output of the first hidden layer, b 1 is the bias vector of the first hidden layer; the output of the output layer is: y'=W o a L +b o , where W o is the weight matrix of the output layer, b o is the bias vector of the output layer, a L is the activation output of the last hidden layer, and y′ is the pressure prediction value;

[0036] S3, adjust the weight of BP neural network by back propagation to calculate the error gradient to optimize the prediction ability of the prediction model: First, calculate the error between the output layer and the hidden layer; the error of the output layer is: δ o=y′-y, where y is the actual pressure value; taking the lth layer as an example, the error of the hidden layer is: δ l =(W l+1 ) T δ (l+1) ⊙f′(z l ), where f′ is the derivative of the activation function and ⊙ represents element-by-element multiplication; then, the partial derivatives are calculated based on the gradient of the error function with respect to the weights of each layer in the BP neural network: Finally, according to the update formula: Update weights and biases, where η is the learning rate;

[0037] S4, repeat steps S1-S3 until the loss function converges or the predetermined number of training times is reached.

[0038] The BP neural network is designed to fit the features extracted by the autoencoder with the final regulated pressure parameters, playing a key role in fusing features into the final prediction. In this embodiment, a GA (genetic algorithm) is used to tune the hyperparameters of the BP (back propagation) neural network. By automatically adjusting hyperparameters, the GA (genetic algorithm) can perform a global search in a complex, high-dimensional hyperparameter space, avoiding local optimality. This improves the efficiency and robustness of the prediction model, resulting in better performance in practical applications.

[0039] The specific calculation process is as follows:

[0040] I, initialize the population, and the encoding of each individual is as follows: x=(a,L,N1,N2,…,N L ), generate P individuals, each of which consists of random values ​​of the above hyperparameters, such as a∈[a min ,a max ],L∈[L min ,L max ], each N i ∈[N min ,N max ];

[0041] II, defined by training the BP neural network through the fitness function and evaluating its performance on the validation set, mainly looking at the negative value of the loss function Fitness(x) = -Validation Loss;

[0042] III. The selection operation uses the roulette wheel selection method to calculate the individual selection probability:

[0043] IV. The crossover operation uses a single-point crossover method to generate a new individual by selecting the crossover point k:

[0044] V, the mutation operation applies mutation to each hyperparameter with a mutation probability of P m : Among them, Δα, ΔL and ΔN i is the random perturbation value;

[0045] VI, the replacement operation uses a full replacement strategy to completely replace individuals with low fitness with the newly generated individuals (or uses a partial replacement strategy to only replace a portion of individuals with the lowest fitness);

[0046] VII. The algorithm runs until it reaches the preset maximum number of iterations T or the fitness threshold Fitness threshold Through the above steps and formulas, the genetic algorithm can effectively optimize the hyperparameters of the BP neural network and improve its performance.

[0047] c. Input the data in the test set into the pre-trained prediction model and output the prediction results, then use the formula Calculate the mean absolute error MAE to evaluate the performance of the prediction model; where y i is the true pressure value, y′ i To predict the pressure value, N represents the number of samples; if the MAE is not greater than the set threshold, the training is completed, otherwise, repeat step b.

[0048] Step three: input the predicted pressure value into the control system, which adjusts the pressure during the grinding and polishing process to achieve accurate optimization of the pressure parameters.

[0049] In this embodiment, the DAE-MMR prediction model is integrated into the control system to predict the optimal value under the current processing state online, and by controlling the solenoid valve, the actual output value of the cylinder is adjusted to realize automatic regulation of the pressure parameters during the processing. At the same time, the control system stores the obtained new state parameters and process parameters (i.e., the detected vibration signal, non-vibration signal, and predicted pressure parameter); thus, after the prediction model has been used for a period of time, the vibration signal, non-vibration signal, and predicted pressure parameter within the time period can be input into the prediction model, and the prediction model can be trained again to obtain a new prediction model, so as to realize regular updating and optimization of the prediction model, thereby continuously improving the accuracy of the prediction model and making it more adaptable to changes in the processing process.

[0050] The above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may be modified or some or all of the technical features thereof may be replaced with equivalents; and such modifications or replacements do not deviate from the essence of the corresponding technical solutions within the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing pressure parameters in the micro-ball polishing process based on multi-source data fusion, characterized in that: The specific steps include: In the first step, multiple sensors are used to collect vibration and non-vibration signals during the micro-ball grinding and polishing process. The collected vibration signals are then processed to extract key features to distinguish the differences in vibration signals at different processing stages. The extracted vibration signal features are then integrated with the non-vibration signal to ensure consistent data dimensions. Step 2: Input the integrated data into the DAE-MMR prediction model to predict the optimal pressure value under the current processing state; the prediction model is built based on an autoencoder and a BP neural network, and is trained with historical data from multiple sources collected by multiple sensors; The specific training process is as follows: a. The vibration signals and non-vibration signals collected by multiple sensors are divided into rough polishing and fine polishing stages according to the time period, and further divided into training and test sets; b. The autoencoder is trained using the training set data, and the encoding part of the autoencoder is extracted. The BP neural network is used for regression prediction training until the loss function converges or the predetermined number of training times is reached; c. Input the data in the test set into the trained prediction model and output the prediction results, then use the formula Calculate the mean absolute error MAE to evaluate the performance of the prediction model; where y i is the true pressure value, y′ i To predict the pressure value, N represents the number of samples; if MAE is not greater than the set threshold, the training is completed, otherwise, repeat step b; Step three: input the predicted pressure value into the control system, which adjusts the pressure during the grinding and polishing process to achieve accurate optimization of the pressure parameters.

2. The method for optimizing pressure parameters in the micro-ball polishing process based on multi-source data fusion according to claim 1, characterized in that: The autoencoder includes multiple neural network hidden layers, and the calculation of each layer is as follows: h = σ(Wx + b); where x is the input data, W is the encoding weight matrix, b is the encoding bias matrix, σ is the activation function, and h is the hidden layer output matrix; the decoding formula of the autoencoder is: x′ = σ'(W′x + b′), where x' is the reconstructed output, W' is the decoding weight matrix, b' is the decoding bias matrix, and σ' is the activation function; in the training step b, the specific process of the autoencoder training is as follows: first, initialize the weight matrices W and W' and the bias matrices b and b' of the encoder and decoder; then, input data x to calculate the hidden layer h, and then calculate the reconstructed output x' through the decoding formula; finally, use the mean square error as the loss function: The autoencoder is trained until the error between the input data x and the reconstructed output x' is minimized.

3. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 2, characterized in that: In the training step b, the specific process of regression prediction training is as follows: first, initialize the weight matrix and bias vector of the BP neural network; then calculate the output of the hidden layer and the output layer; taking the first layer as an example, the output of the hidden layer is: 1 =W 1 x+b 1 , a 1 =f(z 1 ), where x is the input feature vector, W 1 is the weight matrix of the first hidden layer, f is the activation function, a 1 is the activation output of the first hidden layer, b 1 is the bias vector of the first hidden layer; the output of the output layer is: y'=W o a L +b o , where W o is the weight matrix of the output layer, b o is the bias vector of the output layer, a L is the activation output of the last hidden layer, and y′ is the pressure prediction value; finally, the error gradient is calculated by back propagation to adjust the weight of the BP neural network to optimize the prediction ability of the prediction model.

4. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 3, characterized in that: The specific adjustment process of the weight of the BP neural network is as follows: First, the error of the output layer and the hidden layer is calculated; the error of the output layer is: δ o =y′-y, where y is the actual pressure value; taking the lth layer as an example, the error of the hidden layer is: δ l =(W l+1 ) T δ (l+1) ⊙f′(z l ), where f' is the derivative of the activation function and ⊙ represents element-by-element multiplication; then, the partial derivatives are calculated based on the gradient of the error function with respect to the weights of each layer in the BP neural network: Finally, according to the update formula: Update weights and biases, where η is the learning rate.

5. The method for optimizing pressure parameters in the micro-ball polishing process based on multi-source data fusion according to claim 4, characterized in that: During the regression prediction training process, GA is used to tune the hyperparameters of the BP neural network.

6. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 1, characterized in that: In the step 1, the main frequency domain features of the vibration signal data are extracted by fast Fourier transform, specifically: first, a low-pass filter is used to obtain the signal in the required frequency domain range, and then signal truncation and period extension are performed to reduce spectrum leakage to extract more accurate frequency domain features. On this basis, a window function is used to reduce the influence of boundary effects on frequency domain feature extraction.

7. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 6, characterized in that: In the step 1, a Butterworth low-pass filter is used to remove high-frequency noise and low-frequency drift in the signal, retaining the main frequency components of the signal. The transfer function of the Butterworth low-pass filter is: Where n is the order, f c is the cutoff frequency, ω C =2πf c is the angular frequency of the cutoff frequency; the truncated signal is expressed as: Where x[n] is a given discrete signal, n0 is the sample point index where truncation starts, and n1 is the sample point index where truncation ends. The periodic extension signal is expressed as: x extended [n] = x[n mod N], where nmod N is the modulo operation, which is the signal length x. cut Periodic repetition on [n]; the mathematical expression of the window function is: Where ω[n] is the value of the window function, n is the sample index point of the window function, N is the number of samples of the window function, α and β are used to balance the main lobe width and side lobe attenuation; the calculation formula of the fast Fourier transform is: X[k] = FFT(x extended [n]), where X[k] is the kth frequency component of the frequency domain signal.

8. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 1, characterized in that: The method uses a deterministic polishing device to grind and polish microspheres; the deterministic polishing device includes a workbench (8) and a polishing tank (2) provided on the workbench (8); a polishing disc (1) is provided in the polishing tank (2), and a polishing head (3) is provided above the polishing tank (2); the polishing disc (1) is driven to rotate by a servo motor; the polishing head (3) is connected to a DD motor (4) and driven to rotate by the DD motor (4); the DD motor (4) is connected to a cylinder (6) and driven up and down by the cylinder (6) The DD motor (4) and the cylinder (6) are provided with a pressure sensor (5); during the grinding and polishing process, the microspheres to be polished are mounted on the polishing disc (1), and polishing liquid is added to the polishing tank (2) so that the microspheres are completely immersed in the polishing liquid; the polishing head (3) is then driven downward by the cylinder (6) until the polishing head (3) is pressed on the microspheres with a certain pressure; then, the servo motor is started to drive the polishing disc (1) to rotate, and the DD motor (4) is started to drive the polishing head (3) to rotate, thereby grinding and polishing the microspheres.

9. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 8, characterized in that: In the step 1, the collected vibration signals include the vibration signals of the polishing plate (1) and the vibration signals of the polishing head (3).

10. The method for optimizing pressure parameters during micro-ball polishing based on multi-source data fusion according to claim 9, characterized in that: Signal slip rings are provided on the polishing disc (1) and the polishing head (3); and acceleration sensors are provided on the signal slip rings for dividing the collected vibration signals into three directions of X, Y and Z axes and transmitting them to a signal acquisition instrument.

Images