A smart control method and system for diamond wire cutting of silicon wafers

Abstract

This application relates to the field of intelligent control technology, and discloses an intelligent control method and system for diamond wire cutting of silicon wafers. The method includes: collecting data through multi-physics monitoring; extracting time-varying features and long-term evolution trends using a temporal convolutional-long short-term memory fusion neural network; weighting and outputting predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability through an attention mechanism; optimizing control parameters based on these values ​​and converting them into drive commands; and updating the model through online incremental learning. This application improves the quality control accuracy and process adaptability of the diamond wire cutting process for silicon wafers.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to an intelligent control method and system for diamond wire cutting of silicon wafers. Background Technology

[0002] Diamond wire cutting technology, as a core process in silicon wafer processing, plays a crucial role in the semiconductor and photovoltaic industries. Existing intelligent control methods for diamond wire cutting of silicon wafers primarily involve deploying sensors on the cutting equipment to collect process parameters. These parameters, such as wire speed and feed rate, are then adjusted using PID controllers or fuzzy controllers based on expert rules. Some advanced methods employ machine learning algorithms such as BP neural networks or support vector machines to establish a mapping relationship between process parameters and cutting quality, achieving optimized parameter control. These methods, to a certain extent, improve the automation level and quality stability of silicon wafer cutting.

[0003] Existing methods often employ a single sensor type or a limited set of monitoring parameters, such as monitoring only a few parameters like linear velocity and tension. This makes it difficult to comprehensively reflect the coordinated evolution of multiple physical fields, including mechanical, thermal, acoustic, and electric fields, during the cutting process. Consequently, the control model's ability to identify the coupled effects of multiple factors is insufficient, failing to accurately capture the transient coupling relationships between cutting force and temperature, and between vibration and acoustic emission. Secondly, traditional machine learning algorithms have limitations when processing high-dimensional nonlinear time-series data. Methods such as BP neural networks and support vector machines struggle to effectively extract local time-varying features at different time scales, particularly regarding instantaneous fluctuations in cutting force and slow temperature increases. The algorithms lack targeted modeling capabilities and cannot remember the gradual changes in diamond wire from sharp to worn or coolant from clean to contaminated, resulting in weak generalization ability. Furthermore, existing control strategies are mostly feedback control based on the current state, lacking the ability to predict future trends in cutting quality. They exhibit lag in response to gradual disturbances such as diamond wire wear and coolant performance degradation, and cannot take compensatory measures in advance. In addition, the multi-parameter collaborative optimization capability is insufficient. The control commands of each actuator are often calculated independently, failing to fully consider the coupling interference of wire speed changes on tension and the influence of feed speed changes on temperature, making it difficult to achieve globally optimal control. Summary of the Invention

[0004] This application provides an intelligent control method and system for diamond wire cutting of silicon wafers. It utilizes a multi-physics field collaborative monitoring system and employs a temporal convolutional-long short-term memory fusion neural network for predictive modeling. Combined with a dual-timescale predictive control framework and a parameter interaction compensation mechanism, it addresses the problems of inaccurate identification of multi-factor coupling, lack of long-term evolution trend prediction, insufficient parameter collaborative optimization capability, and weak early warning capability for abnormal operating conditions in existing technologies. This application significantly improves the quality control accuracy and process adaptability of the diamond wire cutting process for silicon wafers.

[0005] In a first aspect, this application provides an intelligent control method for diamond wire cutting of silicon wafers, the intelligent control method for diamond wire cutting of silicon wafers comprising:

[0006] Step S1: Collect data from the cutting work area using a multi-physics collaborative monitoring system to obtain a multi-physics feature vector sequence;

[0007] Step S2: The multi-physics feature vector sequence is predicted and modeled by a temporal convolution-long short-term memory fusion neural network. The temporal convolutional layer uses a causal convolution structure and multi-level dilated convolution to extract the local time-varying features and transient coupling modes of the multi-physics parameters. The long short-term memory layer captures the long-term evolution trend of diamond wire wear and coolant performance degradation. The attention mechanism layer calculates the dynamic weights of each physical field feature and calculates the weighted sum to obtain the context vector. The fully connected layer outputs the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability.

[0008] Step S3: Calculate and optimize the control parameters based on the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability to obtain optimized control parameters;

[0009] Step S4: Convert the optimized control parameters into actuator drive commands for multi-parameter collaborative control, and update the temporal convolutional-long short-term memory fusion neural network through online incremental learning and adjust the execution control strategy based on abnormal operating conditions.

[0010] Secondly, this application provides an intelligent control system for diamond wire cutting of silicon wafers, the intelligent control system for diamond wire cutting of silicon wafers comprising:

[0011] The data acquisition module is used to acquire data from the cutting work area through a multi-physics field collaborative monitoring system, and obtain a multi-physics field feature vector sequence.

[0012] The modeling module is used to predict and model the multi-physics feature vector sequence through a temporal convolution-long short-term memory fusion neural network. The temporal convolutional layer uses a causal convolution structure and multi-level dilated convolution to extract the local time-varying features and transient coupling modes of the multi-physics parameters. The long short-term memory layer captures the long-term evolution trend of diamond wire wear and coolant performance degradation. The attention mechanism layer calculates the dynamic weights of each physical field feature and calculates the weighted sum to obtain the context vector. The fully connected layer outputs the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability.

[0013] The calculation module is used to perform control parameter optimization calculations based on the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability to obtain optimized control parameters.

[0014] The execution module is used to convert the optimized control parameters into actuator drive instructions for multi-parameter coordinated control, and to update the temporal convolutional-long short-term memory fusion neural network through online incremental learning and to adjust the execution control strategy based on abnormal operating conditions.

[0015] The technical solution provided in this application acquires multi-physics feature vector sequences by using a multi-physics collaborative monitoring system to collect data from the cutting work area. This overcomes the limitations of existing technologies that rely on a single sensor type or limited monitoring parameters. During the cutting process, mechanical field parameters, thermal field parameters, acoustic field parameters, and electric field parameters are collected simultaneously. A precision clock synchronization protocol is used for timestamp calibration to ensure strict spatiotemporal alignment of the data from each physical field. This collaborative monitoring system can comprehensively capture the complex dynamic characteristics under the coupling of multiple factors during the cutting process, providing sufficient feature dimensions for subsequent prediction and modeling. Compared with existing technologies that only monitor a few parameters such as linear velocity and tension, multi-physics collaborative monitoring significantly enhances the control system's ability to comprehensively perceive the state of the cutting process. It can accurately capture the transient coupling relationships between cutting force and temperature, and between acoustic field and vibration field, laying an information foundation for achieving high-precision control. This application employs a temporal convolutional-long short-term memory (LSTM) fusion neural network to predict and model multi-physics feature vector sequences. This network architecture is specifically designed for the multi-timescale characteristics and long-term evolution trends of the diamond wire cutting process. The temporal convolutional layer ensures that the prediction relies solely on historical data and conforms to temporal causality through a causal convolutional structure. Multi-level dilated convolutions cause the dilation rate of each residual block to increase exponentially. The first residual block extracts rapidly changing local time-varying features such as instantaneous fluctuations in cutting force, while subsequent residual blocks extract long-term evolutionary local features such as slow temperature increases. This achieves multi-level feature extraction from short to long timescales. The LTM layer maintains long-term memory of important information through the synergistic effect of forget gates, input gates, and output gates, enabling it to capture changes in the diamond wire's state from sharp to worn. The attention mechanism layer adaptively allocates the contribution of different physical field parameters by calculating the dynamic weights of each physical field feature to the performance degradation process of the coolant from clean to contaminated. When the cutting force suddenly increases or the temperature rises sharply at a certain moment, the attention weight at that moment automatically increases to highlight important features. The silicon wafer surface roughness prediction value, warpage prediction value and cutting anomaly probability output by the fully connected layer provide quantitative prediction information for the future cutting quality evolution trend of the control system. Compared with the static mapping model based on traditional algorithms such as BP neural network or support vector machine in the prior art, the temporal convolution-long short-term memory fusion network of this application has significant advantages in multi-time scale feature extraction, long-term memory ability and dynamic weight allocation, and can accurately identify the coupling effect of multiple factors and predict future trends. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of one embodiment of the intelligent control method for diamond wire cutting of silicon wafers in this application.

[0018] Figure 2 This is a schematic diagram of the multi-level dilated convolution feature extraction process of the temporal convolutional layer in an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of the dual-timescale predictive control parameter optimization process in an embodiment of this application. Detailed Implementation

[0020] This application provides an intelligent control method and system for diamond wire cutting of silicon wafers. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent control method for diamond wire cutting of silicon wafers in this application includes:

[0022] Step S1: Collect data from the cutting work area using a multi-physics collaborative monitoring system to obtain a multi-physics feature vector sequence;

[0023] Specifically, a multi-physics collaborative monitoring system is established in the cutting working area of ​​the diamond wire cutting equipment to obtain comprehensive information about the cutting process. This monitoring system includes sensors in four physical domains: mechanical field, thermal field, acoustic field, and electric field. In the cutting contact area, the triaxial components of the cutting force, the diamond wire tension value, and the triaxial signal of vibration acceleration are collected to form mechanical field parameters. In the cutting interface, the temperature field distribution matrix, the inlet and outlet temperatures of the coolant are collected, and the temperature gradient is calculated to form thermal field parameters. In the cutting working area, acoustic emission signals are collected to form acoustic field parameters. The current and power of the spindle motor and the feed motor are collected to form electric field parameters. The time stamp calibration of each physical field parameter is performed through a clock synchronization protocol to ensure that the time deviation is less than 1 microsecond. The physical field parameters are organized into a multi-physics feature vector sequence according to the time sequence.

[0024] Step S2: The multi-physics feature vector sequence is predicted and modeled by a temporal convolution-long short-term memory fusion neural network. The temporal convolutional layer uses a causal convolution structure and multi-level dilated convolution to extract the local time-varying features and transient coupling modes of the multi-physics parameters. The long short-term memory layer captures the long-term evolution trend of diamond wire wear and coolant performance degradation. The attention mechanism layer calculates the dynamic weights of each physical field feature and calculates the weighted sum to obtain the context vector. The fully connected layer outputs the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability.

[0025] Specifically, the temporal convolutional-long short-term memory fusion neural network processes multi-physics feature vector sequences. The temporal convolutional layer adopts a causal convolutional structure to ensure that the prediction at the current moment depends only on historical data and not on future data. Multi-level dilated convolution is performed through multiple residual blocks, with the dilation rate of each residual block set to 1, 2, 4, and 8, increasing exponentially. Dilated convolution expands the receptive field by inserting intervals between convolutional kernel elements. When the dilation rate of the first residual block is 1, the receptive field covers three consecutive time steps; when the dilation rate of the second residual block is 2, the receptive field spans five time steps; and when the dilation rate of the fourth residual block is 8, the receptive field... Spanning 17 time steps, convolutional layers with different expansion rates extract local time-varying features at different time scales, such as instantaneous fluctuations in cutting force, short-term changes in vibrational acceleration, and slow temperature increases. Simultaneously, it captures transient coupling patterns between the mechanical and thermal fields, and between the acoustic and vibrational fields. The long short-term memory (LSTM) layer contains multiple layers of bidirectional LSM units. Each unit determines how much historical information to discard through a forget gate, how much new information to receive through an input gate, and how much hidden state information to output through an output gate. The forget gate is calculated as follows: the forget weight equals the sigmoid function applied to the forget gate weight matrix multiplied by the current input... The input vector is a concatenation of the hidden state from the previous time step. The input and output gates are calculated in a similar manner. Through the gating mechanism, the long short-term memory layer can maintain the long-term memory to capture the evolution process of the diamond wire gradually wearing down from a sharp state, resulting in a slow increase in cutting force, and the decay process of the coolant gradually becoming contaminated from a clean state, resulting in a decrease in cooling efficiency and an increase in temperature. The attention mechanism layer calculates the attention weights for the hidden states of each time step output by the long short-term memory layer through a query-key-value mapping mechanism. The query vector and the key vector are multiplied by a dot product to obtain the attention score. After the score is normalized by softmax, the attention weight is obtained. The magnitude of the attention weight reflects the degree of influence of the physical field characteristics of the corresponding time step on the future cutting quality. The larger attention weight at the moment of sudden increase in cutting force indicates that the moment makes a significant contribution to the quality prediction. The attention weight and the value vector are weighted and summed to obtain the context vector that integrates the information of each time step. The fully connected layer performs nonlinear mapping on the context vector and outputs the silicon wafer surface roughness prediction value, which reflects the microscopic unevenness of the silicon wafer surface. The warpage prediction value reflects the macroscopic deformation of the silicon wafer. The cutting anomaly probability reflects the likelihood of abnormal working conditions such as diamond wire breakage or silicon wafer breakage in the future.

[0026] Step S3: Calculate and optimize the control parameters based on the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability to obtain the optimized control parameters;

[0027] Specifically, control parameters are optimized based on the prediction results. When the probability of cutting anomalies exceeds a set threshold of 0.15, an anomaly risk is identified. In the fast-timescale control layer, the deviation between the current cutting force and the baseline value is calculated. If the deviation exceeds 5 Newtons, the diamond wire linear speed adjustment is calculated according to the relationship that the linear speed adjustment is equal to a proportional coefficient multiplied by the cutting force deviation. The proportional coefficient is set to 0.02 meters per second per Newton. Simultaneously, the deviation between the peak vibration acceleration and the baseline value is calculated. If the deviation exceeds 2 meters per square second, the tension adjustment is calculated according to the relationship that the tension adjustment is equal to a proportional coefficient multiplied by the vibration deviation. If the maximum temperature exceeds 120 degrees Celsius, the coolant flow rate is increased according to the relationship that the flow rate increment is equal to a proportional coefficient multiplied by the over-temperature value. In the slow-timescale optimization layer, a multi-objective optimization function is constructed. This function includes a term representing the square of the deviation between the predicted surface roughness value and the target value of 0.3 micrometers. The squared deviation of the predicted warpage from the target value of 10 micrometers, the material removal rate, and the other two terms are multiplied by weighting coefficients of 0.5, 0.3, and 0.2 respectively and then summed. The optimization variables are the diamond wire speed, feed rate, and coolant flow rate. The constraints include a wire speed range of 8 to 15 meters per second, a feed rate range of 0.1 to 0.5 millimeters per second, and a flow rate range of 2 to 8 liters per minute. The particle swarm optimization algorithm searches for the parameter combination that minimizes the objective function within the constraints. The algorithm initializes 30 particles, each representing a set of parameter values. The particles update their speed and position based on their own historical best position and the group's historical best position. After 20 iterations, the Pareto optimal solution set is obtained. From the solution set, the solution that satisfies the condition that the predicted surface roughness is less than 0.4 micrometers and the predicted warpage is less than 15 micrometers is selected as the optimization control parameter.

[0028] Step S4: Convert the optimized control parameters into actuator drive commands for multi-parameter collaborative control, and update the temporal convolutional-long short-term memory fusion neural network through online incremental learning and adjust the execution control strategy based on abnormal operating conditions.

[0029] Specifically, the optimized control parameters are converted into actuator drive commands. The diamond wire speed is converted into a spindle motor speed adjustment via an incremental PID controller. Since changes in wire speed alter the inertial force of the diamond wire, thus affecting tension, the rate of change of wire speed is calculated and multiplied by a coupling coefficient of 12 N / s / m to obtain the wire speed-tension coupling disturbance compensation. The optimized tension is subtracted from this compensation to obtain the compensated target tension, which is then converted into the drive pressure of the pneumatic tension regulating device via a feedforward-feedback composite controller. Changes in feed rate alter the volume of material cut per unit time, thus affecting cutting heat generation. The feed rate change is calculated and multiplied by a coupling coefficient of 25 degrees Celsius / s / mm to predict the temperature change after 5 seconds. The temperature change is multiplied by 0.08 liters per minute per degree Celsius to convert it into a coolant flow compensation. The optimized coolant flow rate is then superimposed with the flow compensation and converted into a frequency adjustment command for the variable frequency pump via a PI controller. Each actuator receives drive commands synchronously within a 1-millisecond communication cycle via a real-time Ethernet bus to achieve multi-parameter coordinated control. After each silicon wafer is cut, the actual surface roughness and warpage are measured using a coordinate measuring machine. The multi-physics feature vector sequence recorded during the silicon wafer cutting process is combined with the actual quality value to form a new sample and added to the experience playback buffer. When the buffer accumulates 50 new samples, 32 new samples are randomly selected from the buffer and mixed with 32 samples randomly selected from the historical training set. A batch process is used to update the parameters of the fully connected layers and long short-term memory layers of the temporal convolutional-long short-term memory fusion neural network using gradient descent, with a learning rate of 0.0001 for 5 rounds. The updated model is then used on the validation set to calculate the surface roughness prediction error and warpage prediction error. When both errors are less than a set threshold, the updated model parameters are deployed to the real-time control system to replace the original model. The Mahalanobis distance between the multiphysics feature vector and the normal operating condition multivariate Gaussian distribution is calculated in real time. This distance is obtained by multiplying the difference between the feature vector and the mean vector by the inverse of the covariance matrix, then multiplying by the transpose of the difference vector, and finally taking the square root. When the Mahalanobis distance exceeds the threshold of 3.5 for 5 consecutive sampling periods, it is considered an abnormal operating condition. The abnormal moment features are input into a one-dimensional convolutional neural network anomaly classifier. The classifier outputs a probability distribution of 6 abnormal patterns. The category with the highest probability is selected as the recognition result. If it is identified as diamond wire wear, the wire speed is reduced by 15% and the feed speed is reduced by 20%. If it is identified as tension fluctuation, it switches to tension closed-loop control mode and increases the controller gain to 1.5 times the original. If it is identified as clamping looseness, an alarm signal is issued and the feed is stopped. If it is identified as coolant contamination, the flow rate is increased to the maximum value and the filtration system is started. If it is identified as abnormal vibration, the wire speed and feed speed are reduced by 20% each and dynamic balance compensation is started. If it is identified as feed jamming, the feed axis is reversed by 0.5 mm and then the forward feed is restarted.

[0030] In one specific embodiment, step S1 includes:

[0031] Triaxial component data of cutting force, triaxial signals of diamond wire tension and vibration acceleration were collected in the cutting contact area between diamond wire and silicon wafer to obtain time series of mechanical field parameters;

[0032] The temperature field distribution matrix, coolant inlet temperature and outlet temperature of the cutting area are collected at the cutting interface and coolant pipeline. The coolant temperature gradient is calculated to obtain the time series of thermal field parameters.

[0033] Acoustic emission signals, as well as the current and power parameters of the spindle motor and feed motor, are collected in the cutting work area to obtain the time series of acoustic field parameters and electric field parameters.

[0034] The time series of mechanical field parameters, thermal field parameters, acoustic field parameters, and electric field parameters are timestamped using a clock synchronization protocol to obtain a multiphysics field feature vector sequence.

[0035] Specifically, in the cutting contact area between the diamond wire and the silicon wafer, a three-dimensional force sensor synchronously collects cutting force component data in three directions. The cutting force component along the diamond wire's movement direction reflects the magnitude of the cutting resistance, the lateral force component perpendicular to the wire's movement direction reflects the cutting stability, and the normal force component perpendicular to the silicon wafer surface reflects the cutting depth pressure. A strain gauge tension sensor collects the real-time tension value of the diamond wire, reflecting the tensile state of the diamond wire. A MEMS accelerometer collects vibration acceleration signals in three axes, reflecting the vibration characteristics of the cutting process. The above three-axis components of the cutting force, tension value, and vibration acceleration signals are arranged in chronological order to form a time series of mechanical field parameters. The two-dimensional temperature field distribution matrix acquired by the infrared thermal imager records the temperature value of each spatial location within the cutting area as a function of time. A spatial resolution of 0.5 mm means that the distance between two adjacent temperature measurement points is 0.5 mm. The platinum resistance temperature sensor acquires the inlet and outlet temperatures of the coolant. The temperature gradient of the coolant is obtained by subtracting the inlet temperature from the outlet temperature. The larger the temperature gradient value, the more cutting heat the coolant absorbs and the better the cooling effect. The temperature field distribution matrix, inlet temperature, outlet temperature, and temperature gradient are arranged in chronological order to form a time series of thermal field parameters. The high-frequency acoustic wave signal acquired by the acoustic emission sensor contains the acoustic characteristics of material deformation and microcrack propagation caused by the contact friction between the diamond wire and the silicon wafer. The spindle motor current and power and the feed motor current acquired by the current transformer and power analyzer reflect the load state of the motor. A sudden increase in current or power indicates an increase in cutting resistance. The acoustic emission signal, motor current, and power parameters are arranged in chronological order to form a time series of acoustic field parameters and a time series of electric field parameters, respectively.

[0036] The clock synchronization protocol calculates and corrects the deviation between each sensor clock and the master clock by exchanging timestamp messages between the master and slave clocks. The calibrated timestamp deviation is less than 1 microsecond, ensuring that the data of different physical field parameters can be accurately correlated at the same time. For example, when the cutting force suddenly increases at the same moment, the changes in vibration acceleration, cutting temperature, acoustic emission intensity, and motor power can be accurately correlated. Time synchronization calibration eliminates the data misalignment problem caused by inconsistent sensor sampling times. The mechanical field parameters, thermal field parameters, acoustic field parameters, and electric field parameters calibrated by timestamps are combined into a feature vector containing all physical field information at each sampling time. The feature vectors of all sampling times are arranged in chronological order to form a multi-physical field feature vector sequence. This sequence completely records the collaborative evolution process of multiple physical fields during the cutting process.

[0037] In one specific embodiment, step S2, the processing procedure of the temporal convolutional layer, includes:

[0038] Organize the sequence of multiphysics feature vectors into input tensors with continuous time steps;

[0039] By sequentially performing causal convolution on multiple residual blocks, the dilation rate of each residual block increases exponentially, thereby extracting local features at different time scales.

[0040] Batch normalization layers and activation functions are connected in each residual block, and the original information is preserved through skip connections to obtain temporal convolutional feature vectors containing local time-varying features and transient coupling modes.

[0041] Specifically, the multiphysics feature vector sequence is organized chronologically into a continuous time-step input tensor. This tensor contains feature data from 32 consecutive time steps, with each time step corresponding to a multiphysics feature vector at a sampling time. The first dimension of the tensor represents the number of time steps (32), and the second dimension represents the number of features at each time step (51), forming a 32×51 two-dimensional input tensor. Causal convolution processing uses a kernel that only performs convolution operations with data from the current time step and previous time steps, excluding data from future time steps, ensuring that the prediction model conforms to temporal causality. The dilation rate of the first residual block is set to 1, indicating that the kernel convolves with data from consecutive adjacent time steps. When the kernel size is 3, it covers data from the current time step and the two previous time steps (a total of 3 time steps). The dilation rate of the second residual block is set to 2, indicating that the kernel convolves with one data point every other time step in the time dimension. When the kernel size is 3, it actually covers 5 time steps. The data range of the steps is as follows: the expansion rate of the third residual block is 4, covering 9 time steps; the expansion rate of the fourth residual block is 8, covering 17 time steps. The expansion rate increases exponentially with 1, 2, 4, and 8, enabling subsequent residual blocks to capture features over a longer time span. The first residual block extracts rapidly changing local features such as instantaneous fluctuations in cutting force and short-term changes in vibration acceleration; the fourth residual block extracts long-term evolving local features such as slow temperature increases and gradual degradation of coolant performance. Multiple residual blocks are processed sequentially to achieve multi-level feature extraction from short time scales to long time scales.

[0042] In each residual block, the feature data output by the convolutional layer first passes through a batch normalization layer. The batch normalization layer calculates the mean and variance of the current batch data for each feature dimension, then standardizes the data by subtracting the mean and dividing by the standard deviation. The standardized data has a mean of 0 and a variance of 1. Batch normalization eliminates the training instability caused by excessive differences in the numerical ranges of different feature dimensions. The standardized data is then input into an activation function, specifically the ReLU function. When the input value is greater than 0, the output equals the input value; when the input value is less than or equal to 0, the output is 0. The ReLU activation function introduces non-linear transformation capabilities into the network, enabling it to fit complex non-linear relationships and skip steps. The skip connections directly add the original feature vector of the residual block input to the feature vector of the residual block output. Skip connections preserve the original input information and avoid the gradient vanishing problem during deep network training. For example, the input feature vector of the first residual block is added to the output feature vector of the first residual block after convolution, batch normalization and activation to obtain the final output of the first residual block. This output is used as the input of the second residual block for further processing. After sequential processing by four residual blocks, a temporal convolution feature vector is obtained. This feature vector contains local time-varying features reflecting the change law of the cutting parameters at different time scales, and contains transient coupling modes reflecting the coupling relationship between the mechanical field and the thermal field, and between the sound field and the vibration field at a certain moment.

[0043] Figure 2 This is a schematic diagram illustrating the multi-level dilated convolution feature extraction process of the temporal convolutional layer in an embodiment of this application. For example... Figure 2 As shown, the input signal (solid line) is processed by a residual block with a dilation rate of 2 to obtain a dilation rate of 2 feature (dashed line). This feature captures the local time-varying characteristics of the input signal at a medium time scale. After being processed by a residual block with a dilation rate of 8, a dilation rate of 8 feature (dotted line) is obtained. This feature captures the evolution trend of the input signal at a long time scale. The receptive field of the dilation rate of 8 feature covers 17 time steps, compared to the receptive field of the dilation rate of 2 feature which covers 5 time steps. It can extract the segmentation process features over a longer time span. As can be seen from the figure, as the dilation rate increases exponentially, the extracted features gradually change from high-frequency rapid fluctuations to low-frequency slow changes. Convolutional layers with different dilation rates work together to achieve multi-level feature extraction from short time scales to long time scales.

[0044] In one specific embodiment, step S2, the processing procedure of the long short-term memory layer, includes:

[0045] The temporal convolutional feature vector is input into a multi-layer bidirectional long short-term memory unit;

[0046] In each long short-term memory unit, the forgetting gate calculates the forgetting weight to control the degree of retention of historical states, the input gate calculates the input weight to control the degree of updating of current information, and the output gate calculates the output weight to control the degree of output of hidden states.

[0047] By leveraging the synergistic effect of forget gates, input gates, and output gates to capture changes in the wear state of diamond wire and the degradation process of coolant performance, long short-term memory features containing long-term evolution trends are obtained.

[0048] Specifically, the temporal convolution feature vector is input into the first layer of bidirectional long short-term memory (LSTM) units. The bidirectional LSM units include forward LSM units and backward LSM units. The forward units process data from the start time to the end time of the time series, and the backward units process data from the end time to the start time. The number of hidden units in the first layer of bidirectional LSM units is set to 128, meaning that each LSM unit outputs a 128-dimensional hidden state vector. The forward and backward hidden state vectors processed by the first layer are concatenated to form a 256-dimensional feature vector. This feature vector is input into the second layer of bidirectional LSM units. The number of hidden units in the second layer is 64. The forward and backward hidden states output by the second layer are concatenated to form a 128-dimensional LSM feature. The forget gate maps input data to values ​​between 0 and 1 using the sigmoid activation function. When the forget weight is close to 1, it means that most of the historical state information is retained, and when the forget weight is close to 0, it means that most of the historical state information is discarded. The calculation process of the forget gate is to concatenate the temporal convolution feature vector of the current time step with the hidden state vector of the previous time step, multiply it by the weight matrix of the forget gate, add a bias term, and then obtain the forget weight through the sigmoid function. The forget weight is then multiplied element-wise with the cell state of the previous time step to obtain the retained historical cell state.

[0049] The input gate calculates input weights to control the degree to which new information updates the cell state at the current time step. The input gate also maps the calculation result to between 0 and 1 through the sigmoid function. When the input weight is close to 1, it means that the current information contributes a lot to the cell state update, and when the input weight is close to 0, it means that the current information contributes a little. The calculation process of the input gate is as follows: the feature vector at the current time step is concatenated with the hidden state at the previous time step, multiplied by the input gate weight matrix plus a bias, and then passed through the sigmoid function to obtain the input weight. At the same time, the concatenated vector is multiplied by the candidate cell state weight matrix plus a bias and then passed through the tanh function to obtain the candidate cell state. The value of the candidate cell state is between -1 and +1. The input weight and the candidate cell state are multiplied element by element to obtain the cell state update amount at the current time step. The retained historical cell state is added to the cell state update amount to obtain the new cell state at the current time step. The output gate calculates the output weights to control how much information the cell state contains as the output hidden state. The output gate calculates the output weights using the sigmoid function. The current time step feature vector is concatenated with the previous time step hidden state, multiplied by the output gate weight matrix, and then passed through the sigmoid function to obtain the output weights. The current time step cell state is mapped to a value between -1 and +1 using the tanh function. The output weights are then multiplied element-wise with the tanh-mapped cell state to obtain the current time step hidden state output.

[0050] The synergistic effect of the forget gate, input gate, and output gate enables the Long Short-Term Memory (LSTM) unit to retain important information when processing long-term sequences. When the diamond wire is sharp, the cutting force is small and stable. As the number of cutting pieces increases, the diamond wire gradually wears down, causing the cutting force to increase slowly. The forget gate initially assigns a large forgetting weight to discard small random fluctuations in the cutting force. When the cutting force shows a continuous upward trend, the input gate assigns a large input weight to update the cell state and record information about increased wear. When the wear reaches a certain level, the output gate outputs a large hidden state value to transmit a signal to subsequent network layers that the diamond wire needs to be replaced. The LSM unit captures the process of the diamond wire changing from sharp to worn state through the cumulative update of the cell state. Similarly, a large temperature gradient in the early stage of coolant use indicates good cooling effect. As silicon powder particles and impurities accumulate in the coolant, the heat conduction efficiency decreases, and the coolant temperature gradient gradually decreases. The LSM unit records the long-term decay trend of the coolant temperature gradient through a gating mechanism. The 128-dimensional feature vector output by the second-layer bidirectional LSM unit contains information on the long-term evolution trend of the diamond wire wear state change and the coolant performance degradation process.

[0051] In one specific embodiment, the computation process of the attention mechanism layer includes:

[0052] Long Short-Term Memory features are mapped using Query-Key-Value to generate query vectors, key vectors, and value vectors.

[0053] Calculate the attention score function based on the query vector and key vector;

[0054] The attention score function is normalized to obtain the attention weights of the physical field features at each time step. The attention weights reflect the degree of contribution of different physical field parameters to the cutting quality.

[0055] The attention weights are summed with the value vector to obtain the context vector;

[0056] The context vector is input into the fully connected layer for nonlinear mapping, and the output is the predicted value of silicon wafer surface roughness, the predicted value of warpage, and the probability of cutting anomalies.

[0057] Specifically, the long short-term memory features are multiplied by three different weight matrices to obtain the query vector, key vector, and value vector. The three weight matrices are learned through network training. The query vector represents the query information that needs to be focused on at the moment, the key vector represents the index information of the features at each time step, and the value vector represents the actual content information of the features at each time step. The dimensions of the three vectors are the same as the dimensions of the long short-term memory features, which is 128 dimensions. The attention score function is calculated using a dot product approach. The query vector is multiplied by the key vector at each time step. A larger dot product value indicates a stronger relevance between the time step and the query. This yields 32 attention scores for each of the 32 time steps. These 32 scores are scaled by dividing by the square root of the key vector dimension to avoid excessively large values. Then, they are normalized using the softmax function. The softmax function performs an exponential operation on all scores and divides them by the sum of the exponents. The sum of the attention weights is 1, and each weight value is between 0 and 1. When the cutting force suddenly increases or the temperature rises sharply at a certain time step, the attention weight value corresponding to that time step is larger, indicating that this moment has a significant impact on the future cutting quality. The attention weights of time steps in the normal and stable cutting phase are smaller.

[0058] The attention weights at each time step are multiplied by the value vectors of the corresponding time steps. Then, the 32 weighted value vectors are summed to obtain the context vector. The context vector integrates the information of all time steps and highlights the features of important time steps according to the attention weights. It has a dimension of 128. The context vector is input into the first fully connected layer, which contains 128 neurons. The 128 elements of the context vector are multiplied by their corresponding weights, summed, and then a bias is added to obtain the 128 output values ​​of the first layer. The output values ​​are processed by the ReLU activation function and then input into the second fully connected layer, which contains 64 neurons. Each neuron performs a weighted summation of the 128 outputs of the first layer, adds a bias, and then performs ReLU activation. The 64 outputs of the second layer are input into the third fully connected layer, which contains 3 neurons corresponding to the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability, respectively. The predicted surface roughness reflects the degree of microscopic unevenness of the silicon wafer surface, with a value range of 0.1 μm to 2.0 μm. The predicted warpage reflects the degree of macroscopic deformation of the silicon wafer, with a value range of 5 μm to 50 μm. The cutting anomaly probability ranges from 0 to 1, indicating the possibility of abnormal conditions such as diamond wire breakage or silicon wafer breakage occurring within the next 10 sampling periods.

[0059] In one specific embodiment, step S3 includes:

[0060] Based on the probability of cutting anomalies, it is determined whether the preset threshold is exceeded. When the threshold is exceeded, the deviation between the current monitoring parameters and the benchmark value is calculated in the fast time scale control layer. The diamond wire speed adjustment, tension adjustment and coolant flow adjustment are calculated according to the proportional relationship to obtain the fast response control parameters.

[0061] Based on the predicted values ​​of silicon wafer surface roughness and warpage, a multi-objective optimization function is constructed in the slow timescale optimization layer. The multi-objective optimization function takes the quality target deviation and material removal rate as optimization objectives, and takes the equipment physical constraints and process safety boundaries as constraints.

[0062] The Pareto optimal solution set is obtained by iterative calculation of the global optimization of diamond wire linear velocity, feed rate and coolant flow rate within the constraints through a swarm intelligence optimization algorithm.

[0063] Solutions that meet the quality threshold requirements are selected from the Pareto optimal solution set, and then combined with fast response control parameters to obtain optimized control parameters.

[0064] Specifically, the cutting anomaly probability output by the fully connected layer is compared with a preset threshold of 0.15. When the cutting anomaly probability exceeds 0.15, an anomaly risk is identified, requiring a rapid response. The fast timescale control layer has a working cycle of 100 milliseconds. The deviation between the current average cutting force and the set reference value is calculated. The reference value is determined based on the silicon wafer material characteristics and cutting process. When the absolute value of the cutting force deviation exceeds 5 Newtons, the cutting force deviation is multiplied by a scaling factor of 0.02 meters per second per Newton to obtain the diamond wire speed adjustment. For example, when the cutting force deviation is positive 8 Newtons, the speed adjustment is positive 0. 0.16 m / s indicates that the linear speed needs to be increased. Calculate the deviation between the peak value of the vibration acceleration at the main frequency and the reference value. When the vibration deviation exceeds 2 m / s², multiply the vibration deviation by a proportionality factor of 3 N / m / s² to obtain the tension adjustment. Calculate the difference between the highest temperature in the cutting area and the set threshold of 120 degrees Celsius. When the highest temperature exceeds the threshold, multiply the over-temperature value by a proportionality factor of 0.05 L / min / degree Celsius to obtain the coolant flow rate adjustment. Combine the linear speed adjustment, tension adjustment, and flow rate adjustment to form a fast-response control parameter for rapid compensation of transient disturbances.

[0065] The slow timescale optimization layer has a working cycle of 5 seconds. A multi-objective optimization function is constructed. The first term is the squared difference between the predicted value of the silicon wafer surface roughness and the target value of 0.3 micrometers. The second term is the squared difference between the predicted value of the warpage and the target value of 10 micrometers. The third term is the material removal rate, which is calculated by multiplying the diamond wire speed and the feed rate to reflect the cutting efficiency. The three terms are multiplied by weight coefficients of 0.5, 0.3, and 0.2 respectively and then summed to form the multi-objective optimization function. The optimization objective is to minimize the deviation of the first two terms and maximize the material removal rate of the third term. The physical constraints of the equipment include a diamond wire speed range of 8 m / s to 15 m / s, a feed rate range of 0.1 mm / s to 0.5 mm / s, a coolant flow rate range of 2 liters / min to 8 liters / min, and process safety boundaries including a cutting force not exceeding 50 Newtons, a tension not exceeding 200 Newtons, and a maximum temperature not exceeding 150 degrees Celsius. Particle swarm optimization (PSO) is used for global optimization. Thirty particles are initialized, each representing a set of values ​​for diamond wire velocity, feed rate, and coolant flow rate. Particles are randomly initialized within constraints. Each particle updates its velocity and position based on its own historical best position and the group's historical best position. The velocity update formula includes inertia, individual cognition, and social cognition terms. The inertia weight decreases linearly from 0.9 to 0.4. Both the individual learning factor and the group learning factor are set to 2.0. After 20 iterations, multiple parameter combinations that minimize the objective function are obtained. These combinations constitute the Pareto optimal solution set.

[0066] From the Pareto optimal solution set, solutions that satisfy the predicted surface roughness value of less than 0.4 micrometers and the predicted warpage value of less than 15 micrometers are selected. If multiple solutions meet the conditions, the solution with the highest material removal rate is selected. The selected diamond wire speed, feed rate, and coolant flow rate are added to the wire speed adjustment and flow rate adjustment in the fast response control parameters. For example, if the selected wire speed is 12 meters per second and the wire speed adjustment in the fast response control parameters is -0.3 meters per second, the two are added together to obtain the optimized wire speed of 11.7 meters per second. The tension is adjusted from the reference tension by the tension adjustment in the fast response control parameters to obtain the optimized tension. The optimized diamond wire speed, feed rate, coolant flow rate, and tension are combined to form the optimized control parameters.

[0067] Figure 3 This is a schematic diagram illustrating the dual-timescale predictive control parameter optimization process in an embodiment of this application. For example... Figure 3As shown, during the 0-3 second period, the anomaly probability (dotted line) is below the threshold of 0.15, and both the linear velocity adjustment (solid line) and feed rate adjustment (dashed line) remain at 0 and require no adjustment. At 3 seconds, the anomaly probability exceeds the threshold, triggering a fast timescale control layer response. The linear velocity adjustment rapidly decreases to -0.3 within 3-3.5 seconds, indicating a reduction in linear velocity to mitigate the anomaly risk. At 5 seconds, the feed rate adjustment decreases to -0.25, indicating a simultaneous reduction in feed rate. During the 3-8 second period, the anomaly probability gradually decreases from 0.2, demonstrating that the control strategy effectively suppresses the development of abnormal conditions. At 8 seconds, the anomaly probability drops below 0.1, and the linear velocity adjustment begins to rise, indicating a restoration of normal cutting parameters. The figure clearly shows the rapid response of the fast timescale control layer to the anomaly probability within a 100-millisecond period and the dynamic adjustment process of the control parameters, demonstrating the forward-looking compensation capability of the dual timescale predictive control framework for transient disturbances.

[0068] In one specific embodiment, step S4 includes:

[0069] The diamond wire speed in the optimized control parameters is converted into spindle drive commands through a closed-loop controller, and the wire speed-tension coupling disturbance compensation is calculated based on the wire speed change rate and the preset coupling coefficient.

[0070] The target tension after compensation is obtained by subtracting the line speed-tension coupling disturbance compensation from the diamond wire tension in the optimized control parameters, and then converted into a driving command for the tension adjustment device by the tension controller.

[0071] The feed rate change is calculated from the feed rate in the optimized control parameters. The temperature change is predicted based on the feed-temperature coupling relationship and converted into coolant flow compensation. Then, a feed drive command is generated.

[0072] The coolant flow rate in the optimized control parameters is superimposed with the coolant flow rate compensation amount and converted into a cooling system adjustment command to achieve multi-parameter coordinated control.

[0073] After each silicon wafer dicing is completed, the actual quality value is collected. The feature sequence of the dicing process and the actual quality value are combined to form a new sample and added to the experience playback buffer. When the accumulated sample reaches a preset number, the new sample is extracted from the buffer and mixed with the historical sample to perform incremental updates on the temporal convolutional-long short-term memory fusion neural network.

[0074] The deviation between the multiphysics field feature vector sequence and the statistical distribution under normal operating conditions is calculated in real time. When the deviation exceeds the threshold, the anomaly type is identified by an anomaly classifier, and the corresponding control strategy is adjusted according to the anomaly type.

[0075] Specifically, the diamond wire speed in the optimized control parameters is input into an incremental PID controller. The controller calculates the deviation between the current and target wire speeds, the integral of the deviation, and the derivative of the deviation. These three values ​​are multiplied by the proportional, integral, and derivative coefficients, respectively, and then summed to obtain the spindle motor speed adjustment. The speed adjustment is converted into a speed-to-current value to generate a spindle drive command. The difference between the current and previous wire speeds is calculated and divided by the time interval to obtain the rate of change of wire speed. The rate of change of wire speed is multiplied by a preset coupling coefficient of 12 N / s / m to obtain the wire speed-tension coupling disturbance compensation. This compensation reflects the disturbance effect of the wire speed change on the tension caused by the change in inertial force of the diamond wire. The wire speed-tension coupling is then subtracted from the diamond wire tension in the optimized control parameters. The disturbance compensation amount yields the target tension, which is then input to the feedforward-feedback composite controller. The feedforward controller looks up the cylinder pressure setpoint of the pneumatic tension regulator based on the target tension. The feedback controller calculates the deviation between the current tension and the target tension and outputs the pressure adjustment amount through the PI controller. The two are then superimposed to generate the drive command for the tension regulator. The feed rate change is obtained by subtracting the current feed rate from the optimized control parameters. This feed rate change is multiplied by the feed-temperature coupling coefficient of 25 degrees Celsius per second per millimeter and then by 5 seconds to obtain the predicted temperature change. The temperature change is multiplied by 0.08 liters per minute per degree Celsius to convert it into coolant flow compensation. The feed rate is then input into the speed-position dual closed loop. The controller generates feed drive commands, adding coolant flow rate compensation to the optimized control parameters to obtain the total flow requirement. This total flow requirement is input to the PI controller, which, based on the deviation between the current flow rate and the target flow rate, outputs the frequency adjustment of the variable frequency pump to generate cooling system adjustment commands. These drive commands are synchronously sent to the spindle motor, tension adjustment device, feed motor, and variable frequency pump within a 1-millisecond communication cycle via a real-time Ethernet bus, achieving multi-parameter coordinated control. After each silicon wafer is cut, the actual surface roughness and warpage are measured using a coordinate measuring machine. The multiphysics feature vector sequence recorded during the wafer cutting process is combined with the actual quality value to form a new sample, which is added to an experience playback buffer with a capacity of 500 samples. When 50 new samples accumulate in the buffer, 32 new samples are randomly selected from the buffer and mixed with 32 samples randomly selected from the historical training set to form a batch. Five rounds of gradient descent are performed to update the parameters of the fully connected layer and the long short-term memory layer of the temporal convolutional-long short-term memory fusion neural network, with the learning rate set to 0.0001. After the update, the model calculates the surface roughness prediction error and warpage prediction error on the validation set. When both errors are less than the set threshold, the updated model parameters are deployed to replace the original model. The Mahalanobis distance is obtained by multiplying the difference between the multiphysics feature vector and the mean vector of the normal working condition multivariate Gaussian distribution, multiplying by the inverse of the covariance matrix, and then multiplying by the transpose of the difference vector. When the Mahalanobis distance exceeds the threshold of 3 for 5 consecutive sampling periods, the Mahalanobis distance is calculated.At 5 o'clock, an abnormal operating condition is identified. The characteristics of the abnormal moment are input into a one-dimensional convolutional neural network anomaly classifier. The classifier outputs the probability distribution of six types of anomalies: diamond wire wear, tension fluctuation, clamping looseness, coolant contamination, abnormal vibration, and feed jamming. The category with the highest probability is selected as the identification result. If diamond wire wear is identified, the wire speed is reduced by 15% and the feed speed by 20%. If tension fluctuation is identified, the system switches to tension closed-loop control mode and increases the controller gain to 1.5 times. If clamping looseness is identified, an alarm is issued and feed is paused. If coolant contamination is identified, the flow rate is increased to the maximum value and the filtration system is activated. If abnormal vibration is identified, the wire speed and feed speed are reduced by 20% each and dynamic balancing compensation is activated. If feed jamming is identified, the feed axis is reversed by 0.5 mm and then the forward feed is restarted.

[0076] The intelligent control method for diamond wire cutting of silicon wafers in this application has been described above. The intelligent control system for diamond wire cutting of silicon wafers in this application is described below. One embodiment of the intelligent control system for diamond wire cutting of silicon wafers in this application includes:

[0077] The data acquisition module is used to acquire data from the cutting work area through a multi-physics field collaborative monitoring system, and obtain a multi-physics field feature vector sequence.

[0078] The modeling module is used to predict and model the multi-physics feature vector sequence through a temporal convolution-long short-term memory fusion neural network. The temporal convolutional layer uses a causal convolution structure and multi-level dilated convolution to extract the local time-varying features and transient coupling modes of the multi-physics parameters. The long short-term memory layer captures the long-term evolution trend of diamond wire wear and coolant performance degradation. The attention mechanism layer calculates the dynamic weights of each physical field feature and calculates the weighted sum to obtain the context vector. The fully connected layer outputs the predicted values ​​of silicon wafer surface roughness, warpage, and cutting anomaly probability.

[0079] The calculation module is used to perform optimization calculations on the control parameters based on the predicted values ​​of silicon wafer surface roughness, warpage, and the cutting anomaly probability, to obtain optimized control parameters.

[0080] The execution module is used to convert the optimized control parameters into actuator drive instructions for multi-parameter coordinated control, and to update the temporal convolutional-long short-term memory fusion neural network through online incremental learning and to adjust the execution control strategy based on abnormal operating conditions.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent control method for cutting silicon wafers with diamond wires, characterized in that, The method comprises: Step S1: data acquisition on the cutting work area by a multi-physical field cooperative monitoring system to obtain a multi-physical field characteristic vector sequence; Step S2: prediction modeling of the multi-physical field characteristic vector sequence by a time series convolution-long short-term memory fusion neural network, wherein a time series convolution layer adopts a causal convolution structure and a multi-level dilated convolution to extract local time-varying characteristics and transient coupling modes of multi-physical field parameters, a long short-term memory layer captures long-term evolution trends of diamond wire wear and cooling liquid performance attenuation, an attention mechanism layer calculates dynamic weights of each physical field characteristic and obtains a context vector by weighted summation, and a full connection layer outputs a silicon wafer surface roughness prediction value, a warping degree prediction value and a cutting abnormality probability; Step S3: control parameter optimization calculation according to the silicon wafer surface roughness prediction value, the warping degree prediction value and the cutting abnormality probability to obtain optimized control parameters; Step S4: conversion of the optimized control parameters into actuator driving instructions for multi-parameter cooperative control, and updating of the time series convolution-long short-term memory fusion neural network through online incremental learning and adjustment of a control strategy based on abnormal working condition identification, including: conversion of a diamond wire speed in the optimized control parameters into a spindle driving instruction through a closed-loop controller, and calculation of a speed-tension coupling disturbance compensation amount based on a speed change rate and a preset coupling coefficient; subtraction of the speed-tension coupling disturbance compensation amount from the diamond wire tension in the optimized control parameters to obtain a compensated target tension, and conversion of the compensated target tension into a tension adjusting device driving instruction through a tension controller; calculation of a feed speed change amount based on the feed speed in the optimized control parameters, prediction of a temperature change amount based on a feed-temperature coupling relationship and conversion of the temperature change amount into a cooling liquid flow compensation amount to generate a feed driving instruction; superposition of the cooling liquid flow in the optimized control parameters and the cooling liquid flow compensation amount to convert into a cooling system adjusting instruction to realize the multi-parameter cooperative control; after each silicon wafer cutting, actual quality values are collected, feature sequences and actual quality values in the cutting process are combined to form new samples which are added to an experience replay buffer, when the accumulated samples reach a preset number, new samples are extracted from the buffer to mix with historical samples to update the time series convolution-long short-term memory fusion neural network incrementally; real-time calculation of a deviation degree between the multi-physical field characteristic vector sequence and a normal working condition statistical distribution, identification of an abnormal type by an abnormality classifier when the deviation degree exceeds a threshold, and execution of a corresponding control strategy adjustment according to the abnormal type. 2.The intelligent control method for cutting silicon wafers with diamond wires according to claim 1, wherein, The step S1 comprises: Collecting three-axis component data of cutting force, diamond wire tension value and three-axis signals of vibration acceleration in a cutting contact area of the diamond wire and the silicon wafer to obtain a mechanical field parameter time sequence; Collecting a cutting region temperature field distribution matrix, cooling liquid inlet temperature and outlet temperature at a cutting interface and a cooling liquid pipeline to calculate a cooling liquid temperature gradient and obtain a thermal field parameter time sequence; Collecting acoustic emission signals, current and power parameters of a spindle motor and a feed motor in the cutting work area to obtain an acoustic field parameter time sequence and an electric field parameter time sequence; The time series of the mechanical field parameters, the time series of the thermal field parameters, the time series of the acoustic field parameters and the time series of the electrical field parameters are timestamped and calibrated through a clock synchronization protocol to obtain the sequence of multi-physical field feature vectors. 3.The intelligent control method for cutting silicon wafers with diamond wires according to claim 1, characterized in that, In the step S2, the processing procedure of the time series convolution layer includes: organizing the sequence of multi-physical field feature vectors into an input tensor of continuous time steps; performing causal convolution processing through multiple residual blocks in sequence, the expansion rate of each residual block being exponentially increased, and local features of different time scales being extracted; connecting a batch normalization layer and an activation function in each residual block and retaining original information through a skip connection to obtain a time series convolution feature vector containing the local time-varying features and the transient coupling mode. 4.The intelligent control method for cutting silicon wafers with diamond wires according to claim 3, characterized in that, In the step S2, the processing procedure of the long short-term memory layer includes: inputting the time series convolution feature vector into multiple layers of bidirectional long short-term memory units; calculating a forgetting weight through a forgetting gate in each long short-term memory unit to control the degree of history state retention, calculating an input weight through an input gate to control the degree of current information update, and calculating an output weight through an output gate to control the degree of hidden state output; capturing the diamond wire wear state change and the cooling liquid performance attenuation process based on the synergistic effect of the forgetting gate, the input gate and the output gate to obtain a long short-term memory feature containing the long-term evolution trend. 5.The intelligent control method for cutting silicon wafers with diamond wires according to claim 4, characterized in that, In the step S2, the calculation procedure of the attention mechanism layer includes: generating a query vector, a key vector and a value vector through Query-Key-Value mapping of the long short-term memory feature; calculating an attention score function based on the query vector and the key vector; performing normalization processing on the attention score function to obtain an attention weight of the physical field feature of each time step, the attention weight reflecting the contribution degree of different physical field parameters to the cutting quality; performing weighted summation of the attention weight and the value vector to obtain the context vector; inputting the context vector into the fully connected layer for nonlinear mapping to output the wafer surface roughness prediction value, the warping degree prediction value and the cutting abnormality probability. 6.The intelligent control method for cutting silicon wafers with diamond wires according to claim 1, wherein, The step S3 includes: judging whether the cutting abnormality probability exceeds a preset threshold based on the cutting abnormality probability, calculating a deviation between the current monitoring parameter and a reference value in a fast time scale control layer when the threshold is exceeded, calculating a diamond wire linear speed adjustment amount, a tension adjustment amount and a cooling liquid flow adjustment amount according to a proportional relationship to obtain a fast response control parameter; constructing a multi-objective optimization function in a slow time scale optimization layer based on the wafer surface roughness prediction value and the warping degree prediction value, the multi-objective optimization function taking a quality target deviation and a material removal rate as optimization objectives and taking device physical constraints and process safety boundaries as constraint conditions; performing global optimization of the diamond wire linear speed, the feed speed and the cooling liquid flow within the constraint range through a swarm intelligence optimization algorithm to obtain a Pareto optimal solution set through iterative calculation; filtering a solution meeting the quality threshold requirement from the Pareto optimal solution set and combining the fast response control parameter to obtain the optimization control parameter.

7. An intelligent control system for cutting silicon wafers with diamond wires, characterized in that, The intelligent control method for realizing the diamond wire cutting silicon wafer as claimed in any one of claims 1 to 6, the intelligent control system of the diamond wire cutting silicon wafer comprises: A collection module is configured to collect data of a cutting work area through a multi-physical field cooperative monitoring system to obtain a multi-physical field feature vector sequence; A modeling module is configured to perform predictive modeling on the multi-physical field feature vector sequence through a time series convolution-long short-term memory fusion neural network, wherein a time series convolution layer adopts a causal convolution structure and a multi-level dilated convolution to extract local time-varying features and transient coupling modes of multi-physical field parameters, a long short-term memory layer captures long-term evolution trends of diamond wire wear and cooling liquid performance attenuation, an attention mechanism layer calculates dynamic weights of each physical field feature and performs weighted summation to obtain a context vector, and a full connection layer outputs a silicon wafer surface roughness prediction value, a warping degree prediction value and a cutting abnormality probability; A calculation module is configured to perform control parameter optimization calculation according to the silicon wafer surface roughness prediction value, the warping degree prediction value and the cutting abnormality probability to obtain optimized control parameters; An execution module is configured to convert the optimized control parameters into execution mechanism driving instructions for multi-parameter cooperative control, and update the time series convolution-long short-term memory fusion neural network through online incremental learning and perform control strategy adjustment based on abnormal working condition identification, including: converting a diamond wire speed in the optimized control parameters into a spindle driving instruction through a closed-loop controller, and calculating a speed-tension coupling disturbance compensation amount based on a speed change rate and a preset coupling coefficient; subtracting the speed-tension coupling disturbance compensation amount from the diamond wire tension in the optimized control parameters to obtain a compensated target tension, and converting the compensated target tension into a tension adjusting device driving instruction through a tension controller; calculating a feed speed change amount based on the feed speed in the optimized control parameters, predicting a temperature change amount based on a feed-temperature coupling relationship and converting the temperature change amount into a cooling liquid flow compensation amount to generate a feed driving instruction; superimposing the cooling liquid flow in the optimized control parameters on the cooling liquid flow compensation amount to convert the cooling liquid flow compensation amount into a cooling system adjusting instruction to realize the multi-parameter cooperative control; collecting an actual quality value after completing silicon wafer cutting, adding a new sample composed of a feature sequence of a cutting process and the actual quality value to an experience playback buffer, and when the cumulative samples reach a preset number, extracting new samples from the buffer to mix with historical samples to perform incremental update on the time series convolution-long short-term memory fusion neural network; and calculating a deviation degree between the multi-physical field feature vector sequence and a normal working condition statistical distribution in real time, identifying an abnormal type through an abnormality classifier when the deviation degree exceeds a threshold value, and performing corresponding control strategy adjustment according to the abnormal type.

8. An intelligent control device for cutting silicon wafers with diamond wires, characterized in that, The computer program is stored in the memory and can be run on the processor, and the processor implements the intelligent control method for the diamond wire cutting silicon wafer as claimed in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is stored in the memory and can be run on the processor, and the processor implements the intelligent control method for the diamond wire cutting silicon wafer as claimed in any one of claims 1 to 6 when executing the computer program.

Images