An automatic cutting fluid circulation control method and system for CNC intelligent manufacturing

Abstract

The application relates to the technical field of mechanical manufacturing automation, and discloses an automatic cutting fluid circulation control method and system for CNC intelligent manufacturing, which comprises the following steps: collecting multi-source working condition data of a CNC machine tool, carrying out noise filtering and quantifying machining load fluctuation to obtain machining load fluctuation indexes; combining historical parameter change data to carry out parameter coupling analysis, clustering grouping and mode matching, and predicting a cutting fluid flow target value and matching a flow regulation sequence through a neural network model; according to the flow regulation sequence and a denoising data set, carrying out power change level division and feed speed sequence matching optimization to obtain a feed speed regulation sequence, and further obtaining a cooling compensation sequence through cooling uniformity analysis; and according to the cooling compensation sequence, the machining load fluctuation indexes and the flow regulation sequence, carrying out correlation analysis and global compensation adjustment to obtain a global regulation and control sequence. The method improves the cooling uniformity of a tool and a workpiece during machining load fluctuation.

Description

Technical Field

[0001] This invention relates to the field of mechanical manufacturing automation technology, and in particular to an automated cutting fluid circulation control method and system for CNC intelligent manufacturing. Background Technology

[0002] Currently, in CNC intelligent manufacturing, cutting fluid circulation control plays a crucial role in ensuring machining quality and efficiency. Because process parameters such as feed rate and depth of cut fluctuate continuously during machining, the demand for cutting fluid flow changes dynamically. Especially when machining complex-shaped parts, sudden changes in local load can cause instantaneous temperature increases. If the flow rate cannot be adjusted in time, it can easily lead to tool overheating and a decrease in surface finish.

[0003] In one existing technology, the operator first pre-sets a set of fixed process parameters, such as feed rate and depth of cut, in the system based on the material type of the workpiece and the diameter of the cutting tool. Based on these parameters, the system calculates and sets a corresponding, fixed cutting fluid supply flow rate by consulting a built-in empirical data table. After the machining process starts, regardless of how the cutting load dynamically changes during actual machining, such as drastic fluctuations in the depth of cut during contour machining, the system's flow control unit strictly maintains this initially set flow rate value, and the pumping system and valve openings remain unchanged until the current machining operation is completed. The entire control process forms an open-loop structure, with no feedback or adjustment link between the flow output and the real-time operating status of the machine tool. Traditional technology relies on a preset fixed flow rate, which cannot detect real-time changes in feed rate and depth of cut during machining, causing flow rate adjustment to lag behind load fluctuations and resulting in uneven cooling.

[0004] Therefore, existing technologies cannot guarantee uniform cooling of the cutting tool and workpiece when the machining load fluctuates. Summary of the Invention

[0005] This invention provides an automated cutting fluid circulation control method and system for CNC intelligent manufacturing to improve the cooling uniformity of tools and workpieces when machining load fluctuates.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an automated cutting fluid circulation control method for CNC intelligent manufacturing, comprising:

[0007] Data on feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool are collected to obtain the raw dataset;

[0008] Based on the original dataset, noise is filtered out to obtain a denoised dataset, and the processing load fluctuation is quantified to obtain a processing load fluctuation index.

[0009] Based on the processing load fluctuation index and combined with the pre-stored historical parameter change data, parameter coupling analysis, clustering and pattern matching are performed to obtain parameter coupling characteristics;

[0010] Based on the parameter coupling characteristics, the target value of the cutting fluid flow rate is predicted and the flow rate adjustment sequence is matched by a pre-built neural network model.

[0011] Based on the flow rate regulation sequence and the denoised dataset, power change levels are classified, feed rate sequence is matched and optimized to obtain the feed rate regulation sequence.

[0012] Based on the feed rate adjustment sequence and the denoised dataset, a cooling uniformity analysis is performed to obtain a cooling compensation sequence;

[0013] Based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence, correlation analysis, global compensation adjustment, path optimization, and flow regulation are performed.

[0014] Secondly, the present invention provides an automated cutting fluid circulation control system for CNC intelligent manufacturing, comprising:

[0015] The data acquisition module is used to collect data on the feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool to obtain the raw dataset.

[0016] The load fluctuation quantization module is used to filter out noise based on the original dataset to obtain a denoised dataset and quantify the processed load fluctuation to obtain a processed load fluctuation index.

[0017] The coupling characteristic analysis module is used to perform parameter coupling analysis, clustering and pattern matching based on the processing load fluctuation index and pre-stored historical parameter change data to obtain parameter coupling characteristics;

[0018] The flow prediction module is used to predict the target value of the cutting fluid flow rate and match the flow adjustment sequence based on the parameter coupling characteristics through a pre-built neural network model.

[0019] The feed rate adjustment module is used to perform power change level classification, feed rate sequence matching and sequence optimization based on the flow rate adjustment sequence and the denoised dataset to obtain the feed rate adjustment sequence;

[0020] The cooling uniformity analysis module is used to perform cooling uniformity analysis based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence.

[0021] The output module is used to perform correlation analysis, global compensation adjustment, path optimization, and flow regulation based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) This invention collects data on feed rate, cutting depth, spindle speed, vibration frequency and temperature change of CNC machine tool, and combines Kalman filtering algorithm for noise filtering and weighted averaging to quantify machining load fluctuations in real time, thereby improving data accuracy and real-time performance.

[0024] (2) Based on the processing load fluctuation index and combined with the pre-stored historical parameter change data, this invention performs parameter coupling analysis and pattern matching through fuzzy logic algorithm, k-nearest neighbor algorithm, Pearson correlation coefficient analysis and Gaussian mixture clustering, and uses a neural network model to predict the target value of cutting fluid flow rate, thereby realizing intelligent prediction and accurate matching of cutting fluid flow rate and effectively responding to dynamic changes in processing parameters.

[0025] (3) This invention optimizes the feed rate adjustment sequence by dividing the power change level, matching the feed rate sequence and analyzing the cooling uniformity, combined with the fast Fourier transform and weighted least squares method, and generates a cooling compensation sequence by using fuzzy logic algorithm and extreme value detection, so as to ensure that the cutting fluid flow rate changes in coordination with the machining parameters, achieve uniform cooling of the tool and the workpiece, and avoid local overheating.

[0026] (4) This invention uses grey relational analysis and weighted moving average to perform correlation analysis and global compensation adjustment on cooling compensation sequence, machining load fluctuation index and flow regulation sequence, outputs global control sequence, realizes multi-parameter collaborative control and global optimization of cutting fluid circulation, and improves the stability and accuracy of machining process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the automated cutting fluid circulation control method for CNC intelligent manufacturing provided in the first embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of the automated cutting fluid circulation control system for CNC intelligent manufacturing provided in the second embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Reference Figure 1 The first embodiment of the present invention provides an automated cutting fluid circulation control method for CNC intelligent manufacturing, comprising the following steps:

[0031] S11: Collect data on the feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool to obtain the raw dataset;

[0032] S12, Based on the original dataset, noise is filtered out to obtain a denoised dataset and the processing load fluctuation is quantified to obtain a processing load fluctuation index;

[0033] S13. Based on the processing load fluctuation index and combined with the pre-stored historical parameter change data, perform parameter coupling analysis, clustering and pattern matching to obtain parameter coupling characteristics;

[0034] S14, Based on the parameter coupling characteristics, the target value of the cutting fluid flow rate is predicted by a pre-built neural network model and the flow rate adjustment sequence is matched;

[0035] S15, based on the flow rate adjustment sequence and the denoised dataset, perform power change level classification, feed rate sequence matching and sequence optimization to obtain the feed rate adjustment sequence;

[0036] S16, Based on the feed rate adjustment sequence and the denoised dataset, perform cooling uniformity analysis to obtain a cooling compensation sequence;

[0037] S17. Based on the cooling compensation sequence, the processing load fluctuation index, and the flow rate adjustment sequence, perform correlation analysis, global compensation adjustment, path optimization, and flow rate adjustment.

[0038] In step S11, data on the feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool are collected to obtain the raw dataset.

[0039] Specifically, sensors installed on the CNC machine tool collect real-time data on feed rate, depth of cut, spindle speed, vibration frequency, and temperature changes, along with cutting fluid flow rate data. These data collectively constitute the raw dataset. Feed rate data originates from the encoder of the machine tool's feed axis, representing the tool's movement speed during machining. Depth of cut data is acquired through a position sensor, reflecting the thickness of the cut layer. Spindle speed data is measured by the spindle encoder, indicating the spindle's rotational speed. Vibration frequency data is captured by an accelerometer, characterizing the machine tool's vibration state during operation. Temperature change data utilizes thermocouples or infrared sensors to monitor the temperature of the machining area. Cutting fluid flow rate data is recorded by a flow meter, recording the cutting fluid supply. The data acquisition process involves converting sensor signals into digital signals via an analog-to-digital converter and storing them as structured data in a time-series format. This step provides the foundational data for subsequent noise filtering and quantification of machining load fluctuations.

[0040] In step S12, noise is filtered out based on the original dataset to obtain a denoised dataset and the processing load fluctuation is quantified to obtain a processing load fluctuation index.

[0041] In one specific implementation, the step of filtering out noise from the original dataset to obtain a denoised dataset and quantifying processing load fluctuations to obtain a processing load fluctuation index includes:

[0042] Based on the original dataset, a denoised dataset is obtained by performing a Kalman filter algorithm.

[0043] Based on the denoised dataset, the processing load fluctuation index is obtained by weighted summation of the denoised dataset using a weighted average algorithm;

[0044] The denoising dataset includes denoising feed rate, denoising depth of cut, denoising spindle speed, denoising vibration frequency, denoising temperature change data, and denoising cutting fluid flow rate.

[0045] Specifically, firstly, noise is filtered out from the original dataset. The feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and coolant flow rate data contained in the original dataset are affected by environmental factors such as electromagnetic interference and mechanical vibration during the acquisition process, resulting in random noise. A Kalman filter algorithm is used to denoise each of these data sequences separately. The algorithm is implemented by taking the true value of a single process parameter (such as feed rate) at each moment as the state variable and the value of the parameter collected by the sensor as the observation variable. The uncertainty of state prediction and sensor measurement is described by setting the process noise covariance matrix and the measurement noise covariance matrix, respectively. The initial values ​​of the process noise covariance matrix and the measurement noise covariance matrix are determined as follows: multiple parameter sequence data of the CNC machine tool during the stable operation phase in the historical machining process are collected. The stable operation phase refers to the period when the fluctuation range of parameters such as feed rate, depth of cut, spindle speed, vibration frequency, temperature change and cutting fluid flow rate is small. Specifically, the fluctuation range is determined by calculating the variance of these parameter sequences in the historical data. The period when the variance value is continuously lower than the first quartile of the overall variance distribution is selected as the stable operation phase. For each parameter sequence, the variance of its value in the stable operation phase is calculated. Then, these variance values ​​are used as diagonal elements to construct a diagonal matrix, and the off-diagonal elements are set to zero, thereby obtaining the initial values ​​of the process noise covariance matrix and the measurement noise covariance matrix.

[0046] The Kalman filter algorithm recursively performs two steps: prediction and update. In the prediction step, based on the state estimate from the previous time step, the predicted state value for the current time step is calculated using the state transition equation. The state transition equation assumes that the system state changes linearly with time, meaning the current predicted state value equals the previous state estimate. In the update step, the predicted state value is corrected using the actual observation value at the current time step, combined with the Kalman gain. The Kalman gain is calculated by multiplying the inverse matrix of the sum of the prediction error covariance and the measurement noise covariance by the prediction error covariance. After correction, the optimal state estimate for the current time step is obtained, which is the denoised data. This process is performed sequentially on all process parameter sequences, ultimately yielding a denoised dataset containing denoised feed rate, denoised depth of cut, denoised spindle speed, denoised vibration frequency, denoised temperature change, and denoised cutting fluid flow rate. This process eliminates random interference, providing a data foundation for subsequent accurate quantification of load fluctuations.

[0047] Then, machining load fluctuations are quantified based on the denoised dataset. Machining load fluctuations are a comprehensive reflection of changes in multiple process parameters. A weighted average algorithm is used to sum the weighted data of denoised feed rate, denoised depth of cut, denoised spindle speed, denoised vibration frequency, and denoised temperature changes in the denoised dataset to calculate a comprehensive machining load fluctuation index. The weight of each parameter reflects its contribution to the machining load.

[0048] The weighting method involves collecting a large amount of historical machining data covering different working conditions. For each of the five parameters—noising feed rate, noising depth of cut, noising spindle speed, noising vibration frequency, and noising temperature change—the correlation coefficient between its numerical change and the spindle motor power change rate is calculated. Spindle motor power data is acquired in real-time by power sensors installed in the CNC machine tool spindle drive system, resulting in a time-series data set. The spindle motor power change rate is calculated by applying the first-order difference method to the time-series data. This involves calculating the difference in power values ​​between adjacent time points in chronological order and dividing this difference by the power value of the previous time point to obtain the power change rate sequence for each time point. The spindle motor power change rate is chosen as the correlation basis because, in the field of mechanical manufacturing, spindle motor power is a core physical quantity that directly reflects cutting load and energy consumption. Its change has a clear positive correlation with the machining load, which can be verified by physical laws and extensive practical experience. Next, normalization is performed to determine the weights. The absolute values ​​of the five correlation coefficients are added together to obtain a sum. Then, the absolute value of the correlation coefficient of each parameter is divided by this sum, and the quotient is the weight of that parameter, thus ensuring that the sum of all weights is one.

[0049] The weighted summation calculation involves using the weights preset through the above method, multiplying each weight by the instantaneous value of the corresponding denoising parameter, and summing all the products to obtain the machining load fluctuation index. This index comprehensively reflects the dynamic changes in load during machining, and its accurate calculation provides input for subsequent parameter coupling analysis and the prediction and control of cutting fluid flow.

[0050] In step S13, based on the processing load fluctuation index and combined with the pre-stored historical parameter change data, parameter coupling analysis, clustering and pattern matching are performed to obtain parameter coupling characteristics.

[0051] In one specific implementation, the step of performing parameter coupling analysis, clustering, and pattern matching based on the processing load fluctuation index and pre-stored historical parameter change data to obtain parameter coupling characteristics includes:

[0052] Based on the aforementioned processing load fluctuation index, the uncertainty of flow demand is analyzed using a fuzzy logic algorithm to obtain the uncertainty value of flow demand;

[0053] The similarity between the pre-stored historical flow demand and the flow demand of the uncertain value is calculated using the k-nearest neighbor algorithm, and the corresponding process parameter change records are queried from the pre-stored historical database based on the flow demand similarity.

[0054] Based on the process parameter variation records, the correlation between parameters is analyzed using the Pearson correlation coefficient analysis method to obtain the parameter coupling degree;

[0055] Based on the coupling degree of the parameter, coupling pattern clustering is performed using Gaussian mixture clustering algorithm to obtain coupled group data;

[0056] The minimum cumulative distance between the coupled grouped data and the original dataset is calculated using a dynamic time warping algorithm as a parameter deviation characteristic.

[0057] Based on the parameter deviation characteristics and the coupled grouped data, the parameter coupling characteristics are obtained by matching in a pre-established coupling feature library using the nearest neighbor algorithm.

[0058] Specifically, firstly, based on the current machining load fluctuation index, the uncertainty of flow rate demand is analyzed using a fuzzy logic algorithm. The implementation of this fuzzy logic algorithm involves using the machining load fluctuation index as an input variable, dividing it into three fuzzy sets: "low," "medium," and "high," according to its numerical value. In this implementation, the boundary values ​​of the membership function rely on historical machining data. This data originates from the recorded values ​​of the machining load fluctuation index during previous machining processes and the corresponding actual machining quality index values ​​after adjusting the cutting fluid flow rate, such as workpiece surface roughness or tool wear rate. The degree of improvement in machining quality is quantified by calculating the changes in machining quality indicators before and after the cutting fluid flow rate adjustment, such as the reduction in workpiece surface roughness or the decrease in tool wear rate.

[0059] The specific process for determining boundary values ​​is as follows: First, a large amount of historical data is collected, and the numerical range of the processing load fluctuation index is divided into multiple equal-width intervals, each containing a certain number of data points. Then, for each interval, the average value of the processing quality improvement corresponding to all data points within that interval is calculated. Next, the sliding window method is used to analyze the variation range of the average processing quality improvement between adjacent intervals. The width of the sliding window is set to cover three consecutive intervals, and the variation range threshold is determined by statistically analyzing the historical data distribution of the variation range of all adjacent intervals, taking the third quartile of the historical variation range value as the threshold. When the variation range of the average improvement of adjacent intervals at the boundary of an interval exceeds the threshold, the boundary point is marked as a critical point. Finally, these critical points are used as the boundary values ​​of the trapezoidal membership function, thereby completing the division of the processing load fluctuation index into three fuzzy sets: "low," "medium," and "high."

[0060] The membership function for each set uses a trapezoidal function, and the determination of its boundary values ​​depends on a historical debugging dataset independent of this control loop. This dataset contains a sequence of machining load fluctuation indicators collected under various typical operating conditions (covering different workpiece materials, tools, and process parameters), as well as cutting fluid flow records manually intervened by experienced operators based on machining conditions (such as observing chip morphology and listening to cutting sounds) to confirm the achievement of good cooling effects. Based on this dataset, the correspondence between the numerical distribution of machining load fluctuation indicators and empirical flow requirements is analyzed to define the boundaries of three fuzzy sets: "low," "medium," and "high." For example, the three quantiles of the numerical distribution of machining load fluctuation indicators (such as the 33rd and 67th quantiles) can be taken as boundary values. Then, inference and defuzzification steps are performed. Inference uses a set of preset fuzzy rules (such as "if the machining load fluctuation indicator is high, then the flow requirement uncertainty is high") to transform clear input values ​​into fuzzy outputs. The preset fuzzy rules are based on the analysis of historical machining data. Historical data are derived from long-term accumulated machining process records, including machining load fluctuation index sequences and corresponding cutting fluid flow rate adjustment effect data. The machining load fluctuation index is calculated through step S12, and the cutting fluid flow rate adjustment effect data uses workpiece surface roughness or tool wear rate as evaluation indicators.

[0061] The rule construction process first involves correlation analysis between the machining load fluctuation index value and the cutting fluid flow demand uncertainty value in historical data. The cutting fluid flow demand uncertainty value is quantified by the degree of fluctuation of the historical cutting fluid flow sequence, for example, by calculating the standard deviation of the historical cutting fluid flow sequence.

[0062] The specific process of correlation analysis is as follows: First, each machining load fluctuation index value and its corresponding cutting fluid flow demand uncertainty value in the historical dataset are grouped into a data point, and all data points constitute the analysis sample. Next, to quantify the overall correlation trend between the two, Pearson correlation coefficient analysis is used. The calculation involves first calculating the average of all historical machining load fluctuation index values ​​and the average of all historical cutting fluid flow demand uncertainty values; then, calculating the deviation between the machining load fluctuation index value and its average value, and the deviation between the cutting fluid flow demand uncertainty value and its average value for each pair of data points. The average of all data points is then calculated after multiplying each pair of deviations to obtain the covariance of the two sequences. Simultaneously, the standard deviations of the machining load fluctuation index value sequence and the cutting fluid flow demand uncertainty value sequence are calculated separately. Finally, the covariance is divided by the product of the two standard deviations to obtain the Pearson correlation coefficient. The magnitude and sign of this coefficient determine the strength and direction of the correlation between the premise and consequence in the rule.

[0063] For example, if the calculated correlation coefficient is a large positive number, a positive relationship is determined between the processing load fluctuation index and the flow demand uncertainty, and a rule is set that "if the processing load fluctuation index is high, then the flow demand uncertainty is high"; if it is a large negative number, a reverse rule is set. The rule base covers all combinations of input and output fuzzy sets. The input variable, the processing load fluctuation index, is divided into three fuzzy sets: "low," "medium," and "high," and the output variable, the flow demand uncertainty, is also divided into three fuzzy sets: "low," "medium," and "high."

[0064] The membership function of each fuzzy set adopts a trapezoidal function. Its boundary values ​​are calculated by collecting a large number of historical machining load fluctuation index values ​​and determining the quantiles of their numerical distributions. For example, after sorting the historical data by value, the values ​​at the third and two-thirds positions are used as the boundary points from "low" to "medium" and from "medium" to "high". Simultaneously, the improvement in machining quality indicators after adjusting the cutting fluid flow rate is considered. For instance, if adjusting the cutting fluid flow rate significantly improves the workpiece surface roughness when the machining load fluctuation index is within a certain range, then the boundary value of that range is used as the boundary of the membership function. The inference process involves inputting clear machining load fluctuation index values, calculating their membership degree in each fuzzy set using the membership function, and then performing fuzzy inference based on the rule base. Each rule generates a fuzzy output, and the outputs of all rules are combined into a comprehensive fuzzy output through a maximum operation. Defuzzification uses the centroid method to convert the fuzzy output into a specific, clear uncertain value for the flow demand. Specifically, the abscissa value corresponding to the centroid of the area enclosed by the membership function curve of the fuzzy output set and the abscissa is calculated as the output. This step quantifies the ambiguity of traffic demand under the current operating conditions, providing a basis for matching similar operating conditions in the future.

[0065] Next, the similarity between the current uncertain flow demand and all historical flow demand data in the pre-stored historical database is calculated using the k-nearest neighbor algorithm. The number of nearest neighbors, k, in the k-nearest neighbor algorithm is determined by cross-validation, testing the matching accuracy of different k values ​​on the historical dataset and selecting the best one. Specifically, the historical dataset is randomly divided into multiple equal subsets. Each subset is used as the test set, and the remaining subsets are used as the training set. For each candidate k value, a k-nearest neighbor model is constructed using the training set data, and the matching accuracy is calculated on the test set. Finally, the k value with the highest average matching accuracy across all test rounds is selected as the algorithm parameter. When calculating similarity, Euclidean distance is used as the metric, which is the linear distance between the current uncertain flow demand and each historical flow demand data point; the smaller the distance, the higher the similarity. Based on the calculated similarity, the process parameter change records corresponding to the k most similar historical records to the current uncertain flow demand are retrieved from the historical database. These records include historical feed rate, depth of cut, spindle speed, vibration frequency, and temperature change data.

[0066] Then, based on the k sets of process parameter change records retrieved, correlation analysis between the parameters is performed using the Pearson correlation coefficient method. The calculation process for the Pearson correlation coefficient is as follows: for any two parameters (e.g., feed rate and depth of cut), first, the mean of each parameter's numerical sequence is calculated; then, the covariance of the two parameter sequences is calculated, which is the average of the products of the deviations of each data point from its respective mean; next, the standard deviation of each parameter sequence is calculated; finally, the covariance is divided by the product of the two standard deviations to obtain the correlation coefficient. The Pearson correlation coefficients between each pair of parameters—feed rate, depth of cut, spindle speed, vibration frequency, and temperature change—are calculated. This coefficient measures the strength and direction of the linear correlation between the two parameters. The calculated correlation coefficients together constitute the parameter coupling degree, which describes the degree of mutual influence and correlation between various process parameters under similar flow requirements.

[0067] Subsequently, based on the obtained parameter coupling data, Gaussian mixture clustering algorithm is used to cluster and group coupling patterns. The number of cluster components in the Gaussian mixture model is determined after evaluating the goodness of fit of the historical coupling data using the Bayesian information criterion. The specific process of goodness of fit evaluation is as follows: Gaussian mixture models with different numbers of cluster components are used to fit the historical coupling data. For each model, its Bayesian information criterion value is calculated. This value comprehensively considers the model's fit to the data and model complexity. The fit is measured by the model's likelihood function value, and the model complexity is penalized by the number of model parameters. Finally, the number of cluster components corresponding to the model with the smallest Bayesian information criterion value is selected. This algorithm aggregates parameter coupling samples with similar coupling characteristics into the same group, and finally outputs coupling group data, thereby identifying different typical processing parameter coupling patterns.

[0068] Next, the dynamic time warping algorithm is used to calculate the minimum cumulative distance on the time axis between the current real-time acquired raw dataset (containing denoised feed rate, denoised cutting depth, denoised spindle speed, denoised vibration frequency, and denoised temperature change) and the representative sequence in the coupled group data, obtained through step S12 denoising. The specific implementation of the dynamic time warping algorithm involves first aligning the current denoised parameter sequence with the representative sequence in the coupled group to construct a distance matrix. Each element in the matrix represents the Euclidean distance between the two sequences at the corresponding time point. Then, a path is found from the lower left corner to the upper right corner of the matrix such that the sum of all elements on the path is minimized. This minimum cumulative distance is the parameter deviation characteristic, reflecting the degree of difference between the current real-time parameter sequence and the historical typical coupling mode in terms of dynamic temporal evolution.

[0069] Finally, based on the calculated parameter deviation characteristics and coupled grouping data, a nearest neighbor algorithm is used to match them in a pre-established coupling feature library. The construction process of this coupling feature library involves collecting a large amount of historical data covering various typical machining scenarios before system deployment. For each group of historical data, the complete calculation process from machining load fluctuation indicators to parameter deviation characteristics is executed. The calculated parameter deviation characteristics are then associated and stored with the corresponding historical parameter coupling characteristics that have been verified as optimal by historical machining results. Here, the criterion for "optimal" is that, in the historical data, when a certain parameter coupling characteristic is used, the corresponding cutting fluid flow control ensures that machining quality indicators (such as workpiece dimensional accuracy or surface quality) remain stable within a preset acceptable range and that the cutting fluid utilization efficiency is highest. These indicators are set based on industry standards or technical specifications provided by machine tool manufacturers and together constitute the coupling feature library.

[0070] During matching, the Euclidean distance between the current parameter deviation characteristic and all historical parameter deviation characteristics in the feature library is calculated. The historical parameter coupling characteristic corresponding to the historical sample with the smallest distance is selected as the final output, i.e., the parameter coupling characteristic. This parameter coupling characteristic comprehensively characterizes the inherent correlation patterns between various process parameters under the current machining state, providing key feature inputs for the subsequent neural network model to accurately predict the cutting fluid flow rate. It is the core link in realizing intelligent matching of cutting fluid flow rate and dynamically changing machining parameters.

[0071] In step S14, based on the parameter coupling characteristics, the target value of the cutting fluid flow rate is predicted by a pre-built neural network model and a flow rate adjustment sequence is matched.

[0072] In one specific implementation, the process of constructing the neural network model includes:

[0073] Obtain historical parameter coupling characteristics and historical cutting fluid flow target values;

[0074] The historical parameter coupling characteristics are input into the input layer of the initially constructed neural network model for training, and the predicted value of cutting fluid flow rate output by the output layer of the neural network model is obtained.

[0075] Substitute the predicted cutting fluid flow rate and the historical target cutting fluid flow rate into the loss function to calculate the loss value;

[0076] The gradient of the output layer of the neural network model is calculated based on the loss value, and the gradient is passed forward layer by layer through the chain rule to calculate the gradient of the parameters of each layer and obtain the gradient data.

[0077] Based on the gradient data and the preset learning rate, the parameters of each layer of the neural network model are updated using the gradient descent method.

[0078] The parameters of each layer are iteratively updated until the number of training iterations of the neural network model is greater than a preset number of iterations, or the loss value of the neural network model is less than a preset loss threshold. At this point, the training is considered complete, and a trained neural network model is obtained.

[0079] Specifically, firstly, the data required for model training is acquired from accumulated historical processing data. This includes historical parameter coupling characteristics calculated in step S13 during multiple historical processing processes, and corresponding cutting fluid flow rates that have been proven to achieve good cooling effects and ensure processing quality in those historical processing processes. These values ​​serve as the historical cutting fluid flow rate target values. The neural network model adopts a feedforward neural network structure. The number of nodes in its input layer is the same as the dimension of the historical parameter coupling characteristics, used to receive this characteristic data. The output layer has one node, used to output the predicted cutting fluid flow rate value. Two hidden layers are set between the input and output layers. The number of hidden layer nodes is determined by optimizing the selection on historical data using cross-validation. Specifically, the dataset of historical parameter coupling characteristics and historical cutting fluid flow rate target values ​​is randomly divided into multiple subsets of equal size. Each subset is used as the validation set, and the remaining subsets are used as the training set. For a set of candidate combinations of hidden layer node numbers, a neural network model is constructed and trained on the training set. The mean square of the prediction error is calculated on the validation set, and the combination of hidden layer node numbers with the smallest mean square of the prediction error is selected as the final setting. The activation function for hidden layer nodes is calculated as follows: for the sum of the inputs to each node in the hidden layer, if the sum is greater than zero, the output is equal to the sum of the inputs; if the sum is less than or equal to zero, the output is zero. The activation function for the output layer is a linear function, calculated as the output value equal to the sum of its inputs, without any nonlinear transformation. The weight and bias parameters in the model are initialized using the Xavier initialization method. Specifically, the weight parameters are randomly sampled from a uniform distribution. The upper bound of this distribution is the square root of a positive six divided by the square root of the sum of the number of nodes in the previous and next layers, and the lower bound is the negative of this upper bound. The bias parameters are initialized to zero.

[0080] During training, a set of historical parameter coupling characteristics is input into the input layer of the neural network model. The data is processed layer by layer through the hidden layers, and finally, a predicted value of cutting fluid flow rate is output by the output layer. Then, this predicted value and the corresponding historical target value of cutting fluid flow rate are substituted into the loss function, which is defined as the square mean of the difference between the predicted value and the target value. Based on the calculated loss value, the gradient of the model parameters is calculated using the error backpropagation algorithm. First, the gradient of the loss function with respect to the output layer is calculated. Then, according to the chain rule, this gradient is multiplied by the rate of change of the activation function of the current layer at its input value (for the hidden layer, the rate of change rule is: if the sum of the inputs of the nodes in the layer is greater than zero, the rate of change is one; if the sum of the inputs is less than or equal to zero, the rate of change is zero; for the output layer, the rate of change is always one). This is passed to the previous layer. At each layer, using the passed gradient, the gradient of the loss function with respect to the weight parameters of the layer (equal to the product of the input data of the layer and the passed gradient) and the gradient of the loss function with respect to the bias parameters of the layer (equal to the passed gradient itself) are calculated, thus obtaining the complete gradient data.

[0081] Based on the calculated gradient data and the preset learning rate, the weight and bias parameters of each layer of the neural network model are updated using gradient descent. The learning rate is a hyperparameter that controls the step size of parameter updates. Its specific value is determined by performing the following steps on historical training data: dividing the historical training data into training and validation subsets; setting a set of candidate learning rate values ​​(e.g., starting from 0.001 and multiplying by 10 to the power of 0.5 each time to obtain an increasing candidate sequence); for each candidate learning rate, training the model using the training subset and observing the decrease in loss value with the number of training iterations on the validation subset; finally, selecting the candidate value that maximizes the decrease in the validation set loss value within a finite number of iterations without drastic fluctuations as the final learning rate. The model training process iteratively updates parameters until a preset stopping condition is met. The stopping condition includes the number of training iterations of the neural network model exceeding a preset maximum number, or the loss value of the neural network model being less than a preset loss threshold. The maximum number of training iterations is set to a sufficiently large value based on computational resources and the required model accuracy. The loss threshold is determined based on convergence behavior analysis of a large amount of historical training data. Specifically, during historical training, the curve of the loss value changing with the number of iterations is observed. When the curve enters a phase where the decrease is significantly flat and the rate of decline remains below a tiny amount (e.g., one ten-thousandth), the loss value at the beginning of this phase is recorded, and its statistical average is taken as the loss threshold. When any of these conditions are met, training is considered complete, resulting in a fully trained neural network model.

[0082] Before inputting the parameter coupling characteristics calculated in real-time in step S13 into the trained neural network model during the model usage phase, the parameter coupling characteristics need to be converted into an input format acceptable to the model. The parameter coupling characteristics originate from the output of step S13 and are specifically a set of numerical values ​​representing the inherent correlation patterns between process parameters. This conversion process involves sequentially combining all the numerical elements contained in the parameter coupling characteristics according to their fixed arrangement order during the calculation in step S13 to form a one-dimensional numerical vector. The dimension of this vector is consistent with the number of nodes in the input layer of the neural network model, and this number of nodes is determined during model construction based on the dimension of the historical parameter coupling characteristics.

[0083] To ensure the input data maintains the same scale as the data used during model training, the vector needs to be normalized. Normalization employs a min-max normalization method, linearly transforming the value of each element in the vector to between zero and one. The minimum and maximum values ​​of the original data required for normalization are derived from the historical parameter coupling characteristic training dataset used during model training. Specifically, these are the minimum and maximum values ​​for each feature dimension in the training dataset, and these parameters are saved to a normalized parameter file after model training is complete. During model usage, these minimum and maximum values ​​are read from this file, and the same normalization calculation is performed on the parameter coupling characteristic vector generated in real time.

[0084] The normalized vector becomes the standard input format for the model. This vector is input into the input layer of the trained neural network model, and the output layer calculates the predicted target value for the cutting fluid flow rate under the current operating conditions. The flow rate adjustment sequence matching process involves searching a pre-stored flow rate adjustment pattern library for the closest corresponding adjustment command sequence based on this target flow rate value. This pattern library is constructed from historical machining data records deemed to have excellent control effects. The criteria for judgment are that, within the corresponding machining cycle, the workpiece machining quality parameters (such as dimensional accuracy and surface roughness) are consistently within the acceptable range specified in the machine tool manufacturer's technical manual, and the cutting fluid utilization efficiency (defined as the ratio of effective cooling time to total supply time) is higher than a benchmark value set according to industry standards.

[0085] The data structure of the flow regulation sequence is a time-ordered list. Each element in the list is a structure containing a timestamp, a target flow rate value, and specific control instructions sent to the actuator to achieve that target flow rate value. These control instructions include the opening percentage of the regulating valves in the circulating delivery system and the setpoint of the drive pump's rotational speed. The final output flow regulation sequence directly guides the actuator in precise flow control. This step utilizes a neural network model to capture the complex nonlinear relationship between parameter coupling characteristics and cutting fluid flow rate, enabling the prediction of cutting fluid flow rate. Combined with historical control experience, it generates an executable regulation sequence, providing conditions for maintaining a balance between cutting fluid supply and demand and achieving uniform cooling under dynamically changing machining loads.

[0086] In step S15, based on the flow rate adjustment sequence and the denoised dataset, power change level classification, feed rate sequence matching, and sequence optimization are performed to obtain the feed rate adjustment sequence.

[0087] In one specific implementation, the step of performing power change level classification, feed rate sequence matching, and sequence optimization based on the flow rate adjustment sequence and the denoised dataset to obtain the feed rate adjustment sequence includes:

[0088] Based on the flow regulation sequence, the power change amplitude between adjacent time periods is calculated using the first-order difference method, and the power change amplitude is classified into levels using a fuzzy logic algorithm to obtain the power change level.

[0089] Based on the power change level, a speed matching sequence is obtained by matching in a pre-established change level-feed speed association database using the nearest neighbor algorithm;

[0090] Based on the denoised feed rate, frequency features are extracted using a fast Fourier transform algorithm, and the Pearson correlation coefficient between the frequency features and the denoised vibration frequency is calculated to obtain the feed rate correction coefficient.

[0091] Based on the feed rate correction coefficient, the speed matching sequence is corrected using the weighted least squares method to obtain the feed rate adjustment sequence.

[0092] Specifically, firstly, the power variation amplitude is calculated and classified based on the spindle motor power data. The spindle motor power data is collected in real time by a power sensor installed in the CNC machine tool spindle drive system to obtain the spindle motor power time series data. The specific calculation process of the first-order difference method is as follows: the power value of the spindle motor power sequence arranged in chronological order is subtracted from the power value of the previous adjacent time point from the power value of the next time point to obtain a series of differences. These differences constitute the true power variation amplitude sequence. Among them, the power variation amplitude needs to be processed to the range of [0,1] through minimum-maximum normalization. The minimum and maximum values ​​required for normalization are derived from the global minimum and global maximum values ​​of the power variation amplitude sequence in historical machining data. This amplitude directly represents the load fluctuation caused by changes in cutting parameters (such as depth of cut and feed rate) during the machining process.

[0093] Subsequently, a fuzzy logic algorithm was used to classify the power change amplitude into levels. The algorithm's implementation involves using the power change amplitude as an input variable and dividing it into ten fuzzy sets based on its magnitude: "extremely small," "very small," "small," "lower," "medium," "higher," "significant," "very significant," "extremely significant," and "drastic." The membership function for each set uses a trapezoidal function. The boundary values ​​are determined by statistically analyzing the distribution of power change amplitude values ​​in historical processing data and selecting the ten quantiles of this numerical sequence as the key boundary points for classifying these ten levels. The reasoning process uses a set of preset fuzzy rules (e.g., "if the power change amplitude belongs to the 'extremely small' set, then the power change level is level one"; "if the power change amplitude belongs to the 'very small' set, then the power change level is level two," and so on up to level ten) to transform the clear input values ​​into fuzzy outputs. Defuzzification is achieved using the centroid method, which calculates the abscissa value corresponding to the centroid of the area enclosed by the membership function curve of the fuzzy output set and the abscissa axis, and uses this value as the clear power change level output.

[0094] Next, a preliminary speed sequence is matched based on the power change level. Using the nearest neighbor algorithm, the feed speed sequence corresponding to the historical record closest to the current power change level is found in a pre-established database of power change level and feed speed correlation, and this sequence is used as the speed matching sequence. The construction process of this correlation database involves collecting historical machining data, including historical power change levels and their corresponding historical feed speed sequences that have been verified as effective in practical applications. The criterion for "effective" is that when using the feed speed sequence, the vibration level of the machining process is maintained below the safety threshold specified in the machine tool manufacturer's technical manual, and the workpiece machining dimensional accuracy meets the requirements of the process drawings. During matching, the Euclidean distance between the current power change level and all historical power change levels in the database is calculated. This distance is calculated as the absolute value of the difference between the two level values, and the feed speed sequence corresponding to the historical record with the smallest distance is output.

[0095] Then, the speed matching sequence is optimized and corrected. Based on the denoised feed speed data in the denoised dataset, frequency domain features are extracted using the Fast Fourier Transform algorithm. The time series data of the denoised feed speed is taken as input and decomposed into the sum of sine and cosine wave components of different frequencies, thus obtaining the frequency domain representation of the sequence. The frequency component with the largest amplitude is extracted as the dominant frequency feature. The linear correlation between this frequency feature and the denoised vibration frequency data in the denoised dataset is calculated using the Pearson correlation coefficient. First, the average values ​​of the frequency feature sequence and the denoised vibration frequency sequence are calculated. Then, the covariance of the two sequences is calculated, which is the average of the products of the deviations of each data point from its respective mean. Next, the standard deviations of the two sequences are calculated separately. Finally, the covariance is divided by the product of the two standard deviations to obtain the correlation coefficient, and its absolute value is taken as the feed speed correction coefficient. This coefficient reflects the correlation strength between feed speed fluctuations and machine tool vibration in the frequency domain.

[0096] Finally, based on the feed rate correction coefficient, the aforementioned speed matching sequence is corrected using weighted least squares. The specific method for setting the weights in weighted least squares is as follows: for each data point in the speed matching sequence, its corresponding weight value is equal to the ratio of the feed rate correction coefficient calculated for the processing time period in which the data point is located to the sum of the feed rate correction coefficients corresponding to all data points in the entire speed matching sequence. Thus, the larger the feed rate correction coefficient, the greater the weight assigned to the data point in the weighted least squares calculation. The calculation process involves finding a correction curve, which is set as a linear function. The slope and intercept parameters of this linear function are determined by minimizing the sum of squares of the differences between all data points in the speed matching sequence and the corresponding points on this curve (where each squared difference is multiplied by the weight corresponding to that point). After generating the determined correction curve, the sequence values ​​corresponding to the original time points in the speed matching sequence are replaced with the new sequence values ​​calculated for the same time points on the correction curve. This newly generated sequence is the final optimized feed rate adjustment sequence. This step translates the intended flow rate adjustment into power change levels, correlates them with historical experience to match the feed rate, and then combines real-time vibration characteristics for frequency domain optimization correction. This ensures that the feed rate adjustment is precisely coordinated with the changes in cutting fluid flow rate and the actual operating status of the machine tool, thereby providing motion parameter guarantees for maintaining the thermal balance and mechanical stability of the machining process and ultimately achieving uniform cooling.

[0097] In step S16, a cooling uniformity analysis is performed based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence.

[0098] In one specific implementation, the step of performing cooling uniformity analysis based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence includes:

[0099] Based on the feed rate adjustment sequence and the noise reduction cutting depth, the cooling compensation requirement is obtained by performing a fuzzy logic algorithm to analyze the compensation requirement.

[0100] Based on the cooling compensation requirements and the noise reduction cutting fluid flow rate, non-uniformity detection is performed by calculating the standard deviation within a preset sliding window to generate a cooling non-uniformity signal;

[0101] Based on the uneven cooling signal, the compensation peak point is determined by the extreme value detection algorithm, and the compensation peak point is fused with the noise reduction spindle speed by the weighted average method to obtain the cooling compensation sequence.

[0102] Specifically, firstly, based on the feed rate adjustment sequence and the denoised cutting depth data in the denoised dataset, a fuzzy logic algorithm is used to analyze the compensation requirements and obtain the cooling compensation requirements. The implementation of this fuzzy logic algorithm involves using the instantaneous rate of change of the feed rate adjustment sequence and the instantaneous value of the denoised cutting depth as input variables. The rate of change of the feed rate adjustment sequence is calculated based on the numerical difference between adjacent time points. Based on the magnitude of the feed rate of change, nine fuzzy sets are defined: "extremely low," "very low," "low," "relatively low," "medium," "relatively high," "high," "very high," and "extremely high." Similarly, based on the magnitude of the denoised cutting depth, nine fuzzy sets are defined: "extremely shallow," "very shallow," "shallow," "relatively shallow," "medium," "relatively deep," "deep," "very deep," and "extremely deep." The membership function of each fuzzy set uses a trapezoidal function, and its boundary values ​​are determined by analyzing historical data. The numerical distributions of the feed rate of change and cutting depth during historical machining processes are statistically analyzed, and the nine quantiles of each numerical sequence are selected as the boundary points of the corresponding fuzzy set. Reasoning is performed using a set of pre-defined fuzzy rules, such as "if the feed rate changes rapidly and the depth of cut is deep, the cooling compensation requirement is extremely high." After reasoning, defuzzification is performed using the centroid method, which calculates the abscissa value corresponding to the centroid of the area enclosed by the membership function curve of the fuzzy output set and the abscissa axis. This value is then used as the clear cooling compensation requirement output. This step quantifies the dynamic adjustment of cooling requirements caused by the combined changes in motion and geometric parameters.

[0103] Next, based on the cooling compensation demand and the denoised cutting fluid flow rate data in the denoised dataset, non-uniformity detection is performed to generate a cooling non-uniformity signal. Non-uniformity detection is achieved by calculating the standard deviation within a preset sliding window. The duration of the sliding window is determined based on the statistical analysis of the durations of typical cooling non-uniformity phenomena in historical data, and the mode of this duration distribution is selected as the window size. The calculation process involves sliding a fixed-length window across the time series of the ratio of cooling compensation demand to denoised cutting fluid flow rate, and calculating the standard deviation of the ratio sequence within each window.

[0104] The standard deviation threshold is set by analyzing historical data. Standard deviation data for windows under conditions of good and uneven cooling in historical processing are collected, and the distribution of these two types of data is analyzed using a box plot method. The implementation of the box plot method is as follows: for each type of data, firstly, its first quartile, third quartile, and interquartile range are calculated, where the interquartile range equals the third quartile minus the first quartile. Then, the lower edge of the data distribution is calculated, which is the first quartile minus 1.5 times the interquartile range. The lower edge value of the box plot for data under uneven cooling conditions is set as the standard deviation threshold. For example, in steel part processing, the standard deviation threshold can be set to 0.1. When the calculated standard deviation within a certain window exceeds this threshold, a cooling unevenness signal is generated. This signal identifies the time period during the processing where uneven coolant distribution may exist.

[0105] Then, based on the cooling unevenness signal, the compensation peak points are determined using an extreme value detection algorithm. This algorithm operates within the time interval identified by the cooling unevenness signal, searching for local maxima in the cooling compensation demand sequence. A local maximum is defined as a value greater than the values ​​of its preceding and following adjacent data points. From all detected local maxima, points whose values ​​exceed an amplitude threshold determined by historical data statistics are selected. This amplitude threshold is set as the first quartile of all local maxima in the historical cooling compensation demand sequence. These selected points are identified as the compensation peak points requiring focused compensation.

[0106] Finally, the compensation peak points are fused with the denoised spindle speeds in the denoised dataset using a weighted average method to obtain a cooling compensation sequence. The weights in the weighted average method are related to the denoised spindle speeds. For each compensation peak point, its weight is equal to the ratio of the denoised spindle speed value at that moment to the sum of the denoised spindle speed values ​​at all corresponding compensation peak points. Thus, the higher the spindle speed, the greater the proportion of the corresponding compensation peak point in the fusion. The weighted average calculation involves multiplying the value of each compensation peak point by its corresponding weight, and then distributing and superimposing all such weighted values ​​over time to form a complete cooling compensation sequence that matches the spindle load. This step analyzes the cooling demand under the combined action of feed rate and depth of cut, detects the non-uniformity of coolant flow, locates the compensation timing, and dynamically allocates the compensation amount based on the spindle speed, generating cooling compensation commands adapted to the machining heat load distribution. This provides the prerequisite for global control in step S17 to achieve uniform cooling of the tool-workpiece contact area.

[0107] In step S17, correlation analysis, global compensation adjustment, path optimization, and flow regulation are performed based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence.

[0108] In one specific implementation, the step of performing correlation analysis, global compensation adjustment, path optimization, and flow regulation based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence includes:

[0109] Based on the cooling compensation sequence and the processing load fluctuation index, a correlation analysis is performed using grey relational analysis to obtain the correlation deviation sequence;

[0110] Based on the aforementioned correlation deviation sequence, the flow regulation sequence is compensated and adjusted using a weighted moving average method to obtain a global regulation sequence;

[0111] Based on the global control sequence, the cutting fluid flow path is optimized using the ant colony algorithm to obtain an optimized path sequence;

[0112] Based on the optimized path sequence, the valves and pumps of the circulating conveying system are controlled to regulate the flow rate.

[0113] Specifically, firstly, based on the cooling compensation sequence and the processing load fluctuation index, a correlation analysis is performed using grey relational analysis to obtain the correlation deviation sequence. The specific calculation process of grey relational analysis is as follows: using the cooling compensation sequence as the reference sequence and the processing load fluctuation index as the comparison sequence, firstly, the absolute difference between the two sequences at each time point is calculated. Then, the maximum and minimum values ​​are found from the absolute differences at all time points. Next, the correlation coefficient at each time point is calculated. This coefficient is equal to the minimum value plus the product of the resolution coefficient and the maximum value, and then divided by the sum of the absolute difference at that time point and the resolution coefficient multiplied by the maximum value. The resolution coefficient was determined by analyzing historical data. Multiple historical cooling compensation sequences and historical processing load fluctuation index sequences were collected, along with an independent verification index that objectively reflects the cooling uniformity effect, such as the standard deviation of the workpiece surface temperature distribution measured by an infrared thermal imager. A set of candidate resolution coefficient values ​​(e.g., between 0.1 and 1, with 0.1 as the interval) was set. For each candidate resolution coefficient, its corresponding historical correlation degree sequence was calculated, and the Pearson correlation coefficient between this correlation degree sequence and the historical workpiece surface temperature distribution standard deviation sequence was calculated. Finally, the candidate resolution coefficient value that maximized the absolute value of the Pearson correlation coefficient was selected as the final value. Finally, the correlation coefficients at all time points were arranged in chronological order to form the correlation degree sequence. The correlation deviation sequence was obtained by calculating the difference between the correlation degree sequence and a benchmark value (set as the median of the historical correlation degree sequence). This correlation analysis quantifies the dynamic correlation strength between cooling compensation demand and processing load fluctuations, and the degree to which it deviates from the normal correlation level.

[0114] Then, based on the correlation deviation sequence, the flow regulation sequence is compensated and adjusted using a weighted moving average method to obtain the global regulation sequence. The specific implementation process of the weighted moving average method is as follows: a moving window is selected, the length of which is determined statistically based on typical regulation cycles in historical processing. The statistical method involves analyzing historical flow regulation sequences and calculating the periodicity of the sequence using an autocorrelation function. The autocorrelation function is calculated by correlating the sequence with its copies after different time lags, identifying the lag times that cause the correlation to reach a local maximum. These lag times are the potential regulation cycles, and the cycle with the highest frequency is selected as the typical regulation cycle, which is then used as the length of the moving window. For each data point in the flow regulation sequence, its corresponding weight is determined by the correlation deviation value at that point. The specific weight allocation rule is that for each data point within a window, the absolute value of the correlation deviation value of each point is taken and divided by the sum of the absolute values ​​of the correlation deviations of all points within the window; the quotient is the weight of that point. Thus, data points with larger absolute values ​​of correlation deviation are assigned higher weights in the moving average calculation. During calculation, for each point in the flow regulation sequence, the values ​​of all data points within its moving window are taken, multiplied by their respective weights, and then these weighted values ​​are summed. This weighted sum replaces the value of the corresponding point in the original flow regulation sequence. After point-by-point weighted moving average processing of the entire flow regulation sequence, the global control sequence is obtained. This step identifies systematic coordination deviations by analyzing the intrinsic relationship between cooling compensation demand and processing load fluctuations. Based on these deviations, the preset flow regulation sequence is dynamically weighted, smoothed, and compensated, so that the final global control sequence can respond to real-time flow predictions while taking into account the dynamic coordination between cooling uniformity and processing load.

[0115] Based on the global control sequence, the cutting fluid flow path is optimized using an ant colony algorithm to obtain an optimized path sequence. This step uses the global control sequence as input data, which contains the time series values ​​of flow regulation commands, representing the global flow regulation demand after weighted moving average processing. Simultaneously, the topological structure data of the cutting fluid flow path is obtained from the machine tool configuration database, including pipe node locations, pipe connection relationships, valve control point coordinates, and the length and flow capacity limits of each path segment. The topological structure data is extracted digitally from machine tool design drawings and constructed as a graph structure model, where nodes represent pipe intersections or outlets, edges represent pipe segments, and the edge weights are initialized to the pipe length.

[0116] The implementation process of the ant colony algorithm is as follows: First, the algorithm parameters are initialized, including the number of ants, the initial pheromone concentration, the pheromone evaporation coefficient, the heuristic factor weights, and the number of iterations. The number of ants is determined by calculating the total number of nodes in the topology, rounding the square root of the total number of nodes to the nearest integer, to ensure the coverage and computational efficiency of path exploration. The initial pheromone concentration is set to a small positive constant, such as 0.1 or 0.5. The pheromone evaporation coefficient is determined through statistical analysis of historical path optimization data. Specifically, multiple sets of historical cutting fluid flow path data are collected, including pipe node locations, pipe connection relationships, valve control point coordinates, and the length and flow capacity limits of each path segment. The data source is the machine tool configuration database. Simultaneously, corresponding processing effect indicators are collected. These indicators are represented by a cooling uniformity score, which is calculated using the standard deviation of the workpiece surface temperature distribution actually measured by an infrared thermal imager. A smaller standard deviation indicates better cooling uniformity and a higher score.

[0117] A grid search method was used to test different evaporation coefficient values ​​within a predetermined range. The predetermined range was set based on statistics of the effective range of the evaporation coefficient in historical ant colony algorithm parameter studies, for example, from 0.1 to 0.9, with 0.1 intervals. For each candidate evaporation coefficient value, the ant colony algorithm was run multiple times, with different initial random states used in each run to average the impact of randomness. When running the ant colony algorithm, historical cutting fluid flow path data and corresponding historical flow demand sequences were input, and optimized path sequences were output. Then, the cooling uniformity score corresponding to the optimized path sequence was calculated. The cooling uniformity score was calculated by simulating the cutting fluid flow based on the path sequence, combining the actual cooling effect in historical processing data, using the standard deviation of the workpiece surface temperature distribution as input, and mapping the standard deviation to the scoring interval through linear transformation. The smaller the standard deviation, the higher the score. The average cooling uniformity score of multiple runs was taken as the performance index of the candidate evaporation coefficient value. Finally, the candidate evaporation coefficient value with the highest average cooling uniformity score was selected as the final value.

[0118] The heuristic factor weights are calculated based on the inverse relationship between pipe length and flow demand. The weight values ​​are determined through historical data regression analysis. Historical pipe length, flow demand, and path efficiency data are collected. Path efficiency is defined as a comprehensive indicator of cooling effect and energy consumption. A linear regression model is established with pipe length and flow demand as independent variables and path efficiency as the dependent variable. The linear regression model is fitted to the data using the least squares method to obtain the coefficients corresponding to pipe length and flow demand. These coefficients are the heuristic factor weight values.

[0119] During path optimization, each ant starts from the initial node (the cutting fluid supply source) and selects the next node according to the state transition rules. The state transition rules comprehensively consider pheromone concentration and heuristic information. The pheromone concentration reflects the quality of historical path selection, while the heuristic information is calculated by the matching degree between the pipeline length and the flow demand in the global control sequence. The matching degree is evaluated by comparing the deviation between the allocated flow of the current path segment and the corresponding demand value in the global control sequence.

[0120] Specifically, for each optional node, a transition probability is calculated, which is proportional to a power of the pheromone concentration and a power of the heuristic factor, with the power values ​​representing the pheromone weight and the heuristic factor weight, respectively. After an ant completes a path, the total path length and flow matching degree are recorded. The flow matching degree is calculated by the average deviation between the allocated flow of each segment of the path and the corresponding demand value in the global control sequence. Pheromone updates include local updates and global updates. Local updates are performed after each step the ant moves, reducing the pheromone concentration on the current edge to encourage exploration of other paths. The global update is performed after all ants have completed their paths, increasing the pheromone concentration only for the optimal path. The optimal path is selected based on the objective function value, which is defined as the weighted sum of the deviation between the total path length and the flow matching degree. By assigning weights, physical quantities of different dimensions are integrated into a dimensionless comprehensive evaluation index for comparing the merits of different paths. The weights are determined through historical data statistics. Historical path performance data, including path length, flow matching degree, and cooling effect indicators, are analyzed. The influence coefficients of path length and flow matching degree on the cooling effect are calculated through multiple regression analysis. The influence coefficients are represented by the standardized coefficients in the regression model. Then, these coefficients are normalized by dividing each coefficient by the sum of all coefficients to obtain the final weight value.

[0121] The iterative process is repeated until a preset number of iterations is reached or the objective function value converges. The preset number of iterations is determined by analyzing historical algorithm data; multiple ant colony algorithm runs are analyzed to select the number of iterations that stabilizes the objective function value. Convergence of the objective function value is determined by monitoring the magnitude of change in the objective function value during consecutive iterations. The threshold for this magnitude is determined by statistically analyzing historical data; the standard deviation of the objective function value changes over historical runs is calculated, and one-tenth of this standard deviation is set as the convergence threshold. During the iteration process, the optimal objective function value obtained by all ants after each iteration is continuously monitored, and the magnitude of change of this optimal value in consecutive iterations is calculated. When the magnitude of change in the objective function value in multiple consecutive iterations (e.g., five times) is less than the preset convergence threshold, the algorithm terminates. At this point, the path with the smallest objective function value in that iteration is selected as the global optimal solution, and its node access order and corresponding traffic allocation scheme are extracted, outputting an optimized path sequence.

[0122] Finally, based on the optimized path sequence, which includes a series of node access sequences and the target flow allocation value at each node, specific control instructions for the cyclic conveying system are generated.

[0123] Specifically, firstly, based on the node order in the optimized path sequence, the identification number of the control valve corresponding to each node and its required target opening percentage are queried from the pre-stored valve control mapping relationship. This valve control mapping relationship is a pre-built database. Its construction process involves obtaining the correspondence between each pipeline node and the specific control valve identification number through the machine tool control system configuration file and pipeline network diagram during the system deployment phase, and recording the maximum flow capacity of each valve in the fully open state; at the same time, the optimal opening data of each valve under different flow requirements in historical processing are collected. This data comes from the debugging parameter table provided by the machine tool manufacturer or the actual opening value recorded when the flow is stable and the cooling effect is good during historical operation. A mapping relationship from the target flow value to the valve opening percentage is established through linear interpolation.

[0124] Secondly, based on the total flow requirement of the entire path in the optimized path sequence, i.e., the sum of the target flow allocation values ​​of all nodes in the sequence, the target speed setpoint for driving the pump is obtained by querying the pre-stored pump speed-flow characteristic curve. This pump speed-flow characteristic curve, provided by the pump manufacturer, is a data table showing the correspondence between pump speed and output flow rate measured experimentally under specific pipeline pressures. During use, the corresponding speed setpoint is obtained from the target total flow value using a linear interpolation method. Finally, the generated control command sequence, containing the valve identification number, target opening percentage, and target pump speed setpoint, is sent to the corresponding valve actuator and pump frequency converter in a fixed data frame format via the machine tool CNC system's fieldbus communication protocol. After the command is executed, the system compares the real-time read feedback signals from the valve position sensor and pump speed with the command value to form a closed-loop control, ensuring precise allocation and adjustment of the cutting fluid flow according to the optimized path sequence. This step achieves precise on-demand allocation of cutting fluid, reduces flow resistance and energy loss, improves cooling uniformity and processing efficiency, and enhances the thermal stability and processing quality under load fluctuations.

[0125] Reference Figure 2 The second embodiment of the present invention provides an automated cutting fluid circulation control system for CNC intelligent manufacturing, comprising:

[0126] The data acquisition module is used to collect data on the feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool to obtain the raw dataset.

[0127] The load fluctuation quantization module is used to filter out noise based on the original dataset to obtain a denoised dataset and quantify the processed load fluctuation to obtain a processed load fluctuation index.

[0128] The coupling characteristic analysis module is used to perform parameter coupling analysis, clustering and pattern matching based on the processing load fluctuation index and pre-stored historical parameter change data to obtain parameter coupling characteristics;

[0129] The flow prediction module is used to predict the target value of the cutting fluid flow rate and match the flow adjustment sequence based on the parameter coupling characteristics through a pre-built neural network model.

[0130] The feed rate adjustment module is used to perform power change level classification, feed rate sequence matching and sequence optimization based on the flow rate adjustment sequence and the denoised dataset to obtain the feed rate adjustment sequence;

[0131] The cooling uniformity analysis module is used to perform cooling uniformity analysis based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence.

[0132] The output module is used to perform correlation analysis, global compensation adjustment, path optimization, and flow regulation based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence.

[0133] It should be noted that the automated cutting fluid circulation control system for CNC intelligent manufacturing provided in this embodiment of the invention is used to execute all process steps of the automated cutting fluid circulation control method for CNC intelligent manufacturing in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0134] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0135] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. An automated cutting fluid circulation control method for CNC intelligent manufacturing, characterized by, include: Data on feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool are collected to obtain the raw dataset; Based on the original dataset, noise is filtered out to obtain a denoised dataset, and the processing load fluctuation is quantified to obtain a processing load fluctuation index. Based on the processing load fluctuation index and combined with the pre-stored historical parameter change data, parameter coupling analysis, clustering and pattern matching are performed to obtain parameter coupling characteristics; Based on the parameter coupling characteristics, the target value of the cutting fluid flow rate is predicted and the flow rate adjustment sequence is matched by a pre-built neural network model. Based on the flow rate regulation sequence and the denoised dataset, power change levels are classified, feed rate sequence is matched and optimized to obtain the feed rate regulation sequence. Based on the feed rate adjustment sequence and the denoised dataset, a cooling uniformity analysis is performed to obtain a cooling compensation sequence; Based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence, correlation analysis, global compensation adjustment, path optimization, and flow regulation are performed.

2. The CNC intelligent manufacturing automated cutting fluid circulation control method of claim 1, wherein, The process of filtering out noise from the original dataset to obtain a denoised dataset and quantifying processing load fluctuations to obtain a processing load fluctuation index includes: Based on the original dataset, a denoised dataset is obtained by performing a Kalman filter algorithm. Based on the denoised dataset, the processing load fluctuation index is obtained by weighted summation of the denoised dataset using a weighted average algorithm; The denoising dataset includes denoising feed rate, denoising depth of cut, denoising spindle speed, denoising vibration frequency, denoising temperature change data, and denoising cutting fluid flow rate.

3. The automated cutting fluid circulation control method for CNC intelligent manufacturing according to claim 2, characterized in that, The parameter coupling characteristics are obtained by performing parameter coupling analysis, clustering, and pattern matching based on the processing load fluctuation index and pre-stored historical parameter change data, including: Based on the aforementioned processing load fluctuation index, the uncertainty of flow demand is analyzed using a fuzzy logic algorithm to obtain the uncertainty value of flow demand; The similarity between the pre-stored historical flow demand and the flow demand of the uncertain value is calculated using the k-nearest neighbor algorithm, and the corresponding process parameter change records are queried from the pre-stored historical database based on the flow demand similarity. Based on the process parameter variation records, the correlation between parameters is analyzed using the Pearson correlation coefficient analysis method to obtain the parameter coupling degree; Based on the coupling degree of the parameter, coupling pattern clustering is performed using Gaussian mixture clustering algorithm to obtain coupled group data; The minimum cumulative distance between the coupled grouped data and the denoised dataset is calculated using a dynamic time warping algorithm as a parameter deviation characteristic. Based on the parameter deviation characteristics and the coupled grouped data, the parameter coupling characteristics are obtained by matching in a pre-established coupling feature library using the nearest neighbor algorithm.

4. The automated cutting fluid circulation control method for CNC intelligent manufacturing according to claim 1, characterized in that, The process of constructing the neural network model includes: Obtain historical parameter coupling characteristics and historical cutting fluid flow target values; The historical parameter coupling characteristics are input into the input layer of the initially constructed neural network model for training, and the predicted value of cutting fluid flow rate output by the output layer of the neural network model is obtained. Substitute the predicted cutting fluid flow rate and the historical target cutting fluid flow rate into the loss function to calculate the loss value; The gradient of the output layer of the neural network model is calculated based on the loss value, and the gradient is passed forward layer by layer through the chain rule to calculate the gradient of the parameters of each layer and obtain the gradient data. Based on the gradient data and the preset learning rate, the parameters of each layer of the neural network model are updated using the gradient descent method. The parameters of each layer are iteratively updated until the number of training iterations of the neural network model is greater than a preset number of iterations, or the loss value of the neural network model is less than a preset loss threshold. At this point, the training is considered complete, and a trained neural network model is obtained.

5. The automated cutting fluid circulation control method for CNC intelligent manufacturing according to claim 2, characterized in that, The step of performing power change level classification, feed rate sequence matching, and sequence optimization based on the flow rate adjustment sequence and the denoised dataset to obtain the feed rate adjustment sequence includes: Based on the flow regulation sequence, the power change amplitude between adjacent time periods is calculated using the first-order difference method, and the power change amplitude is classified into levels using a fuzzy logic algorithm to obtain the power change level. Based on the power change level, a speed matching sequence is obtained by matching in a pre-established change level-feed speed association database using the nearest neighbor algorithm; Based on the denoised feed rate, frequency features are extracted using a fast Fourier transform algorithm, and the Pearson correlation coefficient between the frequency features and the denoised vibration frequency is calculated to obtain the feed rate correction coefficient. Based on the feed rate correction coefficient, the speed matching sequence is corrected using the weighted least squares method to obtain the feed rate adjustment sequence.

6. The automated cutting fluid circulation control method for CNC intelligent manufacturing according to claim 2, characterized in that, The step of performing cooling uniformity analysis based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence includes: Based on the feed rate adjustment sequence and the noise reduction cutting depth, the cooling compensation requirement is obtained by performing a fuzzy logic algorithm to analyze the compensation requirement. Based on the cooling compensation requirements and the noise reduction cutting fluid flow rate, non-uniformity detection is performed by calculating the standard deviation within a preset sliding window to generate a cooling non-uniformity signal; Based on the uneven cooling signal, the compensation peak point is determined by the extreme value detection algorithm, and the compensation peak point is fused with the noise reduction spindle speed by the weighted average method to obtain the cooling compensation sequence.

7. The automated cutting fluid circulation control method for CNC intelligent manufacturing according to claim 1, characterized in that, The step of performing correlation analysis, global compensation adjustment, path optimization, and flow regulation based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence includes: Based on the cooling compensation sequence and the processing load fluctuation index, a correlation analysis is performed using grey relational analysis to obtain the correlation deviation sequence; Based on the aforementioned correlation deviation sequence, the flow regulation sequence is compensated and adjusted using a weighted moving average method to obtain a global regulation sequence; Based on the global control sequence, the cutting fluid flow path is optimized using the ant colony algorithm to obtain an optimized path sequence; Based on the optimized path sequence, the valves and pumps of the circulating conveying system are controlled to regulate the flow rate.

8. An automated cutting fluid circulation control system for CNC intelligent manufacturing, characterized in that, include: The data acquisition module is used to collect data on the feed rate, depth of cut, spindle speed, vibration frequency, temperature change, and cutting fluid flow rate of the CNC machine tool to obtain the raw dataset. The load fluctuation quantization module is used to filter out noise based on the original dataset to obtain a denoised dataset and quantify the processed load fluctuation to obtain a processed load fluctuation index. The coupling characteristic analysis module is used to perform parameter coupling analysis, clustering and pattern matching based on the processing load fluctuation index and pre-stored historical parameter change data to obtain parameter coupling characteristics; The flow prediction module is used to predict the target value of the cutting fluid flow rate and match the flow adjustment sequence based on the parameter coupling characteristics through a pre-built neural network model. The feed rate adjustment module is used to perform power change level classification, feed rate sequence matching and sequence optimization based on the flow rate adjustment sequence and the denoised dataset to obtain the feed rate adjustment sequence; The cooling uniformity analysis module is used to perform cooling uniformity analysis based on the feed rate adjustment sequence and the denoised dataset to obtain a cooling compensation sequence. The output module is used to perform correlation analysis, global compensation adjustment, path optimization, and flow regulation based on the cooling compensation sequence, the processing load fluctuation index, and the flow regulation sequence.

Images