Dynamic regulation and control method and control system for cutting fluid
By combining multimodal neural networks and point support vector machines, cutting fluid performance can be monitored and predicted in real time, and control parameter instructions can be generated. This solves the problems of insufficient proportion adjustment and performance monitoring in cutting fluid management solutions, and improves cutting accuracy and processing stability.
Patent Information
- Application Number
- CN202510823286.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-09-09
AI Technical Summary
Existing cutting fluid management solutions are unable to adjust the ratio of oil and water in real time, resulting in reduced cutting accuracy, insufficient performance monitoring and maintenance, and difficulty in adapting to complex and changing cutting requirements.
By obtaining the machining process plan, generating the target cutting fluid performance parameter set, and using multimodal neural networks and point support vector machines for real-time monitoring and prediction, a cutting fluid control parameter instruction set is generated, including the water-based/oil-based additive replenishment rate and circulation filtration flow, to build a closed-loop control system.
The dynamic matching accuracy between cutting fluid performance and processing technology requirements has been improved, the applicability of cutting fluid to complex processing scenarios has been enhanced, and the stability and processing quality of the cutting process have been guaranteed.
Smart Images
Figure CN120610593A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of dynamic control of cutting fluid, and in particular to a dynamic control method and control system for cutting fluid. Background Art
[0002] In modern machining, cutting fluids, as a key process medium, perform multiple functions, including reducing cutting zone temperature, inhibiting tool wear, and improving chip removal efficiency, directly impacting machining accuracy and workpiece surface quality. However, over the course of service, their performance can deteriorate due to the consumption, oxidative decomposition, or adsorption of functional additives (lubricants, extreme pressure agents, etc.), causing their effective concentration to deviate from their initial value. Furthermore, metal debris, abrasive particles, and environmental pollutants generated during machining continue to accumulate in the circulation system, causing dynamic performance degradation.
[0003] To maintain stable cutting fluid performance, the industry generally uses a circulating filtration system for purification, supplemented by regular concentration monitoring and additive replenishment. However, in the current manufacturing environment, the same system often needs to alternate between high-speed cutting and high-precision cutting operations, and the performance requirements of the cutting fluid for these two operations differ significantly: high-speed cutting generates extremely high cutting heat, requiring the cutting fluid to achieve excellent heat dissipation by leveraging the high specific heat capacity and latent heat of vaporization of the water phase, thus requiring a higher water-based ratio; high-precision cutting places stringent demands on processing stability and surface finish, relying on oily components or additives to play a dominant role in improving lubricity and extreme pressure properties, thus requiring a higher oil-based ratio.
[0004] The current cutting fluid management scheme generally has the following problems: (1) The formula is rigid. A single formula cannot adjust the ratio of oil and water or the emphasis of additives in real time according to different processing modes; (2) Performance monitoring and maintenance are insufficient. The real-time and accurate monitoring of key indicators such as concentration and pH value is missing, resulting in delayed or excessive additive replenishment.
[0005] In summary, the cutting fluid produced by existing technology is difficult to adapt to the current complex and changeable cutting requirements, resulting in reduced cutting accuracy. Summary of the Invention
[0006] The content of this application is used to briefly introduce concepts that will be described in detail in the detailed description section below. The content of this application is not intended to identify key features or essential features of the technical solution for which protection is sought, nor is it intended to limit the scope of the technical solution for which protection is sought.
[0007] As a first aspect of the present application, in order to solve the technical problems mentioned in the background technology, the present application provides a method for dynamically controlling a cutting fluid, comprising the following steps:
[0008] Step 1: Obtain a machining process plan and, based on the machining process plan, generate a target cutting fluid performance parameter set required for each machining time interval;
[0009] The target cutting fluid performance parameter set includes: a water-based concentration target threshold range, an oil-based concentration target threshold range, a metal particle concentration limit range, a pH value target threshold range, and a viscosity target threshold range;
[0010] Step 2: Obtaining cutting fluid monitoring information; the monitoring information includes temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information;
[0011] Step 3: Preprocess the real-time condition monitoring data and construct a multi-channel time series matrix;
[0012] Step 4: Extract the fused temporal attention features from the multi-channel time series matrix to generate a state evolution feature vector that characterizes the dynamic changes of the cutting fluid state;
[0013] Step 5: Use point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval;
[0014] Step 6: Determine whether the expected cutting fluid performance parameter set is equal to the target cutting fluid performance parameter set corresponding to the machining time interval in the machining process plan; if they are equal, repeat steps 2 to 6; otherwise, execute step 7;
[0015] Step 7: Extract the performance deviation quantitative index between the expected cutting fluid performance parameter set and the target cutting fluid performance parameter set; generate the cutting fluid control parameter instruction set based on the performance deviation quantitative index; the cutting fluid control parameter instruction set includes the water-based additive replenishment rate, the oil-based additive replenishment rate, the circulation filtration system flow rate and the functional solvent addition rate.
[0016] The dynamic control method for cutting fluids provided in this application utilizes a multimodal neural network to extract evolutionary feature vectors representing the dynamic changes in the cutting fluid's state from monitored information such as cutting fluid temperature, viscosity, pH, and metal particle concentration. This vector accurately captures the temporal characteristics of the cutting fluid's property changes during the machining process. By integrating temporal attention with a point support vector machine, high-precision predictions of future cutting fluid performance parameters are achieved, and the predicted results are matched and analyzed in real time with the target parameter set required for the machining process. When performance deviations are detected, the system automatically generates a set of control parameter instructions, including the water-based / oil-based additive replenishment rate, the circulating filtration flow rate, and the functional solvent addition rate, thereby establishing a closed-loop control system of "real-time monitoring-feature modeling-prediction comparison-precise control." This method significantly improves the dynamic matching accuracy between cutting fluid performance and machining process requirements, and can adaptively adjust according to the real-time evolution trend of cutting fluid properties during actual production. This effectively enhances the cutting fluid's applicability to complex machining scenarios, fundamentally ensures the stability of the cutting process, and provides reliable technical support for improving machining quality and production efficiency.
[0017] During the recycling of cutting fluid, complex operating conditions such as high temperature and high pressure in the tool cutting area, fluid impact in the circulation pipeline, and impurity deposition in the fluid reservoir can easily lead to abnormal values such as jumps and spikes in the data collected by temperature, viscosity, pH value, and metal particle concentration sensors. If these abnormal data are directly used in cutting fluid state analysis and performance prediction, they will interfere with the judgment of the cutting fluid's true state, causing the cutting fluid performance parameters perceived by the multimodal neural network to be inaccurate, thereby affecting the accuracy and timeliness of the cutting fluid's dynamic control, making it difficult to ensure the stable operation of the cutting process and the processing quality.
[0018] Step 2 includes the following steps:
[0019] Step 21: Install temperature, viscosity, pH value, and metal particle concentration sensors in the tool cutting area, the middle section of the circulation pipeline, the liquid storage tank, the return pipeline, and the downstream position of the filter device of the cutting fluid circulation system;
[0020] Step 22: synchronously collecting the temperature, viscosity, pH value and metal particle concentration data of the cutting fluid at a preset frequency to generate temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information;
[0021] Step 23: Use median filtering to filter the temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information.
[0022] The technical solution provided in this application realizes comprehensive and synchronous collection of multi-dimensional data of cutting fluid by arranging sensors in the cutting fluid circulation system, and uses a median filtering algorithm to process the collected temperature, viscosity, pH value and metal particle concentration time series information to effectively eliminate abnormal values and noise interference in the data.
[0023] Furthermore, step 3 includes the following steps:
[0024] Step 31: Normalizing the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information to obtain a temperature time series, a viscosity time series, a pH value time series, and a metal particle concentration time series;
[0025] Step 32: Align the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series at the same time;
[0026] Step 33: Arrange the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series from top to bottom to generate a two-dimensional array, which is a time series matrix;
[0027] The first row in the time series matrix is the temperature time series, the second row is the viscosity time series, the third row is the pH value time series, and the fourth row is the metal particle concentration time series.
[0028] In the technical solution provided in the present application, by integrating temperature, viscosity, pH value, and metal particle concentration into a time series matrix, the correlation characteristics between temperature, viscosity, pH value and metal particle concentration can be better extracted, and the corresponding rules can be found from the time series changes. In addition, the temperature, viscosity, pH value and metal particle concentration are all normalized to reduce the influence of the digital levels between temperature, viscosity, pH value and metal particle concentration, so that when extracting information from the time series matrix, it is not easily affected by outliers.
[0029] During the cutting fluid's circulation process, complex operating conditions such as high temperature and pressure in the tool cutting zone, fluid impact in the circulation pipeline, and impurity deposition in the fluid reservoir can easily cause abnormal values such as jumps and spikes in the data collected by temperature, viscosity, pH, and metal particle concentration sensors. If these abnormal data are directly used in cutting fluid state analysis and performance prediction, they will interfere with the judgment of the cutting fluid's true state and cause the cutting fluid performance parameters perceived by the multimodal neural network to be inaccurate.
[0030] Furthermore, in step 4, the attention-temporal extraction model is used to extract the temporal attention features to generate the state evolution feature vector;
[0031] Among them, the attention-temporal extraction model extracts the attention information of different channels from the multi-channel time series matrix and performs temporal fusion to generate temporal attention features.
[0032] Furthermore, the attention-temporal extraction model includes:
[0033] The attention mechanism extraction module is used to input a multi-channel time series matrix, extract attention mechanisms at different scales, and generate attention matrix features;
[0034] The temporal feature extraction module is integrated to extract long and short memory information from the attention matrix features to generate temporal attention features;
[0035] The regression information output module converts the temporal attention features into state evolution feature vectors based on the RuLe activation function.
[0036] This technical solution introduces an attention-time series extraction model, combining the attention mechanism with time series feature extraction. It can deeply explore the dynamic dependencies between different monitoring parameters in the multi-channel time series matrix and accurately capture the changing patterns of each parameter in the time dimension. By extracting and fusing time series attention features, the generated state evolution feature vector can comprehensively and accurately characterize the dynamic changes in the state of the cutting fluid. It not only achieves the efficient integration of multi-dimensional parameter information, but also highlights the influence weight of key parameters on the state of the cutting fluid at different time points, so that the dynamic control of the cutting fluid can more accurately predict the trend of changes in cutting fluid performance and adjust the control parameters in a timely manner.
[0037] When extracting the attention mechanism, the attention mechanism extraction network in the prior art generally integrates the attention mechanism weights into the convolutional layer. This attention extraction method requires a lot of computing resources to train the attention weights in practice. Therefore, when the number of samples is small, it is difficult to accurately obtain a high-precision attention extraction network. To this end, this application provides the following technical solutions:
[0038] Furthermore, the attention mechanism extraction module includes:
[0039] There are three convolutional layers, and the multi-channel time series matrix is input into the three convolutional layers to generate the first feature map, the second feature map, and the third feature map;
[0040] A first processing network performs matrix transformation and transposition transformation on the first feature map to generate a first transformation feature;
[0041] A second processing network performs a matrix transformation on the second feature map to generate a second transformed feature;
[0042] a third processing network, performing matrix transformation on the third feature map to generate a third transformed feature;
[0043] Activate the network, multiply the first transformed feature and the second transformed feature and activate them with the activation function to generate a spatial feature map;
[0044] Restore the network, multiply the control feature with the third transformation feature and perform matrix transformation to restore it to a multi-channel attention matrix, and add the multi-channel attention matrix to the multi-channel time series matrix to generate the attention matrix feature.
[0045] In the technical solution provided by the present application, when extracting the attention mechanism, the input multi-channel time series matrix is first convolved with different weights and bias terms. The first transformation feature is then integrated into the second transformation feature and the third transformation feature at different stages until it is integrated into the original multi-channel time series matrix. Throughout the entire process, information with original characteristics is continuously added, so that the information can be continuously refined, and the attention matrix feature finally output is an enhanced version of the multi-channel time series matrix, which can give more information to important areas in the multi-channel time series matrix and effectively highlight these key areas. During the training process of the network structure of the present application, the bias terms and weights of the three convolutional layers are different, and the feature processing routes output by the three convolutional layers are inconsistent. During training, the differences between different bias terms and weights can be quickly increased, thereby increasing the model convergence speed.
[0046] The fusion time series feature extraction module includes:
[0047] The temporal input layer performs temporal differentiation on the attention matrix features to generate n differential features;
[0048] The first LSTM layer includes n LSTM cores, and n differential features are input into n LSTM cores respectively;
[0049] The second LSTM layer includes n LSTM cores, and n differential features are respectively input into the n LSTM cores. The first LSTM layer is forward propagated, and the second LSTM layer is backward propagated. The LSTM cores at the same position in the first and second LSTM layers generate a time series feature.
[0050] The fully connected layer collects all the time series features and connects them;
[0051] The temporal output layer generates temporal attention features based on the activation function output of the temporal features.
[0052] Furthermore, step 5 includes the following steps:
[0053] Step 5: Use the point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval.
[0054] Step 5 includes the following steps:
[0055] Step 51: Collect a number of training samples to form a training sample set; each training sample includes a temporal attention feature and an expected cutting fluid performance parameter set;
[0056] Step 52: Build a support vector machine model;
[0057] ;
[0058] Among them, x represents the temporal attention feature, sign() represents the sign function, K represents the kernel function, b represents the bias term, i represents the index of the training sample, S represents the set of samples, α represents the Lagrange multiplier, and y i Represents the training sample x i Realistic set of cutting fluid performance parameters;
[0059] Step 53: Cross-train the support vector machine model;
[0060] Step 54: Input the temporal attention features into the trained support vector machine model to generate an expected cutting fluid performance parameter set.
[0061] In the technical solution provided in this application, the support vector machine is used to map the temporal attention features and the expected cutting fluid performance parameter set to each other, which can effectively ensure the accuracy of the mapping and increase the estimated accuracy of the expected cutting fluid performance parameter set.
[0062] The present invention also provides a cutting fluid dynamic control system, which uses the cutting fluid dynamic control method to generate a cutting fluid control parameter instruction set. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] The drawings constituting a part of this application are used to provide a further understanding of this application and make other features, purposes and advantages of this application more apparent. The drawings and descriptions of the exemplary embodiments of this application are used to explain this application and do not constitute an improper limitation on this application.
[0064] In addition, throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the elements and components are not necessarily drawn to scale.
[0065] In the attached figure:
[0066] Figure 1 This is a flow chart of the dynamic control method of cutting fluid.
[0067] Figure 2 Schematic diagram of the structure of the attention mechanism extraction module.
[0068] Figure 3Schematic diagram of the structure of the fusion time series feature extraction module. DETAILED DESCRIPTION
[0069] The following will describe embodiments of the present application in more detail with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are for illustrative purposes only and are not intended to limit the scope of protection of the present application.
[0070] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.
[0071] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0072] Example 1:
[0073] refer to Figures 1-3 A cutting fluid dynamic control method of the present invention comprises the following steps:
[0074] Step 1: Obtain a machining process plan and, based on the machining process plan, generate a target cutting fluid performance parameter set required for each machining time interval;
[0075] The target cutting fluid performance parameter set includes: a water-based concentration target threshold range, an oil-based concentration target threshold range, a metal particle concentration limit range, a pH value target threshold range, and a viscosity target threshold range.
[0076] Specifically, a machining process plan can be obtained in advance, including information such as the part type, process steps, equipment, and machining parameters. Based on the machining steps and parameters in the process plan, the functional requirements of the cutting fluid at different machining stages can be analyzed. For example, rough machining focuses on cooling and chip removal, while finishing machining emphasizes lubrication and surface quality improvement.
[0077] In practice, industry standards and technical specifications can be referenced to clarify the specific range of cutting fluid performance parameters under different processing requirements. In this way, the target cutting fluid performance parameter set can be obtained according to the processing technology plan.
[0078] The water-based concentration mainly affects the cooling ability of the cutting fluid. The higher the speed during cutting, the higher the water-based concentration needs to be. The oil-based concentration mainly affects the lubricating ability of the cutting fluid and is mainly used to control the accuracy of processing. The metal particle concentration, pH value and viscosity generally need to be controlled within a fixed range and do not change. This part is mainly used to characterize the life of the cutting fluid. Generally speaking, the pH value and viscosity can be accurately controlled by adjusting the addition rate of solvent water, while the metal particle concentration cannot be controlled. When the metal particle concentration exceeds the threshold, the cutting fluid cannot be used and needs to be replaced directly.
[0079] In this solution, the cutting equipment has a built-in cutting fluid circulation system that filters the cutting fluid. However, the faster the filtration rate, the less effective it is at removing metal particles. Therefore, controlling the cutting fluid circulation rate can also adjust the metal particle concentration.
[0080] Step 2: Obtain cutting fluid monitoring information; the monitoring information includes temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information.
[0081] Step 2 includes the following steps:
[0082] Step 21: Install temperature, viscosity, pH value, and metal particle concentration sensors in the tool cutting area, the middle section of the circulation pipeline, the liquid storage tank, the return pipeline, and the downstream position of the filter device of the cutting fluid circulation system;
[0083] Step 22: synchronously collecting the temperature, viscosity, pH value and metal particle concentration data of the cutting fluid at a preset frequency to generate temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information;
[0084] Based on the machining process's requirements for timely monitoring of cutting fluid performance changes, a suitable preset frequency (e.g., once per second, once per minute, etc.) is set. Each sensor simultaneously initiates a data acquisition program at the preset frequency, periodically measuring the cutting fluid's temperature, viscosity, pH value, and metal particle concentration. Each acquired data set is recorded chronologically, generating time-stamped information on temperature, viscosity, pH value, and metal particle concentration.
[0085] Step 23: Use median filtering to filter the temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information.
[0086] Median filtering is primarily used to remove noise. It's a well-known technique, and its specific implementation is omitted here. The general approach is to set up several windows, such as one window of three or five data points. The data within each window is sorted, and the median value is used to replace the original data within the window. This sliding window is then used to traverse the entire data, eliminating isolated noise points and making the time series information of each parameter smoother and more accurate.
[0087] Step 3: Preprocess the real-time condition monitoring data and construct a multi-channel time series matrix.
[0088] Furthermore, step 3 includes the following steps:
[0089] Step 31: Normalize the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information respectively to obtain a temperature time series, a viscosity time series, a pH value time series, and a metal particle concentration time series.
[0090] Temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information are all time-related series.
[0091] For example, the temperature time series information is: 31, 32, 33, 34, ...; the viscosity time series information is: 2, 3, 2, ...; the pH value time series information is: 6, 6, 7, 6, ...; and the metal particle concentration information is: 10, 11, 12, 13, .... The data at the same position in the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information have the same corresponding time. Normalizing the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information separately means normalizing each data in the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information. The specific formula for normalization will not be explained here.
[0092] After normalization, we obtain the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series. In these time series, the data at the same position corresponds to the same time and has the same value range.
[0093] Step 32: Align the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series at the same time.
[0094] Since the data at the same position in the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series correspond to the same time, it is only necessary to arrange the temperature time series, viscosity time series, pH value time series, and metal particle concentration in the same way.
[0095] Step 33: Arrange the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series from top to bottom to generate a two-dimensional array, which is a time series matrix; the first row of the time series matrix is the temperature time series, the second row is the viscosity time series, the third row is the pH value time series, and the fourth row is the metal particle concentration time series.
[0096] In a time series matrix, the number of rows does not change, while the number of columns increases as information is updated.
[0097] Step 4: Extract and fuse temporal attention features from the multi-channel time series matrix to generate a state evolution feature vector that characterizes the dynamic changes in the cutting fluid state. In step 4, the attention-temporal extraction model is used to extract temporal attention features to generate the state evolution feature vector. The attention-temporal extraction model extracts attention information from different channels in the multi-channel time series matrix and performs temporal fusion to generate the temporal attention features.
[0098] The attention-temporal extraction model consists of an attention mechanism extraction module, a fused temporal feature extraction module, and a regression information output module. The attention mechanism extraction module inputs a multi-channel time series matrix, extracts attention mechanisms at different scales, and generates attention matrix features. The fused temporal feature extraction module extracts long-short memory information from the attention matrix features to generate temporal attention features. The regression information output module converts the temporal attention features into state evolution feature vectors based on the RuLe activation function.
[0099] Specifically, the attention mechanism extraction module includes:
[0100] There are three convolutional layers, and the multi-channel time series matrix is input into the three convolutional layers to generate the first feature map, the second feature map, and the third feature map;
[0101] A first processing network performs matrix transformation and transposition transformation on the first feature map to generate a first transformation feature;
[0102] A second processing network performs a matrix transformation on the second feature map to generate a second transformed feature;
[0103] a third processing network, performing matrix transformation on the third feature map to generate a third transformed feature;
[0104] Activate the network, multiply the first transformed feature and the second transformed feature and activate them with the activation function to generate a spatial feature map;
[0105] Restore the network, multiply the control feature with the third transformation feature and perform matrix transformation to restore it to a multi-channel attention matrix, and add the multi-channel attention matrix to the multi-channel time series matrix to generate the attention matrix feature.
[0106] In the attention mechanism extraction module, the core of extracting attention matrix features is the convolutional layer. Each convolutional layer has a 1*1 convolution kernel and different bias and weight parameters. The weight parameters of the convolutional layer and the bias are continuously adjusted during the training process.
[0107] The fusion time series feature extraction module includes:
[0108] The temporal input layer performs temporal differentiation on the attention matrix features to generate n differential features.
[0109] The generation process of differential features is as follows:
[0110] ;
[0111] Represents the features input to the LSTM layer, x t represents the original attention matrix features, and S represents the fusion weight;
[0112] ;
[0113] Among them, x k represents the target signal, y represents the set of different label signals, the total number of signals in the set of different label signals y, n represents the signal length, j represents the index of different label signals, and i represents the index of the time step. Indicates that the target signal is at x k The value at y ij represents the value of the jth signal at time i in the set of different label signals, represents the target differential signal, Indicates the difference between different tag signals.
[0114] The use of differential features can increase the differences between the data input to the LSTM layer, making it easier to focus on temporal features during forward and backward propagation. Traditional fusion weights cannot adapt to the changing pattern of attention matrix features. The S distance dynamically adjusts the fusion weights according to the similarity between the target signal and the interference signal, improving the generalization ability. Therefore, when the cutting process changes, the model can have good convergence ability during training.
[0115] The first LSTM layer includes n LSTM cores, and n differential features are input into n LSTM cores respectively;
[0116] The second LSTM layer includes n LSTM cores, and n differential features are respectively input into the n LSTM cores. The first LSTM layer is forward propagated, and the second LSTM layer is backward propagated. The LSTM cores at the same position in the first and second LSTM layers generate a time series feature.
[0117] The fully connected layer collects all the time series features and connects them;
[0118] The temporal output layer generates temporal attention features based on the activation function output of the temporal features.
[0119] The loss function of the attention-temporal extraction model in the technical solution provided by this application is as follows:
[0120] ;
[0121] N represents the total number of training data, i represents the index of the training sample, represents a focusing parameter greater than zero, represents the true label, Denotes the predicted category, d i Indicates the similarity between the i-th training sample and the standard sample, g i It represents the similarity between the predicted label of the i-th training sample and the standard label, and the similarity is cosine similarity.
[0122] Specifically, a standard sample and a standard label are pre-set, and a predicted label of a training sample is obtained when each training sample is trained. ,Will Calculating the similarity with the standard label can get g i , we can calculate the training sample and the standard sample to get d i , it can be foreseen that if the corresponding prediction labels change synchronously with the changes in training samples, then The value of will tend to evolve in a smaller direction. If the training sample changes, but the predicted label changes little, or does not change, then It will increase. The loss function in this application requires the loss function to develop in the direction of minimum. In this way, the use of the loss function will guide the synchronization of differentiated information between the training samples and the predicted labels, and thus in practice, reducing the number of annotations can also achieve better training results. At the same time, this application sets a focus parameter greater than zero. For samples that are already well predicted and close to the true label, their contribution to the loss is reduced, and the weight of difficult-to-distinguish samples is retained: for samples that are inaccurately predicted by the model (far from the true label), their impact on the loss is maintained.
[0123] Step 5: Use the point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval.
[0124] Step 5 includes the following steps:
[0125] Step 5: Use the point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval.
[0126] Step 5 includes the following steps:
[0127] Step 51: Collect a number of training samples to form a training sample set; each training sample includes a temporal attention feature and an expected cutting fluid performance parameter set;
[0128] Step 52: Build a support vector machine model;
[0129] ;
[0130] Among them, x represents the temporal attention feature, sign() represents the sign function, K represents the kernel function, b represents the bias term, i represents the index of the training sample, S represents the set of samples, α represents the Lagrange multiplier, and y i Represents the training sample x i Realistic set of cutting fluid performance parameters;
[0131] Step 53: Cross-train the support vector machine model;
[0132] When sufficient training samples are provided, training a support vector machine is an existing technology, and the specific training process will not be described here.
[0133] Step 54: Input the temporal attention features into the trained support vector machine model to generate an expected cutting fluid performance parameter set.
[0134] Step 6: Determine whether the expected cutting fluid performance parameter set is equal to the target cutting fluid performance parameter set corresponding to the machining time interval in the machining process plan; if they are equal, repeat steps 2 to 6; otherwise, execute step 7;
[0135] Step 7: Extract the performance deviation quantitative indicators between the expected cutting fluid performance parameter set and the target cutting fluid performance parameter set, and generate the cutting fluid control parameter instruction set based on the performance deviation quantitative indicators; the cutting fluid control parameter instruction set includes: water-based additive replenishment rate, oil-based additive replenishment rate, circulation filtration system flow rate and functional solvent addition rate.
[0136] The cutting fluid performance parameter set includes: a target threshold range for water-based concentration, a target threshold range for oil-based concentration, a limit range for metal particle concentration, a target threshold range for pH value, and a target threshold range for viscosity. Thus, the performance deviation quantification indicator is the difference between the water-based concentration, oil-based concentration, metal particle concentration, pH value, and viscosity targets. Given these differences, cutting fluid control parameter instructions can be generated based on demand, quickly bringing the expected cutting fluid performance parameter set closer to the target cutting fluid performance parameter set.
[0137] Example 2:
[0138] A cutting fluid dynamic control system adopts the above-mentioned cutting fluid dynamic control method to generate a cutting fluid control parameter instruction set.
[0139] The above descriptions are merely some preferred embodiments of the present application and illustrate the technical principles employed. Those skilled in the art should understand that the scope of the invention described in the embodiments of the present application is not limited to technical solutions formed by specific combinations of the aforementioned technical features, but also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present application.
Claims
1. A method for dynamically controlling cutting fluid, characterized in that: The cutting fluid dynamic control method includes: Step 1: Obtain a machining process plan and generate a target cutting fluid performance parameter set required for each machining time interval based on the machining process plan; The target cutting fluid performance parameter set includes: a water-based concentration target threshold range, an oil-based concentration target threshold range, a metal particle concentration limit range, a pH value target threshold range, and a viscosity target threshold range; Step 2: Obtaining cutting fluid monitoring information; the monitoring information includes temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information; Step 3: Preprocess the real-time condition monitoring data and construct a multi-channel time series matrix; Step 4: Extract the fused temporal attention features from the multi-channel time series matrix to generate a state evolution feature vector that characterizes the dynamic changes of the cutting fluid state; Step 5: Use point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval; Step 6: Determine whether the expected cutting fluid performance parameter set is equal to the target cutting fluid performance parameter set corresponding to the machining time interval in the machining process plan; if they are equal, repeat steps 2 to 6; otherwise, execute step 7; Step 7: Extract the performance deviation quantitative index between the expected cutting fluid performance parameter set and the target cutting fluid performance parameter set; generate the cutting fluid control parameter instruction set based on the performance deviation quantitative index; the cutting fluid control parameter instruction set includes the water-based additive replenishment rate, the oil-based additive replenishment rate, the circulation filtration system flow rate and the functional solvent addition rate.
2. The cutting fluid dynamic control method according to claim 1, characterized in that: Step 2 includes the following steps: Step 21: Install temperature, viscosity, pH value and metal particle concentration sensors in the tool cutting area, the middle section of the circulation pipeline, the liquid storage tank, the return pipeline and the downstream position of the filter device of the cutting fluid circulation system; Step 22: synchronously collecting the temperature, viscosity, pH value and metal particle concentration data of the cutting fluid at a preset frequency to generate temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information; Step 23: Use median filtering to filter the temperature time series information, viscosity time series information, pH value time series information and metal particle concentration information.
3. The cutting fluid dynamic control method according to claim 2, characterized in that: Step 3 includes the following steps: Step 31: Normalizing the temperature time series information, viscosity time series information, pH value time series information, and metal particle concentration information to obtain a temperature time series, a viscosity time series, a pH value time series, and a metal particle concentration time series; Step 32: Align the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series at the same time; Step 33: Arrange the temperature time series, viscosity time series, pH value time series, and metal particle concentration time series from top to bottom to generate a two-dimensional array, which is a time series matrix; The first row in the time series matrix is the temperature time series, the second row is the viscosity time series, the third row is the pH value time series, and the fourth row is the metal particle concentration time series.
4. The cutting fluid dynamic control method according to claim 1, characterized in that: In step 4, the attention-temporal extraction model is used to extract the temporal attention features to generate the state evolution feature vector; Among them, the attention-temporal extraction model extracts the attention information of different channels from the multi-channel time series matrix and performs temporal fusion to generate temporal attention features.
5. The cutting fluid dynamic control method according to claim 4, characterized in that: The attention-temporal extraction model includes: The attention mechanism extraction module is used to input a multi-channel time series matrix, extract attention mechanisms at different scales, and generate attention matrix features; The temporal feature extraction module is integrated to extract long and short memory information from the attention matrix features to generate temporal attention features; The regression information output module converts the temporal attention features into state evolution feature vectors based on the RuLe activation function.
6. The cutting fluid dynamic control method according to claim 5, characterized in that: The attention mechanism extraction module includes: There are three convolutional layers, and the multi-channel time series matrix is input into the three convolutional layers to generate the first feature map, the second feature map, and the third feature map; A first processing network performs matrix transformation and transposition transformation on the first feature map to generate a first transformed feature; A second processing network performs a matrix transformation on the second feature map to generate a second transformed feature; a third processing network, performing matrix transformation on the third feature map to generate a third transformed feature; Activate the network, multiply the first transformed feature and the second transformed feature and activate them with the activation function to generate a spatial feature map; Restore the network, multiply the control feature with the third transformation feature and perform matrix transformation to restore it to a multi-channel attention matrix, and add the multi-channel attention matrix to the multi-channel time series matrix to generate the attention matrix feature.
7. The cutting fluid dynamic control method according to claim 6, characterized in that: The fusion time series feature extraction module includes: The temporal input layer performs temporal differentiation on the attention matrix features to generate n differential features; The first LSTM layer includes n LSTM cores, and n differential features are input into n LSTM cores respectively; The second LSTM layer includes n LSTM cores, and n differential features are respectively input into the n LSTM cores. The first LSTM layer is forward propagated, and the second LSTM layer is backward propagated. The LSTM cores at the same position in the first and second LSTM layers generate a time series feature. The fully connected layer collects all the time series features and connects them; The temporal output layer generates temporal attention features based on the activation function output of the temporal features.
8. The cutting fluid dynamic control method according to claim 7, characterized in that: Step 5 includes the following steps: Step 5: Use point support vector machine to perform feature mapping and correlation analysis on the state evolution feature vector and the target cutting fluid performance parameter set to predict the expected cutting fluid performance parameter set in the future specified time interval; Step 5 includes the following steps: Step 51: Collect a number of training samples to form a training sample set; each training sample includes a temporal attention feature and an expected cutting fluid performance parameter set; Step 52: Build a support vector machine model; ； Among them, x represents the temporal attention feature, sign() represents the sign function, K represents the kernel function, b represents the bias term, i represents the index of the training sample, S represents the set of samples, α represents the Lagrange multiplier, and y i Represents the training sample x i Realistic set of cutting fluid performance parameters; Step 53: Cross-train the support vector machine model; Step 54: Input the temporal attention features into the trained support vector machine model to generate an expected cutting fluid performance parameter set.
9. A cutting fluid dynamic control system, characterized in that: The cutting fluid dynamic control method according to any one of claims 1 to 8 is used to generate a cutting fluid control parameter instruction set.