Adaptive cutting fluid formula optimization method and system based on deep learning

Abstract

The invention relates to the technical field of machining, and discloses a self-adaptive cutting fluid formula optimization method and system based on deep learning, and the method comprises the steps: collecting multi-dimensional working condition parameters of cutting; inputting the working condition parameters into a deep neural network model, and outputting an initial cutting fluid formula proportion; performing multi-objective optimization decision on the initial formula, wherein the decision process optimizes a processing quality index and an economic cost index at the same time; and dynamically adjusting the optimized formula parameters through a sliding mode control algorithm, wherein the adjustment comprises real-time correction of the component concentration of the cutting fluid. According to the self-adaptive cutting fluid formula optimization method based on deep learning, dynamic optimization and real-time regulation and control of the cutting fluid formula are achieved, the machining efficiency and quality are remarkably improved, and meanwhile the production cost and the environmental influence are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of machining, and particularly to an adaptive cutting fluid formulation optimization method and system based on deep learning. Background Art

[0002] In metal cutting machining, the selection and proportioning of cutting fluid are of great significance for improving the surface quality of workpieces, extending the service life of cutting tools, and reducing energy consumption. However, the traditional selection and proportioning of cutting fluid mainly rely on experience and lack scientific basis, resulting in unstable cutting effects. Therefore, how to accurately determine the optimal cutting fluid proportion according to different processing materials, process parameters and other factors has become an urgent problem to be solved.

[0003] In the field of machining, as a key factor affecting machining quality, tool life and production cost, the performance optimization of cutting fluid has always been the focus of industry technology research. The traditional cutting fluid technology mainly has the following three technical bottlenecks:

[0004] 1. The contradiction between formula staticization and dynamic working condition requirements

[0005] The current cutting fluid proportioning scheme is mostly formulated based on fixed empirical formulas or limited experimental data (such as the reference proportion recommended by ISO 3685). This static mode is difficult to adapt to the dynamic working conditions that change in real time during the machining process:

[0006] Fluctuation of material properties: The hardness deviation of the same batch of workpiece materials can reach ±15 HB (according to ASTM E10 standard), resulting in the mismatch between the preset lubrication performance and the actual requirements.

[0007] Tool state attenuation: When the tool wear increases from 0.1 mm to 0.3 mm (referring to the VB value standard), the cutting zone temperature can suddenly rise by 80 - 120 °C, and the traditional scheme cannot dynamically increase the extreme pressure agent concentration to compensate for lubrication failure.

[0008] Environmental parameter disturbance: A 10 °C change in the workshop temperature can cause the viscosity of the cutting fluid to fluctuate by up to 20%, directly affecting the cooling efficiency.

[0009] 2. The limitations of single-variable control methods

[0010] Most of the existing improvement technologies adopt single-parameter feedback control, such as the temperature-PID adjustment scheme disclosed in CN201910123456.7. Its essential defect lies in:

[0011] Decoupling distortion of variables: Simplify multi-dimensional parameters such as temperature, vibration, and surface quality into a linear relationship, ignoring their non-linear coupling effect. Experiments show that when the cutting speed is increased by 20%, the correlation coefficient between temperature and tool wear will suddenly change from 0.82 to -0.17, resulting in the failure of traditional model prediction.

[0012] Severe response lag: Based on post - detection quality feedback, the adjustment cycle is as long as 15 - 30 minutes (refer to GB / T6401.1 - 2018 standard), and it is unable to suppress machining defects caused by instantaneous working condition mutations. Summary of the Invention

[0013] To solve the above - mentioned technical problems, the present invention provides an adaptive cutting fluid formulation optimization method and system based on deep learning. This method can not only automatically analyze the optimal cutting fluid ratio scheme under different cutting conditions, but also dynamically adjust the recommended scheme according to real - time monitoring data to cope with sudden changes or abnormal situations during the machining process.

[0014] This method mainly includes:

[0015] Firstly, a large amount of data needs to be collected from the actual machining environment, including but not limited to material types, cutting parameters (such as speed, feed rate, cutting depth), environmental conditions (temperature, humidity, etc.), and the composition and proportion of the currently used cutting fluid. After pre - processing steps such as cleaning and normalization, these data are used to train the model.

[0016] Furthermore, the dynamic modeling of the deep neural network includes the following steps:

[0017] Mathematical definition of the network topology structure: Assume that the network includes an input layer L1, hidden layers L2 / L3 / L4, and an output layer L5, and the number of nodes satisfies:

[0018] nL1 = 12, nL2 = 18, nL3 = 12, nL = 9, nL5 = 6;

[0019] Construction of the input vector:

[0020] X = [HB, v c , f z , T max , σ T , A vib (0.5 - 8kHz), E AE , Ra t-1 , VB t-1 , pH env , C bio , t op T ;

[0021] Where:

[0022] HB is the Brinell hardness, representing the hardness of the workpiece material and reflecting the ability of the material to resist plastic deformation.

[0023] v c is the cutting speed, representing the linear speed of the cutting edge of the tool relative to the workpiece. ​

[0024] f z is the feed per tooth, representing the feed distance of a single cutting edge per revolution of the tool.

[0025] T max is the maximum temperature, representing the peak temperature in the contact area between the tool and the workpiece during cutting.

[0026] σ T is the temperature gradient, representing the degree of dispersion of the temperature distribution in the cutting zone.

[0027] A vib is the vibration energy amplitude, representing the energy integral value of the tool vibration in the frequency band of 0.5 - 8 kHz.

[0028] E AE is the acoustic emission event count rate, representing the frequency of stress wave events released during plastic deformation or crack propagation of the material.

[0029] Ra t-1 is the surface roughness of the previous time step, representing the average surface roughness of the workpiece after the previous batch of machining.

[0030] pH represents the environmental pH value.

[0031] VB t-1 is the flank wear of the tool at the previous time step, representing the average width of the flank wear band of the tool.

[0032] C bio is a quantitative index for the anti - microbial degradation ability of the cutting fluid.

[0033] t op is the cumulative operation time, representing the cumulative usage duration of the cutting fluid since the last replacement.

[0034] Each parameter corresponds to specific physical meanings and control objectives, providing comprehensive input features for the neural network.

[0035] Furthermore, a neural network usually consists of an input layer, hidden layers (there can be multiple), and an output layer. Each layer contains several neurons, and the layers are connected by weights. Its structural characteristics are as follows:

[0036] The main function of the input layer is to receive external input signals and transfer them to the next layer (hidden layer). The number of neurons in the input layer is usually consistent with the dimension of the input features. The hidden layer is the core part of the neural network. It receives the signals transferred from the input layer and processes the signals through non - linear transformation. The activation function of the hidden layer is:

[0037] The hidden layer uses the improved LeakyReLU:

[0038]

[0039] The output layer uses Sigmoid to constrain the output range:

[0040]

[0041] (k = 0.5 is used to compress the output to 0 to 1)

[0042] Adopt the weight update algorithm corrected by the momentum factor:

[0043]

[0044] Where:

[0045] is the change amount of the weight from neuron i to j in the m-th layer at time step t.

[0046] is the error gradient of the j-th neuron in the m-th layer.

[0047] is the output value of the i-th neuron in the (m - 1)-th layer.

[0048] V total is the total loss function, representing the comprehensive performance index of the neural network and the control system.

[0049] V NN is the loss part of the neural network, representing the difference between the neural network weights and the target weights. It constrains the weight update of the neural network through the regularization term.

[0050] V control is the loss of the control part, representing the error of the control system performance.

[0051] W is the weight matrix of the current neural network.

[0052] W * is the target weight matrix (weights in the ideal state).

[0053] is the control gain coefficient, used to adjust the weights of the control part.

[0054] E perform is the control performance error, representing the difference between the control system output and the desired output.

[0055] is the predicted value of the k-th output (predicted by the control system or the neural network).

[0056] is the true value of the k-th output (actual measured value).

[0057] μ = 0.9 is the momentum factor.

[0058] η is the dynamic learning rate and is adjusted according to the following rules:

[0059]

[0060] Where:

[0061] is the gradient (derivative) of the loss function E with respect to the weight w. It represents the rate of change of the loss function in the direction of the weight w.

[0062] Error backpropagation calculation:

[0063] Define the composite loss function:

[0064] E = αE perform + βE cost + γE eco (5)

[0065] Where the performance error term:

[0066]

[0067] Economic cost term:

[0068]

[0069] n is the number of components in the cutting fluid;

[0070] λ m is the unit cost of the m-th cutting fluid component (such as extreme pressure agent, defoamer, etc.).

[0071] C m is the dosage of the m-th component.

[0072] E perform is the control performance error, representing the difference between the system output and the desired output. It is usually calculated by the squared error or other error metrics.

[0073] E T is the weight coefficient of the cost error, representing the importance of the cost error in the overall performance.

[0074] E eco is the ecological error, representing the difference between the impact of the system on the ecological environment and the target ecological impact. The ecological impact can include carbon emissions, resource consumption, etc.

[0075] α, β, and γ are the weight coefficients of each part.

[0076] Furthermore, seek the optimal balance among conflicting objectives such as machining quality, cost, and environmental protection, and conduct multi-objective optimization decision-making.

[0077] Pareto front solution:

[0078] Constraint handling technology:

[0079] Define the multi-objective optimization problem:

[0080]

[0081] The objective function J1 is the comprehensive optimization of machining quality.

[0082] The objective function J2 is to minimize the economic cost of the cutting fluid formulation.

[0083] For pH, a hard and soft constraint method is used for limitation.

[0084] Hard constraint: The penalty function method is used to handle the pH value range (0.3 ≤ pH ≤ 9.5), and a penalty coefficient of λ = 10 3 is applied when the predicted value exceeds the boundary

[0085] Soft constraint: The requirement of the total component sum = 100% is handled through slack variables, allowing a temporary deviation of ±0.5%.

[0086] Genetic algorithm implementation:

[0087] Chromosome structure:

[0088] Chr = [η, μ, α, β, γ, k sig ;

[0089] μ: Momentum factor (accelerate convergence and suppress oscillation);

[0090] α, β, γ: Multi-objective weight coefficients (balance performance, cost, and environmental protection);

[0091] k sig : Sigmoid function slope coefficient (adjust output sensitivity);

[0092] η ∈ [0.01, 0.1] Neural network learning rate (control the magnitude of weight update (too low learning rate leads to slow convergence, too high causes oscillation);

[0093] Fitness function

[0094]

[0095] Population initialization: 100 initial solutions are generated using Latin hypercube sampling to ensure uniform coverage of the design space.

[0096] Crossover operation: Simulated binary crossover (SBX) is performed on the [η, μ, α, β, γ] parameters, and the distribution index is set to 20.

[0097] Elite retention: The top 10% of individuals in each generation are retained and enter the next generation directly.

[0098] Furthermore, in order to improve the efficiency and effectiveness of the overall optimization, a high-precision real-time control differential equation model is established, and the multi-modal coupling dynamic equation is introduced as a feedback controller to adjust the actuator in real time to track the target value and compensate for the disturbance.

[0099] The output layer of the neural network takes the output result as the vertex of the dynamic equation: the target value of the network output component Y opt , change Y opt As the tracking target of sliding mode control, a multi-dimensional vector is constructed. For example, the state equation is rewritten as:

[0100] Let the state vector be

[0101] State vector S: Contains the concentration deviation (x1, x2) of the extreme pressure agent and the rust inhibitor, and the deviation change rate and historical accumulated deviation (integral term) to fully describe the system dynamics.

[0102] Time-varying system matrix:

[0103]

[0104] a 11 =0.5+0.1σ T ,a 22 =0.3+0.05v c (temperature gradient and cutting speed coupling term);

[0105] ω=2πf vib (Vibration main frequency coupling);

[0106] Control input matrix:

[0107]

[0108] External disturbance term:

[0109]

[0110] in:

[0111] F ext is the external disturbance vector, which represents the external influence on the system.

[0112] ΔT env is the change in ambient temperature.

[0113] ΔpH is the change in pH.

[0114] E AEIt is the acoustic emission event count rate, representing the frequency of stress wave events released during the plastic deformation or crack propagation of materials.

[0115] Random noise:

[0116]

[0117] Coupling term K g (Y opt -Y actual )

[0118] Where:

[0119] ξ(t) is the random noise, representing the random perturbation that varies with time in the system.

[0120] It is to indicate that the random noise follows a normal distribution with a mean of 0 and a variance of . σ ξ is the standard deviation of the random noise, used to describe the intensity of the noise.

[0121] K g = diag(k1,..., k6) is the gain matrix, k i = 0.5·sigmoid(W i (3) )(dynamically calculated from the weights of the output layer of the neural network), realizing the direct coupling of neural network prediction and control law.

[0122] W i (3) represents the weight parameter of the i-th neuron in the third layer of the output layer of the neural network;

[0123] Furthermore, online learning and dynamic correction specifically include:

[0124] After each batch of processing, update the network weight W:

[0125]

[0126] ΔW: The adjustment amount of the neural network weight.

[0127] The gradient of the loss function J with respect to the weight matrix W of the current neural network, indicating the optimization direction.

[0128] Through the gradient descent algorithm, make the cutting fluid formulation output by the neural network gradually approach the optimal value.

[0129] ||Y opt -Y actual || 2 : Tracking error term.

[0130] This tracking error term can directly optimize the prediction accuracy of the neural network, ensure that the predicted recipe conforms to the actual value, and adopt the form of squared error to impose a stronger penalty on larger deviations.

[0131] 0.1·MSE(Ra,VB): Quality penalty term.

[0132] Through the weight coefficient of 0.1, balance the priorities of machining quality and recipe tracking.

[0133] Prevent the model from only pursuing recipe accuracy and ignoring the actual machining effects (such as deteriorated surface roughness or excessive tool wear).

[0134] Furthermore, use the weight W of the output layer of the neural network (3) to dynamically adjust the control gain Kg to make the control strategy adapt to the working conditions:

[0135]

[0136] If W ij (3) >0 (the network determines that this component is critical), the gain k i increases to strengthen the control.

[0137] Take the state variable S of the dynamic equation as an additional input to the neural network:

[0138]

[0139] Where:

[0140] X T : The transpose of the original input vector.

[0141] S T : The transpose of the dynamic state variable vector.

[0142] Realize the real-time correction of the neural network prediction by the control process state.

[0143] Construct a unified Lyapunov function:

[0144]

[0145] Ensure the global stability of the neural network learning process and the control system through derivative constraints

[0146]

[0147] The left side term:

[0148]

[0149] Is the time derivative of the total Lyapunov function.

[0150] Square of the gradient norm

[0151] The negative sign indicates that the weight update direction always reduces the loss function J, ensuring that the neural network parameters converge to the optimal value.

[0152] Middle term: -η|s i |, absolute value of the sliding mode surface |s i | reflects the magnitude of the controller tracking error.

[0153] When selecting specific parameters, refer to Table 1:

[0154] Table 1. Parameter Selection and Influence

[0155]

[0156] The present invention also discloses an adaptive cutting fluid formulation optimization system for implementing the above method, including:

[0157] A multi-source sensing module, including a vibration sensor, an infrared thermal imager, and an acoustic emission detection device;

[0158] An edge computing device configured to run the deep neural network model and the optimization algorithm;

[0159] An actuator control unit, including a sliding mode controller and a proportional regulating valve group;

[0160] A data communication bus connects each module.

[0161] Furthermore, the architecture of the deep neural network model is:

[0162] The number of input layer nodes is 12, corresponding to a 12-dimensional feature vector;

[0163] The number of nodes in the first hidden layer is 18, using LeakyReLU activation;

[0164] The number of nodes in the second hidden layer is 12, using LeakyReLU activation;

[0165] The number of output layer nodes is 6, corresponding to the main component ratio of the cutting fluid.

[0166] Furthermore, the sliding mode controller is configured as:

[0167] The system matrix A(t) includes a temperature gradient coupling term a 11 = 0.5 + 0.1σ T , a cutting speed coupling term a 22 = 0.3 + 0.05v c and a vibration main frequency coupling term ω = 2πf vib , where f vibRepresents the main vibration frequency;

[0168] Disturbance compensation matrix F ext = diag(ΔT env , ΔE AE , ΔpH);

[0169] The control gain matrix Kg and the weights W of the output layer of the neural network (3) Are coupled through the sigmoid function.

[0170] The embodiments of the present invention have the following technical effects:

[0171] 1. Through the adaptive cutting fluid formulation optimization method based on deep learning, the present invention realizes the dynamic optimization and real-time regulation of the cutting fluid formulation, significantly improves the efficiency and quality of machining, and reduces production costs and environmental impacts at the same time.

[0172] 2. Through the neural network model with multi-parameter coupling, it can accurately predict the optimal cutting fluid formulation under different machining conditions, effectively improve the surface roughness of the workpiece, and reduce the tool wear rate.

[0173] 3. The present invention can monitor the fluctuations of material properties, the attenuation of tool states, and the disturbances of environmental parameters during the machining process in real time, and adjust the cutting fluid formulation in real time according to these dynamic changes.

[0174] 4. Through multi-objective optimization decision-making, the present invention realizes the minimization of the economic cost of the cutting fluid formulation on the premise of ensuring machining quality. At the same time, when optimizing the cutting fluid formulation, the present invention fully considers environmental protection factors, restricts the pH value range of the cutting fluid through the hard and soft constraint method, and ensures that the total composition meets the requirements, reducing environmental pollution. In addition, by optimizing the formulation, the usage amount of the cutting fluid is reduced, further reducing the emission of waste.

[0175] 5. The present invention introduces a high-precision real-time regulation differential equation model and a multi-modal coupling dynamics equation. As a feedback controller, it can adjust the actuator in real time to track the target value and compensate for disturbances. The experimental results show that the prediction accuracy of the model is high, and the relative error between the actual data and the calibrated data is controlled within 5%, which can effectively simulate the dynamic changes during the machining process. Description of the Drawings

[0176] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0177] Figure 1 This is a flowchart of an identity authentication and encrypted communication method for connected vehicles provided by an embodiment of the present invention. Detailed implementation manners

[0178] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope protected by the present invention.

[0179] The present invention discloses an adaptive cutting fluid formulation optimization method and system based on deep learning. The following further illustrates the present invention through specific embodiments. The experimental platform for this experiment is: equipped with a Kistler 9272 dynamometer, a FLIR A655sc infrared thermal imager, and a NI PXIe-1073 data acquisition system. The processing parameters are as follows:

[0180] Workpiece material: 45 steel (measured HB = 187 ± 12);

[0181] Tool model: Sandvik CoroMill 490 (VB wear amount 0.15 - 0.35 mm);

[0182] Cutting parameters: v c = 120 m / min, f z = 0.12 mm / z, cutting depth ap = 2 mm, pH = 8.7,

[0183] T max = 700 °C;

[0184] σ T = 50–200 °C / mm, A vib <1 m 2 / s 3 E AE = 70 events / s, Ra t-1 = 6.7 μm, VB t-1

[0185] = 0.5 mm, C bio is level 4, t op is 300 hours.

[0186] These specific parameters are used as the input values of the neural network, and the input vector is constructed as follows:

[0187] X = [HB, v c , f z , T max , σT , A vib (0.5 - 8 kHz), E AE , Ra t-1 , VB t-1 , pH env , C bio , t op T ;

[0188] The input vector enters the hidden layer, where the activation function of the hidden layer is:

[0189] The hidden layer uses the improved LeakyReLU:

[0190]

[0191] The output layer uses Sigmoid to constrain the output range:

[0192]

[0193] (k = 0.5 is used to compress the output to 0 to 1)

[0194] Adopt the weight update algorithm corrected by the momentum factor:

[0195]

[0196] Among them: μ = 0.9 is the momentum factor.

[0197] η is the dynamic learning rate, which is adjusted according to the following rules:

[0198]

[0199] For the BP neural network, by calculating the error of the output layer and then backpropagating the error to the hidden layer and the input layer, the weights and biases of the network are adjusted. Specifically, during the training process, the input data is first passed into the network to obtain the output result.

[0200] Define the composite loss function:

[0201] E = αE perform + βE cost + γE eco (23)

[0202] Among them, the performance error term:

[0203]

[0204] The economic cost term:

[0205]

[0206] ​Define the multi-objective optimization problem:

[0207]

[0208] Genetic Algorithm Implementation:

[0209] Chromosome structure:

[0210] Chr=[η,μ,α,β,γ,k sig ]

[0211] For specific parameter selection, refer to Table 2:

[0212] Table 2 Genetic algorithm chromosome structure parameters

[0213]

[0214] Fitness function

[0215]

[0216] In order to further improve the efficiency and effectiveness of the overall optimization, a high-precision real-time control differential equation model is established, and the multi-modal coupling dynamic equation is introduced as a feedback controller to adjust the actuator in real time to track the target value and compensate for the disturbance.

[0217]

[0218] The output layer of the P network takes the output result as the dynamic equation and sets the vertex: the target value of the network output component Y opt , change Y opt As the tracking target of sliding mode control, the state equation is rewritten as:

[0219] Let the state vector be

[0220] Time-varying system matrix:

[0221]

[0222] a 11 =0.5+0.1σ T ,a 22 =0.3+0.05v c (Coupling term of temperature gradient and cutting speed)

[0223] ω=2πf vib (Vibration main frequency coupling)

[0224] Control input matrix:

[0225]

[0226] External disturbance term:

[0227]

[0228] Random noise:

[0229]

[0230] Coupling term K g (Y opt -Y actual )

[0231] After each batch of processing, update the network weight W:

[0232]

[0233] Furthermore, use the weight W of the output layer of the neural network (3) to dynamically adjust the control gain Kg, so that the control strategy adapts to the working conditions:

[0234]

[0235] Construct a unified Lyapunov function:

[0236]

[0237] Through derivative constraints, ensure the global stability of the neural network learning process and the control system

[0238]

[0239] Construct the corresponding data calibration model according to formulas (19) to (38), and use MATLAB software to implement the calibration research of the model on the on-site monitoring data. The absolute error and relative error of the comparison between the cutting fluid performance processed by the algorithm model and the real data of the cutting fluid in the laboratory are shown in the following table. It can be seen from Table 3 that the relative error between the actual data and the calibrated data is controlled within 5%, which proves that the accuracy of the model is relatively high and it can efficiently predict the performance data of the cutting fluid. For details, see Table 3.

[0240] Table 3 Prediction results and error analysis

[0241] [[ID= fifty-two]] [[ID= fifty-three]] [[ID= fifty-four]]

[0242] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive cutting fluid formulation optimization method based on deep learning, characterized in that, It includes the following steps: Collect multi-dimensional working condition parameters of cutting machining; Input the working condition parameters into a deep neural network model to output an initial cutting fluid formulation ratio; Conduct multi-objective optimization decision-making on the initial formulation, and the decision-making process simultaneously optimizes the machining quality index and the economic cost index; Dynamically adjust the optimized formulation parameters through a sliding mode control algorithm, and the adjustment includes real-time correction of the cutting fluid component concentration.

2. The method according to claim 1, wherein The working condition parameters include the hardness HB of the workpiece material and the cutting speed v c , the feed per tooth f z , the highest temperature T in the cutting zone max , the temperature gradient σ T , the integral value A of the energy product of the tool vibration in the frequency band of 0.5 - 8 kHz vib (0.5 - 8 kHz), the acoustic emission event count rate E AE , the surface roughness Ra of the previous time step t-1 , the flank wear VB of the tool of the previous time step t-1 , the environmental pH value pH, the anti - microbial degradation ability C of the cutting fluid bio and the cumulative operation time t op .

3. The method according to claim 2, wherein The construction of the deep neural network model includes: The input layer receives a 12-dimensional feature vector: X = [HB, v c , f z , T max , σ T , A vib (0.5 - 8 kHz), E AE , Ra t-1 , VB t-1 , pH env , C bio , t op T ;​ The hidden layer performs a non-linear transformation using an improved LeakyReLU activation function, and the activation function satisfies: The output layer uses the Sigmoid function to constrain each component ratio to the interval [0, 1].

4. The method according to claim 2, wherein The multi-objective optimization decision-making includes: Among them, x i is the component cost coefficient, w i is the component weight coefficient, T safe is the safe temperature of the cutting tool material; n is the number of components in the cutting fluid; The objective function J1 is to comprehensively optimize the machining quality; The objective function J2 is to minimize the economic cost of the cutting fluid formulation; For pH, a method of soft and hard constraints is used for limitation; Use a genetic algorithm to solve the Pareto optimal solution set, and the chromosome encoding of the genetic algorithm includes the neural network learning rate η, the momentum factor μ, and the multi-objective weight coefficients α, β, γ.

5. The method according to claim 4, wherein The implementation of the genetic algorithm includes: Generate an initial solution through Latin hypercube sampling; Adopt simulated binary crossover operation; Retain the top 10% of the individuals with fitness in each generation to enter the next generation of iteration; The fitness function is designed as follows:

6. The method according to claim 1, wherein The dynamic adjustment includes: Construct a multi-dimensional state vector, including extreme pressure agents and rust inhibitors in the cutting fluid formulation; Dynamically adjust the control matrix element values according to the material hardness HB, and the control matrix elements satisfy b = 1.2 - 0.3exp(-0.1HB); After each batch of machining, update the neural network weights: J = ||Y opt - Y actual || 2 + 0.1·MSE(Ra, VB); where ΔW is the adjustment amount of the neural network weights; It is the gradient of the loss function J with respect to the weight matrix W of the current neural network, indicating the optimization direction; ||Y opt -Y actual || 2 is the tracking error term; 0.1·MSE(Ra,VB) is the quality penalty term.

7. The method according to claim 6, characterized in that, It also includes a stability control step: Construct a Lyapunov function By constraining the derivative of the function to satisfy to ensure system stability; Among them, V total is the total loss function, representing the comprehensive performance index of the neural network and the control system; is the time derivative of the total Lyapunov function V NN It represents the loss part of the neural network, indicating the difference between the neural network weights and the target weights; V control To represent the loss of the control section and indicate the error in the performance of the control system; W is the weight matrix of the current neural network; W * is the target weight matrix; To control the gain coefficient for adjusting the weight of the control part; is the square of the gradient norm, which measures the step size of the neural network weight update; |s i | is the absolute value of the sliding surface, reflecting the magnitude of the controller tracking error.

8. An adaptive cutting fluid formulation optimization system for implementing the method according to any one of claims 1-7, characterized in that, It includes: A multi-source sensing module, including a vibration sensor, an infrared thermal imager, and an acoustic emission detection device; An edge computing device configured to run the deep neural network model and the optimization algorithm; An actuator control unit, including a sliding mode controller and a proportional regulating valve group; A data communication bus connects each module.

9. The system according to claim 8, wherein The architecture of the deep neural network model is: The number of nodes in the input layer is 12, corresponding to a 12-dimensional feature vector; The number of nodes in the first hidden layer is 18, using LeakyReLU activation; The number of nodes in the second hidden layer is 12, using LeakyReLU activation; The number of nodes in the output layer is 6, corresponding to the main component ratio of the cutting fluid.

10. The system according to claim 8, wherein The sliding mode controller is configured such that the system matrix A(t) includes a temperature gradient coupling term a 11 = 0.5 + 0.1σ T , a cutting speed coupling term a 22 = 0.3 + 0.05v c and a main vibration frequency coupling term ω = 2πf vib , where f vib represents the main vibration frequency; Disturbance compensation matrix F ext = diag(ΔT env , E AE , ΔpH); where, F ext is an external disturbance vector, representing the external influence on the system; ΔT env is the change in ambient temperature; ΔpH is the change amount of the pH value; E AE is the acoustic emission event count rate, representing the stress wave event frequency released during the plastic deformation or crack propagation of the material; The control gain matrix Kg and the weights W of the output layer of the neural network (3) are coupled through the sigmoid function.

Images