Intelligent design method for large language model machining process fusing multi-modal industrial data

Abstract

The embodiment of the invention provides a multi-modal industrial data fused large language model machining process intelligent design method, which comprises the following steps: carrying out preprocessing and feature extraction on multi-modal industrial data to generate a standardized input sequence; constructing a semantic association relationship between the cutting parameters and the process performance based on a large language model; cutting parameters are adjusted based on the real-time cutting force and the vibration data and fed back to the model for iterative optimization; and an optimal processing scheme is generated and implemented in combination with a historical process database and a real-time monitoring result. Through the scheme of the embodiment of the invention, the problem of how to dynamically regulate and control the cutting parameters according to the real-time cutting force and vibration data collected by the multi-mode sensor so as to solve the problem of unbalanced tool wear in the machining process can be solved;

Description

TECHNICAL FIELD

[0001] The application relates to a large language model intelligent design method for machining process by fusing multi-modal industrial data. BACKGROUND

[0002] The application provides a large language model intelligent design method for machining process by fusing multi-modal industrial data, which aims to improve the understanding and decision-making ability of the large language model in the field of machining by integrating various types of industrial data such as images, texts, and sensor signals, so as to realize the intelligent design and optimization of the machining process. In view of the problem of how to dynamically adjust the cutting parameters according to the real-time cutting force and vibration data collected by multi-modal sensors to solve the problem of uneven tool wear, the method needs to analyze the real-time data and adaptively adjust the machining parameters to ensure the stability of the machining process and prolong the tool life. SUMMARY

[0003] Therefore, the application provides a large language model intelligent design method for machining process by fusing multi-modal industrial data, which at least partially solves the problems in the prior art.

[0004] A large language model intelligent design method for machining process by fusing multi-modal industrial data, comprising: preprocessing and feature extraction of multi-modal industrial data to generate standardized input sequences; constructing semantic association between cutting parameters and process performance based on a large language model; adjusting cutting parameters based on real-time cutting force and vibration data and feeding back to the model for iterative optimization; generating an optimal machining scheme by combining a historical process database and real-time monitoring results and implementing the scheme.

[0005] Preferably, the adjustment of cutting parameters based on real-time cutting force and vibration data further comprises: sliding window mean calculation of the real-time collected cutting force signal F(t) to obtain the average cutting force F_avg; Fourier transform of the vibration acceleration signal a(t) to obtain the frequency energy E_freq; using the formula T = k1 F_avg + k2 E_freq to judge the current machining state, wherein k1 and k2 are adjustment coefficients, F_avg is the average cutting force, and E_freq is the energy of high-frequency vibration; if T>T_threshold, the feed speed f is corrected downward by 10%.

[0006] Preferably, the adjusting the cutting parameters based on the real-time cutting force and vibration data further comprises: synchronously collecting the cutting force peak P_peak and the maximum vibration acceleration A_max at the current time from the multi-modal sensor; establishing a cutting state classification rule: if P_peak>P_upper and A_max>A_upper, it is considered to be a serious wear state; based on the following condition judgment formula: C = (P_peak / P_nom) (A_max / A_nom), where P_nom is the upper limit of normal cutting force, and A_nom is the upper limit of normal vibration acceleration; if C>5, automatically switch to the speed reduction strategy, and start the tool wear warning mechanism.

[0007] Preferably, the adjusting according to the cutting state further comprises: adaptively filtering the cutting force waveform of the continuous n cycles to obtain a smooth cutting force sequence F_n after removing noise interference; defining an abnormal state based on the kurtosis coefficient K_f of the cutting force sequence: K_f>K_upper is considered to be abnormal fluctuation; adjusting the parameters through a dynamic weight function W = α (F_n F_mean) + β (ΔF), wherein α and β are weight coefficients, F_mean is the current average cutting force, and ΔF is the cutting force variation amplitude; when W>W_threshold, reduce the cutting speed v and adjust the tool deflection angle θ.

[0008] Preferably, the adjusting according to the dynamic weight function further comprises: comparing the current cutting force F_curr with the optimal cutting force F_ref in the historical similar process to obtain the cutting force difference ΔF = |F_curr F_ref|; calculating the main frequency proportion r_freq = A_freq_high / A_freq_total of the amplitude spectrum A_amp; using a decision tree structure to judge the process adaptability: if ΔF<ΔF_threshold and r_freq<r_upper, the original parameters are allowed to be maintained; otherwise, introduce a fuzzy control algorithm to generate a compensation instruction for adjusting the spindle speed ω and the feed rate f.

[0009] Preferably, the determining process adaptability based on the decision tree structure further comprises: collecting time step Δt and temperature field distribution T_grid of the current process; calculating process stability index S_stable = 1 (σ_F / μ_F) (σ_a / μ_a), wherein σ_F is the standard deviation of cutting force, μ_F is the mean value, σ_a is the standard deviation of vibration acceleration, and μ_a is the mean value; constructing a conditional formula: if S_stable < S_threshold and Δt > t_window, triggering dynamic compensation logic; At this time, the least square method is used to fit the cutting force change trend, and the compensation coefficient γ = ΔF / (T_grid × v) is obtained.

[0010] Preferably, the introducing the least square method to fit the cutting force change trend further comprises: selecting cutting force F_i (i = 1,...,m) in the last m cycles as input data; defining the model function as F_i = β_0 + β_1 t_i + β_2 t_i², wherein t_i represents the cycle sequence; solving the regression coefficients β_0, β_1 and β_2 by the least square method; judging the cutting force trend according to whether β_2 is greater than zero, if β_2 > 0 and β_1 > 0, executing the feed rate reduction operation, otherwise keeping the current process scheme.

[0011] Preferably, the prediction analysis based on the model function further comprises: comparing the fitted trend curve with the target cutting force F_target, calculating the error e_t = |F_trendF_target|; setting the tolerance range ε, when e_t > ε and the error growth rate Δe_t / Δt > 0, triggering the secondary optimization process; using genetic algorithm to optimize the cutting speed v and the feed rate f, the objective function is minimize [α e_t + β (v v_initial)], wherein α and β are weight coefficients; if the optimization result meets the constraint condition, updating the current parameter group, and writing the new process scheme into the process database.

[0012] Preferably, the optimizing the cutting parameters based on the genetic algorithm further comprises: Initialize the parameter coding of each individual in the population as [v, f, θ]; Set selection pressure P_sel, crossover rate p_c, mutation rate p_m; Evaluate the fitness value based on the penalty function, and give a larger penalty value if the parameters do not meet the physical constraints or equipment limitations; Select excellent individuals to enter the next round of evolution, and record the best scheme as the optimal solution of this process.

[0013] Preferably, the evaluation of individual advantages and disadvantages based on the penalty function further comprises: Define the physical constraint conditions: v_min ≤ v ≤ v_max, f_min ≤ f ≤ f_max, θ_min ≤ θ≤ θ_max; If the constraint range is exceeded, the fitness value is reduced by a penalty coefficient p_pen; Add an energy consumption coefficient C_energy = γ v The weighted term of f, and γ is the energy loss coefficient; Through the mixed selection strategy combined with the elite reservation strategy, the optimal solution is ensured to be continuously converged to the feasible region.

[0014] The embodiment of the present disclosure provides a large language model intelligent design method for machining process based on fusion of multi-modal industrial data, comprising: preprocessing and feature extraction of multi-modal industrial data to generate standardized input sequences; constructing semantic association relationship between cutting parameters and process performance based on large language model; adjusting cutting parameters based on real-time cutting force and vibration data, and feeding back to the model for iterative optimization; combining historical process database and real-time monitoring results to generate optimal machining scheme and implement. Through the scheme of the embodiment of the present disclosure, the problem of 1. How to dynamically control the cutting parameters according to the real-time cutting force and vibration data collected by the multi-modal sensor to solve the problem of uneven tool wear in the machining process. BRIEF DESCRIPTION OF DRAWINGS

[0015] In the drawings, like reference numerals refer to same or similar components throughout the several views. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments in accordance with the present disclosure and should not be considered limiting the scope of the present disclosure.

[0016] Figure 1 is a flowchart of a large language model intelligent design method for machining process based on fusion of multi-modal industrial data; Figure 2is a further flowchart of adjusting cutting parameters based on real-time cutting force and vibration data; Figure 3 is a further flowchart of adjusting cutting parameters based on real-time cutting force and vibration data; Figure 4 is a further flowchart of adjusting based on cutting state; Figure 5 is a further flowchart of adjusting based on dynamic weight function; Figure 6 is a further flowchart of judging process adaptability based on decision tree structure; Figure 7 is a further flowchart of introducing least square method to fit cutting force variation trend; Figure 8 is a further flowchart of predicting analysis based on model function; Figure 9 is a further flowchart of optimizing cutting parameters based on genetic algorithm; Figure 10 is a further flowchart of evaluating individual based on penalty function; DETAILED DESCRIPTION In the following, only certain example embodiments are simply described. As those skilled in the art will recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. The drawings and description are therefore to be considered in an illustrative, rather than a restrictive, sense.

[0017] Next, referring to the accompanying drawings, a large language model-based intelligent design method for fusing multi-modal industrial data is described. This method fuses data from various information sources such as sensors, historical process databases, and operator experience to build an intelligent design system for machining processes. The multi-modal industrial data is preprocessed and feature extracted to generate standardized input sequences. First, a large amount of raw data is collected from cutting force sensors, vibration sensors, temperature sensors, and other devices, including but not limited to time series data, image information, and process record documents. Through denoising, normalization, and missing value filling, etc. preprocessing operations, redundant information and inconsistencies are eliminated, and data quality is improved. At the same time, in the feature extraction stage, spectral analysis, time domain statistical analysis, and image feature extraction algorithms are used to extract key features that can effectively reflect the processing state, and different modal data is converted into a sequence format that the model can understand. For example, a certain enterprise collects hundreds of sets of actual machining data through cutting force and vibration sensors installed on the numerical control machine tool, and standardizes the maximum cutting force, main frequency vibration frequency, and temperature difference in each set of data, generating training samples suitable for large language models. This approach not only improves data consistency, but also provides a solid foundation for subsequent semantic association construction.

[0018] Based on the large language model, the semantic association between cutting parameters and process performance is constructed. This step relies on deep learning and natural language processing techniques, especially the ability to handle structured and unstructured data. The large language model is used to mine and express the complex relationship between cutting parameters (such as speed, feed rate, cutting depth) and process performance indicators (such as surface roughness, tool life, energy consumption). This is not just a simple regression analysis, but a multi-dimensional mapping closer to human understanding logic. Through learning from massive data, the model can establish a concept association similar to a semantic network, so that the selection of cutting parameters is no longer a single reference to the process manual, but a comprehensive judgment that dynamically adapts to the current machine tool state and material characteristics. In one embodiment, researchers used a BERT model fine-tuned on industry datasets to predict the quality performance of workpieces under different cutting conditions, and adjusted the weight of input word vectors to reflect the interaction between parameters. The results show that the model has high accuracy in predicting machining quality.

[0019] Based on real-time cutting force and vibration data, the cutting parameters are adjusted and fed back to the model for iterative optimization. The core of this process lies in the real-time perception ability of the system and the feedback closed-loop mechanism. When the system receives new cutting data, it will immediately embed it into the trained large language model, quickly evaluate the effectiveness of the existing parameters and suggest the best modification scheme. For example, in a specific manufacturing workshop scenario, an abnormally high vibration level occurs during the cutting process of the machine tool. At this time, the system detects this data, automatically triggers the cutting force and vibration analysis module, and inputs the new parameters into the model to generate a more optimal combination of feed rate and cutting speed. Subsequently, the system can issue the adjusted parameters to the control device, while recording this decision as supplementary data for subsequent optimization, forming a continuously improving closed loop. In addition, the model can also find potential trends and give early warnings of possible tool wear based on the comparison of historical data and real-time parameters, avoiding production losses caused by unexpected downtime.

[0020] The optimal processing scheme is generated and implemented in combination with the historical process database and real-time monitoring results. This stage integrates information from historical experience and current conditions, uses the recommended scheme output by the model, and considers the relationship between cost, efficiency, and quality to select the most suitable operation strategy. At the same time, since the large language model has the ability to understand and reason, the system can automatically generate detailed operation instructions or suggestion texts for technical personnel to refer to. Specifically, during the implementation of this technology by an automobile parts manufacturer, a batch of parts prone to tool wear was identified through the comparison of historical data and current processing status, and the process path was actively adjusted, changing the original high-speed processing mode to a segmented slow-speed processing mode, significantly extending the tool life. Ultimately, the factory achieved higher one-time pass rate and significantly reduced tool replacement frequency and maintenance costs. Such system deployment not only enhances the stability and controllability of the machining process, but also further promotes the development of intelligent manufacturing towards a data-driven direction, bringing considerable economic benefits and technological progress space to enterprises.

[0021] Next, the adjustment of cutting parameters based on real-time cutting force and vibration data of the present application is further described. First, the sliding window mean calculation is performed on the real-time collected cutting force signal F(t) to obtain the average cutting force F_avg. This step can effectively suppress noise interference and more accurately reflect the trend of the current processing state by averaging the cutting force data within a short period. For example, in a turning process, if the cutting force fluctuates greatly, the sliding window processing can obtain a stable average value, improving the reliability of the data.

[0022] Secondly, the Fourier transform is performed on the vibration acceleration signal a(t) to obtain the frequency domain energy E_freq. The Fourier transform converts the time domain signal into the frequency domain, which can extract the main frequency components and energy distribution of the vibration signal, and is helpful to identify abnormal conditions such as tool wear or workpiece loosening. In an embodiment, by analyzing the energy change in the high frequency band, the adverse development trend of the tool state can be discovered early.

[0023] Subsequently, the formula T = k1 F_avg + k2 E_freq is used to determine the current machining state, where k1 and k2 are adjustment coefficients, and the recommended value is between 0.5 and 1.5 to balance the influence weight of cutting force and vibration. F_avg represents the average cutting force, with the unit of Newton; E_freq represents the energy of high-frequency vibration, with the unit of volt square. T is a comprehensive score index for evaluating the machining stability. If T>T_threshold, it means that the current state is unstable, and the parameters need to be adjusted.

[0024] Finally, if T>T_threshold, the feed speed f is adjusted downward by 10%. This step can effectively alleviate the mechanical impact caused by excessive cutting, improve the machining surface quality, and reduce tool wear. Specifically, when machining high-strength metal materials, the system realizes the optimization and stability guarantee of the cutting process by dynamically adjusting the feed speed, significantly improving the overall process efficiency and equipment life.

[0025] Next, the present application based on real-time cutting force and vibration data to adjust the cutting parameters is further described. First, the cutting force peak P_peak and the maximum vibration acceleration A_max at the current time are synchronously collected from the multi-modal sensor. This step is used to obtain real-time dynamic data in the machining process, providing basic information for subsequent analysis. For example, in an embodiment, by installing strain gauges and accelerometers on the machine tool spindle and the workpiece, the physical changes of the tool during the cutting process are captured in real time.

[0026] Then the cutting state classification rule is established: if P_peak>P_upper and A_max>A_upper, it is considered to be a serious wear state. This rule is used to identify abnormal conditions that occur during machining, where P_upper and A_upper are threshold values set according to the limits of the tool when it is working normally. Specifically, P_upper is recommended to be set to 1.2 times the upper limit of the normal cutting force P_nom, and A_upper is 1.5 times A_nom, to avoid misjudgment.

[0027] Then the following condition judgment formula is used: C = (P_peak / P_nom) The formula (A_max / A_nom) represents the upper limit of normal cutting force, typically a stable value within a reasonable range, such as 200N; A_nom represents the upper limit of normal vibration acceleration, generally set to 0.8g. The purpose of this formula is to comprehensively consider the deviations of both cutting force and vibration dimensions. A larger C value indicates a more unstable state. The optimal value is set to 5; exceeding this value indicates that the system is in a potentially risky state.

[0028] Finally, if C > 5, the system automatically switches to a speed reduction strategy and activates a tool wear warning mechanism. This effectively prevents decreased machining quality or equipment damage caused by excessive tool wear. Specifically, when the C value exceeds the limit, the system automatically reduces the spindle speed and issues an alarm signal so that operators can handle the situation promptly. This real-time adjustment mechanism significantly improves production efficiency and safety, and reduces downtime and tool wear.

[0029] Next, the adjustment based on the cutting state of the present invention will be further described. Adaptive filtering is performed on the cutting force waveform of n consecutive cycles to remove noise interference, resulting in a stable cutting force sequence F_n. The purpose is to extract the true cutting force data to avoid the influence of environmental or equipment noise on subsequent judgments. This process utilizes algorithms to identify and remove outliers, ensuring the accuracy and reliability of the data analysis results. For example, in machining, sensors may be subject to external vibration interference; adaptive filtering can effectively improve the purity of the cutting force signal.

[0030] An abnormal state is defined based on the kurtosis coefficient K_f of the cutting force sequence: when K_f is greater than K_upper, abnormal fluctuations are considered to have occurred, and data distributions with excessively high peak values ​​indicate instability during the cutting process. This method can quickly identify anomalies and trigger further adjustment mechanisms. In one embodiment, K_upper is set to 3.5 to represent the normal range, and when the calculated K_f reaches or exceeds this value, the system determines that an anomaly exists.

[0031] Through the dynamic weighting function W = α (F_n - F_mean) + β (ΔF) is used for parameter adjustment, where α and β are weighting coefficients, typically ranging from 0.1 to 1.0, with 0.6 and 0.4 being preferred. F_mean is the current average cutting force, and ΔF is the variation in cutting force. The formula integrates absolute error and relative change, making the adjustment more comprehensive and precise. Specifically, if the cutting force fluctuates drastically, the formula will amplify its weight, indicating that early intervention and control are necessary.

[0032] When W exceeds W_threshold, the cutting speed v is reduced and the tool deflection angle θ is adjusted to reduce material wear and improve machining quality. For example, under high cutting force, appropriate speed reduction and angle change can relieve pressure and maintain stable machining. This technical solution can improve the intelligence and adaptability of the system, reduce the failure rate, and optimize the process design efficiency.

[0033] Next, the adjustment according to the dynamic weight function of the present application is described. First, compare the current cutting force F_curr with the optimal cutting force F_ref in the historical similar process to get the cutting force difference ΔF = |F_curr - F_ref|. Where F_curr represents the actual cutting force measured in the current process, unit Newton (N), usually between 500 to 1500 N, F_ref represents the maximum stable cutting force recorded when successfully machining this type of workpiece before. The value of ΔF reflects the deviation between the current parameters and the ideal state. The setting of this formula aims to judge whether the process is within a reasonable range by quantifying the deviation.

[0034] Next, calculate the main frequency proportion r_freq = A_freq_high / A_freq_total of the amplitude spectrum A_amp. A_freq_high represents the energy intensity of the main frequency component in the amplitude spectrum, and A_freq_total is the total energy of all frequency components. The larger r_freq indicates the higher concentration of system vibration, which may cause instability. This parameter is used to evaluate the vibration level in the machining process and help identify factors that may cause abnormalities.

[0035] Then, use the decision tree structure to judge the process adaptability: if ΔF<ΔF_threshold and r_freq<r_upper, allow to maintain the original parameters; otherwise, introduce a fuzzy control algorithm to generate compensation instructions for adjusting the spindle speed ω and feed rate f. ΔF_threshold is set according to material properties and equipment performance, generally between 100-200 N, while r_upper is usually set to 0.3 to 0.4. Exceeding this threshold indicates a risk that needs to be adjusted.

[0036] For example, when machining aluminum alloy parts, suppose the system detects that ΔF reaches 180 N and r_freq is 0.45, both exceeding the preset threshold. At this time, the fuzzy controller will automatically reduce the spindle speed and appropriately reduce the feed rate to stabilize the machining process and ensure product quality. This solution improves process stability, reduces the risk of abnormal downtime, and optimizes machining efficiency, with obvious practical application value.

[0037] Next, the decision tree structure-based process adaptability judgment of the present application is further described. First, the time step Δt and temperature field distribution T_grid of the current process are collected. This step is used to obtain key dynamic parameters in the machining process, where Δt represents the sampling interval time, usually ranging from 0.01 to 0.1 seconds, with an optimal value of 0.05 seconds, and T_grid reflects the thermal distribution state of the workpiece or tool surface, usually covering a range of 20°C to 80°C. For example, recording the temperature distribution at each time during cutting can be used to analyze the influence of temperature change on material removal rate.

[0038] Second, the process stability index S_stable = 1 / (σ_F / μ_F) (σ_a / μ_a) is calculated, where σ_F and μ_F are the standard deviation and mean of cutting force, and σ_a and μ_a are the standard deviation and mean of vibration acceleration. This formula is used to quantify the stability of the machining process, and when the ratio of standard deviation to mean is small, the data is more concentrated and the system is more stable. For example, in the steady cutting state, σ_F / μ_F is close to 0.1 and σ_a / μ_a is close to 0.05, and at this time S_stable is greater than the threshold value.

[0039] Third, the conditional formula is constructed: if S_stable < S_threshold and Δt > t_window, the dynamic compensation logic is triggered. S_threshold is set according to experience, usually taking 1.5; t_window can be set to 0.05 seconds to ensure that the compensation mechanism is activated only in the non-steady state. For example, when large fluctuations are detected, the compensation logic is immediately enabled.

[0040] Finally, the least squares method is used to fit the cutting force variation trend to obtain the compensation coefficient γ = ΔF / (T_grid × v), where ΔF is the cutting force variation and v is the cutting speed. The purpose of this formula is to adjust the compensation coefficient according to the temperature rise and cutting parameters to improve cutting efficiency and quality.

[0041] This technical solution can evaluate the process adaptability in real time, improve the machining stability, and optimize the process design parameters.

[0042] Next, the introduction of the least squares fitting method to fit the cutting force trend is described further. The cutting force F_i (i=1,...,m) in the last m period is selected as the input data to capture the change rule of the cutting force. Here F_i represents the cutting force value measured at different time points, with units of Newton (N), and the range varies according to the equipment capacity, preferably m=20 to ensure the sufficiency of the data. For example, during the machining process, the cutting force data of the past three weeks is obtained through the sensor.

[0043] The model function is defined as F_i = β_0 + β_1 t_i + β_2 t_i², where t_i represents the cycle sequence, indicating the i-th time node, generally starting from 1 and increasing, such as 1, 2, 3...m, to reflect the trend of change over time. β_0 is the constant term, representing the initial cutting force level; β_1 is the linear term coefficient, reflecting the rate of change; β_2 is the quadratic term coefficient, used to measure the curve trend. The formula is set because the cutting force may change nonlinearly, and the quadratic model can more accurately describe its development path.

[0044] The regression coefficients β_0, β_1, β_2 are solved by the least squares method, so that the squared error sum of the model output and the actual cutting force is minimized. This optimization method can provide stable and reliable parameter estimates, ensuring the accuracy of the model prediction. For example, in a certain processing, the system calculates β_2=0.3, β_1=1.5, indicating that the cutting force shows an upward trend and there is an overload risk.

[0045] According to whether β_2 is greater than zero, if β_2>0 and β_1>0, the feed rate reduction operation is performed, otherwise the current process scheme is retained. This decision mechanism aims to avoid equipment damage or processing precision decline due to continuously increasing cutting force, and improve production safety and efficiency. Specifically, when β_2=0.4 and β_1=2.1 are detected, the system automatically reduces the feed speed to keep the process in a stable running state.

[0046] This technical solution uses mathematical modeling to dynamically analyze the cutting force, realizes intelligent adjustment, improves process stability and prolongs equipment life, and at the same time enhances the comprehensive processing capability of the large language model for multi-source data.

[0047] Next, the model function-based prediction analysis of the present application is further described. First, compare the fitted trend curve with the target cutting force F_target, calculate the error e_t = |F_trend - F_target|. Where F_trend represents the cutting force prediction value obtained by the model function, F_target is the target cutting force expected to be achieved in process design, e_t is the absolute deviation between the actual value and the target value. This step measures the difference between the model output and the expectation by quantifying the error, thereby providing feedback information to support subsequent optimization operations. For example, during a machining process, the system detects that the deviation between F_trend and F_target exceeds the set range.

[0048] Secondly, set the tolerance range ε, when e_t>ε and the error growth rate Δe_t / Δt>0, trigger the secondary optimization process. ε is the maximum error threshold allowed, generally set in the range of 5% to 15%, depending on the material properties and equipment precision. Δe_t / Δt represents the rate of change of error with time, if it continues to increase, it means that the model adaptation ability decreases, and the parameters need to be adjusted again. For example, in continuous machining, it is observed that the error continues to rise, that is, the optimization mechanism is activated to prevent process failure.

[0049] Next, use genetic algorithm to optimize cutting speed v and feed rate f, the objective function is minimize [α e_t+ β (v - v_initial)], where α and β are weight coefficients, usually taken between 0.5 and 2.0 to balance error and process stability. v_initial represents the initial cutting speed, the purpose is to control the parameter deviation during optimization, avoid the risk brought by extreme value. For example, when milling steel, adjust v and f to ensure that the error is reduced while ensuring efficiency and safety.

[0050] Finally, if the optimization result meets the constraint condition, update the current parameter set, and write the new process scheme into the process database. This process makes the process design continuously learn and accumulate the best practice, and improves the production consistency. For example, the system archives the newly obtained optimal parameter combination, providing data reference for subsequent machining of similar parts.

[0051] This technical solution effectively improves the process design precision and self-adaptation ability, reduces the cost of human intervention and trial and error, and enhances the intelligent level of the machining process.

[0052] Next, the genetic algorithm-based optimization of cutting parameters of the present application is further described. First, the parameter coding of each individual in the population is initialized as [v, f, θ], where v represents the cutting speed (unit: m / min), usually between 50-200; f represents the feed rate (unit: mm / r), generally taking a value of 0.1-0.3; θ represents the cutting depth (unit: mm), usually ranging from 1-5. This coding is used to represent different combinations of cutting parameters and is given genetic information to participate in subsequent evolution. For example, in a certain machining process, set the initial value of v as 100, f as 0.2, and θ as 2 as part of the initial population, which is used to simulate possible processing schemes.

[0053] The selection pressure P_sel, the crossover rate p_c, and the mutation rate p_m are set, P_sel controls the selection probability of superior individuals, which is usually set to 1.5-2.0; p_c controls the probability of gene exchange, which is usually taken as 0.8-0.9; p_m controls the random disturbance probability, which is set between 0.01-0.1. These parameters control the convergence speed and search ability of the algorithm, so that it can quickly find better solutions in a multi-dimensional parameter space. In one embodiment, if the selection pressure is large, it is beneficial to retain superior individuals and improve optimization efficiency; otherwise, it increases diversity and prevents premature convergence.

[0054] The fitness value is evaluated based on the penalty function, and if the parameters do not meet the physical constraints or equipment limitations, a larger penalty value is given. For example, when machining aluminum alloy, if the set cutting speed exceeds the maximum allowable range of the equipment, or the feed rate exceeds the carrying capacity of the machine tool, this parameter combination will be given a high penalty value, reducing the possibility of being selected. This helps to ensure that the generated parameters meet the actual working conditions and improve the feasibility of the optimization results.

[0055] The best individual is selected to enter the next round of evolution, and the best scheme is recorded as the optimal solution of this process. Through the iterative process, the parameter combination is gradually improved, and finally the most optimal machining scheme in terms of energy consumption, machining precision, production efficiency, etc. is found. Specifically, after a certain process optimization is completed, the system can output the optimal cutting parameters, significantly improving the machining quality and equipment utilization, reducing the scrap rate and energy consumption, and improving the overall intelligent manufacturing level.

[0056] Next, the evaluation of individuals based on the penalty function of the present application is further described. First, define the physical constraints: v_min ≤ v ≤ v_max, f_min ≤ f ≤ f_max, θ_min ≤ θ ≤ θ_max. This step limits the reasonable range of individual parameters by setting the upper and lower limits of cutting speed (v), feed rate (f) and cutting depth (θ). Among them, v_min and v_max represent the minimum and maximum cutting speed allowed, generally between 50 to 200 m / min; f_min and f_max represent the minimum and maximum feed rate, usually between 0.05 to 0.3 mm / r; θ_min and θ_max represent the minimum and maximum cutting depth, commonly ranging from 0.1 to 3 mm. The selection of these parameters is determined according to the machining material and tool performance to prevent the generation of unrealistic or dangerous process parameters. For example, in one embodiment, for the machining of aluminum alloy materials, the cutting speed is set to 80 to 150 m / min to balance efficiency and tool life.

[0057] Second, if the constraints are exceeded, the fitness value is reduced by a penalty coefficient p_pen. This step aims to reduce the overall fitness of the out-of-range parameters by penalizing them, guiding the optimization process towards a reasonable and feasible solution space. p_pen is a fixed value set by humans to reflect the degree of negative effect after exceeding the limit, usually set to a range of 10~50, which can be adjusted according to the strength of different constraints. For example, when the feed rate selected by an individual is 0.4 mm / r, exceeding the default maximum value, the fitness value will be reduced by p_pen, making it relatively inferior and reducing the probability of being selected.

[0058] Third, add the energy consumption coefficient C_energy = γ v f weighted term to the fitness evaluation, where γ is the energy loss coefficient. The meaning of this formula is to introduce energy consumption as another consideration factor to improve the economic and feasibility of the overall optimization result. γ is usually taken as 0.01~0.1, set according to the energy consumption characteristics of the machine tool, the higher the value, the greater the energy loss per unit power. For example, in the design of a high-speed numerical control milling machine, set γ=0.05, so that high cutting speed and large feed rate bring higher energy consumption, thereby encouraging the algorithm to find energy-saving combinations.

[0059] Finally, a hybrid selection strategy combined with an elite preservation strategy is used to ensure that the optimal solution does not converge to the feasible region. The hybrid selection strategy combines the advantages of roulette selection and ranking selection to improve search diversity, while the elite preservation strategy directly preserves the best individuals in the population, enhancing the stability and convergence speed of the algorithm. This combination of strategies effectively balances exploration and exploitation, making the optimization process more efficient and accurate. For example, in the generation of multi-modal data fusion machining schemes, after multiple iterations, the model can quickly lock in the optimal process parameters that meet the requirements of machining quality, time, and energy consumption. This technical solution has the beneficial effects of improving process design rationality, ensuring machining safety, and reducing production costs.

[0060] The method for intelligent design of machining process of a large language model fused with multi-modal industrial data includes: first, collecting and preprocessing industrial data from multi-modal sensors, converting complex and diverse data into input sequences of a unified format through feature extraction and standardization processing; then, based on the strong semantic understanding ability of the large language model, a deep semantic association between cutting parameters and process performance is established, so that the system can understand and predict the influence of different cutting parameter combinations on machining quality; then, combined with real-time collected cutting force and vibration data, the cutting parameters are dynamically adjusted, so that the model has the ability of self-adaptive adjustment; finally, the knowledge in the historical process database and the real-time monitoring results are used to generate and optimize the final machining scheme, and the scheme is applied to actual production.

[0061] In solving the technical problem of how to dynamically regulate cutting parameters based on real-time cutting force and vibration data collected by multi-modal sensors to solve the problem of uneven tool wear in the machining process, the present application adopts a method of real-time data analysis and model feedback iteration. By high-frequency collection and analysis of real-time cutting force and vibration signals, the system can determine whether the tool is in an abnormal wear state or under uneven stress during the current machining process. Using the strong semantic association ability of the large language model, the system can not only identify the similarity between the current machining conditions and past experiences, but also predict better cutting parameter settings to reduce local overload and uneven wear. At the same time, the system feeds the adjusted parameters back to the model as new inputs, forming a closed-loop optimization mechanism to continuously improve the accuracy of model prediction and control, thereby achieving balanced control of tool wear and improving machining efficiency and product quality.

[0062] The method, program, system, device, etc. of the embodiments of the present application can be executed or implemented in a single or multiple networked computers, or can be practiced in a distributed computing environment. In the embodiments of the present application, in these distributed computing environments, tasks can be performed by remote processing devices connected through a communication network.

[0063] Those skilled in the art should understand that the embodiments of the present specification can be provided as a method, a system or a computer program product. Therefore, those skilled in the art can conceive that the implementation of the functional modules / units or controllers and the related method steps illustrated by the above-mentioned embodiments can be realized by software, hardware or a combination of software and hardware.

[0064] Unless explicitly stated, the actions or steps of the methods, programs recited in the embodiments of the present application do not have to be performed in a specific order and still achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0065] In this document, multiple embodiments of the present application are described, but for the sake of brevity, the description of each embodiment is not exhaustive and identical or similar features or parts between different embodiments can be omitted. In this document, "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means applicable to at least one embodiment or example according to the present application, but not all embodiments. The above terms do not necessarily mean referring to the same embodiment or example. Those skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of different embodiments or examples without contradiction.

[0066] The exemplary systems and methods of the present application have been specifically illustrated and described herein through the use of the above-mentioned embodiments, which are merely examples of the best modes of implementing the systems and methods. Those skilled in the art can understand that various changes can be made to the embodiments of the systems and methods described herein without departing from the spirit and scope of the present application defined in the appended claims.

Claims

1. A method for intelligent design of machining processes using a large language model that integrates multimodal industrial data, characterized in that: include: Multimodal industrial data is preprocessed and features are extracted to generate standardized input sequences; Semantic relationships between cutting parameters and process performance are constructed based on a large language model; The cutting parameters are adjusted based on real-time cutting force and vibration data, and then fed back to the model for iterative optimization. The optimal processing plan is generated and implemented by combining historical process databases and real-time monitoring results.

2. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 1, is characterized in that... The adjustment of cutting parameters based on real-time cutting force and vibration data further includes: The average cutting force F_avg is obtained by calculating the mean of the real-time acquired cutting force signal F(t) using a sliding window method. The frequency domain energy E_freq is obtained by performing a Fourier transform on the vibration acceleration signal a(t); Using the formula T = k1 F_avg + k2 E_freq determines the current machining state, where k1 and k2 are adjustment coefficients, F_avg is the average cutting force, and E_freq is the energy of the high-frequency vibration. If T > T_threshold, then adjust the feed rate f downwards by 10%.

3. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 1, is characterized in that... The adjustment of cutting parameters based on real-time cutting force and vibration data further includes: The peak cutting force P_peak and the maximum vibration acceleration A_max at the current moment are synchronously acquired from the multimodal sensor. Establish a cutting state classification rule: if P_peak > P_upper and A_max > A_upper, then it is considered a severe wear state; The formula is based on the following condition: C = (P_peak / P_nom) (A_max / A_nom), where P_nom is the upper limit of normal cutting force and A_nom is the upper limit of normal vibration acceleration; If C > 5, the tool wear warning mechanism will be automatically switched to a speed reduction strategy and activated.

4. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 3, is characterized in that... The adjustment based on the cutting state further includes: An adaptive filter is applied to the cutting force waveform for n consecutive periods to remove noise interference, resulting in a stable cutting force sequence F_n. An abnormal state is defined based on the kurtosis coefficient K_f of the cutting force sequence: when K_f > K_upper, abnormal fluctuations are considered to have occurred. Through the dynamic weighting function W = α (F_n F_mean) + β (ΔF) is used to adjust parameters, where α and β are weighting coefficients, F_mean is the current average cutting force, and ΔF is the variation range of the cutting force. When W > W_threshold, reduce the cutting speed v and adjust the tool angle θ.

5. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 4, is characterized in that... The adjustment based on the dynamic weight function further includes: The current cutting force F_curr is compared with the optimal cutting force F_ref in similar historical processes to obtain the cutting force difference ΔF = |F_curr F_ref|; Calculate the dominant frequency percentage of amplitude spectrum A_amp: r_freq = A_freq_high / A_freq_total; The decision tree structure is used to determine process adaptability: if ΔF < ΔF_threshold and r_freq < r_upper, then the original parameters can be maintained; Otherwise, a fuzzy control algorithm is introduced to generate compensation commands to adjust the spindle speed ω and feed rate f.

6. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 5, is characterized in that... The process adaptability assessment based on the decision tree structure further includes: Collect the time step Δt and temperature field distribution T_grid for the current process; The process stability index S_stable = 1 (σ_F / μ_F) (σ_a / μ_a) is calculated, where σ_F is the standard deviation of the cutting force, μ_F is the mean, σ_a is the standard deviation of the vibration acceleration, and μ_a is the mean. Construct the conditional formula: If S_stable < S_threshold and Δt > t_window, then trigger the dynamic compensation logic; At this point, the least squares method is used to fit the trend of the cutting force change, and the compensation coefficient γ = ΔF / (T_grid ×v) is obtained.

7. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 6, is characterized in that... The introduction of the least squares method to fit the cutting force variation trend further includes: Select the cutting force F_i (i=1,...,m) within the most recent m cycles as the input data; Define the model function as F_i = β_0 + β_1 t_i + β_2 t_i², where t_i represents a periodic sequence; The regression coefficients β_0, β_1, and β_2 are solved using the least squares method. The cutting force trend is determined based on whether β_2 is greater than zero. If β_2 > 0 and β_1 > 0, the feed rate is reduced; otherwise, the current process plan is retained.

8. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 7, is characterized in that... The predictive analysis based on model functions further includes: Compare the fitted trend curve with the target cutting force F_target, and calculate the error e_t = |F_trend F_target|; Set a tolerance range ε. When e_t > ε and the error growth rate Δe_t / Δt > 0, trigger the secondary optimization process. A genetic algorithm is used to optimize the cutting speed v and feed rate f, with the objective function being minimize [α]. e_t + β [(vv_initial)], where α and β are weighting coefficients; If the optimization result meets the constraints, the current parameter group is updated, and the new process scheme is written to the process database.

9. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 8, is characterized in that... The optimization of cutting parameters based on genetic algorithm further includes: The parameters of each individual in the population are initialized as [v, f, θ]. Set the selection pressure P_sel, crossover rate p_c, and mutation rate p_m; The fitness value is evaluated based on a penalty function. If the parameters do not meet physical constraints or device limitations, a larger penalty value is assigned. Select the best individuals to proceed to the next round of evolution, and record the best solution as the optimal solution for this process.

10. The intelligent design method for machining processes using a large language model that integrates multimodal industrial data, as described in claim 9, is characterized in that... The evaluation of individual merits based on the penalty function further includes: Define the physical constraints: v_min ≤ v ≤ v_max, f_min ≤ f ≤ f_max, θ_min ≤ θ ≤ θ_max; If the constraint is exceeded, the fitness value will be subtracted by a penalty coefficient p_pen; An energy consumption coefficient C_energy = γ is added during fitness evaluation. v The weighted term of f, where γ is the energy loss coefficient; By combining a hybrid selection strategy with an elite retention strategy, we can ensure that the optimal solution continuously converges to the feasible region.

Images