Rotor slip root cause traceability regulation and control method based on large model supervised fine tuning
By constructing a slippage text knowledge base and using quantized low-rank adapter technology to perform supervised fine-tuning of the large language model, the problem of accuracy in tracing the root causes and controlling rotor slippage was solved, and the accurate tracing and effective control of rotor slippage faults were achieved.
Patent Information
- Application Number
- CN202510901390.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-01
- Publication Date
- 2025-11-14
AI Technical Summary
Existing technologies struggle to accurately identify and regulate the coupling relationships between various factors that cause rotor slippage, resulting in a lack of targeted fault tracing and control measures.
A slippage text knowledge base is constructed and a supervised fine-tuning of a large language model is performed using quantized low-rank adapter technology. By leveraging the semantic understanding and complex pattern recognition capabilities of the large model, the root cause of rotor slippage is identified and control measures are formulated.
It improves the accuracy and reliability of tracing the root causes of rotor slippage, and can concisely summarize the analysis of slippage causes, control measures and supplementary explanations, thus achieving targeted control.
Smart Images
Figure CN120952135A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of high-speed rotating bearing slippage technology containing flexible rotors, and mainly to a root cause control method for rotor slippage based on supervised fine-tuning of a large model. Background Technology
[0002] Statistics show that time-varying slippage accounts for more than 30% of bearing failure causes, becoming a major factor restricting the development of aero-engines towards ultra-high speed, low friction, and lightweight designs. Research indicates that the flexible rotor in an aero-engine undergoes significant deformation when moving within the resonance region, thereby altering the slippage state of the bearings sequentially distributed along the flexible shaft. Specifically, the contact between the bearing raceway and rolling elements weakens and becomes unstable within the resonance region, resulting in a reduction in the traction force exerted by the raceway on the rolling elements. Therefore, this invention, based on existing research on individual bearing slippage, further considers rotor eccentricity and the phenomenon of "resonance slippage," conducting root cause analysis and control research on rotor slippage.
[0003] Studies have shown that rotor slippage is often caused by a combination of factors. Without understanding the coupling relationships between these factors, it is difficult to determine the root cause of slippage failures, let alone develop targeted control measures. Customized large language models for specific domains have demonstrated powerful capabilities in fault tracing and control decision-making. Considering the potential for illusions and outdated and limited expertise in large models, numerous studies have focused on supervised fine-tuning to improve the accuracy and reliability of knowledge acquisition. In general, fine-tuning large language models is a highly effective method for improving their performance and adding or removing desirable behaviors. This invention fully utilizes the powerful semantic understanding capabilities, rich knowledge reserves, complex pattern recognition capabilities, and probability-driven contextual association and reasoning abilities of large models to achieve root cause tracing and control of rotor slippage. Summary of the Invention
[0004] Based on the problems existing in the above-mentioned background technology, the purpose of this invention is to provide a method for supervised fine-tuning of rotor slippage root cause tracing and control based on large model. By constructing a slippage text knowledge base and combining it with quantized low-rank adapter technology to perform supervised fine-tuning of large model, reliable, credible and accurate results of rotor slippage root cause tracing and control can be obtained.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A method for controlling the root causes of rotor slippage based on supervised fine-tuning of a large model includes the following steps:
[0007] Step S1: Construct a slip text knowledge base based on slip parameter division rules, slip domain cue words, and the organized pre-declarations. The pre-declarations emphasize that the large language model considers the output content format requirements and the correlation analysis of coupling factors.
[0008] Step S2: Divide the slippage text knowledge base constructed in step S1 into a training set and a test set, and each data entry contains three parts: slippage cause analysis, slippage control measures, and supplementary explanation.
[0009] Step S3: Use the slippery text dataset constructed in step S2 and combine it with the quantized low-rank adapter technique to perform supervised fine-tuning of the large language model.
[0010] Step S4: Match the collected rotor operating parameters with the slippage parameter classification rules to obtain the corresponding rotor operating parameter level, and input it into the large language model after supervised fine-tuning in step S3, so as to determine the root cause of slippage and formulate targeted control measures.
[0011] Furthermore, step S1 includes the following sub-steps:
[0012] Step S1.1: Formulate a classification rule for the level of slippage parameters based on the factors affecting rotor slippage. Classify the slippage behavior into three levels: severe, moderate and slight according to the degree of influence of different parameter ranges, and clarify the parameter range corresponding to different levels.
[0013] Step S1.2: Arrange and combine the slippage influence parameters that were divided into three levels in step S1.1 to construct different slippage conditions;
[0014] Step S1.3: Construct slippage area prompt words based on the parameter levels divided in step S1.1;
[0015] Step S1.4: Organize the pre-declaration from two aspects: output content format requirements and coupling factor correlation analysis;
[0016] In step S1.5, the slippage conditions constructed in step S1.2, the slippage domain cue words constructed in step S1.3, and the pre-declarations organized in step S1.4 are input into the large language model to obtain the rotor slippage text knowledge base.
[0017] Furthermore, in step S1.1, the factors affecting rotor slippage include: clearance, surface roughness, assembly preload, assembly interference, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, speed, and eccentricity.
[0018] Furthermore, in step S1.3, the slippage area prompts include the following technical association rules:
[0019] (1) When the speed parameter is in the first specific range and is combined with other parameters in the severe level, it is associated with resonant slippage.
[0020] (2) When the speed parameter is higher than the second specific threshold, it is associated with high-speed light-load slippage.
[0021] (3) When the surface roughness parameter is at a severe level, its impact analysis needs to be correlated with changes in lubrication state and friction coefficient.
[0022] (4) When the acceleration parameter is at a severe level, its impact analysis needs to be correlated with changes in lubrication state and changes in the lubricating oil film formation process;
[0023] (5) The eccentric parameter has the effect of reducing the rotor slippage rate, but when it is at a severe level, it will cause the bearing vibration to intensify and lead to slippage damage and the formation of foreign matter.
[0024] (6) When the eccentricity parameter is at a severe level, the increase in the lubricating oil flow parameter will be associated with the increased activity of foreign matter and the deterioration of the internal contact state of the bearing, thereby aggravating the bearing slippage.
[0025] (7) Under the condition of resonance slippage, the priority of slippage control measures is assigned as adjusting other parameters over speed parameters.
[0026] In some embodiments of the present invention, the slippage warning words are as follows:
[0027] Rotor resonance will only occur when the speed is between 8000 r / min and 12000 r / min. If other parameters are at a severe level at this time, it will cause the bearing to "resonance slippage" and the bearing slippage rate will increase significantly.
[0028] When the speed is greater than 12000 r / min, the bearing is prone to high-speed light-load slippage. This is because the centrifugal force causes the drag force received by the rolling element to be less than its motion resistance, thus resulting in slippage.
[0029] If the surface roughness is at a severe level, the cause of slippage should be analyzed from the perspective of the influence of surface roughness on lubrication and friction coefficient.
[0030] If the acceleration is severe, the cause of slippage should be analyzed from the perspective of the impact of acceleration on the lubrication state and the formation process of the lubricating oil film.
[0031] Appropriate eccentricity can increase bearing contact load and thus reduce bearing slippage rate. However, severe eccentricity will aggravate bearing vibration and cause bearing slippage damage, which will form foreign matter.
[0032] When the eccentricity is severe, a larger lubricating oil flow will increase the activity of foreign matter inside the bearing, affecting the internal contact state of the bearing and exacerbating bearing slippage.
[0033] Considering the operability of different parameters, when "resonance slippage" occurs, other parameters should be adjusted first, and the speed should be adjusted last.
[0034] "Resonance slippage" can be suppressed and the bearing slippage rate reduced by measures such as critical speed avoidance, variable preload control, and active damping injection.
[0035] Furthermore, in step S1.4, the output response structure of the slippery text knowledge base is defined by a pre-declaration regarding the format requirements of the output content, as follows:
[0036] Answer in the following format:
[0037] Because of the mechanism
[0038] Therefore, it might be necessary to:
[0039] 1. <Solution 1>: <Specific Solution 1>, <Mechanism Basis 1>
[0040] 2. <Solution 2>: <Specific Solution 2>, <Mechanism Basis 2>
[0041] 3. <Solution 3>: <Specific Solution 3>, <Mechanism Basis 3>
[0042] <Mechanism> indicates the consequences of this slippage cause, and it is necessary to fully consider all slippage causes and the complex relationships between them;
[0043] The "Solutions" section represents the solutions obtained after fully considering the "Mechanism," listing the top three most likely and effective solutions in a brief description.
[0044] The specific solutions need to include the specific operational details of the solutions and their corresponding mechanistic basis.
[0045] Meanwhile, the organization's pre-declaration regarding the correlation analysis of coupling factors emphasizes that the output response of the slippage text knowledge base must include an analysis of the coupling relationships between multiple slippage influencing factors.
[0046] In some embodiments of the present invention, the pre-declaration is organized from two aspects: output content format requirements and coupling factor correlation analysis, as follows:
[0047] First, the organization's pre-declaration regarding output content formatting requirements is as follows:
[0048] "The answer must strictly follow the format below:"
[0049] Because of the mechanism
[0050] Therefore, it might be necessary to:
[0051] 1. <Solution 1>: <Specific Solution 1>, <Mechanism Basis 1>
[0052] 2. <Solution 2>: <Specific Solution 2>, <Mechanism Basis 2>
[0053] 3. <Solution 3>: <Specific Solution 3>, <Mechanism Basis 3>
[0054] <Mechanism> indicates the consequences of this slippage cause, and it is necessary to fully consider all slippage causes and the complex relationships between them;
[0055] The "Solutions" section represents the solutions obtained after fully considering the "Mechanism," listing the top three most likely and effective solutions in a brief description.
[0056] The specific solutions must include the specific operational details of the solutions and their corresponding mechanistic basis. Data cannot be fabricated; the answers must be based on complete facts.
[0057] Then, the organization's pre-declaration regarding the correlation analysis of coupling factors is "Please analyze the mutual influence between parameters step by step."
[0058] Furthermore, in step S2, the rotor slippage text knowledge base constructed in step S1 is divided into a training set and a test set. The slippage cause analysis part of each data point comprehensively considers the correlation between multiple slippage coupling influencing factors. The slippage control measures part gives the three most likely and effective solutions for the slippage causes obtained from the analysis. The supplementary explanation part gives the scientific basis for root cause tracing and control.
[0059] Furthermore, step S3 includes the following sub-steps:
[0060] Step S3.1: Supervised fine-tuning of the large language model is performed using the quantized low-rank adapter technique. The forward propagation mechanism of the quantized low-rank adapter is defined as follows:
[0061]
[0062] in, d o line k o The large model pre-training weight matrix; This represents the initialization of the large model pre-training weight matrix; This represents the projection matrix, which contains trainable parameters and is initialized using Gaussian randomization. Let r represent the projection matrix, which contains trainable parameters and is initialized with all zeros; o Represents the low-rank increment matrix ΔW o The rank of the ... o < <min(d o ,k o );x o h represents the feature vector input to the quantized low-rank adapter. o This represents the output feature vector of the quantization low-rank adapter;
[0063] The forward computation framework for the quantized low-rank adapter is defined as follows:
[0064]
[0065] Among them, Y BF16 This represents the output of the low-rank quantization adapter with BFloat16 precision. and X represents the first-order and second-order quantization constants, respectively; BF16 The input features represent BFloat16 precision; W NF4 This represents a 4-bit standard floating-point quantization weight; and Both represent low-rank projection matrices; doubleDequant(·) represents double-weighted dequantization operation;
[0066] Step S3.2: Based on the slippery text dataset constructed in step S2 and the quantized low-rank adapter technique proposed in step S3.1, supervised fine-tuning of the large language model is performed, and the cross-entropy loss of supervised fine-tuning is minimized.
[0067] Furthermore, step S4 includes the following sub-steps:
[0068] Step S4.1: Collect specific values of rotor operating parameters including clearance, surface roughness, assembly preload, assembly interference, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, speed and eccentricity.
[0069] Step S4.2: Match the rotor operating condition parameter values collected in step S4.1 with the parameter level classification rules defined in step S1 to obtain the corresponding rotor operating condition parameter levels.
[0070] Step S4.3: Input the rotor operating condition parameter levels matched in step S4.2 and the statement "Provide corresponding mechanisms and solutions based on the combination of parameters provided by the user" into the supervised fine-tuned large language model obtained in step S3.
[0071] In step S4.4, based on the input from step S4.3, the supervised fine-tuned large language model will output a slippage source control text containing three parts: slippage cause analysis, slippage control measures, and supplementary explanation. The slippage cause analysis part corresponds to the identified root cause of slippage, and the slippage control measures part corresponds to the control measures formulated for the root cause of slippage.
[0072] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0073] (1) In the process of supervising the construction of the slip text knowledge base of the large model, domain experts can evaluate the slip knowledge by integrating their own experience, feelings and ideas and other potential information, and selectively screen the fine-tuning dataset; further supervising the large model with the screened dataset can enable the large model to learn the valuable potential thinking of humans, which is difficult to express and summarize effectively.
[0074] (2) The large language model, combined with its rich knowledge reserves and associative reasoning ability, can deeply integrate the summarized slippage domain cue words into the existing slippage knowledge, thereby greatly leveraging the advantages of the existing slippage domain knowledge; in particular, the finely tuned large model can accurately understand and master the conditions for the occurrence of the "resonance slippage" phenomenon and formulate reasonable control measures.
[0075] (3) There are intricate coupling relationships among the 11 parameters that affect rotor slippage faults. Even experienced domain experts cannot construct all the state parameter propagation chains. This invention, by leveraging the powerful context association and reasoning capabilities of the large model and introducing a pre-declaration on the correlation analysis of coupling factors, can enable the large model to fully explore the coupling relationships between multiple slippage influencing factors, and thus identify the root cause of slippage among many slippage influencing factors.
[0076] (4) By constructing structured slippage text knowledge, the accuracy of supervised fine-tuning of large models can be significantly improved. Based on the collected rotor operating parameters, the large model after supervised fine-tuning can concisely and clearly summarize the three parts of slippage cause analysis, slippage control measures and supplementary explanations, thereby determining the root cause of slippage and formulating targeted control measures. Attached Figure Description
[0077] Figure 1 Flowchart of the rotor slippage root cause control method based on large model supervised fine-tuning provided by the present invention;
[0078] Figure 2 The structural diagram of the high-speed flexible rotor slippage test bench provided by the present invention;
[0079] Figure 3This embodiment of the invention comprehensively considers the coupling effects of slippage factors and the slippage text knowledge generated by prompt words;
[0080] Figure 4 This is the training dataset for supervised fine-tuning of a customized large model for rotor slippage in this embodiment of the invention;
[0081] Figure 5 This is the cross-entropy loss change curve of the supervised fine-tuning process of the large model in this embodiment of the invention;
[0082] Figure 6 This refers to the similarity output of the large model after supervised fine-tuning in this embodiment of the invention.
[0083] Figure 7 This is the probability density function for the output similarity of the large model after supervised fine-tuning in this embodiment of the invention;
[0084] Figure 8 This refers to the similarity between the output of the large model before supervised fine-tuning in this embodiment of the invention.
[0085] Figure 9 The results of the rotor slippage root cause control obtained from the large model after supervised fine-tuning in this embodiment of the invention are shown. Detailed Implementation
[0086] The invention will now be further described with reference to the accompanying drawings.
[0087] See Figures 1 to 9 This embodiment provides a method for controlling the root cause of rotor slippage based on supervised fine-tuning of a large model. The process of this method is as follows: Figure 1 As shown. This embodiment uses the following... Figure 2 The algorithm performance is verified using the high-speed flexible rotor slippage test bench shown as an example. The specific steps include the following:
[0088] Step S1: Construct a slip text knowledge base based on slip parameter division rules, slip domain prompts and the organized pre-declarations. The pre-declaration requires the large language model to focus on output content format requirements and coupling factor correlation analysis.
[0089] Specifically, in this embodiment, step S1 includes:
[0090] Step S1.1: Develop parameter level classification rules for the factors affecting rotor slippage, and classify the impact of different parameter ranges on slippage behavior into three levels: severe, moderate and slight.
[0091] Specifically, the 11 parameters closely related to rotor slippage behavior—clearance, surface roughness, assembly preload, assembly interference, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, velocity, and eccentricity—are uniformly divided into three levels, and the parameter ranges corresponding to different levels are clearly defined as shown in Table 1.
[0092] Table 1. Rules for classifying rotor slippage parameter levels
[0093] grade gap Surface roughness Assembly preload Assembly interference serious >0.114mm >0.1μm 0-155Nm 0-0.002mm moderate 0.111-0.114mm 0.08-0.1μm >180Nm >0.01mm slight 0-0.111mm 0-0.08μm 155-180Nm 0.002-0.01mm grade Lubricating oil flow Lubricating oil viscosity Lubricating oil temperature load serious 0-1.25L / min <![CDATA[0-27.6mm 2 / s]]> >70℃ 0-2kN moderate 1.25-3.5L / min <![CDATA[>35.3mm 2 / s]]> <15℃ 2-4kN slight >3.5L / min <![CDATA[27.6-35.3mm 2 / s]]> 15-70℃ >4kN grade acceleration speed eccentric —— serious <![CDATA[>9.42rad / s 2 ]]> >12000r / min >3g —— moderate <![CDATA[7.85-9.42rad / s 2 ]]> 8000-12000r / min 0-1g —— slight <![CDATA[0-7.85rad / s 2 ]]> 0-8000r / min 1-3g ——
[0094] Step S1.2: Arrange and combine the 11 slippage influencing parameters that were divided into three levels in step S1.1 to construct 177,147 different slippage conditions.
[0095] Step S1.3, combining the parameter levels defined in step S1.1, construct the following prompt words for rotor slippage:
[0096] Rotor resonance will only occur when the speed is between 8000 r / min and 12000 r / min. If other parameters are at a severe level at this time, it will cause the bearing to "resonance slippage" and the bearing slippage rate will increase significantly.
[0097] When the speed is greater than 12000 r / min, the bearing is prone to high-speed light-load slippage. This is because the centrifugal force causes the drag force received by the rolling element to be less than its motion resistance, thus resulting in slippage.
[0098] If the surface roughness is at a severe level, the cause of slippage should be analyzed from the perspective of the influence of surface roughness on lubrication and friction coefficient.
[0099] If the acceleration is severe, the cause of slippage should be analyzed from the perspective of the impact of acceleration on the lubrication state and the formation process of the lubricating oil film.
[0100] Appropriate eccentricity can increase bearing contact load and thus reduce bearing slippage rate. However, severe eccentricity will aggravate bearing vibration and cause bearing slippage damage, which will form foreign matter.
[0101] When the eccentricity is severe, a larger lubricating oil flow will increase the activity of foreign matter inside the bearing, affecting the internal contact state of the bearing and exacerbating bearing slippage.
[0102] Considering the operability of different parameters, when "resonance slippage" occurs, other parameters should be adjusted first, and the speed should be adjusted last.
[0103] "Resonance slippage" can be suppressed and the bearing slippage rate reduced by measures such as critical speed avoidance, variable preload control, and active damping injection.
[0104] Step S1.4 involves organizing the pre-declaration from two aspects: output content format requirements and coupling factor correlation analysis, as detailed below:
[0105] First, the organization's pre-declaration regarding output content formatting requirements is as follows:
[0106] "The answer must strictly follow the format below:"
[0107] Because of the mechanism
[0108] Therefore, it might be necessary to:
[0109] 1. <Solution 1>: <Specific Solution 1>, <Mechanism Basis 1>
[0110] 2. <Solution 2>: <Specific Solution 2>, <Mechanism Basis 2>
[0111] 3. <Solution 3>: <Specific Solution 3>, <Mechanism Basis 3>
[0112] <Mechanism> indicates the consequences of this slippage cause, and it is necessary to fully consider all slippage causes and the complex relationships between them;
[0113] The "Solutions" section represents the solutions obtained after fully considering the "Mechanism," listing the top three most likely and effective solutions in a brief description.
[0114] The specific solutions must include the specific operational details of the solutions and their corresponding mechanistic basis. Data cannot be fabricated; the answers must be based on complete facts.
[0115] Then, the organization's pre-declaration regarding the correlation analysis of coupling factors is "Please analyze the mutual influence between parameters step by step."
[0116] Step S1.5: Input the 177,147 slip conditions constructed in step S1.2, the slip domain cue words constructed in step S1.3, and the pre-declarations organized in step S1.4 into the DeepSeek-R1 large language model to obtain a knowledge base of 177,147 rotor slip texts.
[0117] Step S2: Divide the slippage text knowledge base constructed in step S1 into two parts: training set and test set. Each data entry contains three parts: slippage cause analysis, slippage control measures, and supplementary explanation.
[0118] Specifically, in this embodiment, step S2 includes:
[0119] The 177,147 rotor slippage text knowledge bases constructed in step S1.5 are divided into a training set and a test set. The training set contains 159,431 text knowledge entries, and the test set contains 17,715 text knowledge entries. Specifically, each data entry includes three parts: slippage cause analysis, slippage control measures, and supplementary explanations. The slippage cause analysis section comprehensively considers the correlation between multiple coupled influencing factors of slippage. The slippage control measures section provides the three most likely and effective solutions for the analyzed slippage causes. The supplementary explanations section provides the scientific basis for root cause tracing and control. Figure 3 It demonstrates the knowledge of slip text generated by a large model that comprehensively considers the coupling effects of slip factors and proprietary cue words.
[0120] Step S3: Use the slippery text dataset constructed in step S2 and combine it with the quantized low-rank adapter technique to perform supervised fine-tuning of the large language model.
[0121] Specifically, in this embodiment, step S3 includes:
[0122] Step S3.1: Supervised fine-tuning of the large language model is performed using the quantized low-rank adapter technique. The forward propagation mechanism of the quantized low-rank adapter is defined as follows:
[0123]
[0124] in, d o line k o The large model pre-training weight matrix; This represents the initialization of the large model pre-training weight matrix; This represents the projection matrix, which contains trainable parameters and is initialized using Gaussian randomization. Let r represent the projection matrix, which contains trainable parameters and is initialized with all zeros; o Represents the low-rank increment matrix ΔW o The rank of the ... o < <min(d o ,k o );x o h represents the feature vector input to the quantized low-rank adapter. o This represents the output feature vector of the quantized low-rank adapter; the forward computation framework of the quantized low-rank adapter is defined as follows:
[0125]
[0126] Among them, Y BF16 This represents the output of the low-rank quantization adapter with BFloat16 precision. and X represents the first-order and second-order quantization constants, respectively; BF16 The input features represent BFloat16 precision; W NF4 This represents a 4-bit standard floating-point quantization weight; and These represent the low-rank projection matrix respectively; doubleDequant(·) represents the double-weighted dequantization operation.
[0127] In this embodiment, to construct a locally deployable and highly accurate customized large-scale model for rotor slippage, the InternLM3-8B open-source model from the Shanghai Artificial Intelligence Laboratory was selected for fine-tuning. The InternLM3-8B model contains approximately 8B parameters and possesses both high precision and low latency in industrial applications. Thanks to its core advantages such as ultra-long context modeling, innovative alignment mechanisms, and rigorous data engineering, the InternLM3-8B model is the preferred platform for deploying large-scale models in vertical domains. During the training of the customized large-scale model, the fine-tuning algorithm using quantized low-rank adapter technology only requires adjusting 26M parameters, with the number of training parameters accounting for only 0.2961% of the total parameters of the large-scale model. This satisfies the lightweight deployment requirements of the large-scale model while effectively avoiding overfitting. Furthermore, 177,147 normalized slippage textual knowledge points were constructed and used for supervised fine-tuning of the InternLM3-8B model. The fine-tuning training dataset contains 159,431 data points, and the test dataset contains 17,715 data points. Figure 4 As shown.
[0128] A customized large-scale model for rotor slippage, built on the InternLM3-8B, was deployed on a server. The server used in this study consisted of an Intel(R) Xeon(R) Gold6230R CPU at 2.10GHz, three Nvidia RTX3090 24GB GPUs, and two RECC DDR4 2933 32GB memory modules. Supervised fine-tuning of the large model, setting five epochs, required nearly 106 hours of continuous training on the server.
[0129] Step S3.2: Based on the slippery text dataset constructed in step S2 and the quantized low-rank adapter technique proposed in step S3.1, supervised fine-tuning of the large language model is performed, and the cross-entropy loss of supervised fine-tuning is minimized. Figure 5 The cross-entropy loss variation curves during the supervised fine-tuning process of a large model are shown. Specifically, this embodiment uses the Qwen3 embedding model to calculate the similarity of the fine-tuning results of the large model, which is represented as follows:
[0130] LLM sim = emb (Large model test output), Q emb (Note: Slippery conditions) > (3)
[0131] Among them, Q emb (·) denotes the rotor slippage text knowledge embedding function; <·,·> denotes the calculation of the inner product of the embedding vectors, which is used to measure the similarity between two slippage texts. Figure 6 This demonstrates the similarity between the output of the large model after supervised fine-tuning and the constructed standard slippage knowledge. To more intuitively analyze the similarity calculation results of the large model output, further statistical analysis was performed on the obtained 17715 similarity values, and the results were plotted as follows: Figure 7 The probability density function curve is shown. Meanwhile, Figure 8 This demonstrates the similarity between the output of the large model before supervised fine-tuning and the constructed standard slippage knowledge. Through comparison... Figure 6 and Figure 8 As can be seen, the average similarity of the large model output before fine-tuning was 65.15%, while the average similarity of the large model output after fine-tuning was 91.85%. This indicates that supervised fine-tuning significantly improved the accuracy of the large model output, i.e., the effect of controlling the root cause of slippage.
[0132] Step S4: Match the collected rotor operating parameters with the slippage parameter classification rules to obtain the corresponding rotor operating parameter level, and input it into the large model after supervised fine-tuning in step S3 to determine the root cause of slippage and formulate targeted control measures.
[0133] Specifically, in this embodiment, step S4 includes:
[0134] Step S4.1 involves collecting specific values for 11 rotor operating parameters, including clearance, surface roughness, assembly preload, assembly interference fit, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, velocity, and eccentricity. Specifically, in this embodiment, the rotor operating parameters collected at a certain moment are: clearance 0.112 mm, surface roughness 0.1 μm, assembly preload 100 Nm, assembly interference fit 0.015 mm, lubricating oil flow rate 3.8 L / min, and lubricating oil viscosity 41.2 mm. 2 / s, lubricating oil temperature 10℃, load 1.3kN, acceleration 5.6rad / s 2 Speed 10000 r / min, eccentricity 2.4g.
[0135] Step S4.2: Match the rotor operating condition parameter values collected in step S4.1 with the parameter level classification rules established in step S1.1 to obtain the corresponding rotor operating condition parameter levels. Specifically, the matched operating condition parameter levels are: clearance at a moderate level, surface roughness at a moderate level, assembly preload at a severe level, assembly interference at a moderate level, lubricating oil flow rate at a slight level, lubricating oil viscosity at a moderate level, lubricating oil temperature at a moderate level, load at a severe level, acceleration at a slight level, speed at a moderate level, and eccentricity at a slight level.
[0136] Step S4.3: Input the rotor operating condition parameter levels matched in step S4.2 and the statement "Provide corresponding mechanisms and solutions based on the combination of parameters provided by the user" into the supervised fine-tuned large language model obtained in step S3.
[0137] In step S4.4, based on the input from step S4.3, the supervised fine-tuned large language model will output a slippage source control text containing three parts: slippage cause analysis, slippage control measures, and supplementary explanation. The slippage cause analysis part corresponds to the identified root cause of slippage, and the slippage control measures part corresponds to the control measures formulated for the root cause of slippage. Figure 9 The results of the root cause control of rotor slippage obtained based on the large model after supervised fine-tuning are presented.
[0138] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for controlling the root cause of rotor slippage based on supervised fine-tuning of a large model, characterized in that: Includes the following steps: Step S1: Construct a slip text knowledge base based on slip parameter division rules, slip domain cue words, and the organized pre-declarations. The pre-declarations emphasize that the large language model considers the output content format requirements and the correlation analysis of coupling factors. Step S2: Divide the slippage text knowledge base constructed in step S1 into a training set and a test set, and each data entry contains three parts: slippage cause analysis, slippage control measures, and supplementary explanation. Step S3: Use the slippery text dataset constructed in step S2 and combine it with the quantized low-rank adapter technique to perform supervised fine-tuning of the large language model. Step S4: Match the collected rotor operating parameters with the slippage parameter classification rules to obtain the corresponding rotor operating parameter level, and input it into the large language model after supervised fine-tuning in step S3, so as to determine the root cause of slippage and formulate targeted control measures.
2. The rotor slippage root cause control method based on supervised fine-tuning of a large model as described in claim 1, characterized in that: Step S1 includes the following sub-steps: Step S1.1: Formulate a classification rule for the level of slippage parameters based on the factors affecting rotor slippage. Classify the slippage behavior into three levels: severe, moderate and slight according to the degree of influence of different parameter ranges, and clarify the parameter range corresponding to different levels. Step S1.2: Arrange and combine the slippage influence parameters that were divided into three levels in step S1.1 to construct different slippage conditions; Step S1.3: Construct slippage area prompt words based on the parameter levels divided in step S1.1; Step S1.4: Organize the pre-declaration from two aspects: output content format requirements and coupling factor correlation analysis; In step S1.5, the slippage conditions constructed in step S1.2, the slippage domain cue words constructed in step S1.3, and the pre-declarations organized in step S1.4 are input into the large language model to obtain the rotor slippage text knowledge base.
3. The rotor slippage root cause control method based on supervised fine-tuning of a large model as described in claim 2, characterized in that: In step S1.1, the factors affecting rotor slippage include: clearance, surface roughness, assembly preload, assembly interference, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, speed, and eccentricity.
4. The rotor slippage root cause control method based on supervised fine-tuning of a large model as described in claim 2, characterized in that: In step S1.3, the slippage area prompts include the following technical association rules: (1) When the speed parameter is in the first specific range and is combined with other parameters in the severe level, it is associated with resonant slippage. (2) When the speed parameter is higher than the second specific threshold, it is associated with high-speed light-load slippage. (3) When the surface roughness parameter is at a severe level, its impact analysis needs to be correlated with changes in lubrication state and friction coefficient. (4) When the acceleration parameter is at a severe level, its impact analysis needs to be correlated with changes in lubrication state and changes in the lubricating oil film formation process; (5) The eccentric parameter has the effect of reducing the rotor slippage rate, but when it is at a severe level, it will cause the bearing vibration to intensify and lead to slippage damage and the formation of foreign matter. (6) When the eccentricity parameter is at a severe level, the increase in the lubricating oil flow parameter will be associated with the increased activity of foreign matter and the deterioration of the internal contact state of the bearing, thereby aggravating the bearing slippage. (7) Under the condition of resonance slippage, the priority of slippage control measures is assigned as adjusting other parameters over speed parameters.
5. The rotor slippage root cause control method based on supervised fine-tuning of a large model according to claim 2, characterized in that: In step S1.4, the output response structure of the slippery text knowledge base is defined by a pre-declaration regarding the format requirements of the output content, as follows: Answer in the following format: Because of the mechanism Therefore, it might be necessary to:
1. <Solution 1>: <Specific Solution 1>, <Mechanism Basis 1> 2. <Solution 2>: <Specific Solution 2>, <Mechanism Basis 2> 3. <Solution 3>: <Specific Solution 3>, <Mechanism Basis 3> <Mechanism> indicates the consequences of this slippage cause, and it is necessary to fully consider all slippage causes and the complex relationships between them; The "Solutions" section represents the solutions obtained after fully considering the "Mechanism," listing the top three most likely and effective solutions in a brief description. The specific solutions need to include the specific operational details of the solutions and their corresponding mechanistic basis. Meanwhile, the organization's pre-declaration regarding the correlation analysis of coupling factors emphasizes that the output response of the slippage text knowledge base must include an analysis of the coupling relationships between multiple slippage influencing factors.
6. The rotor slippage root cause control method based on supervised fine-tuning of a large model as described in claim 1, characterized in that: In step S2, the rotor slippage text knowledge base constructed in step S1 is divided into a training set and a test set. The slippage cause analysis part of each data point comprehensively considers the correlation between multiple slippage coupling influencing factors. The slippage control measures part gives the three most likely and effective solutions for the slippage causes obtained from the analysis. The supplementary explanation part gives the scientific basis for root cause tracing and control.
7. The rotor slippage root cause control method based on supervised fine-tuning of a large model according to claim 1, characterized in that: Step S3 includes the following sub-steps: Step S3.1: Supervised fine-tuning of the large language model is performed using the quantized low-rank adapter technique. The forward propagation mechanism of the quantized low-rank adapter is defined as follows: in, d o line k o The large model pre-training weight matrix; This represents the initialization of the large model pre-training weight matrix; This represents the projection matrix, which contains trainable parameters and is initialized using Gaussian randomization. Let r represent the projection matrix, which contains trainable parameters and is initialized with all zeros; o Represents the low-rank increment matrix ΔW o The rank of the ... o < <min(d o ,k o );x o h represents the feature vector input to the quantized low-rank adapter. o This represents the output feature vector of the quantization low-rank adapter; Meanwhile, the forward computation framework for the quantized low-rank adapter is defined as follows: Among them, Y BF16 This represents the output of the low-rank quantization adapter with BFloat16 precision. and X represents the first-order and second-order quantization constants, respectively; BF16 The input features represent BFloat16 precision; W NF4 This represents a 4-bit standard floating-point quantization weight; and Both represent low-rank projection matrices; doubleDequant(·) represents double-weighted dequantization operation; Step S3.2: Based on the slippery text dataset constructed in step S2 and the quantized low-rank adapter technique proposed in step S3.1, supervised fine-tuning of the large language model is performed, and the cross-entropy loss of supervised fine-tuning is minimized.
8. The rotor slippage root cause control method based on supervised fine-tuning of a large model according to claim 1, characterized in that: Step S4 includes the following sub-steps: Step S4.1: Collect specific values of rotor operating parameters including clearance, surface roughness, assembly preload, assembly interference, lubricating oil flow rate, lubricating oil viscosity, lubricating oil temperature, load, acceleration, speed and eccentricity. Step S4.2: Match the rotor operating condition parameter values collected in step S4.1 with the parameter level classification rules defined in step S1 to obtain the corresponding rotor operating condition parameter levels. Step S4.3: Input the rotor operating condition parameter levels matched in step S4.2 and the statement "Provide corresponding mechanisms and solutions based on the combination of parameters provided by the user" into the supervised fine-tuned large language model obtained in step S3. In step S4.4, based on the input from step S4.3, the supervised fine-tuned large language model will output a slippage source control text containing three parts: slippage cause analysis, slippage control measures, and supplementary explanation. The slippage cause analysis part corresponds to the identified root cause of slippage, and the slippage control measures part corresponds to the control measures formulated for the root cause of slippage.