A method and system for predicting air compressor failure based on large models

Abstract

The present invention discloses a method for predicting air compressor failure based on a large model, which involves electronic digital data processing, wherein the collected air compressor data is classified, and each level of data is subjected to abnormal analysis processing in turn to obtain the interaction influence value, matching degree and associated support value between the data at each level; the data at each level, the interaction influence degree, matching degree and associated support value are normalized and textually analyzed through the large model to convert them into target prompt words; then, the target prompt word is subjected to abnormal score calculation to obtain an abnormal score, and the abnormal score is compared with a set threshold value T A Comparisons are made to obtain anomaly detection results. This invention discloses a large-scale model-based air compressor fault prediction system. This invention provides efficient, reliable, and interpretable fault prediction and decision support for intelligent air compressor operation and maintenance, significantly improving the intelligent level of equipment management.

Description

Technical Field

[0001] The present invention relates to electronic digital data processing, and more particularly to a large model-based air compressor fault prediction method and system. Background Art

[0002] In recent years, with the rapid accumulation of industrial equipment operating data and the continuous development of artificial intelligence technology, fault prediction technologies based on big data and deep learning have gradually emerged. Traditional fault detection methods often rely on manual experience and simple threshold settings, which are difficult to cope with the complex operating environment and changing working conditions of industrial equipment. Existing technologies focus more on how to effectively integrate multi-source data such as time series data, environmental parameters, and sensor outputs, and identify fault patterns through machine learning algorithms. However, traditional models have limitations in handling heterogeneous data, real-time response, and interpretability, making them difficult to meet the high standards of predictive maintenance for equipment required by intelligent manufacturing. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide an air compressor fault prediction method and system based on a large model, which mainly solves the problems of difficulty in multi-source data fusion, limited anomaly detection and interpretation capabilities, poor model adaptability, data scarcity and weak generalization ability.

[0004] The present invention provides a large-scale model-based air compressor fault prediction method, which includes:

[0005] The data classification processing step is used to classify the collected air compressor data and perform abnormal analysis on each level of data in turn to obtain the interaction impact value, matching degree and correlation support value between the data at each level;

[0006] The LLM abnormality judgment and interpretation step is used to normalize and textually analyze the data at all levels, interaction influence, matching degree and association support through the large model to convert them into target prompt words; then the target prompt word is subjected to abnormality score calculation to obtain an abnormality score, and the abnormality score is compared with the set threshold T A Compare and get the anomaly detection results.

[0007] Traditional methods struggle to effectively integrate diverse and heterogeneous data, resulting in one-sided predictions. However, this invention leverages multimodal reprogramming technology to achieve deep integration of multiple data sources, including time series, sensor data, and text records. Furthermore, existing technologies are insufficiently interpretable for complex failure modes. This invention combines semantic graphs with adaptive prompt design to provide transparent analysis of anomaly causes.

[0008] As a further improvement, the classified air compressor data is divided into primary data, secondary data and tertiary data, wherein:

[0009] The primary data is the core data, which is used to analyze whether there is any abnormality in the air compressor. If so, the interaction impact analysis is performed on the secondary data;

[0010] Secondary data is auxiliary data, which is used to calculate the interaction effect value with primary data based on its characteristic mean and variance;

[0011] The third-level data is associated historical data, which is used to analyze the matching degree between the data and the historical failure mode, and calculate the correlation support value based on the matching degree.

[0012] Furthermore, the analysis of whether the air compressor is abnormal is carried out in the following manner:

[0013] Perform time series analysis or trend detection on the time series of the primary data to obtain an abnormal time series; calculate a residual ∈(t) based on the time series of the primary data and its abnormal time series, and compare the residual ∈(t) with a preset threshold T to determine whether there is an abnormality in the core data of the air compressor:

[0014]

[0015] in, X(t) represents the time series of the primary data; Indicates abnormal timing.

[0016] Furthermore, the interaction impact value is calculated by the following formula:

[0017] I X,Y =w1*g1(μ Y ,X(t))+w2*g2(σ Y ,X(t));

[0018] Among them, I X,Y is the interaction effect value; w1 and w2 are feature weights; Indicates the cumulative impact of the integral calculation feature mean on the operating data within a specific time window;

[0019] Indicates the rate of change of the primary data; μ Y is the characteristic mean; σ Y is the variance; X(t) represents the time series of the primary data; the sequence of the secondary data is represented by Y={y1,y2,...,y n}.

[0020] Furthermore, the feature mean is calculated by the following formula:

[0021]

[0022] The variance is calculated by the following formula:

[0023]

[0024] Furthermore, the correlation support value is calculated by the following formula:

[0025]

[0026] Where S is the correlation support value; COV(R Z,H ,M) represents covariance; R Z,H represents the matching degree between the third-level data Z and the historical failure mode H; M represents the failure mode; σ R and σ M are the standard deviation of matching degree and failure mode, respectively.

[0027] As a further improvement, the method also includes optimization and data enhancement steps for dynamically adjusting the feature weights and prediction strategies of the large model based on the anomaly scores described herein.

[0028] Furthermore, the optimization and data enhancement steps specifically include:

[0029] The first step is to optimize the feature weights based on the anomaly score:

[0030]

[0031] Among them, w i (t) is the feature weight of the i-th feature at time t; α represents the learning rate; Represents the feature-to-anomaly score Contribution function of i Represents the features extracted from the air compressor data;

[0032] Step 2: Based on the optimized feature weights, the prompt generation module in the large model automatically adjusts the task guidance content:

[0033]

[0034] Among them, w i is the optimized feature weight; I X,Y is the interaction effect value; R Z,H Indicates the matching degree; S is the correlation support value;

[0035] The third step is to provide training samples for the optimized feature weights through GANs.

[0036] Traditional models have limited adaptive adjustment capabilities when faced with changing equipment operating conditions. This invention dynamically optimizes prediction strategies through a real-time weight adjustment mechanism. To address data scarcity, this invention introduces generative adversarial networks (GANs) to generate synthetic fault data, improving model training and the ability to detect unknown faults, thereby significantly enhancing the accuracy and reliability of fault prediction.

[0037] Furthermore, the method also includes model output and visualization steps.

[0038] An air compressor fault prediction system based on a large model, comprising:

[0039] The data classification processing module is used to classify the collected air compressor data and perform abnormal analysis on each level of data in turn to obtain the interaction impact value, matching degree and correlation support value between the data at each level;

[0040] The LLM anomaly judgment and interpretation module is used to normalize and textually analyze the data at all levels, interaction influence, matching degree and association support through a large model to convert them into target prompt words; then the target prompt words are anomaly scored to obtain anomaly scores, and the anomaly scores are compared with the set threshold T A Compare and get the anomaly detection results.

[0041] Beneficial effects

[0042] The advantages of the present invention are:

[0043] 1. Explanation and operability: This invention not only detects anomalies but also provides detailed cause analysis and optimization suggestions, providing direct action guidance for operation and maintenance personnel.

[0044] 2. Data-driven intelligence: The present invention adopts hierarchical processing and fusion of multimodal data, which enables the model to fully utilize multi-source information and improve the reliability and comprehensiveness of predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the flow of the air compressor fault prediction method of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with the embodiments, but this does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0047] See Figure 1The present invention presents a large model-based air compressor fault prediction method, which includes data hierarchical processing, LLM anomaly determination and interpretation, optimization and data enhancement, and model output and visualization. The LLM stands for Large Language Model, also referred to as the large model. These four steps are described in detail below.

[0048] Regarding the data classification and processing steps, this solution uses various sensors to acquire the data shown in Table 1. This data collection over a period of time includes 13 types of air compressor data: ambient temperature, compressor operating load, pipe diameter, leak detector data, air compressor station exhaust status, condensate discharge status, customer on-site inspection forms, customer troubleshooting records, pressure sensors, flow meters, electricity meters, and dew point. We perform classification based on the characteristics and importance of this data. The specific data classification is shown in Table 1.

[0049]

[0050] Table 1 Data classification

[0051] Primary data is core data and requires real-time monitoring and feature extraction. Time series analysis can employ an autoregressive moving average model, trend detection can employ a Kalman filter, and threshold setting can employ statistical thresholds. These methods are not limited; they suffice to achieve the desired results. Using pressure data and air compressor operating load as an example, the following formula illustrates this.

[0052] Assume that the first-level data is x1, x2, ..., x n ,

[0053]

[0054] Where X(t)={x1,x2,...,x n} represents the time series of any one of the primary data of pressure, flow and load; It represents the abnormal time series obtained after time series analysis or trend detection of X(t).

[0055] Anomaly detection is obtained by comparing the residual ∈(t) with the threshold T:

[0056]

[0057] in,

[0058] For example, by monitoring pressure changes in an air compressor system in real time, time series analysis (such as LSTM or ARIMA) can be used to identify abnormal fluctuations. Abnormal pressure may indicate system leaks or compression efficiency issues. Promptly identifying anomalies can prevent equipment from operating under overpressure or low-pressure shutdowns. Load curves can be used to analyze whether the equipment is operating within the normal load range. Trend detection methods (such as the Holt-Winters method) can be used to analyze load trends and detect overload or underload conditions. These methods can determine the load status of the equipment. Overload may lead to increased motor losses, while underload may indicate system leaks or insufficient demand. Combining this information, if frequent and irregular fluctuations in the operating load curve are detected, further verification through abnormal changes in pressure data can determine whether the air compressor has internal leaks.

[0059] For secondary data, it is auxiliary data, taking temperature data, exhaust status and condensate discharge as examples. The sequence of secondary data is represented as: Y[n] = {y1, y2, ..., y n},

[0060] Its characteristic mean and variance are expressed as:

[0061]

[0062] The interaction effect value between the secondary data and the primary data is calculated based on the characteristic mean and variance of the secondary data:

[0063] I X,Y =f(μ Y ,σ Y ,X(t))=w1*g1(μ Y ,X(t))+w2*g2(σ Y ,X(t)).

[0064] Among them, w1 and w2 are weight parameters used to balance the influence of mean and variance.

[0065] It represents the cumulative impact of the integral calculation mean on the operating data within a specific time window. It can be used to measure the potential impact of high temperature on the increase in equipment pressure.

[0066] This method quantifies the impact of an unstable environment on the dynamic response of a device by multiplying the environmental variance by the rate of change of the device data. It is suitable for identifying fluctuations in device operation caused by environmental changes.

[0067] Finally, the abnormal correlation is obtained by the following formula:

[0068]

[0069] Among them, T YIt is a threshold set based on historical device data or expert experience, and is used to distinguish the degree of correlation between normal and abnormal conditions.

[0070] For temperature data, the stability of the air compressor operating temperature is analyzed through statistical analysis (such as mean, variance, etc.) and time series detection (such as moving average), and abnormal conditions of too high or too low temperature are identified. Too high temperature may affect compression efficiency and cause the equipment overheating protection to be triggered, while too low temperature may affect condensate discharge. For exhaust status and condensate discharge data, the effectiveness of the equipment cooling and drainage system is judged by using temperature, humidity and dew point data, combined with the exhaust and condensate discharge status of the air compression station. Use threshold detection to set the normal range of exhaust and condensate. Poor exhaust or improper condensate discharge may cause the internal temperature of the system to rise or moisture to accumulate, affecting the normal operation of the air compressor. Combining the above information, through the analysis of ambient temperature data and exhaust status, it can be found that the exhaust system is not smooth during high temperature periods, resulting in excessively high air compressor temperatures. Combined with pressure abnormalities, it can be inferred that the equipment cooling system is faulty.

[0071] Level 3 data, as associated historical data, provides valuable insights into current faults through analysis of inspection records and maintenance data. The following describes how to process level 3 data.

[0072] First, the matching and correlation calculation for the three-level data are as follows:

[0073] R Z,H =Match(Z,H).

[0074] where R Z,H It represents the matching degree between the tertiary data Z and the historical fault pattern H. Match() can be a distance metric (such as cosine similarity, Euclidean distance) or a statistical matching function, which matches the current detected data with the historical pattern to identify whether there is a known fault pattern.

[0075] Second, perform semantic analysis and fault mode extraction on the third-level data:

[0076] M=SenmanticAnalysis(Z).

[0077] Here, M represents the fault mode extracted from the tertiary data. Semantic analysis techniques are used to identify key fault descriptions from text data such as inspection sheets and maintenance records. SenmanticAnalysis() can use models such as TF-IDF and BERT to extract key information and labels from the text.

[0078] Third, the matching degree R Z,H Combined with the fault mode M, the correlation support value is generated:

[0079]

[0080] Among them, S is the correlation support value, which provides background support for the current anomaly detection. Z,H ,M) represents the covariance, σ R and σ M are the standard deviation of matching degree and failure mode, respectively.

[0081] Regarding the LLM abnormality judgment and interpretation step, in this step, the LLM is used to comprehensively judge the various results and data calculated in the previous step. Here, the LLM can be a proprietary model for a specific vertical field (air compressor) or a general large model (recommended parameter size of 13B or more).

[0082] In this step, multi-source data integration is first performed, that is, before LLM abnormality judgment, the results of data processing at all levels are integrated to ensure that LLM can fully utilize the various indicators calculated previously. The specific method is as follows:

[0083] Input=PreProcess[X,Y,Z,I X,Y ,R Z,H ,S].

[0084] Among them, X, Y, and Z are primary, secondary, and tertiary data respectively; I X,Y is the interaction effect value between the core data and the auxiliary data, which has been calculated in the previous step; R Z,H is the degree of matching between the third-level data and the historical failure pattern; S represents the associated support value. PreProcess() represents the process of normalizing and textualizing the above data information and converting it into appropriate prompts.

[0085] Then, LLM analyzes the anomaly: LLM receives the integrated input, identifies the anomaly pattern through reasoning and analysis, and makes an anomaly judgment based on the interaction effect, matching degree, and associated support value, namely:

[0086]

[0087] in, is the output of LLM, i.e. the detected anomaly score, which takes into account various interaction and support indicators. A Compare and make an abnormal judgment.

[0088] Regarding LLM anomaly analysis, it's important to note that if the selected LLM is a general-purpose model, this step requires the use of a domain support library as an external knowledge base for the model. This allows for analysis and output of the LLM anomaly analysis results using search-enhanced generation techniques. If the selected LLM is a specialized model, no external knowledge base is required; in this scenario, the default model is knowledge of the air compressor anomaly detection domain.

[0089] Let's take a specific example: Based on the above, by processing the three levels of data, the interaction effect value I of pressure and temperature is obtained. X,Y is 12, and the background data matches R Z,H is 0.85, and the associated support value S is 0.7. LLM combines these results and the detected anomaly score is 0.92, exceeding the threshold T A 0.7. Causal analysis shows that the current anomaly, primarily caused by high pressure due to high temperature, is highly correlated with historical condensate drainage failures. The final explanation output is: "High pressure anomaly detected, primarily due to poor condensate drainage. The correlation support is high. Inspection of the drainage system is recommended."

[0090] Regarding the optimization and data enhancement step. This step aims to dynamically adjust the model's attention weight and prediction strategy based on the current anomaly score to improve the model's ability to respond to abnormal data. The specific steps are:

[0091] Step 1: Real-time optimization: In the data classification processing step and the LLM anomaly judgment and interpretation step, each feature has a different impact on anomaly detection, so the weight of each feature needs to be dynamically adjusted according to the detection results. Let each feature in the data classification processing step and the LLM anomaly judgment and interpretation step be F i , whose weight is w i , both are initially set to 1, and the optimization strategy is as follows:

[0092]

[0093] w i (t) is the weight of the i-th feature at the current moment; α represents the learning rate, which is used to control the adjustment amplitude of the weight; Represents feature F i Scoring anomalies The contribution function determines the direction and magnitude of feature weight adjustment. g() can be the Pearson correlation coefficient formula.

[0094] Step 2: Adaptive Prompt Generation and Task Guidance: Based on the optimized feature weights, the prompt generation module automatically adjusts the task guidance content to more accurately guide the LLM prediction process. Specifically:

[0095]

[0096] In this step w i The optimized feature weights ensure that the generated prompts can better guide the model to focus on important features. For example, it can focus on the impact of high temperature on load and detect abnormal fluctuations.

[0097] Step 3: GANs data enhancement: Provide more corresponding training samples for the adjusted feature weights to improve the learning ability of the model. The formula is as follows:

[0098]

[0099] Among them, the generator G is not only based on the noise z, but also based on the feature weight w i Generate samples, focusing on simulating abnormal data corresponding to high-weight features. Through the above strategy, focus on generating abnormal samples corresponding to high-weight features. For example, if the weight of high temperature w 温度 Increased, GANs prioritizes generating fault samples related to high temperature.

[0100] Regarding model output and visualization steps, the output includes anomaly detection results, causes, and optimization suggestions. This visualization can be presented through charts: graphically displaying the device's operating status, anomaly location, and severity. Real-time monitoring is also provided, such as dynamically updated dashboards showing changes in key indicators. Reports can also be generated, providing detailed text analysis and action suggestions to facilitate rapid decision-making.

[0101] The present invention also discloses an air compressor fault prediction system based on a large model, comprising:

[0102] The data classification processing module is used to classify the collected air compressor data and perform abnormal analysis on each level of data in turn to obtain the interaction impact value, matching degree and correlation support value between the data at each level;

[0103] The LLM anomaly judgment and interpretation module is used to normalize and textually analyze the data at all levels, interaction influence, matching degree and association support through a large model to convert them into target prompt words; then the target prompt words are anomaly scored to obtain anomaly scores, and the anomaly scores are compared with the set threshold T A Compare and get the anomaly detection results.

[0104] Traditional methods struggle to effectively integrate diverse and heterogeneous data, resulting in incomplete predictions. However, this invention leverages multimodal reprogramming technology to achieve deep integration of multiple data sources, including time series, sensor data, and text records. Furthermore, existing technologies are insufficiently interpretable for complex failure modes. This invention combines semantic graphs with adaptive prompt design to provide transparent analysis of anomaly causes.

[0105] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention. These modifications and improvements will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for predicting air compressor failure based on a large model, characterized in that: The method includes: The data classification processing step is used to classify the collected air compressor data and perform abnormal analysis on each level of data in turn to obtain the interaction impact value, matching degree and correlation support value between the data at each level; The LLM abnormality judgment and interpretation step is used to normalize and textually analyze the data at all levels, interaction influence, matching degree and association support through the large model to convert them into target prompt words; then the target prompt words are calculated for abnormality scores to obtain abnormality scores, and the abnormality scores are compared with the set thresholds. T A Compare and obtain abnormality detection results; The classified air compressor data is divided into primary data, secondary data and tertiary data, wherein: The primary data is the core data, which is used to analyze whether there is any abnormality in the air compressor. If so, the interaction impact analysis is performed on the secondary data; Secondary data is auxiliary data, which is used to calculate the interaction effect value with primary data based on its characteristic mean and variance; Level 3 data is associated historical data, which is used to analyze the matching degree between it and historical failure patterns and calculate the correlation support value based on the matching degree; The specific analysis method for analyzing whether there is an abnormality in the air compressor is as follows: Perform time series analysis or trend detection on the time series of the primary data to obtain abnormal time series; calculate the residual ∈(t) based on the time series of the primary data and its abnormal time series, and compare the residual ∈(t) with the preset threshold T Compare and determine whether there is any anomaly in the core data of the air compressor: ； in, ; X(t) Represents the time series of primary data; Indicates abnormal timing; The interaction effect value is calculated by the following formula: ； in, I X,Y is the interaction effect value; w 1 and w 2 is the feature weight; , represents the cumulative impact of the integral calculation feature mean on the operating data within a specific time window; ; Indicates the rate of change of the primary data; μ Y is the characteristic mean; σ Y is the variance; X(t) represents the time series of the primary data; the sequence of the secondary data is represented by Y= { y 1 ,y 2 ,...,y n }; The characteristic mean is calculated by the following formula: ； The variance is calculated by the following formula: ； The relevance support value is calculated by the following formula: ； in, S is the correlation support value; COV( R Z,H , M ) represents the covariance; R Z,H Represents three-level data Z and historical failure modes H The degree of match between M Indicates the failure mode; σ R and σ M are the standard deviation of matching degree and failure mode, respectively.

2. The air compressor fault prediction method based on a large model according to claim 1 is characterized in that: The method also includes optimization and data enhancement steps for dynamically adjusting the feature weights and prediction strategies of the large model based on the anomaly scores described herein.

3. The air compressor fault prediction method based on a large model according to claim 1 is characterized in that: The optimization and data enhancement steps specifically include: The first step is to optimize the feature weights based on the anomaly score: ； in, w i ( t ) is the i-th feature in t The feature weight at the moment; α represents the learning rate; Represents the feature-to-anomaly score Contribution function of F i Represents the features extracted from the air compressor data; Step 2: Based on the optimized feature weights, the prompt generation module in the large model automatically adjusts the task guidance content: ； in, w i is the optimized feature weight; I X,Y is the interaction effect value; R Z,H Indicates the degree of matching; S is the association support value; The third step is to provide training samples for the optimized feature weights through GANs.

4. The air compressor fault prediction method based on a large model according to claim 1 is characterized in that: The method also includes model output and visualization steps.

5. A large-model-based air compressor fault prediction system using the large-model-based air compressor fault prediction method according to any one of claims 1 to 4, characterized in that: include: The data classification processing module is used to classify the collected air compressor data and perform abnormal analysis on each level of data in turn to obtain the interaction impact value, matching degree and correlation support value between the data at each level; The LLM anomaly judgment and interpretation module is used to normalize and textually analyze the data at all levels, interaction influence, matching degree and association support through a large model to convert them into target prompt words; then the target prompt words are anomaly scored to obtain anomaly scores, and the anomaly scores are compared with the set thresholds. T A Compare and get the anomaly detection results.

Images