A prediction model construction method and system for personalized equipment operation and maintenance

Abstract

The present invention provides a prediction model construction method and system for personalized equipment operation and maintenance, which relates to the field of data processing technology. The method comprises: collecting equipment operation data and environmental data, extracting the change characteristics of equipment operation status, performing correlation analysis on environmental data based on historical equipment failure records, performing normalization processing respectively, performing time series analysis, generating a joint feature matrix, dividing the joint feature matrix into time segments of fixed windows in chronological order, classifying them into normal state and abnormal state, performing feature contribution analysis, generating a feature weight matrix, recombining equipment features and environmental factors in a time segment feature data set according to feature weights, generating fused feature data layer by layer, performing supervised learning training on the fused feature data, constructing an equipment prediction model, and outputting the equipment prediction model as a prediction model for personalized equipment operation and maintenance. The present invention improves the personalized equipment operation and maintenance effect.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a prediction model building method and system for personalized equipment operation and maintenance. Background Art

[0002] In the prior art, prediction models for equipment operation and maintenance usually rely on historical operation data for analysis, and predict possible failures or abnormalities of equipment by building statistical models or machine learning models. These technologies mainly include prediction methods based on regression analysis, time series analysis, and classification algorithms. Equipment operation and maintenance data is usually collected in real time by sensors and analyzed as input data after preprocessing. For example, for large industrial equipment, the collected vibration, temperature, and current data are used to build a prediction model to determine the health status of the equipment, thereby achieving preventive maintenance.

[0003] However, in practical applications, existing technologies face the problem of data adaptability, that is, the generalization ability of a single model may be insufficient for a variety of equipment types and operation and maintenance environments, making it difficult to accurately predict the status of equipment in different scenarios. For example, in the operation and maintenance of wind turbines, the operating environment, such as differences in climatic conditions and geographical locations, may cause significant changes in data distribution. Existing models may find it difficult to effectively capture these characteristics, resulting in reduced prediction accuracy and the inability to promptly warn of potential failures, increasing the risk of equipment failures. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for building a prediction model for personalized equipment operation and maintenance, aiming to solve the problems mentioned in the background technology.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, a prediction model construction method for personalized equipment operation and maintenance is provided, the method comprising:

[0007] Collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, where the equipment operation data includes the equipment's operating status parameters, and the environmental data includes external influencing factors related to the equipment's operating scenario;

[0008] Extract the changing characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, perform correlation analysis on the environmental data, screen out the main influencing factors, and generate a key environmental factor data set;

[0009] Preprocessing the preliminary equipment characteristic set and the key environmental factor data set respectively to generate normalized equipment characteristic data and normalized key environmental factor data;

[0010] Performing time series analysis based on normalized equipment feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between equipment features and key environmental factors;

[0011] The joint feature matrix is ​​divided into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set.

[0012] According to the known running status labels of the equipment, the time segment feature data set is classified, and each time segment feature data is divided into a normal state feature set and an abnormal state feature set to generate a classified feature data set;

[0013] Perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label;

[0014] According to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, and fused feature data is generated layer by layer. The fused feature data is then trained with supervised learning to build a device prediction model.

[0015] The equipment prediction model is output as a prediction model for personalized equipment operation and maintenance. The prediction model can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

[0016] Preferably, the change characteristics of the equipment operation status are extracted from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, the environmental data is subjected to correlation analysis to screen out the main influencing factors and generate a key environmental factor data set, including:

[0017] Extract equipment operating status parameters from the initial data set, analyze the change trend of the parameters in time series, calculate the change amplitude, change frequency and stability, extract the preliminary characteristics of the equipment status, and form a preliminary equipment feature set;

[0018] Analyze events related to environmental factors based on historical equipment failure records and determine environmental impact events;

[0019] Based on the timestamp matching of environmental impact events and environmental data, the static correlation between environmental impact events and environmental factors is calculated; the calculation formula of the static correlation is:

[0020] ,in, It is a static correlation, reflecting the characteristics of the device Environmental factors The linear relationship and Respectively The device characteristics and environmental factor values ​​at each time point, and Device characteristics Environmental factors The mean of and Device characteristics Environmental factors The standard deviation of and is the adjustment coefficient, is the Pearson correlation coefficient, , is the number of time points;

[0021] Verify the static correlation, eliminate interference factors caused by data anomalies or noise, and obtain a static correlation data set;

[0022] According to the static correlation data set, environmental factors whose static correlation is higher than a preset static correlation threshold are screened to form a key environmental factor data set.

[0023] Preferably, a time series analysis is performed based on the normalized device feature data and the normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between the device feature and the key environmental factor, including:

[0024] According to the normalized equipment feature data and normalized key environmental factor data, the time series matching relationship between equipment features and environmental factors is analyzed, a dynamic correlation function is constructed, and the dynamic correlation degree is calculated; the calculation formula of the dynamic correlation degree is:

[0025] ,in, For device features Environmental factors The dynamic correlation between For the The weight coefficient of each time point, , and is the weight coefficient, is the standard deviation ratio of equipment characteristics to environmental factors, is the weight coefficient;

[0026] According to a preset dynamic correlation threshold, the dynamic correlation degree is screened to obtain a dynamic correlation degree data set;

[0027] According to the dynamic correlation data set, each pair of equipment characteristics and key environmental factors is taken as an entry of the matrix to generate a joint feature matrix.

[0028] Preferably, the joint feature matrix is ​​divided into time segments of fixed windows in chronological order, and each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set, including:

[0029] According to the length and step size of the preset time window, the joint feature matrix is ​​sliced ​​in time order to divide the matrix into multiple time window sub-matrices, each of which contains a combination of device features and key environmental factors within a set of time periods;

[0030] Perform integrity check on each time window sub-matrix to ensure that each sub-matrix contains enough data entries to reflect the device status change, and obtain a complete time window sub-matrix;

[0031] The complete time window sub-matrices are numbered and stored to form a time segment feature dataset.

[0032] Preferably, according to the known operation status labels of the equipment, the time segment feature data set is classified, each time segment feature data is divided into a normal state feature set and an abnormal state feature set, and a classified feature data set is generated, including:

[0033] Extracting operation status tag information from the historical operation records of the equipment, where the operation status tag information includes a normal state and an abnormal state;

[0034] Match the running status label with the time segment feature data set through the timestamp to establish a mapping relationship between the time segment and the running status label;

[0035] Filter the data in the time segment feature data set by time period, classify the data mapped to the normal state label into the normal state feature set, and classify the data mapped to the abnormal state label into the abnormal state feature set;

[0036] The internal consistency of the normal state feature set and the abnormal state feature set is verified, and the abnormal data that does not conform to the corresponding operating state feature distribution is eliminated to generate a classification feature data set.

[0037] Preferably, feature contribution analysis is performed on each time segment feature data in the classification feature data set, the contribution of device features and key environmental factors to the classification results is calculated, and a feature weight matrix is ​​generated, wherein each entry of the feature weight matrix indicates the degree of influence of the feature on a specific classification label, including:

[0038] According to the characteristic data of each time segment in the classification characteristic data set, the mean difference and distribution overlap of the equipment characteristics and key environmental factors in normal and abnormal states are calculated respectively;

[0039] According to the mean difference and distribution overlap, the influence of each feature on the classification accuracy is evaluated to generate a feature contribution score. The calculation formula for the feature contribution score is:

[0040] ,in, For a set of device features and environmental factors Contribution score of is the weight value, reflecting the characteristics of the device Environmental factors The strength of association, and is the adjustment coefficient;

[0041] According to the preset contribution threshold, the feature contribution scores are screened and normalized to generate a feature weight matrix;

[0042] In the feature weight matrix, each entry includes the weight value of a specific device feature or key environmental factor to the classification label.

[0043] Preferably, according to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, fused feature data is generated layer by layer, and supervised learning training is performed on the fused feature data to build a device prediction model, including:

[0044] According to the weight values ​​in the feature weight matrix, the device features and key environmental factors in the time segment feature data set are weighted and calculated to generate a weighted feature data set;

[0045] According to the weighted feature data set, in a hierarchical and progressive feature fusion method, the device features and environmental factors are combined layer by layer according to the dynamic correlation degree to generate a fused feature data set;

[0046] Use the classification feature dataset as supervision information to train the fusion feature dataset and adjust the model parameters of the fusion feature;

[0047] Construct an equipment prediction model and output the equipment prediction model, which can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

[0048] In a second aspect, a prediction model building system for personalized equipment operation and maintenance is provided, the system comprising:

[0049] A data collection module is used to collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, wherein the equipment operation data includes the equipment operation status parameters, and the environmental data includes external influencing factors related to the equipment operation scenario;

[0050] The feature extraction module is used to extract the change characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, the environmental data is analyzed for correlation, the main influencing factors are screened out, and a key environmental factor data set is generated;

[0051] A normalization processing module is used to pre-process the preliminary equipment feature set and the key environmental factor data set respectively to generate normalized equipment feature data and normalized key environmental factor data;

[0052] A correlation analysis module, used to perform time series analysis based on normalized device feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between the device feature and the key environmental factor;

[0053] A data segmentation module is used to segment the joint feature matrix into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set.

[0054] A classification feature generation module is used to classify the time segment feature data set according to the known operation status labels of the equipment, divide each time segment feature data into a normal state feature set and an abnormal state feature set, and generate a classification feature data set;

[0055] The feature weight matrix generation module is used to perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label;

[0056] A fusion feature generation module is used to recombine the device features and environmental factors in the time segment feature data set according to the feature weight matrix and generate fusion feature data layer by layer;

[0057] The model generation module is used to perform supervised learning training on the fused feature data, build a device prediction model, and output the device prediction model as a prediction model for personalized device operation and maintenance.

[0058] The above solution of the present invention includes at least the following beneficial effects:

[0059] The present invention improves the generalization ability of the model. In the prior art, the prediction model for equipment operation and maintenance usually relies on historical operation data for analysis, and predicts possible failures or abnormalities of the equipment by building a statistical model or a machine learning model. These technologies mainly include prediction methods based on regression analysis, time series analysis, and classification algorithms. However, in practical applications, a single model may not have sufficient generalization ability when faced with diverse equipment types and operation and maintenance environments, and it is difficult to accurately predict the status of the equipment in different scenarios. For example, in the operation and maintenance of wind turbines, differences in climatic conditions and geographical locations may cause significant changes in data distribution. Existing models are difficult to effectively capture these characteristics, resulting in a decrease in prediction accuracy.

[0060] The present invention introduces dynamic correlation analysis between environmental factors and equipment characteristics, so that the model can more comprehensively consider the complex interactive relationship between the equipment operating status and the external environment. By jointly analyzing environmental factors such as temperature, humidity, atmospheric pressure, etc., and equipment characteristics such as vibration, temperature, current, etc., a richer feature set is generated, which enhances the model's adaptability to different environmental conditions. This not only improves the generalization ability of the model, but also ensures its stability and accuracy in different application scenarios.

[0061] The present invention can capture potential hysteresis effects. Traditional prediction models often only focus on the changing trends of the equipment operation data itself, ignoring the impact of environmental factors on the equipment status, especially those with hysteresis. For example, extreme weather conditions may have a significant impact on equipment performance after hours or even days. Existing technologies have difficulty capturing this non-instant but real causal relationship, which affects the accuracy of the prediction.

[0062] The present invention calculates the dynamic correlation between equipment characteristics and key environmental factors, takes into account the correlation under different delays, and reveals the complex interaction pattern between the two that changes over time. This improvement enables the model to more accurately capture potential hysteresis effects, warn of potential risk signals in advance, and provide strong support for preventive maintenance. For example, in the operation and maintenance of wind turbines, dynamic correlation analysis can identify which environmental conditions, such as low temperature and high humidity, may lead to an increased risk of equipment failure in the next few days, so that corresponding preventive measures can be taken.

[0063] The present invention enhances the effectiveness of feature selection. The feature selection methods in the prior art are mainly based on static correlation analysis, which fails to fully consider the changing trend of features in different time periods and their impact on classification results. This may cause some important features to be ignored, while insignificant features are retained, reducing the prediction performance of the model.

[0064] The present invention introduces a feature contribution scoring mechanism to evaluate the influence of each feature on classification accuracy and generate a feature weight matrix. This method not only takes into account the mean difference and distribution overlap of features in normal and abnormal states, but also introduces variance differences to measure the fluctuation of features in different states. In this way, the importance of each feature can be evaluated more comprehensively to ensure that the final selected feature set has stronger discrimination and representativeness. This not only improves the classification accuracy of the model, but also enhances its robustness and generalization ability.

[0065] The present invention improves real-time monitoring and intelligent management. Although the existing technology can achieve a certain degree of equipment status prediction, it is still insufficient in real-time monitoring and intelligent management. Due to the lack of in-depth analysis of environmental factors, the existing model is difficult to provide timely and accurate risk warning information, which increases the risk of equipment failure.

[0066] The present invention combines environmental factors with equipment features to generate fused feature data, and conducts supervised learning training to construct an equipment prediction model. The model can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data. In this way, real-time monitoring and intelligent management of equipment status can be achieved, potential faults can be warned in advance, maintenance costs can be reduced, and the operating efficiency and reliability of the equipment can be improved. In addition, by continuously optimizing feature selection and model parameter adjustment, the prediction performance of the model can be further improved to meet the needs of different application scenarios.

[0067] In summary, the present invention constructs a more flexible and adaptable prediction model by introducing dynamic correlation analysis between environmental factors and equipment characteristics, which significantly improves the accuracy and reliability of equipment operation and maintenance prediction. This not only solves the data adaptability problem existing in the prior art, but also provides strong support for achieving efficient fault prediction and risk warning, which helps to promote the intelligent development of equipment operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flowchart of a method for building a prediction model for personalized equipment operation and maintenance provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0070] like Figure 1As shown, an embodiment of the present invention proposes a prediction model construction method for personalized equipment operation and maintenance, the method comprising:

[0071] Collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, where the equipment operation data includes the equipment's operating status parameters, and the environmental data includes external influencing factors related to the equipment's operating scenario;

[0072] Extract the changing characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, perform correlation analysis on the environmental data, screen out the main influencing factors, and generate a key environmental factor data set;

[0073] Preprocessing the preliminary equipment characteristic set and the key environmental factor data set respectively to generate normalized equipment characteristic data and normalized key environmental factor data;

[0074] Performing time series analysis based on normalized equipment feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between equipment features and key environmental factors;

[0075] The joint feature matrix is ​​divided into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set.

[0076] According to the known running status labels of the equipment, the time segment feature data set is classified, and each time segment feature data is divided into a normal state feature set and an abnormal state feature set to generate a classified feature data set;

[0077] Perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label;

[0078] According to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, and fused feature data is generated layer by layer. The fused feature data is then trained with supervised learning to build a device prediction model.

[0079] The equipment prediction model is output as a prediction model for personalized equipment operation and maintenance. The prediction model can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

[0080] In the embodiment of the present invention, the device operation data and environmental data are collected in real time from multiple sensors and environmental monitoring devices, and these data are associated by timestamp to form an initial data set. The device operation data includes but is not limited to the device's operating state parameters such as temperature, pressure, vibration frequency, etc.; the environmental data covers external influencing factors such as temperature, humidity, atmospheric pressure, etc.

[0081] To ensure the integrity and accuracy of the data, all data must be strictly aligned with the timestamp. Timestamp alignment not only ensures that the data at the same moment can accurately correspond, but also provides a solid foundation for subsequent analysis. In practical applications, this step is crucial because it directly affects the prediction accuracy of the final model. For example, in an industrial environment, the state change of equipment is often affected by environmental conditions, and this influence may have a certain lag. Therefore, accurate timestamp matching can effectively capture these potential correlations, thereby improving the accuracy of fault prediction.

[0082] For the equipment operation status parameters, the time series analysis method is used to count the change amplitude, change frequency and stability, and quantify the equipment status characteristics. Taking the vibration parameters of a certain industrial equipment as an example, after calculating its mean, variance and extreme value range, it can be found that its vibration frequency fluctuates significantly before the failure, providing an early warning basis for potential problems.

[0083] In addition, based on historical fault records, we further analyze events related to environmental factors, such as the probability of equipment failure under specific humidity or extreme temperature conditions. Through timestamp matching, we combine environmental data with equipment status data, calculate static correlation, and screen out environmental factors with a high correlation with faults, thus forming a key environmental factor data set. This method helps improve the quality of data and the representativeness of features.

[0084] The preliminary equipment feature set and key environmental factor data are cleaned and normalized to ensure that the feature data are within the same numerical range.

[0085] This processing method solves the impact of scale differences between features on model training. For example, the vibration feature value range of a device is 0.1-10, while the ambient temperature feature value range is 20-50. After normalization, the values ​​of all features are converted to a unified range, making model training more stable and efficient.

[0086] Based on the normalized device feature data, the dynamic correlation between device features and key environmental factors is calculated to generate a joint feature matrix. Each entry in the matrix reflects the correlation between a specific device feature and a key environmental factor within a specific time range. For example, by analyzing the dynamic changes between vibration and temperature, the irregular fluctuation of the vibration frequency of the device at high temperature is found, indicating the possibility of potential failure. The generation of this joint feature provides multi-dimensional correlation information for subsequent model training.

[0087] The joint feature matrix is ​​divided into fixed window time segments in chronological order. Each segment contains the combination of equipment features and environmental factors in a specific time period. For example, for a certain device, a 10-minute time window can be selected to form a sub-dataset that reflects short-term state changes. The division of time segments helps the model capture the operating state characteristics in a short period of time more accurately and improves the ability to respond to dynamic changes.

[0088] Using the device's operating status label, normal or abnormal, the time segment feature data is divided into a normal state feature set and an abnormal state feature set. This classification method provides a clear goal for subsequent supervised learning. For example, by marking the abnormal state of the device in a certain time period, the optimization direction of the model in anomaly detection can be clarified, thereby significantly improving the classification accuracy.

[0089] By calculating the impact of equipment features and key environmental factors on classification results, a feature weight matrix is ​​generated. For example, if the operating status of a device is mainly affected by temperature and vibration, these features will have a higher weight value in the weight matrix. This weight analysis highlights the role of key features while weakening the interference of secondary features on model results, making the model more accurate.

[0090] According to the feature weight matrix, the device features and environmental factors in the time segment feature data are fused layer by layer according to the weights to generate high-dimensional feature data. Subsequently, the prediction model is trained through supervised learning methods. This multi-level feature fusion method fully explores the complex relationship between device features and environmental factors. For example, after the humidity and vibration features of a certain device are fused, the fault prediction accuracy can be significantly improved.

[0091] Ultimately, the model can output prediction results of equipment status and potential risk warning information based on real-time input equipment operation data and environmental data. For example, based on real-time collected vibration data, the probability of bearing wear in the equipment in the future can be predicted, providing accurate warnings for operation and maintenance personnel. This real-time prediction capability significantly improves the intelligence level of equipment operation and maintenance, and reduces the risk of failure and maintenance costs.

[0092] More specifically, the preliminary equipment feature set and the key environmental factor data set are preprocessed respectively to generate normalized equipment feature data and normalized key environmental factor data, including:

[0093] Data cleaning: missing values ​​and outlier detection are performed on the preliminary equipment feature set and key environmental factor data set. Missing values ​​can be filled by interpolation or using the mean / median; outliers can be identified by box plot method or Z-score method, and can be removed or corrected according to the actual situation.

[0094] Normalization: convert all feature values ​​to between 0 and 1.

[0095] Verify the normalization effect: Ensure that the data distribution after normalization is reasonable to avoid extreme values ​​affecting model training. Visualize the data distribution before and after normalization to check whether it meets expectations.

[0096] In a preferred embodiment of the present invention, the change characteristics of the equipment operation status are extracted from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, the environmental data is subjected to correlation analysis to screen out the main influencing factors and generate a key environmental factor data set, including:

[0097] Extract equipment operating status parameters from the initial data set, analyze the change trend of the parameters in time series, calculate the change amplitude, change frequency and stability, extract the preliminary characteristics of the equipment status, and form a preliminary equipment feature set;

[0098] Analyze events related to environmental factors based on historical equipment failure records and determine environmental impact events;

[0099] Based on the timestamp matching of environmental impact events and environmental data, the static correlation between environmental impact events and environmental factors is calculated; the calculation formula of the static correlation is:

[0100] ,in, It is a static correlation, reflecting the characteristics of the device Environmental factors The linear relationship and Respectively The device characteristics and environmental factor values ​​at each time point, and Device characteristics Environmental factors The mean of and Device characteristics Environmental factors The standard deviation of and is the adjustment coefficient, is the Pearson correlation coefficient, , is the number of time points;

[0101] Verify the static correlation, eliminate interference factors caused by data anomalies or noise, and obtain a static correlation data set;

[0102] According to the static correlation data set, environmental factors whose static correlation is higher than a preset static correlation threshold are screened to form a key environmental factor data set.

[0103] In the embodiment of the present invention, by analyzing the change trend of the equipment operating status parameters, characteristic indicators such as change amplitude, change frequency and stability are extracted. These characteristics quantify the performance of the equipment status in different time periods. For example, the stability change of vibration parameters can indicate the potential wear of the equipment. This process effectively eliminates irrelevant or low-correlation features, ensuring the high quality of the model input data.

[0104] Based on historical fault records, we analyze fault events related to environmental data, match environmental impact events with environmental data through timestamps, calculate the static correlation between environmental factors and equipment status, and screen out environmental factors (such as temperature, humidity, etc.) that have a significant impact on equipment operation. This not only improves the model's adaptability to environmental data, but also reduces the interference of redundant data on model performance.

[0105] After the static correlation was verified by data anomaly and noise elimination, the accuracy and robustness of the screening results were ensured. The key environmental factor data set finally obtained significantly improved the validity of the data and provided a solid foundation for subsequent dynamic correlation analysis and predictive modeling.

[0106] More specifically, the equipment operating status parameters are extracted from the initial data set, and the change trend of the parameters is analyzed in time series. The change amplitude, change frequency and stability are statistically analyzed to extract the preliminary characteristics of the equipment status and form a preliminary equipment feature set, which specifically includes:

[0107] Parameter extraction: Extract equipment operating status parameters such as temperature, pressure, vibration frequency, etc. from the data collected by sensors in real time.

[0108] Time series analysis: Perform time series analysis on each device feature to count its change amplitude, change frequency and stability.

[0109] Change: Calculates the difference between the maximum and minimum values.

[0110] Frequency of change: Calculate the number of significant changes that occur per unit time.

[0111] Stability: Calculate the standard deviation or variance to assess the fluctuation of the parameters.

[0112] Feature quantification: Quantify the above statistical results into specific values ​​to form a preliminary device feature set.

[0113] For example, for the vibration frequency of a machine, you can calculate its mean, variance, maximum and minimum difference and other statistics to evaluate the performance of the equipment in different time periods.

[0114] More specifically, based on historical equipment failure records, analyze events related to environmental factors and determine environmental impact events, including:

[0115] Collect historical failure records of the equipment, especially those that clearly mark the cause of the failure.

[0116] Analyze whether these failures are related to specific environmental conditions, such as extreme temperature, high humidity, etc. Match the timestamps in the historical failure records with the timestamps of the environmental data to find out the environmental conditions in the period before and after the failure occurred.

[0117] For example, if a piece of equipment fails after a cold spell, you would want to examine the temperature data for that time period.

[0118] In a preferred embodiment of the present invention, time series analysis is performed based on normalized device feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between device features and key environmental factors, including:

[0119] According to the normalized equipment feature data and normalized key environmental factor data, the time series matching relationship between equipment features and environmental factors is analyzed, a dynamic correlation function is constructed, and the dynamic correlation degree is calculated; the calculation formula of the dynamic correlation degree is:

[0120] ,in, For device features Environmental factors The dynamic correlation between For the The weight coefficient of each time point, , and is the weight coefficient, is the standard deviation ratio of equipment characteristics to environmental factors, is the weight coefficient;

[0121] According to a preset dynamic correlation threshold, the dynamic correlation degree is screened to obtain a dynamic correlation degree data set;

[0122] According to the dynamic correlation data set, each pair of equipment characteristics and key environmental factors is taken as an entry of the matrix to generate a joint feature matrix.

[0123] In the embodiment of the present invention, the calculation of dynamic correlation analyzes the matching relationship between device features and environmental factors in the time series, and mines the change trend and mutual influence of the two in the time dimension. For example, the immediate impact of temperature change on vibration frequency can be intuitively expressed through dynamic correlation, thereby enhancing the model's understanding of time-varying characteristics.

[0124] The joint feature matrix expresses the dynamic association between device features and key environmental factors in matrix form, with each entry representing the description of the association of a specific feature within a time segment. Compared with analyzing device or environmental features separately, the matrix effectively integrates the time series information of device status and environmental changes, providing richer input data for the prediction model.

[0125] The dynamic correlation screening process eliminates low-correlation or irrelevant data by setting the dynamic correlation threshold, so that the generated feature matrix can better reflect the dynamic characteristics of the equipment in a complex operating environment, thereby improving the adaptability and generalization ability of the model in diversified operation and maintenance scenarios.

[0126] In a preferred embodiment of the present invention, the joint feature matrix is ​​divided into time segments of fixed windows in chronological order, and each time segment corresponds to a group of device features and key environment features, forming a time segment feature data set, including:

[0127] According to the length and step size of the preset time window, the joint feature matrix is ​​sliced ​​in time order to divide the matrix into multiple time window sub-matrices, each of which contains a combination of device features and key environmental factors within a set of time periods;

[0128] Perform integrity check on each time window sub-matrix to ensure that each sub-matrix contains enough data entries to reflect the device status change, and obtain a complete time window sub-matrix;

[0129] The complete time window sub-matrices are numbered and stored to form a time segment feature dataset.

[0130] More specifically, according to the dynamic correlation data set, each pair of device features and key environmental factors is used as an entry in the matrix to generate a joint feature matrix, which specifically includes:

[0131] The rows of the joint feature matrix represent the device features , columns represent key environmental factors The entries of the matrix represent the device characteristics Key environmental factors Dynamic correlation value The dimension of the joint feature matrix M is set to i×j, that is, the number of rows is the number of device features i, and the number of columns is the number of environmental factors j.

[0132] In the embodiment of the present invention, the joint feature matrix is ​​divided into fixed window time segments in chronological order, and each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set. This process involves the selection and slicing of time windows, the purpose of which is to better capture the trend and pattern of device status changes over time.

[0133] The joint feature matrix is ​​sliced ​​according to the preset time window length and step size. The time window length determines the amount of data contained in each slice, while the step size controls the overlap between adjacent slices. Choosing the right time window length and step size is very important, as it requires balancing two requirements: on the one hand, ensuring that each slice contains enough data entries to reflect device state changes; on the other hand, avoiding excessive computational burden due to excessive redundant information.

[0134] In order to ensure the integrity of each time segment, each time window sub-matrix is ​​also checked to ensure that it contains a complete combination of device characteristics and environmental factors. If there are missing data or outliers in a time window, appropriate processing is required, such as interpolation filling or excluding the segment. This ensures the accuracy and reliability of subsequent analysis. After completing the above steps, each time window sub-matrix is ​​numbered and stored to form the final time segment feature data set.

[0135] This method can decompose continuous time series data into multiple independent time segments, each of which contains a combination of equipment characteristics and environmental factors within a specific time period. This not only simplifies the data processing process, but also provides more targeted input for subsequent classification and prediction tasks. For example, in equipment failure prediction, by analyzing the changes in characteristics within different time segments, potential risk signals can be discovered earlier, so that preventive measures can be taken.

[0136] More specifically, each time window sub-matrix is ​​checked for integrity to ensure that each sub-matrix contains enough data entries to reflect the device state changes, and a complete time window sub-matrix is ​​obtained, including:

[0137] Each time window sub-matrix is ​​checked for completeness to ensure that it contains the complete combination of device characteristics and environmental factors.

[0138] If there are missing data or outliers within a time window, appropriate processing is required, such as interpolation filling or removing the segment.

[0139] Make sure that each time window sub-matrix contains at least a certain number of data entries to ensure that it can fully reflect the device state changes.

[0140] In a preferred embodiment of the present invention, according to the known operation status label of the device, the time segment feature data set is classified, each time segment feature data is divided into a normal state feature set and an abnormal state feature set, and a classified feature data set is generated, including:

[0141] Extracting operation status tag information from the historical operation records of the equipment, where the operation status tag information includes a normal state and an abnormal state;

[0142] Match the running status label with the time segment feature data set through the timestamp to establish a mapping relationship between the time segment and the running status label;

[0143] Filter the data in the time segment feature data set by time period, classify the data mapped to the normal state label into the normal state feature set, and classify the data mapped to the abnormal state label into the abnormal state feature set;

[0144] The internal consistency of the normal state feature set and the abnormal state feature set is verified, and the abnormal data that does not conform to the corresponding operating state feature distribution is eliminated to generate a classification feature data set.

[0145] In the embodiment of the present invention, first, the operation status label information is extracted from the historical operation records of the equipment. These labels usually include two categories: normal state and abnormal state. The operation status label is matched with the time segment feature data set through the timestamp to establish a one-to-one mapping relationship. For example, if the operation status of the equipment in a certain time segment is marked as normal, all feature data of the segment will be classified into the normal state feature set; conversely, if it is marked as abnormal, it will be classified into the abnormal state feature set.

[0146] In order to ensure the accuracy of the classification results, it is also necessary to verify the internal consistency of the normal state feature set and the abnormal state feature set. Specifically, it will check whether each feature set conforms to the feature distribution law of the corresponding operating state. If there is any non-compliance, it is necessary to further investigate the cause and make adjustments. Doing so can not only improve the accuracy of classification, but also enhance the robustness and generalization ability of the model.

[0147] Finally, through the above steps, a classification feature dataset is obtained, in which each time segment is clearly marked as normal or abnormal. Such a classification feature dataset provides valuable data support for subsequent feature contribution analysis and model training. By comparing and analyzing the features under different states, we can more clearly understand which features play a key role in distinguishing normal and abnormal states, thereby optimizing model performance.

[0148] More specifically, the operation status label is matched with the time segment feature data set through the timestamp to establish a mapping relationship between the time segment and the operation status label, including:

[0149] Operation status tag extraction: Extract operation status tag information from the historical operation records of the equipment. These tags usually include normal status, such as "normal operation" and abnormal status, such as "fault" or "warning". Tags should contain clear timestamp information for subsequent matching.

[0150] Timestamp matching: Exactly match the timestamp in the extracted running status label with the timestamp in the time segment feature dataset. For each time segment, find the corresponding running status label. For example, if the time range of a time segment is [t1, t2], find out whether there is a clear running status label in this time period. If multiple labels cover the same time segment, select the label closest to the center point of the time segment as the representative.

[0151] Mapping relationship establishment: Through the above matching process, a one-to-one correspondence between time segments and operating status labels is established. Each time segment is associated with a specific operating status label, normal or abnormal. Ensure the accuracy of the mapping relationship and avoid misclassification due to inconsistent timestamps.

[0152] More specifically, the data in the time segment feature data set is filtered by time period, and the data mapped to the normal state label is classified into the normal state feature set, and the data mapped to the abnormal state label is classified into the abnormal state feature set, which specifically includes:

[0153] Filter by time period: According to the mapping relationship, filter the data in the time segment feature data set by time period. Classify the data mapped to the normal state label into the normal state feature set, and classify the data mapped to the abnormal state label into the abnormal state feature set. Ensure that the data entries in each feature set have clear operating status labels to facilitate subsequent analysis.

[0154] Internal consistency verification: Internal consistency verification is performed on the normal state feature set and the abnormal state feature set respectively. Check whether each feature set conforms to the characteristic distribution law of the corresponding operating state. For example, some feature values ​​should remain stable under normal conditions, but may show significant fluctuations under abnormal conditions. Use statistical methods such as mean, variance, distribution graph, etc. to evaluate the internal consistency of the feature set and eliminate abnormal data that does not conform to the corresponding operating state characteristic distribution. For data entries that are suspected to be abnormal, their contextual information, such as the state labels of adjacent time segments, can be further checked to confirm their true state.

[0155] Generate a classification feature dataset: After screening and verification, a classification feature dataset is finally obtained, in which each time segment is clearly marked as normal or abnormal. Such a classification feature dataset provides valuable data support for subsequent feature contribution analysis and model training.

[0156] More specifically, internal consistency verification is performed on the normal state feature set and the abnormal state feature set, abnormal data that does not conform to the corresponding operating state feature distribution is eliminated, and a classification feature data set is generated, which specifically includes:

[0157] Feature distribution analysis: Calculate the basic statistics of each feature in normal and abnormal states, such as mean, variance, maximum, minimum, etc. Draw a histogram, box plot, or density plot to intuitively display the distribution of feature values. For time series features, draw a time series graph to observe its changing trend and periodicity.

[0158] Through these analyses, we can preliminarily understand the typical performance of each feature under normal and abnormal conditions, and provide a reference for subsequent consistency verification.

[0159] Normal state consistency check: Calculate the mean and standard deviation of each feature in the normal state. Define a reasonable confidence interval and treat feature values ​​outside this interval as potential abnormal data. Use the Z-score method to further filter out abnormal values. Alternatively, use Kullback-Leibler divergence (KL divergence) or Jensen-Shannon distance (JS distance) to measure the difference between the feature distribution of a single time segment and the overall normal state feature distribution. If the difference exceeds the preset threshold, the data of the time segment is considered to be inconsistent with the normal state feature distribution.

[0160] Consistency check of abnormal state feature set: Similar to the normal state feature set, calculate the mean and standard deviation of each feature in the abnormal state. Define a reasonable confidence interval and treat feature values ​​outside this interval as potential abnormal data. Use the Z-score method to further filter out abnormal values. Alternatively, use KL divergence or JS distance to measure the difference between the feature distribution of a single time segment and the overall abnormal state feature distribution. If the difference exceeds the preset threshold, the data of the time segment is considered to be inconsistent with the abnormal state feature distribution.

[0161] Purpose of removing abnormal data: Data entries that are obviously deviated from the normal or abnormal state feature distribution are directly removed from the feature set. If only some features of some data entries are abnormal, you can choose to keep the data entry, but mark its abnormal features so that you can pay attention to it in subsequent analysis. For missing values ​​caused by removing abnormal data, you can fill them through interpolation or use data from adjacent time segments to ensure the integrity of the feature set.

[0162] Generate classification feature data sets: Renumber the normal state feature sets and abnormal state feature sets after consistency verification to ensure that each time segment has a unique identifier. Merge the normal state feature sets and abnormal state feature sets into a complete classification feature data set for subsequent processing and analysis. Record the statistical information and verification results of each feature set, including the number of abnormal data removed, the verification method used, etc., for subsequent review and optimization.

[0163] Assume that you are processing the operation and maintenance data of an industrial device. The specific steps are as follows:

[0164] Feature distribution analysis: Calculate the mean and variance of features such as temperature, pressure, and vibration frequency under normal and abnormal conditions, and draw corresponding histograms and time series graphs.

[0165] Internal consistency check: Z-score analysis was performed on the temperature features in the normal state feature set, and it was found that the temperature values ​​of some time segments were much higher than the normal range, which were marked as potential outliers. KL divergence was used to measure the distribution differences of vibration frequency features in different time segments, and several time segments that did not conform to the normal state feature distribution were identified.

[0166] Eliminate abnormal data: Directly eliminate the time segments marked as potential abnormal values ​​to ensure the purity of the normal state feature set. For missing values ​​caused by eliminating abnormal data, use the average value of adjacent time segments for interpolation filling.

[0167] Generate a classification feature data set: renumber the normal state feature set and abnormal state feature set after consistency verification, and merge them into a complete classification feature data set. Record key information in the verification process, such as the number of abnormal data removed and the verification method used, to ensure the transparency and traceability of the data set.

[0168] In a preferred embodiment of the present invention, feature contribution analysis is performed on each time segment feature data in the classification feature data set, the contribution of device features and key environmental factors to the classification results is calculated, and a feature weight matrix is ​​generated, wherein each entry of the feature weight matrix indicates the degree of influence of the feature on a specific classification label, including:

[0169] According to the characteristic data of each time segment in the classification characteristic data set, the mean difference and distribution overlap of the equipment characteristics and key environmental factors in normal and abnormal states are calculated respectively;

[0170] According to the mean difference and distribution overlap, the influence of each feature on the classification accuracy is evaluated to generate a feature contribution score. The calculation formula for the feature contribution score is:

[0171] ,in, For a set of device features and environmental factors Contribution score of is the weight value, reflecting the characteristics of the device Environmental factors The strength of association, and is the adjustment coefficient;

[0172] According to the preset contribution threshold, the feature contribution scores are screened and normalized to generate a feature weight matrix;

[0173] In the feature weight matrix, each entry includes the weight value of a specific device feature or key environmental factor to the classification label.

[0174] In the embodiment of the present invention, the feature contribution analysis calculates the specific impact of device features and key environmental factors on the classification results, normal or abnormal. For example, certain features such as high humidity may significantly increase the probability of failure. Through the contribution score, such key features can be assigned a higher weight to enhance the model's sensitivity to them.

[0175] Through screening and normalization, the feature weight matrix reduces the dependence on low-contribution features and reduces the negative impact of redundant features on the model's computing performance and classification accuracy, allowing the model to focus more on important features that affect the device status.

[0176] The generation of the weight matrix not only improves the model's predictive performance, but also enhances the model's interpretability, allowing operators to intuitively understand the impact of specific features, such as vibration amplitude or ambient temperature, on the equipment status. This flexibility enables the model to be personalized and optimized for different equipment and operation and maintenance environments, improving its practicality and reliability in actual applications.

[0177] In a preferred embodiment of the present invention, according to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, fused feature data is generated layer by layer, and supervised learning training is performed on the fused feature data to construct a device prediction model, including:

[0178] According to the weight values ​​in the feature weight matrix, the device features and key environmental factors in the time segment feature data set are weighted and calculated to generate a weighted feature data set;

[0179] According to the weighted feature data set, in a hierarchical and progressive feature fusion method, the device features and environmental factors are combined layer by layer according to the dynamic correlation degree to generate a fused feature data set;

[0180] Use the classification feature dataset as supervision information to train the fusion feature dataset and adjust the model parameters of the fusion feature;

[0181] Construct an equipment prediction model and output the equipment prediction model, which can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

[0182] In an embodiment of the present invention, according to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, fused feature data are generated layer by layer, and supervised learning training is performed on the fused feature data to construct a device prediction model.

[0183] Specifically, according to the weight values ​​in the feature weight matrix, the device features and key environmental factors in the time segment feature data set are weighted to generate a weighted feature data set. In this way, the features that contribute more to the classification results can be highlighted, while the influence of minor features can be weakened, thereby improving the classification accuracy of the model.

[0184] Next, in a hierarchical feature fusion method, the device features and environmental factors are combined layer by layer according to the dynamic correlation to generate a fused feature data set. This process is similar to feature extraction and representation learning in neural networks. Through gradual abstraction and generalization, a high-level feature representation is finally formed. These fused features not only retain the main information of the original features, but also enhance the correlation and complementarity between features, providing more effective input for model training.

[0185] Use the classification feature dataset as supervision information to train the fusion feature dataset and adjust the model parameters of the fusion feature. Specifically, you can choose a variety of supervised learning algorithms, such as support vector machine (SVM), random forest (Random Forest) or deep neural network (DNN), to build a prediction model. During the training process, the model parameters are continuously optimized to maximize the classification performance and prediction accuracy.

[0186] Finally, a device prediction model is constructed and output as a prediction tool for personalized device operation and maintenance. This prediction model can generate device status prediction results and potential risk warning information based on real-time input device operation data and environmental data. In this way, real-time monitoring and intelligent management of device status can be achieved, potential failures can be warned in advance, maintenance costs can be reduced, and the operating efficiency and reliability of equipment can be improved.

[0187] More specifically, according to the weighted feature data set, in a hierarchical and progressive feature fusion method, the device features and environmental factors are combined layer by layer according to the dynamic correlation degree to generate a fused feature data set, which specifically includes:

[0188] The device features and environmental factors are combined layer by layer according to the dynamic correlation to generate a high-level feature representation.

[0189] The first layer: Combine the features with high direct correlation to form preliminary fusion features.

[0190] The second layer: Based on the initial fusion of features, more indirectly related features are introduced, gradually abstracted and generalized, and finally a high-level feature representation is formed.

[0191] For example, the first layer can combine two highly correlated environmental factors, temperature and humidity, and the second layer can add device features such as vibration frequency to form a more complex feature combination.

[0192] This layer-by-layer combination approach is similar to feature extraction and representation learning in neural networks, which enhances the correlation and complementarity between features and provides more effective input for model training.

[0193] More specifically, the classification feature dataset is used as supervision information to train the fusion feature dataset and adjust the model parameters of the fusion feature, including:

[0194] Use the classification feature dataset as supervision information: Use the classification feature dataset as supervision information to train the fusion feature dataset. Select an appropriate supervised learning algorithm, such as support vector machine (SVM), random forest (RandomForest), or deep neural network (DNN), to build a prediction model. During the training process, continuously optimize the model parameters to maximize the classification performance and prediction accuracy.

[0195] Adjust model parameters of fused features: Adjust model hyperparameters such as regularization coefficients and learning rates through methods such as cross-validation and grid search. Monitor the model training process to ensure that it does not overfit or underfit. For example, use the validation set to evaluate the generalization ability of the model and select the best model parameter configuration. Finally, build a device prediction model and output the model as a prediction tool for personalized device operation and maintenance.

[0196] Model evaluation and deployment: Use the test set to evaluate the performance of the model to ensure that it performs well on unseen data. Deploy the model to the actual application environment, receive equipment operation data and environmental data in real time, and generate equipment status prediction results and potential risk warning information. Through continuous monitoring and feedback mechanisms, continuously optimize the model to improve its prediction performance and reliability.

[0197] The present invention also provides a prediction model construction system for personalized equipment operation and maintenance, the system comprising:

[0198] A data collection module is used to collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, wherein the equipment operation data includes the equipment operation status parameters, and the environmental data includes external influencing factors related to the equipment operation scenario;

[0199] The feature extraction module is used to extract the change characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, the environmental data is analyzed for correlation, the main influencing factors are screened out, and a key environmental factor data set is generated;

[0200] A normalization processing module is used to pre-process the preliminary equipment feature set and the key environmental factor data set respectively to generate normalized equipment feature data and normalized key environmental factor data;

[0201] A correlation analysis module, used to perform time series analysis based on normalized device feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between the device feature and the key environmental factor;

[0202] A data segmentation module is used to segment the joint feature matrix into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set.

[0203] A classification feature generation module is used to classify the time segment feature data set according to the known operation status labels of the equipment, divide each time segment feature data into a normal state feature set and an abnormal state feature set, and generate a classification feature data set;

[0204] The feature weight matrix generation module is used to perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label;

[0205] A fusion feature generation module is used to recombine the device features and environmental factors in the time segment feature data set according to the feature weight matrix and generate fusion feature data layer by layer;

[0206] The model generation module is used to perform supervised learning training on the fused feature data, build a device prediction model, and output the device prediction model as a prediction model for personalized device operation and maintenance.

[0207] It should be noted that the system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0208] The embodiment of the present invention further provides a computing device, including: a processor, a memory storing a computer program, and when the computer program is executed by the processor, the method described above is executed. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0209] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the above method. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0210] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for constructing a prediction model for personalized equipment operation and maintenance, characterized in that: The method comprises: Collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, where the equipment operation data includes the equipment's operating status parameters, and the environmental data includes external influencing factors related to the equipment's operating scenario; Extract the changing characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, perform correlation analysis on the environmental data, screen out the main influencing factors, and generate a key environmental factor data set; Preprocessing the preliminary equipment characteristic set and the key environmental factor data set respectively to generate normalized equipment characteristic data and normalized key environmental factor data; Performing time series analysis based on normalized equipment feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between equipment features and key environmental factors; The joint feature matrix is ​​divided into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set. According to the known running status labels of the equipment, the time segment feature data set is classified, and each time segment feature data is divided into a normal state feature set and an abnormal state feature set to generate a classified feature data set; Perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label; According to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, and fused feature data is generated layer by layer. The fused feature data is then trained with supervised learning to build a device prediction model. The output device prediction model is a prediction model for personalized equipment operation and maintenance. The prediction model can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

2. The method for constructing a prediction model for personalized equipment operation and maintenance according to claim 1, characterized in that: Extract the change characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, perform correlation analysis on the environmental data, screen out the main influencing factors, and generate a key environmental factor data set, including: Extract equipment operating status parameters from the initial data set, analyze the change trend of the parameters in time series, calculate the change amplitude, change frequency and stability, extract the preliminary characteristics of the equipment status, and form a preliminary equipment feature set; Analyze events related to environmental factors based on historical equipment failure records and determine environmental impact events; Based on the timestamp matching of environmental impact events and environmental data, the static correlation between environmental impact events and environmental factors is calculated; the calculation formula of the static correlation is: ,in, It is a static correlation, reflecting the characteristics of the device Environmental factors The linear relationship and Respectively The device characteristics and environmental factor values ​​at each time point, and Device characteristics Environmental factors The mean of and Device characteristics Environmental factors The standard deviation of and is the adjustment coefficient, is the Pearson correlation coefficient, , is the number of time points; Verify the static correlation, eliminate interference factors caused by data anomalies or noise, and obtain a static correlation data set; According to the static correlation data set, environmental factors whose static correlation is higher than a preset static correlation threshold are screened to form a key environmental factor data set.

3. A method for constructing a prediction model for personalized equipment operation and maintenance according to claim 2, characterized in that: Based on the normalized equipment feature data and the normalized key environmental factor data, a time series analysis is performed to generate a joint feature matrix, where each entry in the matrix is ​​generated by calculating the dynamic correlation function between the equipment feature and the key environmental factor, including: According to the normalized equipment feature data and normalized key environmental factor data, the time series matching relationship between equipment features and environmental factors is analyzed, a dynamic correlation function is constructed, and the dynamic correlation degree is calculated; the calculation formula of the dynamic correlation degree is: ,in, For device features Environmental factors The dynamic correlation between For the The weight coefficient of each time point, , and is the weight coefficient, is the standard deviation ratio of equipment characteristics to environmental factors, is the weight coefficient; According to a preset dynamic correlation threshold, the dynamic correlation degree is screened to obtain a dynamic correlation degree data set; According to the dynamic correlation data set, each pair of equipment characteristics and key environmental factors is taken as an entry of the matrix to generate a joint feature matrix.

4. A method for constructing a prediction model for personalized equipment operation and maintenance according to claim 3, characterized in that: The joint feature matrix is ​​divided into fixed window time segments in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set, including: According to the length and step size of the preset time window, the joint feature matrix is ​​sliced ​​in time order to divide the matrix into multiple time window sub-matrices, each of which contains a combination of device features and key environmental factors within a set of time periods; Perform integrity check on each time window sub-matrix to obtain the complete time window sub-matrix; The complete time window sub-matrices are numbered and stored to form a time segment feature dataset.

5. A method for constructing a prediction model for personalized equipment operation and maintenance according to claim 4, characterized in that: According to the known running status labels of the equipment, the time segment feature data set is classified, and each time segment feature data is divided into a normal state feature set and an abnormal state feature set to generate a classified feature data set, including: Extracting operation status tag information from the historical operation records of the equipment, where the operation status tag information includes a normal state and an abnormal state; Match the running status label with the time segment feature data set through the timestamp to establish a mapping relationship between the time segment and the running status label; Filter the data in the time segment feature data set by time period, classify the data mapped to the normal state label into the normal state feature set, and classify the data mapped to the abnormal state label into the abnormal state feature set; The internal consistency of the normal state feature set and the abnormal state feature set is verified, and the abnormal data that does not conform to the corresponding operating state feature distribution is eliminated to generate a classification feature data set.

6. A method for constructing a prediction model for personalized equipment operation and maintenance according to claim 5, characterized in that: Perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label, including: According to the characteristic data of each time segment in the classification characteristic data set, the mean difference and distribution overlap of the equipment characteristics and key environmental factors in normal and abnormal states are calculated respectively; According to the mean difference and distribution overlap, the influence of each feature on the classification accuracy is evaluated to generate a feature contribution score. The calculation formula for the feature contribution score is: ,in, For a set of device features and environmental factors Contribution score of is the weight value, reflecting the characteristics of the device Environmental factors The strength of association, and is the adjustment coefficient; According to the preset contribution threshold, the feature contribution scores are screened and normalized to generate a feature weight matrix; In the feature weight matrix, each entry includes the weight value of a specific device feature or key environmental factor to the classification label.

7. A method for constructing a prediction model for personalized equipment operation and maintenance according to claim 6, characterized in that: According to the feature weight matrix, the device features and environmental factors in the time segment feature data set are recombined according to the feature weights, and fused feature data is generated layer by layer. The fused feature data is then trained with supervised learning to build a device prediction model, including: According to the weight values ​​in the feature weight matrix, the device features and key environmental factors in the time segment feature data set are weighted and calculated to generate a weighted feature data set; According to the weighted feature data set, in a hierarchical and progressive feature fusion method, the device features and environmental factors are combined layer by layer according to the dynamic correlation degree to generate a fused feature data set; Use the classification feature dataset as supervision information to train the fusion feature dataset and adjust the model parameters of the fusion feature; Construct an equipment prediction model and output the equipment prediction model, which can generate equipment status prediction results and potential risk warning information based on real-time input equipment operation data and environmental data.

8. A prediction model building system for personalized equipment operation and maintenance, characterized in that: Applied to the method according to any one of claims 1 to 7, the system comprises: A data collection module is used to collect equipment operation data and environmental data, and associate the equipment operation data with the environmental data by timestamp to form an initial data set, wherein the equipment operation data includes the equipment operation status parameters, and the environmental data includes external influencing factors related to the equipment operation scenario; The feature extraction module is used to extract the change characteristics of the equipment operation status from the initial data set to form a preliminary equipment feature set. At the same time, based on the historical equipment failure records, the environmental data is analyzed for correlation, the main influencing factors are screened out, and a key environmental factor data set is generated; A normalization processing module is used to pre-process the preliminary equipment feature set and the key environmental factor data set respectively to generate normalized equipment feature data and normalized key environmental factor data; A correlation analysis module, used to perform time series analysis based on normalized device feature data and normalized key environmental factor data to generate a joint feature matrix, wherein each entry in the matrix is ​​generated by calculating a dynamic correlation function between the device feature and the key environmental factor; A data segmentation module is used to segment the joint feature matrix into time segments of fixed windows in chronological order. Each time segment corresponds to a combination of a set of device features and key environmental features to form a time segment feature data set. A classification feature generation module is used to classify the time segment feature data set according to the known operation status labels of the equipment, divide each time segment feature data into a normal state feature set and an abnormal state feature set, and generate a classification feature data set; The feature weight matrix generation module is used to perform feature contribution analysis on each time segment feature data in the classification feature data set, calculate the contribution of device features and key environmental factors to the classification results, and generate a feature weight matrix, where each entry in the feature weight matrix indicates the degree of influence of the feature on a specific classification label; A fusion feature generation module is used to recombine the device features and environmental factors in the time segment feature data set according to the feature weight matrix and generate fusion feature data layer by layer; The model generation module is used to perform supervised learning training on the fused feature data, build a device prediction model, and output the device prediction model as a prediction model for personalized device operation and maintenance.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images