An online real-time fault diagnosis and predictive maintenance platform for industrial big data

Abstract

The application discloses an online real-time fault diagnosis and predictive maintenance platform for industrial big data, and particularly relates to the technical field of fault monitoring, and is used for solving the phenomenon that the existing machine learning model may have inconsistent prediction output in the actual industrial application process; the platform collects and records real-time monitoring data from multiple industrial equipment sources, uses multiple pre-trained machine learning models to predict the data after preprocessing, and performs conflict detection on the model prediction output; when the prediction inconsistency is confirmed, the platform performs nonlinear comparison on the information entropy and phase characteristics based on the historical operation scene and current working condition data, identifies abnormal phenomena caused by insufficient model adaptability, and dynamically adjusts the weights of the models accordingly, so that accurate identification of the equipment fault type and prediction of the remaining useful life are realized, and the accuracy of fault prediction and the reliability of system operation are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault monitoring, and more particularly, to an online real-time fault diagnosis and predictive maintenance platform for industrial big data. BACKGROUND

[0002] At present, online real-time fault diagnosis and predictive maintenance of industrial data mainly adopts technical means such as data acquisition, data preprocessing, model training and fault prediction to monitor and analyze the running state of industrial equipment; in order to improve the accuracy of fault prediction, a plurality of machine learning models are usually used, including deep learning network, neural network and multi-task learning model, to model and predict the real-time monitoring data of industrial equipment.

[0003] However, since each machine learning model is trained on different data sets or specific working conditions, it may appear inconsistent in the prediction output in the actual application process, that is, part of the machine learning model can accurately reflect the equipment state under certain working conditions, but it produces misleading under other working conditions, which leads to the deviation of the overall prediction result from the actual situation, which can be called negative transfer, which will cause difficulty in ensuring the consistency and reliability of the prediction result when multiple models are fused. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an online real-time fault diagnosis and predictive maintenance platform for industrial big data to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] An online real-time fault diagnosis and predictive maintenance platform for industrial big data, comprising:

[0007] A data acquisition module: receiving and recording real-time monitoring data from multiple industrial equipment sources;

[0008] A multi-model management module: storing and scheduling a plurality of machine learning models to process real-time monitoring data of industrial equipment;

[0009] A model conflict detection module: analyzing the prediction output of each machine learning model on the same real-time monitoring data of industrial equipment, and identifying the prediction result conflict that occurs under different working conditions;

[0010] A negative transfer identification module: after confirming the existence of the prediction result conflict, based on the difference analysis of the historical running scene and the current industrial equipment working condition, it is determined whether there is negative transfer caused by the error of the machine learning model adapting to the working condition;

[0011] Dynamic weight adjustment module: reassign different weight parameters to each machine learning model when there is negative transfer, so that the machine learning model more suitable for the current working condition obtains a greater proportion;

[0012] Fault prediction module: identify the type of equipment failure based on the adjusted weight parameters, and predict the remaining useful life of the equipment.

[0013] In a preferred embodiment, real-time monitoring data from multiple industrial equipment is received and recorded, including:

[0014] Real-time monitoring data from multiple industrial equipment is received, including temperature, pressure, vibration, current, voltage, and equipment status.

[0015] The received real-time monitoring data of each industrial equipment is preprocessed and recorded.

[0016] In a preferred embodiment, multiple machine learning models are stored and scheduled to process industrial equipment real-time monitoring data, including:

[0017] Multiple machine learning models for processing industrial equipment real-time monitoring data are stored. The machine learning models are pre-trained to determine the corresponding structure parameters and weight parameters before storage. The storage step saves each machine learning model and its parameters in a stable storage medium.

[0018] According to the pre-set scheduling criteria, the stored multiple machine learning models are scheduled. The scheduling criteria include sorting and selecting multiple machine learning models according to the characteristics of industrial equipment real-time monitoring data, processing delay requirements and model prediction performance. The scheduling step calls the selected machine learning model in turn to process the preprocessed and recorded industrial equipment real-time monitoring data.

[0019] In a preferred embodiment, the prediction outputs of each machine learning model on the same industrial equipment real-time monitoring data are analyzed, and the prediction result conflicts that occur under different working conditions are identified, including:

[0020] Obtain the prediction output of multiple pre-trained machine learning models. The prediction output is generated based on the preprocessed and recorded industrial equipment real-time monitoring data.

[0021] Systematically compare and analyze the prediction data output by each machine learning model within the same time period. According to the preset threshold and consistency determination standard, identify the prediction result conflicts that occur under different industrial equipment working conditions.

[0022] Different industrial equipment working conditions include changes in equipment load, environmental temperature fluctuations, and changes in equipment operating status.

[0023] In a preferred embodiment, after confirming the existence of a prediction result conflict, it is determined whether there is negative transfer caused by the error of the machine learning model adapting to the working condition based on the analysis of the difference between the historical running scene and the current industrial equipment working condition, including:

[0024] The historical running scene data and the current industrial equipment working condition data are respectively subjected to signal processing to obtain their respective multi-scale information entropy distributions, and the multi-scale information entropy distributions reflect the complexity characteristics of the data at multiple time scales. Then, the dynamic entropy coupling coefficient is calculated by calculating the mutual information difference between the information entropy distributions of the historical running scene data and the current industrial equipment working condition data at each scale, and the dynamic entropy coupling coefficient represents whether there is a significant deviation in information transmission between the two groups of data.

[0025] The historical running scene data and the current industrial equipment working condition data are respectively subjected to signal decomposition and instantaneous phase extraction operations to obtain their respective instantaneous phase sequences, and the multi-scale phase consistency index is calculated by comparing the instantaneous phase sequences obtained at each time scale, and the multi-scale phase consistency index reflects the consistency degree of the two groups of data in terms of instantaneous phase change.

[0026] The lower limit of the dynamic entropy coupling coefficient and the threshold value of the multi-scale phase consistency index are set, and when the dynamic entropy coupling coefficient is lower than the lower limit of the dynamic entropy coupling coefficient and the multi-scale phase consistency index is lower than the threshold value of the multi-scale phase consistency index, it is determined that there is a negative transfer phenomenon caused by the failure of the machine learning model to adapt to the current industrial equipment working condition. Otherwise, it is determined that there is no negative transfer phenomenon caused by the failure of the machine learning model to adapt to the current industrial equipment working condition.

[0027] In a preferred embodiment, the calculation step of the dynamic entropy coupling coefficient is:

[0028] Two groups of data are obtained, including historical running scene data and current industrial equipment working condition data, and the Shannon entropy value sequences of the two groups of data at multiple time scales are obtained.

[0029] The dynamic entropy coupling coefficient is determined by calculating the mutual information difference between the entropy distributions of the two groups of data at each time scale:

[0030] The total amount of mutual information between the two groups of data is as follows: Wherein, represents the total mutual information of the two groups of data at all M time scales, I m represents the mutual information between the two groups of data at the mth time scale, M represents the total number of time scales, and m represents the index of different time scales.

[0031] The average Shannon entropy values of the two groups of data are calculated respectively, and are denoted as <H (X) > and <H(Y) >;

[0032] The dynamic entropy coupling coefficient is defined as: wherein, Θ represents the dynamic entropy coupling coefficient, max(<H (X) >,<H (Y) >) represents the larger value of the average Shannon entropy values in the two groups of data.

[0033] In a preferred embodiment, the calculation step of the multi-scale phase consistency index is:

[0034] Obtaining the intrinsic mode function sequence decomposed from the historical running scene data and the intrinsic mode function sequence decomposed from the current industrial equipment working condition data;

[0035] For each intrinsic mode function, the instantaneous phase information thereof is extracted by Hilbert transform;

[0036] The instantaneous phase difference is calculated under each intrinsic mode function, and the instantaneous phase difference is the absolute difference value of the instantaneous phase of the historical running scene data and the current industrial equipment working condition data;

[0037] The multi-scale phase consistency index is calculated by using a nonlinear similarity function.

[0038] In a preferred embodiment, when there is negative transfer, different weight parameters are re-assigned to each machine learning model, so that the machine learning model more suitable for the current working condition obtains a greater proportion, including:

[0039] When there is a negative transfer phenomenon caused by the error of the machine learning model adapting to the working condition, the matching degree between the prediction output generated by each pre-trained machine learning model and the preprocessed recorded real-time monitoring data of the industrial equipment is compared according to the dynamic entropy coupling coefficient and the multi-scale phase consistency index, and the weight parameters of each machine learning model in the overall prediction are re-adjusted, wherein the machine learning model with a higher matching degree between the prediction output and the actual data obtains a higher weight parameter, and the machine learning model with a lower matching degree obtains a lower weight parameter.

[0040] In a preferred embodiment, based on the adjusted weight parameters, the equipment fault type is identified and the remaining useful life of the equipment is predicted, including:

[0041] Based on the re-adjusted weight parameters, the prediction outputs generated by each machine learning model are weighted and fused to obtain a comprehensive prediction result; wherein the weighted fusion is realized by adding the products of each prediction output and its corresponding weight parameter;

[0042] The comprehensive prediction result is analyzed to identify the equipment fault type and determine the equipment running state and the cause of the fault;

[0043] Based on historical equipment operation data, the remaining useful life of the equipment is predicted, and the future operation cycle and maintenance plan of the equipment are determined.

[0044] The technical effect and advantage of the online real-time fault diagnosis and predictive maintenance platform for industrial big data of the present application are as follows:

[0045] 1. By using nonlinear quantitative comparison of historical operation scene data and current industrial equipment working condition data in multi-scale information entropy distribution and instantaneous phase characteristics, it can be more accurately judged whether the prediction deviation caused by insufficient adaptability of the machine learning model appears under the actual working condition, so as to realize accurate identification of abnormal phenomena; by deeply analyzing the complexity of data in multiple time scales and using signal processing and statistical analysis method to extract information entropy and phase consistency index, the evaluation results obtained can fully reflect the subtle differences in data characteristics between different working conditions, thereby providing a strong basis for the reliability of the prediction results.

[0046] 2. The device operation state monitoring, prediction output analysis and data comparison links form a rigorous closed-loop process, so that after detecting the inconsistency of the prediction output, targeted analysis can be quickly started, and the weight of each machine learning model is dynamically adjusted according to the evaluation index. The adjustment process fully considers the applicability of each machine learning model under the current working condition, so that the machine learning model with higher matching degree with the actual operation data in the prediction output obtains a larger proportion, and those machine learning models with poor adaptability correspondingly reduce their influence, thereby optimizing the overall prediction result. Through this multi-link, nonlinear information fusion and weight distribution scheme, the links are tightly coupled, and the accuracy and stability of fault type identification and remaining useful life prediction are significantly improved through synergistic effect, effectively reducing the prediction error caused by data feature mismatch, and ensuring the real-time operation and efficient maintenance of the system in complex industrial environment. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The structural diagram of the online real-time fault diagnosis and predictive maintenance platform for industrial big data of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0049] Embodiment: Figure 1The application provides an online real-time fault diagnosis and predictive maintenance platform for industrial big data, which comprises the following steps:

[0050] A data acquisition module receives and records real-time monitoring data from multiple industrial equipment sources.

[0051] A multi-model management module stores and schedules multiple machine learning models to process real-time monitoring data of industrial equipment.

[0052] A model conflict detection module analyzes the prediction outputs of each machine learning model on the same real-time monitoring data of industrial equipment and identifies prediction conflicts that occur under different working conditions.

[0053] A negative transfer identification module determines whether there is negative transfer caused by incorrect adaptation of machine learning models to working conditions based on the difference analysis of historical running scenarios and the current working conditions of industrial equipment after confirming the existence of prediction conflicts.

[0054] A dynamic weight adjustment module reassigns different weight parameters to each machine learning model when there is negative transfer, so that the machine learning model that is more suitable for the current working condition obtains a greater proportion.

[0055] A fault diagnosis and predictive maintenance module identifies the fault type of the equipment based on the adjusted weight parameters and predicts the remaining useful life of the equipment.

[0056] The real-time monitoring data from multiple industrial equipment sources are received and recorded, including:

[0057] The real-time monitoring data from multiple industrial equipment sources are received and recorded, including:

[0058] The industrial equipment includes but is not limited to machine tools, fans, pumps, compressors, transformers, boilers, turbines and other equipment running in the industrial production process, and the real-time monitoring data includes but is not limited to temperature, pressure, vibration, current, voltage, speed, flow, load, equipment running state and other data.

[0059] The industrial equipment transmits real-time monitoring data through industrial field bus, wireless transmission network, industrial Ethernet or other communication methods, and the data adopts different communication protocols, including but not limited to Modbus, PROF INET, EtherCAT, CAN, OPC UA and the like. After receiving the data, different protocol formats are parsed to convert them into a unified structured data format.

[0060] The real-time monitoring data of each industrial equipment received are preprocessed and recorded.

[0061] The preprocessing includes but is not limited to the following steps: removing abnormal values generated by transmission interference in the communication process, and filtering out short pulse interference data by time window method; the sampling frequencies of different industrial equipment may be different, and the data from different sources are aligned according to the time stamp to ensure the synchronization of the data; when the data of a sensor is lost due to network loss or signal interference, interpolation, time series regression or prediction method based on historical data is used to fill in the missing values to maintain the integrity of the data; part of the analog signal data is converted into digital signal, and unit conversion is carried out according to the equipment calibration parameters to ensure that all data are in the same unit system.

[0062] The recording mode includes but is not limited to: storing according to time sequence to ensure the time continuity of the data; storing according to equipment category to ensure that subsequent queries and analysis can be based on equipment dimension; using industrial database or distributed storage system for data storage to meet the requirements of high concurrency reading and long-term storage; setting data storage strategy, archiving or compressing the data that has not been used for a long time to optimize the storage space.

[0063] Access permission is set for the recorded real-time monitoring data, and a standardized interface is provided for subsequent fault diagnosis, predictive maintenance and data analysis calling. The interface supports multiple data query methods, including searching by equipment, time and parameter type, to ensure the availability of the data.

[0064] Multiple machine learning models are stored and scheduled to process the real-time monitoring data of the industrial equipment, including:

[0065] The multiple machine learning models for processing the real-time monitoring data of the industrial equipment are stored. The machine learning models are pre-trained to determine the corresponding structure parameters and weight parameters before storage. The storage step saves each machine learning model and its parameters in a stable storage medium.

[0066] First, the multiple machine learning models for processing the real-time monitoring data of the industrial equipment are pre-trained. During the pre-training process, each machine learning model is trained according to the pre-processed real-time monitoring data of the industrial equipment to determine the corresponding structure parameters and weight parameters.

[0067] The structure parameters reflect the hierarchical structure, node configuration and connection method of the internal network of the model, and the weight parameters reflect the connection strength between the nodes.

[0068] After the pre-training is completed, each machine learning model and its determined structure parameters and weight parameters are saved to a stable storage medium, which includes but is not limited to a database, a file storage system or other storage devices with long-term data saving function.

[0069] During the storage process, the parameter data of each machine learning model is checked, and the corresponding training version, timestamp and other necessary information are recorded to ensure the integrity and consistency of the stored data.

[0070] According to the pre-set scheduling criteria, the stored multiple machine learning models are scheduled, and the scheduling criteria include sorting and selecting the multiple machine learning models according to the characteristics of the industrial equipment real-time monitoring data, the processing time delay requirements and the model prediction performance. The scheduling step sequentially calls the selected machine learning models to process the pre-processed and recorded industrial equipment real-time monitoring data.

[0071] After the storage of the machine learning model is completed, the stored multiple machine learning models are called for processing according to the pre-set scheduling criteria.

[0072] The pre-set scheduling criteria include but are not limited to sorting and selecting the multiple machine learning models according to the characteristics of the industrial equipment real-time monitoring data, the data processing time delay requirements and the prediction performance of each machine learning model in the pre-training stage.

[0073] Specifically, the scheduling criteria determine the priority calling order of the models by evaluating the accuracy, response time and stability of each machine learning model obtained in the pre-training process. Then, according to the determined priority order, each selected machine learning model is called in turn. Each called machine learning model receives the pre-processed and recorded industrial equipment real-time monitoring data in the previous step for processing and generates the corresponding prediction output.

[0074] During the scheduling process, it is ensured that each machine learning model uses the same data format and time identifier as in the pre-processing stage when processing data, so as to ensure that the output data is consistent in format, time continuity and data synchronization.

[0075] The prediction outputs of each machine learning model on the same industrial equipment real-time monitoring data are analyzed, and the prediction result conflicts that occur under different working conditions are identified, including:

[0076] The prediction outputs of multiple pre-trained machine learning models are obtained, and the prediction outputs are generated based on the pre-processed and recorded industrial equipment real-time monitoring data.

[0077] During the pre-training phase, multiple machine learning models are trained using real-time industrial equipment monitoring data obtained after pre-processing steps such as data cleaning, alignment, completion, and conversion. This allows them to determine their respective structural parameters and weighting parameters. After pre-training, each machine learning model generates a corresponding prediction output based on the pre-processed and recorded real-time industrial equipment monitoring data. This real-time industrial equipment monitoring data includes parameters such as temperature, pressure, vibration, current, voltage, and equipment operating status. The pre-processing phase ensures data format, time series consistency, and numerical accuracy.

[0078] Systematically compare and analyze the prediction data output by each machine learning model within the same time period; identify conflicts in prediction results under different industrial equipment operating conditions based on preset thresholds and consistency judgment criteria.

[0079] Among them, different industrial equipment operating conditions include changes in equipment load, ambient temperature fluctuations, and changes in equipment operating status.

[0080] After obtaining the prediction outputs generated by each machine learning model for the real-time monitoring data of the same industrial equipment in the same time period, these prediction output data are systematically compared and analyzed for differences. This comparative analysis uses a quantitative comparison method, the basic idea of ​​which is to calculate the difference between the prediction outputs of any two machine learning models, and introduce the following formula: ΔP (i，j) =|P i -P j |; where ΔP (i，j) represents the absolute difference between the predicted outputs generated by machine learning model number i and machine learning model number j for the real-time monitoring data of industrial equipment in the same time period; P i represents the predicted output generated by the i-th machine learning model for the data; P j represents the predicted output generated by the j-th machine learning model for this data.

[0081] In the above formula, i and j are the indexes representing different machine learning models, and i and j are not equal; at the same time, a ΔP is pre-set. (i，j) The corresponding preset threshold is the upper limit of the allowable error of the prediction output obtained based on the statistical analysis of historical data. If for any pair of machine learning models, the calculated ΔP (i，j) When the corresponding preset threshold is exceeded, it is considered that there is obvious inconsistency in the prediction outputs of different machine learning models for the real-time monitoring data of the same industrial equipment within this time period.

[0082] Among them, the consistency ratio is used to measure the consistency of the model prediction results: Among them, R (i，j)represents the consistency ratio of the prediction results of machine learning model sequence number i and machine learning model sequence number j in the current time period; max(P i , P j ) represents taking the larger value of the two machine learning model prediction results to avoid the denominator approaching zero in the normalization calculation.

[0083] If ΔP (i，j) exceeds its corresponding preset threshold, and R (i，j) is lower than the consistency determination standard R th , it is determined that the prediction results of the machine learning model in the time period are in conflict.

[0084] Here, the setting of the corresponding preset threshold of ΔP (i，j) includes but is not limited to the average deviation of statistical historical operation data, the equipment load change range, the environmental temperature fluctuation amplitude, and the equipment operation state change; R th is the minimum threshold of the consistency ratio, usually set to 0.85-0.95, adjusted according to historical data statistics.

[0085] After confirming the existence of prediction result conflict, it is determined whether there is negative transfer caused by the error of the machine learning model adapting to the working condition based on the difference analysis of the historical running scene and the current industrial equipment working condition, including:

[0086] The historical running scene data and the current industrial equipment working condition data are respectively subjected to signal processing to obtain their respective multi-scale information entropy distributions, and the multi-scale information entropy distribution reflects the complexity characteristics of the data at multiple time scales; then, the dynamic entropy coupling coefficient is calculated by calculating the mutual information difference between the information entropy distributions of the historical running scene data and the current industrial equipment working condition data at each scale. The dynamic entropy coupling coefficient represents whether there is a significant deviation in information transmission between the two groups of data.

[0087] The signal processing operation includes noise reduction, smoothing, and multi-scale decomposition processing, and the purpose is to extract the complexity characteristics of the data at different time scales. After the above processing, the Shannon entropy value sequences of the two groups of data at multiple time scales are obtained, which reflect the uncertainty of the data at each scale. For clarity, define:

[0088] Let represent the Shannon entropy distribution sequence of the historical running scene data at M different time scales after signal processing; let represent the Shannon entropy distribution sequence of the current industrial equipment working condition data at the same M time scales after signal processing.

[0089] In the above definition, represents the Shannon entropy value of the historical running scene data at the mth time scale; Hm represents the Shannon entropy value of the current industrial equipment working condition data at the mth time scale; M represents the total number of time scales, which depends on the number of scales selected in the multi-scale decomposition process; m represents the index of different time scales.

[0090] The Shannon entropy (H) is calculated here according to the formula wherein E represents the Shannon entropy value; p r represents the probability of the occurrence of the rth state in the data distribution at a certain time scale; R represents the total number of states at the time scale; r represents the index of a state at a certain time scale in the data distribution.

[0091] After obtaining the multi-scale Shannon entropy distribution, the dynamic entropy coupling coefficient is determined by calculating the mutual information difference between the entropy distributions of the two groups of data at each time scale, so as to quantify the coupling degree of the two groups of data in terms of information transmission. For this purpose, first, the total mutual information between the two groups of data is defined as follows: wherein, represents the total mutual information between the historical running scene data and the current industrial equipment working condition data at all M time scales; I m represents the mutual information between the two groups of data at the mth time scale.

[0092] I m The calculation formula is as follows: wherein, represents the probability of the simultaneous occurrence of the uth state and the vth state in the joint probability distribution of the historical running scene data and the current industrial equipment working condition data at the mth time scale; represents the marginal probability of the uth state of the historical running scene data at the mth time scale; represents the marginal probability of the vth state of the current industrial equipment working condition data at the mth time scale; U and V respectively represent the total number of states that the historical running scene data and the current industrial equipment working condition data can appear at the mth time scale; u represents the index of a state of the historical running scene data at a certain time scale; v represents the index of a state of the current industrial equipment working condition data at a certain time scale.

[0093] Subsequently, for the purpose of normalizing the mutual information, the average Shannon entropy values of the historical running scene data and the current industrial equipment working condition data are calculated respectively, and are denoted as <H (X) > and <H (Y) > respectively.

[0094] The dynamic entropy coupling coefficient is defined as: wherein Θ represents the dynamic entropy coupling coefficient, which is used to quantify whether there is a significant deviation in the information entropy distribution between the two groups of data; max(<H (X) > and <H (Y)>) represents the larger value of the average Shannon entropy value of the historical running scene data and the current industrial equipment working condition data, and is used for normalizing mutual information; the lower the dynamic entropy coupling coefficient value, the greater the deviation of the two groups of data in information transmission.

[0095] The historical running scene data and the current industrial equipment working condition data are respectively subjected to signal decomposition and instantaneous phase extraction operations to obtain respective instantaneous phase sequences. The instantaneous phase sequences obtained at each time scale are compared, and a multi-scale phase consistency index is calculated. The multi-scale phase consistency index reflects the consistency degree of the two groups of data in instantaneous phase change.

[0096] To further reflect the difference between the two groups of data in dynamic characteristics, the historical running scene data and the current industrial equipment working condition data are respectively subjected to signal decomposition operations. The operation adopts an empirical mode decomposition method to decompose the original data into a plurality of intrinsic mode functions. Let the intrinsic mode function sequence obtained by decomposing the historical running scene data be denoted as and the intrinsic mode function sequence obtained by decomposing the current industrial equipment working condition data be denoted as wherein, represents the kth intrinsic mode function of the historical running scene data after signal decomposition, and the function varies with time t; represents the kth intrinsic mode function of the current industrial equipment working condition data after signal decomposition; K represents the total number of intrinsic mode functions, reflecting the inherent vibration mode of the data; k represents the index of different intrinsic mode functions.

[0097] For each intrinsic mode function, its instantaneous phase information is extracted by Hilbert transform. Let the instantaneous phase of the historical running scene data under the kth intrinsic mode function be denoted as and the instantaneous phase of the current industrial equipment working condition data under the kth intrinsic mode function be denoted as

[0098] wherein, is the instantaneous phase of the kth intrinsic mode function of the historical running scene data at time t; is the instantaneous phase of the kth intrinsic mode function of the current industrial equipment working condition data at time t.

[0099] Under each intrinsic mode function, the instantaneous phase difference is defined as: wherein, ΔΦ k (t) represents the absolute difference value of the instantaneous phase of the historical running scene data and the current industrial equipment working condition data under the kth intrinsic mode function;

[0100] In order to quantitatively describe the consistency of the two groups of data in each modality, a nonlinear similarity function is used to calculate the multi-scale phase consistency index. The multi-scale phase consistency index is defined as: Wherein, Λ represents the multi-scale phase consistency index, which is used to measure the consistency of the historical running scene data and the current industrial equipment working condition data in the instantaneous phase; β is a positive adjustment constant, which is used to regulate the sensitivity of the exponential function, and the unit is the unit of inverse phase difference; <ΔΦ k (t) > T represents the average value of <ΔΦ k (t) in the predetermined time interval T under the kth intrinsic mode function.

[0101] The lower limit of the dynamic entropy coupling coefficient and the threshold of the multi-scale phase consistency index are set. When the dynamic entropy coupling coefficient is lower than the lower limit of the dynamic entropy coupling coefficient and the multi-scale phase consistency index is lower than the threshold of the multi-scale phase consistency index, it is determined that there is a negative transfer phenomenon caused by the failure of the machine learning model to adapt to the current industrial equipment working condition; otherwise, it is determined that there is no negative transfer phenomenon caused by the failure of the machine learning model to adapt to the current industrial equipment working condition.

[0102] A lower limit Θ0 of the dynamic entropy coupling coefficient (wherein Θ0 is a positive real number) is set as one of the determination criteria. When the calculated Θ is less than Θ0, it is preliminarily indicated that there is a significant deviation in the information distribution of the two groups of data.

[0103] A threshold Λ0 of the multi-scale phase consistency index (wherein Λ0 is a positive real number) is set. When Λ is less than Λ0, it indicates that there is a significant difference in the phase characteristics between the historical running scene data and the current industrial equipment working condition data.

[0104] When and only when Θ < Θ0 and Λ < Λ0, it is determined that there is a negative transfer caused by the error of the machine learning model in adapting to the working condition; otherwise, it is determined that there is no negative transfer caused by the error of the machine learning model in adapting to the working condition.

[0105] Θ0 is the lower limit of the dynamic entropy coupling coefficient set in advance, which is used to represent the minimum coupling degree allowed in information transmission of the two groups of data; Λ0 is the threshold of the multi-scale phase consistency index set in advance, which is used to represent the minimum consistency level allowed in the instantaneous phase of the two groups of data.

[0106] Each step is operated based on the pre-processed and recorded real-time monitoring data of the industrial equipment, ensuring the uniformity of data sources and the consistency of processing methods; through the above multi-angle parameter analysis, the subtle differences in information entropy and instantaneous phase characteristics between the historical running scene and the current industrial equipment working condition data can be fully revealed, and whether there is a negative transfer phenomenon caused by the failure of the machine learning model to adapt to the current industrial equipment working condition can be accurately determined.

[0107] When there is negative transfer, different weight parameters are re-assigned to each machine learning model, so that the machine learning model more suitable for the current working condition obtains a greater proportion, including:

[0108] When there is negative transfer due to the error of the machine learning model adapting to the working condition, the matching degree between the prediction output generated by each pre-trained machine learning model and the real-time monitoring data of the industrial equipment recorded in the pre-processing is compared according to the dynamic entropy coupling coefficient and the multi-scale phase consistency index, and the weight parameters of each machine learning model in the overall prediction are re-adjusted, wherein the machine learning model with a higher matching degree between the prediction output and the actual data obtains a higher weight parameter, and the machine learning model with a lower matching degree obtains a lower weight parameter.

[0109] For each machine learning model, the matching score between its prediction output and the pre-processed recorded data is calculated, and the formula is: wherein Q Ω represents the matching score; Ω represents the number of machine learning models; Ξ represents the time point index of data collection in the time series; N' represents the total number of time points; Y Ω (Ξ) represents the prediction output of the Ωth machine learning model at time point Ξ; Z Ω (Ξ) represents the pre-processed recorded data corresponding to time point Ξ; Λ Q represents a positive factor for controlling sensitivity, which is a positive number parameter for adjusting the error influence degree between the prediction output and the pre-processed recorded data.

[0110] The weight parameter of each machine learning model can be determined by the following formula: wherein P Ω represents the weight parameter of the machine learning model; Ψ represents a positive emphasis factor for adjusting the influence degree of the matching score, which is a positive number parameter for enhancing the influence of the difference in the matching score of each machine learning model on the weight allocation effect; N" represents the total number of machine learning models participating in the calculation.

[0111] When and only when the negative transfer identification module determines that there is negative transfer, the weight parameters of each machine learning model are updated according to the above formula, so that the model with a higher matching score obtains a larger weight, and the model with a lower matching score obtains a smaller weight.

[0112] Based on the adjusted weight parameters, the device fault type is identified and the remaining useful life of the device is predicted, including:

[0113] Based on the re-adjusted weight parameters, the prediction outputs generated by each machine learning model are weighted and fused to obtain a comprehensive prediction result; wherein the weighted fusion is realized by multiplying and adding each prediction output and its corresponding weight parameter.

[0114] Based on the adjusted weight parameters, the prediction outputs generated by each pre-trained machine learning model in the same time period are weighted and fused. Specifically, each machine learning model produces a prediction output for the same batch of pre-processed and recorded real-time monitoring data of the industrial equipment, and each prediction output of each model is multiplied by its corresponding adjusted weight parameter, and then the products are sequentially accumulated to form a comprehensive prediction result.

[0115] For example, if several machine learning models are involved in prediction, the prediction result generated by each machine learning model at a certain time is multiplied by its pre-adjusted weight, and then all the products are added to obtain an overall prediction index.

[0116] The comprehensive prediction result is analyzed to identify the equipment failure type, and the equipment operating state and the cause of the failure are determined.

[0117] After obtaining the comprehensive prediction result, first, the comprehensive prediction result is analyzed in detail to identify the equipment failure type and determine the equipment operating state and the cause of the failure. Specifically, this embodiment uses data trend analysis, outlier detection, and historical data comparison methods to compare the changes in the comprehensive prediction result in each time period.

[0118] For example, when the comprehensive prediction result shows a sudden deviation from the conventional fluctuation range in a certain time period, by comparing with the pre-processed and recorded actual operating data of the equipment, it can be preliminarily determined that the equipment corresponding to the deviation phenomenon has a risk of failure. Further, by comparing the failure records corresponding to similar deviation phenomena in historical operating scenarios, this embodiment can identify the specific failure type, such as mechanical part wear, overload, or temperature anomaly, and determine the current operating state of the equipment and the possible cause of the failure.

[0119] Based on the historical equipment operating data, the remaining useful life of the equipment is predicted, and the future operating cycle and maintenance plan of the equipment are determined.

[0120] After the failure type is identified, the remaining useful life of the equipment is predicted based on the historical equipment operating data. The specific steps include: first, collecting and organizing the operating time and maintenance cycle data of the equipment in different failure states in history; second, by analyzing the changes in operating indicators before and after the occurrence of equipment failure, a corresponding relationship between failure occurrence and equipment remaining life is established; third, according to the corresponding relationship, the current equipment state is predicted to determine the future operating cycle of the equipment; finally, combined with the actual operating conditions of the equipment, a reasonable maintenance plan and repair cycle is developed.

[0121] It is worth noting that "fault diagnosis and predictive maintenance" in the "fault diagnosis and predictive maintenance module" is a short form of "fault diagnosis and predictive maintenance".

[0122] The above formulas are all de-dimensioned to calculate the numerical values, the formulas are obtained by collecting a large amount of data to simulate a formula closest to the actual situation, and the preset parameters and threshold values in the formulas are set by a person skilled in the art according to the actual situation.

[0123] The above embodiments can be implemented wholly or partially by software, hardware, firmware, or any other combination. When implemented by software, the above embodiments can be implemented wholly or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0124] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and module can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0125] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed ones can be indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0126] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0127] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0128] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0129] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0130] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An online real-time fault diagnosis and predictive maintenance platform for industrial big data, characterized by: include: Data acquisition module: receives and records real-time monitoring data from multiple source industrial equipment; Multi-model management module: stores and schedules multiple machine learning models to process real-time monitoring data of industrial equipment; Model conflict detection module: Analyzes the prediction outputs of various machine learning models for the same industrial equipment real-time monitoring data and identifies conflicts in prediction results under different operating conditions; Negative transfer identification module: After confirming that there is a conflict in the prediction results, it determines whether there is negative transfer caused by the machine learning model's incorrect adaptation to the operating conditions based on the difference analysis between the historical operating scenarios and the current industrial equipment operating conditions, including: Signal processing is performed on historical operating scenario data and current industrial equipment operating condition data to obtain their respective multi-scale information entropy distributions, which reflect the complexity characteristics of the data at multiple time scales. Then, the dynamic entropy coupling coefficient is calculated by calculating the mutual information difference between the information entropy distributions of historical operating scenario data and current industrial equipment operating condition data at each scale. The dynamic entropy coupling coefficient indicates whether there is a significant deviation in the information transmission between the two sets of data. Signal decomposition and instantaneous phase extraction were performed on historical operating scenario data and current industrial equipment operating condition data to obtain their respective instantaneous phase sequences. The instantaneous phase sequences obtained at different time scales were compared and a multi-scale phase consistency index was calculated. The multi-scale phase consistency index reflects the degree of consistency between the two sets of data in terms of instantaneous phase changes. The lower limit of the dynamic entropy coupling coefficient and the threshold of the multi-scale phase consistency index are set. When the dynamic entropy coupling coefficient is lower than the lower limit of the dynamic entropy coupling coefficient and the multi-scale phase consistency index is lower than the threshold of the multi-scale phase consistency index, it is determined that there is a negative migration phenomenon caused by the failure of the machine learning model to adapt to the current working conditions of the industrial equipment; otherwise, it is determined that there is no negative migration phenomenon caused by the failure of the machine learning model to adapt to the current working conditions of the industrial equipment; Dynamic weight adjustment module: When negative transfer occurs, different weight parameters are reallocated to each machine learning model, so that the machine learning model that is more suitable for the current working conditions receives a larger proportion; Fault diagnosis and pre-maintenance module: Identifies equipment failure types based on adjusted weight parameters and predicts the remaining useful life of the equipment.

2. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: Receive and record real-time monitoring data from multiple sources of industrial equipment, including: Receive real-time monitoring data from multiple industrial devices, including temperature, pressure, vibration, current, voltage, and equipment status; The received real-time monitoring data of each industrial equipment is pre-processed and recorded.

3. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: Store and schedule multiple machine learning models to process real-time monitoring data from industrial equipment, including: Storing multiple machine learning models used to process real-time monitoring data of industrial equipment. The machine learning models are pre-trained before storage to determine corresponding structural parameters and weight parameters. The storage step saves each machine learning model and its parameters in a stable storage medium. According to pre-set scheduling criteria, multiple stored machine learning models are scheduled. The scheduling criteria include sorting and selecting multiple machine learning models based on the characteristics of real-time monitoring data of industrial equipment, processing delay requirements and model prediction performance. The scheduling step calls the selected machine learning models in sequence to process the pre-processed and recorded real-time monitoring data of industrial equipment.

4. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: Analyze the prediction outputs of various machine learning models for the same real-time monitoring data of industrial equipment and identify conflicts in prediction results under different operating conditions, including: Obtaining prediction outputs from multiple pre-trained machine learning models, generated based on pre-processed and recorded real-time monitoring data from industrial equipment; Systematically compare and analyze the prediction data output by various machine learning models within the same time period; identify conflicts in prediction results under different industrial equipment operating conditions based on preset thresholds and consistency criteria; Among them, different industrial equipment operating conditions include changes in equipment load, ambient temperature fluctuations, and changes in equipment operating status.

5. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: The calculation steps of the dynamic entropy coupling coefficient are: Obtain two sets of data, including historical operation scenario data and current industrial equipment operating condition data, and obtain Shannon entropy value sequences of the two sets of data at multiple time scales; The dynamic entropy coupling coefficient is determined by calculating the mutual information difference between the entropy distributions of the two sets of data at each time scale: The total amount of mutual information between the two sets of data is as follows: ;in, Indicates that the two sets of data are The total mutual information at each time scale is Indicates in The mutual information between two sets of data at a time scale is represents the total number of time scales, Indices representing different time scales; Calculate the average Shannon entropy of the two sets of data respectively, and record them as and ; The dynamic entropy coupling coefficient is defined as: ;in, represents the dynamic entropy coupling coefficient, It represents the larger value of the average Shannon entropy in the two sets of data.

6. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: The calculation steps of the multi-scale phase consistency index are: Obtaining an intrinsic mode function sequence obtained by decomposing historical operating scenario data and an intrinsic mode function sequence obtained by decomposing current industrial equipment operating condition data; For each intrinsic mode function, its instantaneous phase information is extracted through Hilbert transform; Calculate the instantaneous phase difference under each intrinsic mode function. The instantaneous phase difference is the absolute difference between the instantaneous phase of the historical operating scenario data and the current industrial equipment operating condition data. The multi-scale phase consistency index was calculated using a nonlinear similarity function.

7. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: When negative transfer occurs, different weight parameters are reallocated to each machine learning model so that the machine learning model that is more suitable for the current working conditions gains a larger proportion, including: When there is a negative migration phenomenon caused by the error of the machine learning model in adapting to the working conditions, the degree of matching between the predicted output generated by each pre-trained machine learning model and the real-time monitoring data of industrial equipment recorded in the preprocessing is compared based on the dynamic entropy coupling coefficient and the multi-scale phase consistency index, and the weight parameters of each machine learning model in the overall prediction are readjusted. Among them, the machine learning model whose predicted output matches the actual data more closely obtains a higher weight parameter, while the machine learning model with a lower degree of matching obtains a lower weight parameter.

8. The online real-time fault diagnosis and predictive maintenance platform for industrial big data according to claim 1 is characterized in that: Identify equipment failure types based on adjusted weight parameters and predict the remaining useful life of the equipment, including: The prediction outputs generated by each machine learning model are weighted and fused based on the readjusted weight parameters to obtain a comprehensive prediction result; the weighted fusion is achieved by multiplying and accumulating each prediction output with its corresponding weight parameter; Analyze the comprehensive prediction results to identify the equipment failure type and clarify the equipment operating status and failure cause; Based on historical equipment operation data, the remaining useful life of the equipment is predicted and the future operation cycle and maintenance plan of the equipment are determined.

Images