Artificial intelligence-based frequency converter fault prediction method and system

Abstract

The application discloses an artificial intelligence-based frequency converter fault prediction method. The method comprises the following steps: first, obtaining a plurality of parameters of the frequency converter corresponding to the sensing data set respectively, and selecting target parameters from the plurality of parameters, and then obtaining a plurality of target data sets corresponding to the target parameters in the historical data; second, constructing and training an integrated learning model, calculating the correlation between any two data groups in each target data set of the target parameters through the model, and combining into a correlation array of the target data set; then, analyzing and processing the correlation array corresponding to each target data set, calculating a plurality of representative vectors of each target data set, and constructing a space feature model based on the representative vectors; finally, collecting a real-time data set of the target parameters in the current time period, calculating a real-time representative vector, and inputting the real-time representative vector into the space feature model, so as to obtain a fault prediction result. The method can effectively improve the accuracy and efficiency of frequency converter fault prediction.

Description

Technical Field

[0001] This application relates to the field of fault prediction technology, and in particular to a method and system for predicting frequency converter faults based on artificial intelligence. Background Technology

[0002] Inverters are key equipment in industrial automation, power systems, and other fields. Their failures can lead to production interruptions, equipment damage, and energy waste. Existing technologies for inverter fault prediction have shortcomings, such as low detection efficiency, poor accuracy, and difficulty in handling complex fault modes. Traditional fault prediction methods often rely on manual experience or single models, making them ill-suited to the diverse and complex operating conditions of inverters. They often only predict known types of faults and have weak generalization ability for unknown or novel faults. A similar prior art is Chinese Patent Publication No. CN118551266A, which proposes a fault diagnosis method for inverters. This method includes: receiving dust concentration data from a dust detector; monitoring the inverter's operating status using a trained digital twin model to obtain operating status data; identifying the dust concentration and operating status data as monitoring data; obtaining potential fault data, performance trend data, and remaining lifespan data of the inverter; establishing a fault tree for the inverter based on the potential fault data, performance trend data, and remaining lifespan data; and using causal reasoning to select the main fault causes from multiple fault causes and outputting a fault diagnosis report for the inverter. This method aims to improve the accuracy of fault diagnosis. A similar prior art is Chinese patent CN119475162A, which proposes a fault detection method for high-voltage frequency converters. This method uses a data acquisition module installed inside the high-voltage frequency converter to collect its operating parameters in real time, relating to the field of data processing technology. This fault detection method for high-voltage frequency converters establishes a fault diagnosis model by extracting data parameters from a database of similar high-voltage frequency converters. Historical fault data of high-voltage frequency converters is then introduced into the fault diagnosis model for training and optimization. The operating parameters of the high-voltage frequency converter requiring fault detection are then input into the fault diagnosis model for processing, outputting the fault diagnosis result. This allows for real-time monitoring of the possibility of equipment failure during data acquisition, rather than relying on human experience to judge fault problems during inspection, thus improving both the efficiency and accuracy of fault detection. Both of the aforementioned patent documents address the problem of frequency converter fault prediction, but the data complexity is relatively high and the accuracy is not high enough during prediction. Summary of the Invention

[0003] This application provides an artificial intelligence-based method and system for predicting inverter faults, the method comprising:

[0004] Based on the historical sensor datasets corresponding to multiple parameters of the frequency converter, a target parameter is selected from the multiple parameters, and multiple target datasets corresponding to the target parameter are obtained.

[0005] Construct and train an ensemble learning model, calculate the correlation between any two data groups based on each of the target datasets according to the target parameters of the ensemble learning model, and combine the correlations into a correlation array of the target datasets;

[0006] By analyzing and processing the association array corresponding to each target dataset, multiple representative vectors corresponding to each target dataset are calculated, and a spatial feature model is constructed based on the multiple representative vectors corresponding to each target dataset.

[0007] Collect the real-time dataset of the target parameters within the current time period, calculate and input the real-time representative vector corresponding to the real-time dataset into the spatial feature model to obtain the fault prediction result.

[0008] As a preferred embodiment of the present invention, the acquisition of multiple target datasets corresponding to the target parameters includes:

[0009] The historical sensor datasets and corresponding fault information corresponding to the multiple parameters of the frequency converter are obtained. The target parameter is selected according to the criticality of each parameter to the fault detection of the frequency converter. Multiple target datasets corresponding to the target parameter are obtained from the historical sensor dataset according to the working mode. The target parameter includes multiple parameters of the frequency converter.

[0010] As a preferred embodiment of the present invention, the acquisition of the related array includes:

[0011] The ensemble learning model is constructed and trained. The ensemble learning model includes multiple sub-models. The numerical combinations of the target parameters collected at the same time in the target dataset are taken as the data groups. Any two data groups are input into the ensemble learning model. Each sub-model in the ensemble model analyzes the two input data groups from different perspectives and outputs the corresponding inference relevance. The correlation between the two data groups is calculated by weighting according to the numerical distribution of each inference relevance. The correlation of any two data groups in the target dataset is used to form the correlation array.

[0012] As a preferred embodiment of the present invention, obtaining the representative vectors of the target dataset and constructing the spatial feature model include:

[0013] Based on the relevant array and mean-removing array corresponding to the target dataset, a representative array is obtained, and the representative array is diagonalized to obtain multiple representative values. The vectors corresponding to the m representative values ​​whose contribution is greater than a set threshold are used as the representative vectors of the target dataset. The value of m is adjusted according to the inverter fault information corresponding to the data group in the target dataset.

[0014] Each representative vector is mapped to a multi-dimensional feature coordinate system, and the inverter fault information corresponding to the target data group is added to the position of the corresponding representative vector to obtain the spatial feature model.

[0015] As a preferred embodiment of the present invention, the real-time representative vector is input into the spatial feature model to obtain the fault prediction result, including:

[0016] Each real-time representative vector corresponding to the real-time dataset is input into the spatial feature model. The spatial feature model obtains a target representative vector whose correlation with the real-time representative vector is within a set range. When the fault information of the target representative vector is normal, the fault information is predicted to be normal; otherwise, the fault information is predicted to be abnormal. When the predicted fault information is normal, the distance between the real-time representative vector and the closest representative vector with abnormal fault information is calculated. The fault information change trend is obtained based on the change of the distance. When the distance is less than the warning distance, a warning message is sent to the user terminal, and the predicted fault information is used as the fault prediction result. The working mode corresponding to the target vector is the same as the working mode of the real-time representative vector.

[0017] As a preferred embodiment of the present invention, before inputting the real-time representative vector corresponding to the real-time dataset into the spatial feature model, the method further includes:

[0018] The historical representative vector corresponding to the historical dataset within a preset time period before the current time period is input into the spatial feature model, and a representative vector with the same working mode as the historical dataset and a correlation greater than a set value is found as a similar representative vector. When the fault information corresponding to the historical representative vector and the similar representative vector is inconsistent or the similar representative vector cannot be found, the historical representative vector replaces the representative vector in the working mode.

[0019] As a preferred embodiment of the present invention, the training of the ensemble learning model includes:

[0020] Multiple sample datasets of the target parameters in different working modes of the strain gauge are selected from the database, and the data groups in each sample dataset are normalized. Any two data groups in the sample dataset are taken as sample data pairs. Each sub-model in the ensemble learning model is trained based on the sample data pairs and label data. The analysis perspectives in the sub-model include at least time perspective, frequency domain perspective and statistical perspective.

[0021] As a preferred embodiment of the present invention, the sub-model in the ensemble learning model is a gradient boosting tree model.

[0022] This invention also provides an artificial intelligence-based inverter fault prediction system for implementing the above-described method, the system comprising:

[0023] The acquisition unit is used to select a target parameter from the multiple parameters based on the historical sensor datasets corresponding to multiple parameters of the frequency converter, and acquire multiple target datasets corresponding to the target parameter.

[0024] A computing unit is used to construct and train an ensemble learning model, calculate the correlation between any two data groups based on each of the target datasets according to the target parameters of the ensemble learning model, and combine the correlations into a correlation array of the target datasets.

[0025] The construction unit is used to analyze and process the correlation array corresponding to each target dataset, calculate multiple representative vectors for each target dataset, and construct a spatial feature model based on the multiple representative vectors corresponding to each target dataset.

[0026] The prediction unit is used to collect the real-time dataset of the target parameter within the current time period, calculate and input the real-time representative vector corresponding to the real-time dataset into the spatial feature model to obtain the fault prediction result.

[0027] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0028] Effect

[0029] This patent proposes an AI-based method and system for predicting inverter faults. This invention offers significant technical advantages. First, through multi-dimensional data processing and the application of an ensemble learning model, the system effectively improves the accuracy of inverter fault prediction. By selecting multiple key parameters of the inverter and combining them with historical sensor data, the system extracts a target dataset and calculates the correlation between each data group, constructing a high-precision correlation array. This process analyzes the data from multiple perspectives, including time, frequency domain, and statistics, through multiple sub-models in the ensemble learning model, enabling the system to more comprehensively understand the data characteristics. Second, the system reduces the interference of redundant data while retaining the most representative feature information by extracting representative vectors and constructing a spatial feature model. By diagonalizing the representative vectors, the representative values ​​with the greatest contribution are extracted, further reducing the dimensionality of the data and improving the computational efficiency of subsequent predictive analysis. Furthermore, this method uses a dynamic update mechanism to process real-time data. When the system detects inconsistencies between historical representative vectors and real-time fault information, it automatically replaces or adds representative vectors, ensuring that the model always accurately reflects the current operating state. This mechanism not only enhances the system's adaptability but also updates the fault prediction model in real time, effectively improving the timeliness and accuracy of prediction results. Most importantly, based on the spatial feature model, the system can quickly identify the most relevant fault modes by comparing real-time data with historical data, providing early warnings of potential fault risks. By setting a warning distance, the system can monitor the changing trends of faults and issue warnings within the warning distance, ensuring the safe operation of the frequency converter. In summary, this invention, by combining integrated learning, spatial feature models, and dynamic update mechanisms, can greatly improve the efficiency and accuracy of frequency converter fault prediction, providing reliable technical support for preventive maintenance of industrial equipment. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the inverter fault detection method based on artificial intelligence in the embodiment;

[0032] Figure 2 This is a flowchart illustrating the method for constructing the spatial feature model in this embodiment;

[0033] Figure 3 This is a flowchart illustrating the method for obtaining fault prediction results in this embodiment;

[0034] Figure 4 This is a structural diagram of the inverter fault detection system based on artificial intelligence in the embodiment. Detailed Implementation

[0035] This application provides an artificial intelligence-based method and system for predicting inverter faults. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0036] For ease of understanding, the specific process of the embodiments of this application is described below, such as... Figure 1 As shown in the figure, an artificial intelligence-based inverter fault prediction method in this application includes:

[0037] Step S1: Based on the historical sensor datasets corresponding to multiple parameters of the frequency converter, select the target parameter and the multiple target datasets corresponding to the target parameter from the multiple parameters;

[0038] Step S2: Construct and train an ensemble learning model, calculate the correlation between any two data groups based on each of the target datasets according to the target parameters of the ensemble learning model, and combine the correlations into a correlation array of the target datasets;

[0039] Step S3: By analyzing and processing the correlation array corresponding to each target dataset, calculate multiple representative vectors corresponding to each target dataset, and construct a spatial feature model based on the multiple representative vectors corresponding to each target dataset;

[0040] Step S4: Collect the real-time dataset of the target parameter within the current time period, calculate and input the real-time representative vector corresponding to the real-time dataset into the spatial feature model to obtain the fault prediction result.

[0041] Furthermore, obtaining the multiple target datasets corresponding to the target parameters includes:

[0042] The historical sensor datasets and corresponding fault information corresponding to the multiple parameters of the frequency converter are obtained. The target parameter is selected according to the criticality of each parameter to the fault detection of the frequency converter. Multiple target datasets corresponding to the target parameter are obtained from the historical sensor dataset according to the working mode. The target parameter includes multiple parameters of the frequency converter.

[0043] Specifically, by acquiring the historical sensor datasets corresponding to the aforementioned multiple parameters of the frequency converter, and based on historical maintenance records and maintenance testing information, the causes of historical faults can be obtained. By analyzing the causes of these historical faults, the criticality of each of the aforementioned multiple parameters to the fault detection of the frequency converter can be obtained, that is, the reference value of the parameters reflecting the fault state of the frequency converter. For example, the output current, voltage, motor speed, mechanical vibration, and temperature of the frequency converter are key parameters reflecting frequency converter abnormalities. Furthermore, multiple target datasets corresponding to the aforementioned target parameters are acquired from the aforementioned historical sensor data according to the operating mode. The aforementioned data sets are arrays composed of the values ​​of the aforementioned target parameters at the same point in time. Through the above technical solution, not only is the impact of redundant parameters on fault prediction reduced, making subsequent analysis more efficient, but the accuracy of fault prediction is also improved.

[0044] Furthermore, the acquisition of the relevant array includes:

[0045] The ensemble learning model is constructed and trained. The ensemble learning model includes multiple sub-models. The numerical combinations of the target parameters collected at the same time in the target dataset are taken as the data groups. Any two data groups are input into the ensemble learning model. Each sub-model in the ensemble model analyzes the two input data groups from different perspectives and outputs the corresponding inference relevance. The correlation between the two data groups is calculated by weighting according to the numerical distribution of each inference relevance. The correlation of any two data groups in the target dataset is used to form the correlation array.

[0046] Specifically, the aforementioned target parameters are the main parameter variables related to inverter fault detection, such as: operating mode, input and output voltage, input and output current, temperature, motor speed, and mechanical vibration data. The aforementioned target dataset is a time-series dataset under the same operating mode. The aforementioned target dataset is large data; for example, the aforementioned historical data is time-series data of the aforementioned target parameters collected every 30 seconds. Multiple sets of the aforementioned target parameter data collected every 7 days are used as the aforementioned target dataset. The inverter fault information corresponding to the aforementioned data sets is stored separately according to the numerical collection time of the aforementioned target parameters. Since the aforementioned ensemble learning model includes multiple aforementioned sub-models, each sub-model analyzes and infers from different angles and feature spaces on the aforementioned two data sets, obtaining the inference correlation between the output of each sub-model and the inference correlation corresponding to the aforementioned two data sets. The inference correlation of the output of each sub-model is weighted, wherein the sum of the differences between each of the aforementioned inference correlations and other inferences output by other models is calculated, and based on each... The sum of the differences corresponding to the inference relevance is weighted and the weighted result is used as the relevance between the two data groups. The sum of the weights of the inference relevances corresponding to different model outputs for the two data groups is 1. A larger sum of differences corresponds to a smaller weight for the inference relevance, indicating a larger deviation between the inference relevance and other inference relevances output by other models, thus lower accuracy and a relatively lower weight. Conversely, a smaller sum of differences indicates a smaller deviation between the inference relevance and other inference relevances output by other models, higher accuracy, and a better reflection of the true relevance between the two data groups, thus a relatively higher weight. Furthermore, the relevances of any two data groups in the target dataset are combined to form the relevance array. This technical solution quantifies the global relevance of data groups in the target dataset, also representing similarity, and captures data features, laying the foundation for obtaining a representative vector of the target dataset.

[0047] Furthermore, the acquisition of representative vectors of the target dataset and the construction of the spatial feature model, such as... Figure 2 As shown, it includes:

[0048] Based on the relevant array and mean-removing array corresponding to the target dataset, a representative array is obtained, and the representative array is diagonalized to obtain multiple representative values. The vector corresponding to the minimum number of representative values ​​whose sum of representative contribution values ​​is greater than a set threshold is used as the representative vector of the target dataset. Here, m representative values ​​are the minimum number that can satisfy the set threshold. The value of m is adjusted according to the inverter fault information corresponding to the data group in the target dataset.

[0049] The representative vector corresponding to each target dataset is mapped to a multi-dimensional feature coordinate system, and the inverter fault information corresponding to the target data group is added to the positions of the corresponding multiple representative vectors to obtain the spatial feature model.

[0050] Specifically, by constructing the aforementioned mean-removed array A, Where B is an n-dimensional, n-column identity array, and e is an n-dimensional, n-column array of all-ones. The following example uses a target dataset containing four data groups. In practical applications, the target dataset may contain more than four data groups. For example, in the aforementioned related array...

[0051] At that time, the elements in the relevant array The correlation between the i-th data group and the j-th data group is represented by the following formula: ,

[0052] The calculation result is Wherein, the values ​​on the diagonal represent the variance of the data set in the feature space, and the values ​​off-diagonal represent the variance of the corresponding data.

[0053] The correlation values ​​between groups are positive (positive) and negative (negative). By diagonalizing the mean-reduced array and performing feature analysis, representative values ​​(also eigenvalues) are obtained. These representative values ​​reflect the importance of the corresponding representative vector within the array. The representative contribution of each value is defined as the ratio of its absolute value to the sum of the absolute values ​​of all representative values. Specifically, when the representative values ​​are a1 = 0.35, a2 = 0.2, a3 = -0.12, and a4 = -0.43, their respective contributions are 31.8%, 18.2%, 10.9%, and 39.1%. With a threshold of 70%, m is set to 2. The eigenvector b1 corresponding to a1 = 0.35 and the eigenvector b1 corresponding to a4 = 0.35 are then compared. The feature vector b2 corresponding to -0.43 is used as the representative vector of the target dataset and also represents the feature data of the target dataset. Each of the above representative vectors is mapped to a multi-dimensional feature coordinate system. The inverter fault information corresponding to the target data group is added to the position of the corresponding representative vector to obtain the spatial feature model. The above technical solution, based on the selection of target parameters according to the degree of influence of each parameter on inverter fault in step S1 to reduce the influence of redundant parameters, also performs dimensionality reduction processing on the target dataset in this step to obtain representative vectors that can reflect the data characteristics of target parameters in the target dataset and construct the corresponding spatial feature model. This not only simplifies the subsequent calculation volume but also improves the subsequent search efficiency.

[0054] Furthermore, the real-time representative vector is input into the spatial feature model to obtain the fault prediction result, such as... Figure 3 As shown, it includes:

[0055] Each real-time representative vector corresponding to the real-time dataset is input into the spatial feature model. The spatial feature model obtains a target representative vector whose correlation with the real-time representative vector is within a set range. When the fault information of the target representative vector is normal, the fault information is predicted to be normal; otherwise, the fault information is predicted to be abnormal. When the predicted fault information is normal, the distance between the real-time representative vector and the closest representative vector with abnormal fault information is calculated. The fault information change trend is obtained based on the change of the distance. When the distance is less than the warning distance, a warning message is sent to the user terminal, and the predicted fault information is used as the fault prediction result. The working mode corresponding to the target vector is the same as the working mode of the real-time representative vector.

[0056] Specifically, by inputting each of the aforementioned real-time representative vectors into the aforementioned spatial feature model, the spatial feature model calculates the aforementioned representative vectors whose distance from the aforementioned real-time representative vectors is within a set range, i.e., the aforementioned target representative vectors whose correlation is within a set range. The operating mode corresponding to the aforementioned target vectors is the same as the operating mode of the aforementioned real-time representative vectors. Since the target parameters of the frequency converter may change in the future time period, and since fault conditions only appear after accumulating to a certain extent during operation, the future time period and the aforementioned set range are corresponding. The longer the future time period is set, the larger the set range, and vice versa. The fault information corresponding to the aforementioned target representative vectors reflects the fault information corresponding to the aforementioned real-time representative vectors. Therefore, when the fault information of the aforementioned target representative vectors is normal, the fault information corresponding to the aforementioned real-time representative vectors is also normal. Thus, the predicted fault information of the frequency converter is also normal. Conversely, when there is an anomaly in the fault information of the aforementioned target representative vectors, the aforementioned... The predicted fault information is abnormal. Since the predicted information may be normal, it could be at a critical point between normal and abnormal states. To obtain abnormal information in a timely manner, the distance to the closest real-time representative vector with abnormal fault information is calculated. The larger the distance, the closer the fault state of the real-time dataset corresponding to the real-time representative vector is to the normal state; conversely, the smaller the distance, the closer the fault state of the real-time dataset corresponding to the real-time representative vector is to the abnormal state. As time progresses, the smaller the distance, the more the inverter's operating state is moving towards an abnormal state; conversely, the inverter's operating state is moving towards or maintaining a normal operating state. When the distance is less than the warning distance, a warning message is sent to the user terminal. Through this technical solution, not only can the inverter's fault information be accurately obtained, but the trend of fault information can also be obtained based on the change in the distance, and a warning message can be issued in a timely manner to take measures to prevent inverter faults from occurring.

[0057] Furthermore, before inputting the real-time representative vector corresponding to the real-time dataset into the spatial feature model, the method further includes:

[0058] The historical representative vector corresponding to the historical dataset within a preset time period before the current time period is input into the spatial feature model, and a representative vector with the same working mode as the historical dataset and a correlation greater than a set value is found as a similar representative vector. When the fault information corresponding to the historical representative vector and the similar representative vector is inconsistent or the similar representative vector cannot be found, the historical representative vector replaces the representative vector in the working mode.

[0059] Specifically, since the correspondence between the target parameter values ​​and fault information changes over time, in order to improve the prediction accuracy of the spatial prediction model, before inputting the real-time representative vector corresponding to the real-time dataset into the spatial feature model, the historical representative vector corresponding to the historical dataset within a preset time period before the current time period is input into the spatial feature model. The historical dataset operates in the same mode as the real-time dataset. The accuracy of the spatial feature vector is verified by searching for similar representative vectors. If the fault information corresponding to the historical representative vector and the corresponding similar representative vector is inconsistent or the similar representative vector cannot be found, it indicates that the accuracy of the spatial feature model in the current operating mode is insufficient. Therefore, the historical representative vector replaces the representative vector in this operating mode. Through the above technical solution, the real-time accuracy of the spatial feature model can be guaranteed, thereby improving the prediction accuracy of inverter fault information.

[0060] Furthermore, the training of the ensemble learning model includes: selecting multiple sample datasets of the target parameter pairs under different operating modes of the frequency converter from the database, normalizing the data groups in each sample dataset, taking any two data groups in the sample dataset as sample data pairs, and training each sub-model in the ensemble learning model based on the sample data pairs and label data, wherein the analysis perspectives in the sub-model include at least a time perspective, a frequency domain perspective, and a statistical perspective.

[0061] Specifically, by selecting multiple sample datasets from the database corresponding to the target parameters of the inverter under different operating modes, and normalizing the data groups in the sample datasets, it is ensured that data from different sensors or of different types can be compared and analyzed on the same scale. A multi-sub-model structure is introduced into the ensemble learning model, with each sub-model analyzing the correlation of data groups from different perspectives, such as time, frequency domain, and statistical perspectives, thereby enhancing the diversity of the model. Each sub-model is trained for specific data features, such as current, voltage, mechanical vibration, and corresponding label data. The aforementioned label data is correlation data. The generalization ability of the model is improved by using a strategy of random sub-samples and random feature selection, thereby improving the accuracy of the ensemble learning model in calculating the correlation between any two data groups in the sample dataset.

[0062] Furthermore, the sub-model in the ensemble learning model is a gradient boosting tree model.

[0063] This invention also provides an artificial intelligence-based inverter fault prediction system for implementing the above-mentioned method, such as... Figure 4 As shown, the system includes:

[0064] The acquisition unit is used to select a target parameter from the multiple parameters based on the historical sensor datasets corresponding to multiple parameters of the frequency converter, and acquire multiple target datasets corresponding to the target parameter.

[0065] The computing unit is used to construct and train an ensemble learning model, calculate the correlation between any two data groups based on the target dataset with the target parameters according to the ensemble learning model, and combine the correlations into a correlation array of the target dataset;

[0066] The construction unit is used to analyze and process the correlation array corresponding to each target dataset, calculate multiple representative vectors for each target dataset, and construct a spatial feature model based on the multiple representative vectors corresponding to each target dataset.

[0067] The prediction unit is used to collect the real-time dataset of the target parameter within the current time period, calculate and input the real-time representative vector corresponding to the real-time dataset into the spatial feature model to obtain the fault prediction result.

[0068] The present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the above-described method.

[0069] In summary, this invention proposes an artificial intelligence-based method and system for predicting inverter faults, which has significant technical advantages. First, through multi-dimensional data processing and the application of an ensemble learning model, the system can effectively improve the accuracy of inverter fault prediction. By selecting multiple key parameters of the inverter and combining them with historical sensor data, the system extracts the target dataset and calculates the correlation between each data group based on these data, constructing a high-precision correlation array. This process analyzes the data from multiple perspectives, including time, frequency domain, and statistics, through multiple sub-models in the ensemble learning model, enabling the system to more comprehensively understand the characteristics of the data. Second, the system reduces the interference of redundant data by extracting representative vectors and constructing spatial feature models, while retaining the most representative feature information. By diagonalizing the representative vectors, the representative values ​​with the greatest contribution are extracted, further reducing the dimensionality of the data and improving the computational efficiency of subsequent predictive analysis. In addition, this method also uses a dynamic update mechanism to process real-time data. When the system detects inconsistencies between historical representative vectors and real-time fault information, it automatically replaces or adds representative vectors, ensuring that the model always accurately reflects the current operating state. This mechanism not only enhances the system's adaptability but also updates the fault prediction model in real time, effectively improving the timeliness and accuracy of prediction results. Most importantly, based on the spatial feature model, the system can quickly identify the most relevant fault modes by comparing real-time data with historical data, providing early warnings of potential fault risks. By setting a warning distance, the system can monitor the changing trends of faults and issue warnings within the warning distance, ensuring the safe operation of the frequency converter. In summary, this invention, by combining integrated learning, spatial feature models, and dynamic update mechanisms, can greatly improve the efficiency and accuracy of frequency converter fault prediction, providing reliable technical support for preventive maintenance of industrial equipment.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0071] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0072] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting a fault of a frequency converter based on artificial intelligence, characterized by, The method comprises: Based on the acquisition of variable frequency converter multiple parameters corresponding to the historical data set, and from the multiple parameters to select the target parameter and the target parameter corresponding to the multiple target data set; Target parameters at least include working mode, input and output voltage, input and output current, temperature, motor speed and mechanical vibration data; Construct and train the integrated learning model, calculate the correlation between any two data groups based on each target data set of the target parameter according to the integrated learning model, and combine the correlation into the correlation array of the target data set; By analyzing and processing each target data set corresponding to the correlation array, calculate the multiple representative vectors corresponding to each target data set, and construct the space feature model based on the multiple representative vectors corresponding to each target data set; Collect the real-time data set of the target parameter in the current time period, calculate and according to the real-time representative vector corresponding to the real-time data set, input the real-time representative vector into the space feature model, obtain the target representative vector with the correlation degree in the set range, and obtain the fault prediction result according to the fault information corresponding to the target representative vector; Wherein, the acquisition of the correlation array comprises: constructing and training the integrated learning model, the integrated learning model comprises multiple sub models, the values of the target parameters collected in the same time in the target data set are combined as the data group, any two data groups are input into the integrated learning model, two data groups input by each sub model in the integrated learning model are analyzed and inferred by different angles, the corresponding inference correlation is output, the correlation between two data groups is calculated according to the numerical distribution of each inference correlation, and the correlation between any two data groups in the target data set is combined into the correlation array, the different angles at least include time angle, frequency domain angle and statistical angle; Wherein, according to the numerical distribution of each inference correlation, the correlation between two data groups is calculated, which comprises: respectively calculating the sum of difference value between each inference correlation and other inference output by other model, and setting weight according to the difference value corresponding to each inference correlation, and taking the weighted result as the correlation between two data groups, wherein the sum of weight of inference correlation corresponding to different model output of two data groups is 1, and the weight corresponding to the inference correlation with larger difference value is smaller; The obtaining of the representative vector of the target data set and the construction of the spatial feature model comprises: obtaining a representative array according to a correlation array and a de-meaning array corresponding to the target data set, performing diagonalization processing on the representative array, obtaining a plurality of representative values, and taking a vector corresponding to m representative values with a contribution greater than a set threshold as the representative vector of the target data set, and adjusting the value of m according to the frequency converter fault information corresponding to the data group in the target data set; mapping each representative vector to a multi-dimensional feature coordinate system, adding the frequency converter fault information corresponding to the target data group to the position of the corresponding representative vector, and obtaining the spatial feature model.

2. The method of claim 1, wherein, The obtaining of a plurality of target data sets corresponding to the target parameter comprises: Obtaining a historical sensor data set corresponding to the plurality of parameters of the frequency converter and corresponding fault information, selecting the target parameter according to the key degree of the frequency converter fault detection of each parameter, and obtaining a plurality of target data sets corresponding to the target parameter from the historical sensor data set according to the working mode, wherein the target parameter comprises a plurality of parameters of the frequency converter.

3. The method of claim 1, wherein, The input of the real-time representative vector into the spatial feature model to obtain a fault prediction result comprises: The input of each real-time representative vector corresponding to the real-time data set into the spatial feature model, the obtaining of a target representative vector with a correlation degree within a set range with the real-time representative vector through the spatial feature model, the prediction of a normal fault information when the fault information of the target representative vector is all normal, otherwise, the prediction of an abnormal fault information, the calculation of a distance between the real-time representative vector and the representative vector with an abnormal fault information and the closest distance when the predicted fault information is normal, the obtaining of a fault information change trend according to the change of the distance, the sending of a warning information to a user terminal when the distance is less than a warning distance, and the taking of the predicted fault information as the fault prediction result, wherein the working mode corresponding to the target representative vector is the same as the working mode of the real-time representative vector.

4. The method of claim 1, wherein, Before the input of the real-time representative vector corresponding to the real-time data set into the spatial feature model, the method further comprises: The input of a historical representative vector corresponding to a historical data set in a preset time period before the current time period into the spatial feature model, the finding of a representative vector with a correlation degree greater than a set value in the same working mode as the historical data set as a similar representative vector, the replacement of the historical representative vector with the representative vector in the working mode when the fault information corresponding to the historical representative vector and the similar representative vector is inconsistent or the similar representative vector cannot be found.

5. The method of claim 1, wherein, The training of the ensemble learning model comprises: The target parameter corresponding to the frequency converter in different working modes is selected from the database, and each data set in the sample data set is normalized. Any two data sets in the sample data set are used as sample data pairs, and each sub-model in the ensemble learning model is trained based on the sample data pairs and label data pairs. The analysis angle in the sub-model includes at least time angle, frequency domain angle and statistical angle.

6. The method of claim 1, wherein, The sub-model in the ensemble learning model is a gradient boosting tree model.

7. An artificial intelligence based frequency inverter failure prediction system for implementing the method as claimed in any one of claims 1-6, characterized by, The system comprises: An acquisition unit is configured to acquire a plurality of parameters of a frequency converter corresponding to a sensing data set, select a target parameter from the plurality of parameters, and acquire a plurality of target data sets corresponding to the target parameter in historical data. The target parameter includes at least working mode, input and output voltage, input and output current, temperature, motor speed and mechanical vibration data. A calculation unit is configured to construct and train an ensemble learning model, calculate the correlation between any two data sets based on the target data set of the target parameter according to the ensemble learning model, and combine the correlation into a correlation matrix of the target data set. A construction unit is configured to analyze and process the correlation matrix corresponding to each target data set, calculate a plurality of representative vectors of each target data set, and construct a spatial feature model based on the plurality of representative vectors corresponding to each target data set. A prediction unit is configured to acquire a real-time data set of the target parameter in a current time period, calculate and obtain a real-time representative vector corresponding to the real-time data set, input the real-time representative vector into the spatial feature model, and obtain a fault prediction result.

8. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement the method of any one of claims 1-6.

Images