Frequency converter fault prediction method and system based on machine learning

Abstract

The invention relates to the field of frequency converter fault detection, and discloses a frequency converter fault prediction method and system based on machine learning, and the method comprises the steps: obtaining multi-dimensional real-time data in the operation process of a frequency converter; constructing a dynamic mapping relation to obtain a basic feature set; generating a time sequence feature vector capable of reflecting the state change of the equipment based on the basic feature set; comparing, analyzing and judging whether the equipment state deviates from a normal operation interval or not based on the historical operation data and the time sequence feature vector, and outputting a state deviation index; performing abnormal fluctuation judgment on the time sequence feature vector; extracting fluctuation amplitude and frequency characteristics of the key indexes to obtain quantitative description data of abnormal fluctuation; inputting the quantitative description data of the abnormal fluctuation into an abnormal prediction model; and generating a coping strategy and a triggering condition of the coping strategy based on the risk prediction result. The method has the advantages that the abnormal state of the frequency converter is recognized in time, and potential risks are predicted.

Description

Technical Field

[0001] The present invention relates to the field of inverter fault detection, and in particular to an inverter fault prediction method and system based on machine learning. Background Art

[0002] As core equipment in the field of industrial automation, inverters play a vital role in ensuring production efficiency and stable equipment operation. Their health status is directly related to the safety and reliability of the entire system.

[0003] However, existing technologies, which mostly rely on superficial parameters such as current or voltage changes, make it difficult to accurately identify early signs of failure. With the advancement of industrial intelligence, predicting inverter failures and taking proactive intervention measures are key to extending equipment life and reducing maintenance costs.

[0004] How to achieve early prediction of inverter failure is a technical problem that technicians in this field need to overcome. Summary of the Invention

[0005] The present invention provides a method and system for predicting inverter faults based on machine learning to at least partially solve the above technical problems.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a method for predicting inverter faults based on machine learning, comprising: Acquire multi-dimensional real-time data during the operation of the inverter and perform standardization processing on the multi-dimensional real-time data to obtain a structured operation data set; the multi-dimensional real-time data includes: current, voltage and vibration amplitude; Mining the deep correlation characteristics between key indicators in the multi-dimensional real-time data, building a dynamic mapping relationship to obtain a basic feature set; Generating a time series feature vector reflecting a change in a device state based on the basic feature set; Based on the comparison and analysis of historical operation data and time series feature vectors, it is determined whether the equipment status deviates from the normal operating range and outputs the status deviation index; If the state deviation index exceeds the preset threshold, the time series feature vector is judged to have abnormal fluctuations; if abnormal fluctuations exist, the fluctuation amplitude and frequency characteristics of the key indicators are extracted to obtain quantitative description data of the abnormal fluctuations; Inputting the quantitative description data of abnormal fluctuations into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability and time window of occurrence of potential risks; Generate response strategies and triggering conditions for the response strategies based on the risk prediction results.

[0007] In an optional embodiment, deep correlation characteristics between key indicators in the multi-dimensional real-time data are mined, and a dynamic mapping relationship is constructed to obtain a basic feature set, including: The Pearson correlation coefficient algorithm is used to calculate the linear correlation strength between the standardized operating parameter sets to identify the key parameter combinations with significant linear correlation; The mutual information entropy algorithm is used to quantify the nonlinear correlation between key parameters in key parameter combinations with significant linear correlation; The dynamic transfer entropy model based on time series is used to analyze the information transfer amount under different time delays to determine the optimal delay time for the causal influence between key parameters; Based on the key parameter combinations with significant linear correlation, the nonlinear correlation between key parameters and the optimal delay time, a dynamic mapping relationship between key parameters is constructed to describe the trend of parameter changes over time and operating conditions; The principal component analysis algorithm is used to extract the first several principal components whose cumulative contribution rate reaches a preset threshold as the basic feature set.

[0008] In an optional implementation, generating a time series feature vector reflecting a change in a device state based on the basic feature set includes: Obtaining a preliminary feature set based on the basic feature set using a pre-established feature screening mechanism; Comparing the correlation matrix between the features in the preliminary feature set to determine an initial core feature set; the core feature set is a key feature subset selected based on significant correlations between features; Based on the initial core feature set, a neural network model is used to perform multi-level abstract processing on the features in the initial core feature set to obtain a deep feature representation; Serialize and decompose deep features in the time dimension to obtain the distribution pattern of deep features at different time points; A feature reconstruction operation is performed according to the distribution pattern to reassemble the decomposed deep features into a continuous time series feature sequence; based on the reorganized time series feature sequence, the change pattern of the deep features in the time dimension is identified; wherein, if the fluctuation of the time series feature sequence in certain time periods exceeds a preset threshold, the segment of the sequence is locally smoothed to obtain a stable time series feature vector.

[0009] In an optional embodiment, determining whether the device state deviates from the normal operating range based on a comparative analysis of historical operating data and a time series feature vector and outputting a state deviation index includes: Based on the feature data set and historical data records, the constructed equipment status perception model is trained to determine the normal interval boundary of the equipment status; The latest time series feature vector is obtained from the real-time operation of the device and compared with the features in the historical data to determine whether the current time series feature vector meets the normal interval boundary. If the current time series feature vector deviates from the normal interval boundary, the degree of deviation is calculated to generate a state deviation index.

[0010] In an optional embodiment, if the state deviation index exceeds a preset threshold, the time series feature vector is judged to be abnormally fluctuating; if abnormal fluctuation exists, the fluctuation amplitude and frequency characteristics of the key indicators are extracted to obtain quantitative description data of the abnormal fluctuation, including: If the state deviation index exceeds the preset threshold range, the time series feature data is preliminarily screened to extract abnormal data points to obtain a preliminary abnormal data set; Decomposing the time series features using a vector analysis method based on the preliminary abnormal data set, separating key parameters to obtain a key parameter set; For a set of key parameters, the fluctuation amplitude and frequency characteristics are calculated; if the fluctuation amplitude exceeds the preset range, it is marked as an abnormal fluctuation point, and a set of abnormal fluctuation points is obtained; Perform time series analysis on the abnormal fluctuation point set to extract the fluctuation frequency characteristics and obtain frequency distribution feature data; if the frequency distribution feature data presents a non-uniform distribution, then combine it with the abnormal fluctuation point set for analysis; The support vector machine algorithm is used to classify abnormal fluctuations to obtain classified abnormal type data; The quantitative results are extracted based on the classified abnormal type data to generate quantitative description data of abnormal fluctuations.

[0011] In an optional embodiment, the quantitative description data of the abnormal fluctuation is input into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability of occurrence and time window of the potential risk, including: Based on the input quantitative description data of abnormal fluctuations, the abnormal fluctuation features are extracted to obtain the abnormal fluctuation data set; Perform trend feature analysis based on the key parameter set in the abnormal fluctuation data set to obtain the change trend and obtain the trend feature vector; inputting the trend feature vector into an anomaly prediction model; Obtain risk probability from the anomaly prediction model and combine it with time window analysis to determine the probability distribution of risk occurrence; According to the risk probability distribution, a weighted average method is used to generate risk prediction results and output the final prediction value; Based on the final prediction value and the time window, the risk warning time point is determined and the risk distribution within the time window is generated.

[0012] In an optional embodiment, generating a response strategy and a triggering condition of the response strategy based on the risk prediction result includes: Obtain the risk occurrence probability value and its corresponding prediction time window from the risk prediction results, and use the classification model to divide the risk level to obtain the risk level division result; Based on the risk level classification results and the risk probability distribution characteristics, a Bayesian network model is used to generate multiple potential risk response strategies to form a strategy set; Extracting the applicable time window of each strategy from the strategy set, combining the predicted time window, and using a time series analysis method to calculate the expected effect of each strategy at different time points to obtain an effect prediction value; If the effect prediction value is greater than the preset effect significance threshold T, the corresponding strategy is included in the candidate intervention strategy set; Based on the candidate intervention strategy set and the risk probability distribution characteristics, the triggering probability of each strategy in the prediction time window is calculated to obtain the triggering probability distribution; If the maximum probability value in the trigger probability distribution is greater than the preset trigger reliability threshold P, the strategy with the highest trigger probability is selected from the candidate strategies as the final intervention strategy; According to the final intervention strategy and the predicted time window combined with the real-time monitoring indicators, a specific intervention trigger condition is generated and a trigger condition description is output.

[0013] In an optional embodiment, the method further includes: Collecting inverter operating status data and control instruction response data before and after each response strategy execution; the control instruction response data includes strategy type, execution time point, parameter adjustment value and control module feedback status; Compare the real-time data before execution with the real-time data after execution to calculate the execution deviation index; The process information of each strategy execution is stored in a strategy execution effect database; the database record fields include: strategy type, trigger probability, execution timestamp, running status data before and after execution, execution deviation index, energy consumption change and strategy effectiveness mark; A reinforcement learning algorithm is used to train the Bayesian network strategy generation model online. This includes: using data from the strategy execution effect database as training samples; using execution success rate, deviation reduction, and energy consumption change as reward functions; updating the state transition probability and selection weight of each strategy node in the Bayesian network; retaining the optimal model version and setting up an automatic rollback mechanism to prevent model performance degradation; In the subsequent risk prediction stage, the updated Bayesian network strategy generation model is called.

[0014] In an optional embodiment, before recombining the deep features into a continuous temporal feature sequence, the method further includes: According to the operating characteristics of the equipment, sliding windows of multiple time scales are set to capture the state change rules at different time scales; the sliding windows of multiple time scales include a first scale window, a second scale window, and a third scale window; The feature subsets extracted at each time scale are organized into feature vectors, where each feature vector contains a timestamp, feature name, feature value, and the time scale identifier to which it belongs; Normalize the feature vectors of each scale; calculate the correlation weights between the features; and perform weighted fusion of features of different time scales according to the weights to generate a unified fused time series feature vector.

[0015] In a second aspect, the present invention provides a frequency converter fault prediction system based on machine learning, comprising: The first processing module is configured to obtain multi-dimensional real-time data during the operation of the inverter and perform standardization processing on the multi-dimensional real-time data to obtain a structured operation data set; the multi-dimensional real-time data includes current, voltage, and vibration amplitude; The second processing module is used to: mine the deep correlation characteristics between the key indicators in the multi-dimensional real-time data, and construct a dynamic mapping relationship to obtain a basic feature set; A third processing module is configured to generate a time series feature vector reflecting a change in a device state based on the basic feature set; The fourth processing module is configured to: determine whether the device state deviates from the normal operating range based on a comparative analysis of the historical operating data and the time series feature vector and output a state deviation index; The fifth processing module is used to: if the state deviation index exceeds the preset threshold, determine the abnormal fluctuation of the time series feature vector; if abnormal fluctuation exists, extract the fluctuation amplitude and frequency characteristics of the key indicators to obtain quantitative description data of the abnormal fluctuation; A sixth processing module is configured to: input the quantitative description data of the abnormal fluctuation into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability of occurrence and time window of the potential risk; The seventh processing module is used to generate a response strategy and a triggering condition of the response strategy based on the risk prediction result.

[0016] Compared with the existing technology, the present invention has at least the following beneficial effects: timely identification of abnormal status of the inverter, prediction of potential risks, and taking corresponding intervention measures, thereby improving the reliability and safety of equipment operation and reducing the probability of failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of a machine learning-based inverter fault prediction method provided by a first embodiment of the present application; Figure 2 is a block diagram of a machine learning-based inverter fault prediction system provided by a second embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described 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, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0019] With reference to Figure 1 The first embodiment of the present application provides a machine learning-based inverter fault prediction method, comprising the following steps: S101, acquiring multi-dimensional real-time data in the running process of an inverter and performing standardization processing on the multi-dimensional real-time data to obtain a structured running data set; the multi-dimensional real-time data includes current, voltage, and vibration amplitude; S102, mining deep-level correlation characteristics between each key indicator in the multi-dimensional real-time data to build a dynamic mapping relationship to obtain a basic feature set; S103, generating a time sequence feature vector reflecting the state change of the device based on the basic feature set; S104, comparing and analyzing the historical running data and the time sequence feature vector to determine whether the device state deviates from the normal running interval and outputting a state deviation indicator; S105, if the state deviation indicator exceeds a preset threshold, performing an abnormal fluctuation judgment on the time sequence feature vector; if there is an abnormal fluctuation, extracting the fluctuation amplitude and frequency characteristics of the key indicators to obtain quantitative description data of the abnormal fluctuation; S106, inputting the quantitative description data of the abnormal fluctuation into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the occurrence probability and time window of the potential risk.

[0020] S107, generating a coping strategy and a trigger condition of the coping strategy based on the risk prediction result.

[0021] Specifically, this application obtains multi-dimensional real-time data such as current, voltage and vibration amplitude during the operation of the inverter and standardizes it, mines the correlation between various indicators to construct a dynamic mapping relationship to form a basic feature set and then generates a time series feature vector, combines historical data comparison to judge the deviation of the equipment status, quantifies the abnormal fluctuations that exceed the threshold, and inputs the time series feature vector into the abnormal prediction model to determine the probability and time window of potential risks, and finally generates a response strategy and trigger conditions, thereby realizing early prediction of inverter failures, improving the timeliness of equipment operation status assessment, and being able to identify potential risks in advance and actively intervene, effectively reducing the probability of failure.

[0022] In one embodiment, deep correlation characteristics between key indicators in the multi-dimensional real-time data are mined, and a dynamic mapping relationship is constructed to obtain a basic feature set, including: The Pearson correlation coefficient algorithm is used to calculate the linear correlation strength between the standardized operating parameter sets to identify the key parameter combinations with significant linear correlation; The mutual information entropy algorithm is used to quantify the nonlinear correlation between key parameters in key parameter combinations with significant linear correlation; The dynamic transfer entropy model based on time series is used to analyze the information transfer amount under different time delays to determine the optimal delay time for the causal influence between key parameters; Based on the key parameter combinations with significant linear correlation, the nonlinear correlation between key parameters and the optimal delay time, a dynamic mapping relationship between key parameters is constructed to describe the trend of parameter changes over time and operating conditions; The principal component analysis algorithm is used to extract the first several principal components whose cumulative contribution rate reaches a preset threshold as the basic feature set.

[0023] Specifically, the Pearson correlation coefficient algorithm is used to calculate the strength of the linear association between each parameter on the standardized set of operating parameters. The Pearson correlation coefficient measures the degree of linear correlation between two variables. Its value ranges from -1 to +1, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation. Larger absolute values ​​indicate stronger associations. This method identifies key parameter combinations with significant linear correlations.

[0024] For the key parameter combinations with significant linear correlations identified, the mutual information entropy algorithm is used to quantify the nonlinear correlation between these parameters. Mutual information entropy is a method to measure the dependence between two random variables and can capture the nonlinear relationship between variables.

[0025] A dynamic transfer entropy model based on time series is used to set the length and step of the sliding window, calculate the information transfer amount under different time delays, and determine the optimal delay time of the causal effect between the key parameters. Transfer entropy is used to measure the amount of information flow from one time series to another.

[0026] Based on the above analysis results, i.e., the combination of key parameters with significant linear correlation, the nonlinear correlation between key parameters, and the optimal delay time, a dynamic mapping relationship between key parameters is constructed. The mapping relationship is used to describe the trend of the parameters over time and operating conditions.

[0027] The principal component analysis algorithm is used to process the above-mentioned data set, and the first several principal components with a cumulative contribution rate reaching a preset threshold are extracted as a basic feature set. Principal component analysis algorithm is a statistical method that converts a set of possibly correlated variables into a set of linearly independent variables called principal components through orthogonal transformation, thereby reducing the dimensionality of the data set without losing important information.

[0028] Through the above scheme, the deep features of the equipment operating state can be deeply mined, and an accurate dynamic mapping relationship can be established, thereby realizing state monitoring and fault prediction.

[0029] In one implementation, a time series feature vector reflecting the change of the equipment state is generated based on the basic feature set, including: A preliminary feature set is obtained based on the basic feature set using a pre-established feature screening mechanism; An initial core feature set is determined by comparing the correlation matrix between the features in the preliminary feature set; the core feature set is a key feature subset selected based on significant correlation between features; A deep feature representation is obtained by using a neural network model to perform multi-level abstract processing on the features in the initial core feature set based on the initial core feature set; The deep features are sequentially decomposed in the time dimension to obtain the distribution pattern of the deep features at different time points; A feature reconstruction operation is performed according to the distribution pattern to recombine the decomposed deep features into a continuous time series feature sequence; based on the recombined time series feature sequence, the change rule of the deep features in the time dimension is identified; if the fluctuation of the time series feature sequence in some time period exceeds a preset threshold, local smoothing processing is performed on the sequence to obtain a stable time series feature vector.

[0030] Specifically, a pre-established feature screening mechanism is used to extract a core feature set from the underlying feature set of the latent pattern. The initial core feature set is determined by comparing the correlation matrix between features, ensuring significant correlations between the selected features to form a subset of key features. For example, a support vector machine algorithm can be used to classify the latent pattern to obtain the final set of core features.

[0031] Based on the initial core feature set, a neural network model (such as convolutional neural network (CNN)) is used to perform multi-level abstract processing on these features to obtain deep feature representation.

[0032] To represent deep features, we apply time series analysis methods to serialize and decompose features along the time dimension, capturing their distribution patterns at different time points. Long Short-Term Memory (LSTM) networks can be used here. For example, by setting an appropriate time step and number of hidden layer units, and inputting feature graph data from the past several hours or minutes, we can capture trends in device status over time.

[0033] Based on this distribution pattern, feature reconstruction is performed to reassemble the decomposed deep features into a continuous time series feature sequence, identifying the main changes in the deep features over time. If the fluctuations in the reorganized time series feature sequence exceed a preset threshold within a certain time period, local smoothing is performed on that segment to obtain a stable time series feature vector.

[0034] In one embodiment, the process of determining whether the device state deviates from the normal operating range and outputting a state deviation index based on a comparative analysis of historical operating data and a time series feature vector includes: Based on the feature data set and historical data records, the constructed equipment status perception model is trained to determine the normal interval boundary of the equipment status; The latest time series feature vector is obtained from the real-time operation of the device and compared with the features in the historical data to determine whether the current time series feature vector meets the normal interval boundary. If the current time series feature vector deviates from the normal interval boundary, the degree of deviation is calculated to generate a state deviation index.

[0035] Specifically, a device status awareness model is constructed by combining historical data records. This model typically uses machine learning algorithms (such as support vector machines (SVMs)) to classify and train feature datasets to determine the boundaries of the normal range for device status. For example, using the past 30 days of normal operation data as a training set, SVM training can determine the decision boundary for normal operation. The decision boundary is the dividing line between different categories. For the device status awareness model, it serves as the boundary that distinguishes normal from abnormal operation.

[0036] The latest time series feature vector acquired during real-time device operation is compared and analyzed with features from historical data to determine whether the current time series feature vector conforms to the normal range boundaries. If the current vector deviates from the normal range boundaries, the degree of deviation is calculated to generate a state deviation indicator. Specifically, the current feature vector is input into a trained device state perception model, and its distance from the decision boundary is calculated. If this distance exceeds a preset threshold, the device state is considered to have deviated, and a state deviation indicator is generated based on this information.

[0037] Once it is confirmed that the equipment status deviates from the normal range, the degree of deviation is further calculated to generate a status deviation index.

[0038] Through the above steps, it is possible to accurately identify whether the state of the inverter deviates from the normal operating range and the specific deviation degree index.

[0039] In one embodiment, if the state deviation indicator exceeds a preset threshold, the time series feature vector is judged to have abnormal fluctuations; if abnormal fluctuations exist, the fluctuation amplitude and frequency characteristics of the key indicator are extracted to obtain quantitative description data of the abnormal fluctuations, including: If the state deviation index exceeds the preset threshold range, the time series feature data is preliminarily screened to extract abnormal data points to obtain a preliminary abnormal data set; Decomposing the time series features using a vector analysis method based on the preliminary abnormal data set, separating key parameters to obtain a key parameter set; For a set of key parameters, the fluctuation amplitude and frequency characteristics are calculated; if the fluctuation amplitude exceeds the preset range, it is marked as an abnormal fluctuation point, and a set of abnormal fluctuation points is obtained; Perform time series analysis on the abnormal fluctuation point set to extract the fluctuation frequency characteristics and obtain frequency distribution feature data; if the frequency distribution feature data presents a non-uniform distribution, then combine it with the abnormal fluctuation point set for analysis; The support vector machine algorithm is used to classify abnormal fluctuations to obtain classified abnormal type data; Extract quantitative results based on the classified abnormal type data and generate quantitative description data of abnormal fluctuations; The quantitative description data are integrated, and the state deviation and deviation range are associated to form a complete abnormal fluctuation analysis result.

[0040] Specifically, if the state deviation indicator exceeds the preset threshold, the time series feature data is first preliminarily screened to extract abnormal data points, thereby forming a preliminary abnormal data set. Based on this preliminary abnormal data set, vector analysis methods are used to decompose the time series features, isolate key parameters, and obtain a key parameter set. For this key parameter set, the fluctuation amplitude and frequency characteristics are calculated. If the fluctuation amplitude of a parameter exceeds the preset range, it is marked as an abnormal fluctuation point, forming an abnormal fluctuation point set. For example, the standard deviation formula is used to measure the fluctuation amplitude, and the fast Fourier transform method is used to extract the frequency characteristics. Time series analysis is performed on the abnormal fluctuation point set to extract the fluctuation frequency characteristics and obtain frequency distribution feature data. If the frequency distribution feature data shows a non-uniform distribution, in-depth analysis is required in conjunction with the abnormal fluctuation point set. In this case, the support vector machine (SVM) algorithm can be used to classify abnormal fluctuations to identify different types of abnormal patterns.

[0041] Based on the classified abnormal type data, quantitative results are extracted to generate quantitative description data of abnormal fluctuations. These quantitative description data are integrated and associated with state deviations and deviation ranges to form a complete abnormal fluctuation analysis result.

[0042] In one embodiment, quantitative description data of abnormal fluctuations are input into an abnormality prediction model; the abnormality prediction model outputs a risk prediction result to determine the probability and time window of occurrence of potential risks, including: Based on the input quantitative description data of abnormal fluctuations, the abnormal fluctuation features are extracted to obtain the abnormal fluctuation data set; Perform trend feature analysis based on the key parameter set in the abnormal fluctuation data set to obtain the change trend and obtain the trend feature vector; inputting the trend feature vector into an anomaly prediction model; Obtain risk probability from the anomaly prediction model and combine it with time window analysis to determine the probability distribution of risk occurrence; According to the risk probability distribution, a weighted average method is used to generate risk prediction results and output the final prediction value; Based on the final prediction value and the time window, the risk warning time point is determined and the risk distribution within the time window is generated.

[0043] Specifically, based on the input quantitative description of abnormal fluctuations, a feature extraction method is used to determine a set of key parameters and generate an abnormal fluctuation dataset. This process includes calculating statistics such as mean, standard deviation, skewness, and kurtosis, and processing the time series data using a sliding window method to capture the key characteristics of abnormal fluctuations.

[0044] Perform trend analysis based on the key parameter set in the abnormal fluctuation data set to identify its changing trend and obtain the trend feature vector. For example, you can use ARIMA (Autoregressive Integrated Moving Average) to fit the standard deviation series and predict the fluctuation trend over a period of time.

[0045] The trend feature vector obtained above is input into the anomaly prediction model. The anomaly prediction model can be built using machine learning algorithms such as Random Forest and XG Boost.

[0046] After obtaining the risk probability from the anomaly prediction model, we combine it with time window analysis to determine the risk probability distribution. The time window here refers to the risk that may occur within a specific time period in the future. Next, we use a weighted average method to generate the risk prediction results and output the final prediction value.

[0047] Based on the final predicted value and the time window, the risk warning time point is determined and the risk distribution within the time window is generated. This not only provides an overall risk score, but also indicates the specific time period in the future when the risk is highest, so that appropriate preventive measures can be taken.

[0048] In one embodiment, generating a response strategy and triggering conditions of the response strategy based on the risk prediction results includes: Obtain the risk occurrence probability value and its corresponding prediction time window from the risk prediction results, and use the classification model to divide the risk level to obtain the risk level division result; Based on the risk level classification results and the risk probability distribution characteristics, a Bayesian network model is used to generate multiple potential risk response strategies to form a strategy set; Extracting the applicable time window of each strategy from the strategy set, combining the predicted time window, and using a time series analysis method to calculate the expected effect of each strategy at different time points to obtain an effect prediction value; If the effect prediction value is greater than the preset effect significance threshold T, the corresponding strategy is included in the candidate intervention strategy set; Based on the candidate intervention strategy set and the risk probability distribution characteristics, the triggering probability of each strategy within the prediction time window is calculated to obtain the triggering probability distribution; If the maximum probability value in the trigger probability distribution is greater than the preset trigger reliability threshold P, the strategy with the highest trigger probability is selected from the candidate strategies as the final intervention strategy; According to the final intervention strategy and the predicted time window combined with the real-time monitoring indicators, a specific intervention trigger condition is generated and a trigger condition description is output.

[0049] Specifically, a risk occurrence probability value and a corresponding prediction time window are obtained from the risk prediction result, and a pre-established classification model is used to divide the risk level to obtain a risk level division result. According to the risk level division result and in combination with a risk probability distribution feature, a Bayesian network model is used to generate a plurality of potential risk coping strategies to form a strategy set.

[0050] An applicable time window of each strategy is extracted from the strategy set, and in combination with the prediction time window, a time series analysis method is used to calculate an expected effect of each strategy at different time points to obtain an effect prediction value.

[0051] If the effect prediction value is greater than a preset effect significance threshold T, the corresponding strategy is included in a candidate intervention strategy set. Based on the candidate intervention strategy set and in combination with the risk probability distribution feature, a trigger probability of each strategy within the prediction time window is calculated to obtain a trigger probability distribution. If a maximum probability value in the trigger probability distribution is greater than a preset trigger reliability threshold P, a strategy with the highest trigger probability is selected from the candidate strategy as a final intervention strategy. According to the final intervention strategy and the prediction time window and in combination with real-time monitoring indicators, a specific intervention trigger condition is generated and a trigger condition description is output. The trigger condition specifies when and how to execute the intervention measure.

[0052] In an implementation manner, the method further includes: Running state data and control instruction response data of the frequency converter are collected before and after each coping strategy is executed; the control instruction response data includes a strategy type, an execution time point, a parameter adjustment value, and a control module feedback state; The real-time data before execution is compared with the real-time data after execution to calculate an execution deviation index; Process information of each strategy execution is stored in a strategy execution effect database; the database record fields include a strategy type, a trigger probability, an execution time stamp, running state data before and after execution, an execution deviation index, an energy consumption change amount, and a strategy effectiveness identifier; A reinforcement learning algorithm is used to perform online training on the Bayesian network strategy generation model; the training includes: data in the strategy execution effect database is input as a training sample; an execution success rate, a deviation reduction amount, and an energy consumption change are used as a reward function; state transition probabilities and selection weights of each strategy node in the Bayesian network are updated; an optimal model version is retained, and an automatic rollback mechanism is set to prevent model performance from decreasing; In a subsequent risk prediction stage, the updated Bayesian network strategy generation model is called.

[0053] Specifically, before and after each response strategy execution, the inverter's operating status data and control command response data are collected. This control command response data includes, but is not limited to, the strategy type, execution time, parameter adjustment values, and control module feedback status. The real-time data before and after execution is compared to calculate an execution deviation index. Control command response data is a dataset containing all relevant information about a specific control command, such as command type, execution time, and results. The execution deviation index measures the difference between the actual results before and after strategy execution and the expected goals, and is used to evaluate the effectiveness of the strategy.

[0054] The progress of each policy execution is stored in a policy execution effect database. This database records fields including, but not limited to, policy type, trigger probability, execution timestamp, pre- and post-execution operational status data, execution deviation indicator, energy consumption change, and policy effectiveness indicator. The policy effectiveness indicator is an identifier that indicates whether a policy has successfully achieved its intended goal and can be used to quickly screen for effective policies.

[0055] In one embodiment, before recombining the deep features into a continuous temporal feature sequence, the method further includes: Based on the operating characteristics of the equipment, sliding windows with multiple time scales are set to capture state change patterns at different time scales. The sliding windows with multiple time scales include a first-scale window, a second-scale window, and a third-scale window. The first-scale window has a window length of 5 minutes to capture sudden abnormal fluctuations; the second-scale window has a window length of 1 hour to identify periodic trend changes; and the third-scale window has a window length of 24 hours to reflect equipment performance degradation trends. The feature subsets extracted at each time scale are organized into feature vectors, where each feature vector contains a timestamp, feature name, feature value, and the time scale identifier to which it belongs; Normalize the feature vectors of each scale; calculate the correlation weights between the features; and perform weighted fusion of features of different time scales according to the weights to generate a unified fused time series feature vector.

[0056] Reference Figure 2 The second embodiment of the present invention provides a frequency converter fault prediction system based on machine learning, comprising: The first processing module is configured to obtain multi-dimensional real-time data during the operation of the inverter and perform standardization processing on the multi-dimensional real-time data to obtain a structured operation data set; the multi-dimensional real-time data includes current, voltage, and vibration amplitude; The second processing module is used to: mine the deep correlation characteristics between the key indicators in the multi-dimensional real-time data, and construct a dynamic mapping relationship to obtain a basic feature set; A third processing module is configured to generate a time series feature vector reflecting a change in a device state based on the basic feature set; The fourth processing module is configured to: determine whether the device state deviates from the normal operating range based on a comparative analysis of historical operating data and the time series feature vector and output a state deviation index; The fifth processing module is used to: if the state deviation index exceeds the preset threshold, determine the abnormal fluctuation of the time series feature vector; if abnormal fluctuation exists, extract the fluctuation amplitude and frequency characteristics of the key indicators to obtain quantitative description data of the abnormal fluctuation; A sixth processing module is configured to: input the quantitative description data of the abnormal fluctuation into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability of occurrence and time window of the potential risk; The seventh processing module is used to generate a response strategy and a triggering condition of the response strategy based on the risk prediction result.

[0057] It should be noted that the inverter fault prediction system based on machine learning provided in an embodiment of the present invention is used to execute all the process steps of the inverter fault prediction method based on machine learning in the above embodiment. The working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0058] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting inverter faults based on machine learning, characterized in that: include: Acquiring multi-dimensional real-time data during the operation of the inverter and performing standardization processing on the multi-dimensional real-time data to obtain a structured operation data set; The multi-dimensional real-time data includes: current, voltage and vibration amplitude; Mining the deep correlation characteristics between key indicators in the multi-dimensional real-time data, building a dynamic mapping relationship to obtain a basic feature set; Generating a time series feature vector reflecting a change in a device state based on the basic feature set; Based on the comparison and analysis of historical operation data and time series feature vectors, it is determined whether the equipment status deviates from the normal operating range and outputs the status deviation index; If the state deviation index exceeds the preset threshold, the time series feature vector is judged to have abnormal fluctuations; if abnormal fluctuations exist, the fluctuation amplitude and frequency characteristics of the key indicators are extracted to obtain quantitative description data of the abnormal fluctuations; Inputting the quantitative description data of abnormal fluctuations into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability and time window of occurrence of potential risks; Generate response strategies and triggering conditions for the response strategies based on the risk prediction results.

2. The inverter fault prediction method based on machine learning according to claim 1, characterized in that: Mining the deep correlation characteristics between key indicators in the multi-dimensional real-time data, building a dynamic mapping relationship to obtain a basic feature set, including: The Pearson correlation coefficient algorithm is used to calculate the linear correlation strength between the standardized operating parameter sets to identify the key parameter combinations with significant linear correlation; The mutual information entropy algorithm is used to quantify the nonlinear correlation between key parameters in key parameter combinations with significant linear correlation; The dynamic transfer entropy model based on time series is used to analyze the information transfer amount under different time delays to determine the optimal delay time for the causal influence between key parameters; Based on the key parameter combinations with significant linear correlation, the nonlinear correlation between key parameters and the optimal delay time, a dynamic mapping relationship between key parameters is constructed to describe the trend of parameter changes over time and operating conditions; The principal component analysis algorithm is used to extract the first several principal components whose cumulative contribution rate reaches a preset threshold as the basic feature set.

3. The inverter fault prediction method based on machine learning according to claim 2, characterized in that: Generating a time series feature vector reflecting a change in a device state based on the basic feature set includes: Obtaining a preliminary feature set based on the basic feature set using a pre-established feature screening mechanism; Comparing the correlation matrix between the features in the preliminary feature set to determine an initial core feature set; the core feature set is a key feature subset selected based on significant correlation between the features; Based on the initial core feature set, a neural network model is used to perform multi-level abstract processing on the features in the initial core feature set to obtain a deep feature representation; Serialize and decompose deep features in the time dimension to obtain the distribution pattern of deep features at different time points; A feature reconstruction operation is performed according to the distribution pattern to reassemble the decomposed deep features into a continuous time series feature sequence; based on the reorganized time series feature sequence, the change pattern of the deep features in the time dimension is identified; wherein, if the fluctuation of the time series feature sequence in certain time periods exceeds a preset threshold, the segment of the sequence is locally smoothed to obtain a stable time series feature vector.

4. The inverter fault prediction method based on machine learning according to claim 3 is characterized in that: Based on the comparison and analysis of historical operating data and time series feature vectors, it is determined whether the equipment status deviates from the normal operating range and the status deviation index is output, including: Based on the feature data set and historical data records, the constructed equipment status perception model is trained to determine the normal interval boundary of the equipment status; The latest time series feature vector is obtained from the real-time operation of the device and compared with the features in the historical data to determine whether the current time series feature vector meets the normal interval boundary. If the current time series feature vector deviates from the normal interval boundary, the degree of deviation is calculated to generate a state deviation index.

5. The inverter fault prediction method based on machine learning according to claim 4 is characterized in that: If the state deviation index exceeds the preset threshold, the time series feature vector is judged to be abnormally volatile; If there is abnormal fluctuation, the fluctuation amplitude and frequency characteristics of key indicators are extracted to obtain quantitative description data of abnormal fluctuation, including: If the state deviation index exceeds the preset threshold range, the time series feature data is preliminarily screened to extract abnormal data points to obtain a preliminary abnormal data set; Decomposing the time series features using a vector analysis method based on the preliminary abnormal data set, separating key parameters to obtain a key parameter set; For a set of key parameters, the fluctuation amplitude and frequency characteristics are calculated; if the fluctuation amplitude exceeds the preset range, it is marked as an abnormal fluctuation point, and a set of abnormal fluctuation points is obtained; Perform time series analysis on the abnormal fluctuation point set to extract the fluctuation frequency characteristics and obtain frequency distribution feature data; if the frequency distribution feature data presents a non-uniform distribution, then combine it with the abnormal fluctuation point set for analysis; The support vector machine algorithm is used to classify abnormal fluctuations to obtain classified abnormal type data; The quantitative results are extracted based on the classified abnormal type data to generate quantitative description data of abnormal fluctuations.

6. The inverter fault prediction method based on machine learning according to claim 5, characterized in that: Inputting the quantitative description data of abnormal fluctuations into the abnormal prediction model; The anomaly prediction model outputs risk prediction results to determine the probability and time window of potential risk occurrence, including: Based on the input quantitative description data of abnormal fluctuations, the abnormal fluctuation features are extracted to obtain the abnormal fluctuation data set; Perform trend feature analysis based on the key parameter set in the abnormal fluctuation data set to obtain the change trend and obtain the trend feature vector; inputting the trend feature vector into an anomaly prediction model; Obtain risk probability from the anomaly prediction model and combine it with time window analysis to determine the probability distribution of risk occurrence; According to the risk probability distribution, a weighted average method is used to generate risk prediction results and output the final prediction value; Based on the final prediction value and the time window, the risk warning time point is determined and the risk distribution within the time window is generated.

7. The inverter fault prediction method based on machine learning according to claim 6, characterized in that: Generate response strategies and triggering conditions of the response strategies based on risk prediction results, including: Obtain the risk occurrence probability value and its corresponding prediction time window from the risk prediction results, and use the classification model to divide the risk level to obtain the risk level division result; Based on the risk level classification results and the risk probability distribution characteristics, a Bayesian network model is used to generate multiple potential risk response strategies to form a strategy set; Extracting the applicable time window of each strategy from the strategy set, combining the predicted time window, and using a time series analysis method to calculate the expected effect of each strategy at different time points to obtain an effect prediction value; If the effect prediction value is greater than the preset effect significance threshold T, the corresponding strategy is included in the candidate intervention strategy set; Based on the candidate intervention strategy set and the risk probability distribution characteristics, the triggering probability of each strategy in the prediction time window is calculated to obtain the triggering probability distribution; If the maximum probability value in the trigger probability distribution is greater than the preset trigger reliability threshold P, the strategy with the highest trigger probability is selected from the candidate strategies as the final intervention strategy; According to the final intervention strategy and the predicted time window combined with the real-time monitoring indicators, a specific intervention trigger condition is generated and a trigger condition description is output.

8. The inverter fault prediction method based on machine learning according to claim 7, characterized in that: The method further comprises: Collecting inverter operating status data and control instruction response data before and after each response strategy execution; the control instruction response data includes strategy type, execution time point, parameter adjustment value and control module feedback status; Compare the real-time data before execution with the real-time data after execution to calculate the execution deviation index; The process information of each strategy execution is stored in a strategy execution effect database; the database record fields include: strategy type, trigger probability, execution timestamp, running status data before and after execution, execution deviation index, energy consumption change and strategy effectiveness mark; A reinforcement learning algorithm is used to train the Bayesian network strategy generation model online. This includes: using data from the strategy execution effect database as training samples; using execution success rate, deviation reduction, and energy consumption change as reward functions; updating the state transition probability and selection weight of each strategy node in the Bayesian network; retaining the optimal model version and setting up an automatic rollback mechanism to prevent model performance degradation; In the subsequent risk prediction stage, the updated Bayesian network strategy generation model is called.

9. The inverter fault prediction method based on machine learning according to claim 8, characterized in that: Before recombining the deep features into a continuous temporal feature sequence, the method further includes: According to the operating characteristics of the equipment, sliding windows of multiple time scales are set to capture the state change rules at different time scales; the sliding windows of multiple time scales include a first scale window, a second scale window, and a third scale window; The feature subsets extracted at each time scale are organized into feature vectors, where each feature vector contains a timestamp, feature name, feature value, and the time scale identifier to which it belongs; Normalize the feature vectors of each scale; calculate the correlation weights between the features; and perform weighted fusion of features of different time scales according to the weights to generate a unified fused time series feature vector.

10. A frequency converter fault prediction system based on machine learning, characterized in that: include: The first processing module is configured to obtain multi-dimensional real-time data during the operation of the inverter and perform standardization processing on the multi-dimensional real-time data to obtain a structured operation data set; the multi-dimensional real-time data includes current, voltage, and vibration amplitude; The second processing module is used to: mine the deep correlation characteristics between the key indicators in the multi-dimensional real-time data, and construct a dynamic mapping relationship to obtain a basic feature set; A third processing module is configured to generate a time series feature vector reflecting a change in a device state based on the basic feature set; The fourth processing module is configured to: determine whether the device state deviates from the normal operating range based on a comparative analysis of historical operating data and the time series feature vector and output a state deviation index; The fifth processing module is used to: if the state deviation index exceeds the preset threshold, determine the abnormal fluctuation of the time series feature vector; if abnormal fluctuation exists, extract the fluctuation amplitude and frequency characteristics of the key indicators to obtain quantitative description data of the abnormal fluctuation; A sixth processing module is configured to: input the quantitative description data of the abnormal fluctuation into an abnormal prediction model; the abnormal prediction model outputs a risk prediction result to determine the probability of occurrence and time window of the potential risk; The seventh processing module is used to generate a response strategy and a triggering condition of the response strategy based on the risk prediction result.

Images