A method for predicting inverter degradation under extreme environments based on causal inference

Abstract

The present invention belongs to the technical field of electronic device degradation prediction and identification, and specifically discloses a method for predicting inverter degradation under extreme environments based on causal inference. The method decomposes the impact of the environment on the working condition into three types of impacts: redundant, unique and synergistic, and quantifies the degree of each type of impact through mutual information, clarifies the specific impact of each environmental factor on the inverter working condition and its degree, and provides a new analysis perspective for the complex relationship between the environment and the working condition under extreme environments; the present invention also performs feature fusion by querying the timing features through environmental features, and splices the fused features with the timing features on the basis of the fused features to further enhance the fused features, effectively captures the impact of the current environment on the working condition, and fully considers the timing effects in the inverter degradation process, thereby significantly improving the prediction accuracy of future working condition data, and providing reliable support for the stable operation of communication systems in extreme environments such as deserts, islands and reefs, high altitudes and extreme cold.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic device degradation prediction and identification, and in particular relates to a method for predicting inverter degradation under extreme environments based on causal inference. Background Art

[0002] Under complex and extreme climatic conditions, such as deserts, high cold, humid or corrosive environments, the degradation of communication equipment is becoming increasingly serious. Common degradation manifestations include data fluctuations, data sudden changes, data drift, etc. These problems not only affect the continuity and reliability of communication services, but may even directly interfere with the execution and decision-making of key tasks, causing potential risks that cannot be ignored. Among the many factors that affect the stability and reliability of communication systems, the performance of inverters, as one of the core electronic components, is particularly affected by environmental changes. Therefore, accurately predicting the degradation state of inverters is the key to ensuring the normal operation of communication systems in extreme environments.

[0003] In the field of electronic device degradation prediction and identification, machine learning and deep learning technologies have been widely used. These methods can mine potential failure modes and make predictions by analyzing historical system degradation data. However, the accuracy and credibility of prediction results are often low when relying solely on historical degradation data for data mining, and the interpretability of existing models is still significantly insufficient. Especially in complex environments, the degradation of core electronic components such as inverters is often limited by environmental factors such as temperature, humidity, and corrosive gases. Modeling based solely on historical degradation data makes it difficult to fully consider the profound impact of the external environment on equipment performance, resulting in poor reliability of prediction results.

[0004] In order to solve this problem, some studies have gradually begun to incorporate environmental factors into the training process of the degradation prediction model. The introduction of environmental factors can not only more comprehensively reveal the performance of the model under different environmental conditions, but also improve the interpretability of the model, thereby improving the accuracy of fault prediction. Especially when it comes to environmentally sensitive electronic devices such as inverters, changes in environmental factors such as temperature, humidity, and air pressure may have a significant impact on their performance. By incorporating these variables into the machine learning model, it is not only possible to effectively predict the performance degradation of the inverter, but also to dynamically adjust to environmental changes during system operation, thereby extending the service life of the equipment.

[0005] However, the current analysis methods of environmental factors mostly use traditional statistical analysis techniques, such as Pearson correlation coefficient and Spearman correlation analysis. Although these methods can reveal the correlation between different variables and provide valuable information for model training, they usually ignore the potential causal relationship between variables and cannot deeply explore the specific impact of each environmental factor in the degradation process. In addition, these traditional methods have relatively weak generalization capabilities in multi-variable and multi-complex environments and cannot effectively cope with complex and changeable working conditions in reality.

[0006] In addition, with the continuous development of technology, real-time acquisition of environmental data has become increasingly feasible. Through an external environmental forecasting system, future environmental data can be provided to the prediction model. These real-time updated environmental data can not only provide more accurate input for the model, but also enhance the ability to predict future changes in working conditions and further improve the accuracy of fault warnings. Especially in scenarios with high-precision requirements, the introduction of future environmental data can significantly improve the timeliness and reliability of predictions, and provide strong guarantees for the long-term stable operation of key equipment such as inverters. Summary of the invention

[0007] The purpose of the present invention is to propose a method for predicting inverter degradation under extreme environments based on causal inference. The method utilizes causal inference and feature fusion technology, combined with historical and future environmental data, to achieve future operating condition prediction and degradation type identification of inverters deployed in extreme environments.

[0008] In order to achieve the above object, the present invention adopts the following technical scheme:

[0009] A method for predicting inverter degradation under extreme environment based on causal inference includes the following steps:

[0010] Step 1. Collect historical environmental and operating condition data of the entire inverter degradation process in extreme climate areas, annotate the inverter's operating status and degradation type, and construct a multi-label degradation dataset that combines the environment and inverter operating conditions;

[0011] Step 2. Build a frequency converter degradation prediction and identification model, which includes the following structure:

[0012] The environmental branch is used to extract the environmental impact characteristics on the working conditions, namely the environmental characteristics, quantify the impact of various environmental factors on the working conditions using causal inference technology, and integrate the impact of various environmental factors on the working conditions into a multi-dimensional environmental impact characteristic matrix;

[0013] Time series branch, used to extract time series features of historical operating condition data;

[0014] A two-branch feature fusion module that maps environmental features to queries and temporal features to keys and values, then fuses local neighborhood features using a cross-attention mechanism and further enhances the fused features through convolution and feature interaction;

[0015] Transformer module, used to predict future operating condition data based on the fusion features output by the dual-branch feature fusion module;

[0016] Multi-label classification module, used to determine whether the inverter is degraded and identify the degradation type based on historical operating condition data and future operating condition data;

[0017] Step 3. Use the multi-label degradation dataset to train the inverter degradation prediction and identification model;

[0018] Step 4. Use the historical operating condition data in the identification phase as the input of the time series branch and the future environmental data as the input of the environmental branch, and use the trained inverter degradation prediction and identification model to predict the inverter degradation.

[0019] In addition, based on the above-mentioned inverter degradation prediction method under extreme environment based on causal inference, the present invention also proposes a computer device, which includes a memory and one or more processors;

[0020] The memory stores executable codes, and when the processor executes the executable codes, the executable codes are used to implement the steps of the above-mentioned method for predicting converter degradation under extreme environment based on causal inference.

[0021] In addition, based on the above-mentioned method for predicting converter degradation under extreme environment based on causal inference, the present invention also proposes a computer-readable storage medium on which a program is stored; when the program is executed by a processor, it is used to implement the steps of the above-mentioned method for predicting converter degradation under extreme environment based on causal inference.

[0022] The present invention has the following advantages:

[0023] As described above, the present invention relates to a method for predicting inverter degradation under extreme environments based on causal inference. The method decomposes the impact of the environment on the working conditions into three types of redundant, unique and collaborative impacts through causal inference, and quantifies the degree of each type of impact through mutual information, clarifies the specific impact of each environmental factor on the inverter working conditions and its degree, and provides a new analytical perspective for the complex relationship between the environment and the working conditions under extreme environments; the method also performs feature fusion by querying the time series features through environmental features, and splices the fused features with the time series features on the basis of the fused features to further enhance the fused features, effectively captures the impact of the current environment on the working conditions, and fully considers the timing effects in the inverter degradation process, thereby significantly improving the prediction accuracy of future working condition data; the method for predicting inverter degradation under extreme environments based on causal inference of the present invention makes full use of historical and future environmental data through causal inference and feature fusion technology, decomposes and quantifies the impact of the environment on the working conditions, thereby effectively improving the prediction accuracy of the future working conditions of the inverter and the recognition accuracy of the degradation type, and provides reliable support for the stable operation of communication system inverters in extreme environments such as deserts, islands and reefs, high altitudes and extreme cold. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The flowchart is a method for predicting inverter degradation under extreme environment based on causal inference in an embodiment of the present invention.

[0025] Figure 2 Schematic diagram of the structure of the inverter degradation prediction and identification model in an embodiment of the present invention.

[0026] Figure 3 This is a working condition data picture of a frequency converter degradation of the data fluctuation type in an embodiment of the present invention.

[0027] Figure 4 This is a working condition data picture of a frequency converter degradation of the data sudden change type in an embodiment of the present invention.

[0028] Figure 5 This is a data picture of the operating condition of a frequency converter degradation of the data drift type in an embodiment of the present invention.

[0029] Figure 6 This is a cause-and-effect decomposition diagram in an embodiment of the present invention.

[0030] Figure 7 Schematic diagram of the structure of a dual-branch feature fusion module in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0032] Example 1

[0033] In this embodiment, a method for predicting inverter degradation under extreme environments based on causal inference is proposed. The method collects degradation data of the communication system inverter under special environmental conditions such as deserts, islands, high altitudes and extreme cold, aiming to analyze the impact of different extreme environments such as drought, corrosion, high temperature, extreme cold, etc. on the performance of the inverter. Environmental data covers factors such as temperature, humidity, salinity, atmospheric pressure, wind speed, wind direction, precipitation, etc., while the operating data of the inverter includes key indicators such as input and output voltage, input and output frequency, power, power consumption, gain, etc. Through in-depth analysis of the causal relationship between environmental data and operating data, the unique effect of environmental factors on the operating conditions of the inverter and their mutual synergistic effects can be effectively decomposed. In addition, in order to further extract the impact of the environment on the operating conditions, the present invention also designs environmental branches and timing branches, focusing on the direct impact of environmental factors on the operating conditions and the timing effect characteristics in the inverter degradation process, respectively. In addition, the method of the present invention introduces future environmental data, combines the impact of current environmental conditions on operating conditions, and through feature fusion and enhancement processing, can achieve accurate prediction of future operating condition data. Based on historical operating condition data and future predicted data, a frequency converter degradation prediction and identification model is further constructed, which can evaluate the degradation state of the current frequency converter based on real-time monitoring of operating conditions and future environmental changes.

[0034] like Figure 1 As shown, a method for predicting inverter degradation under extreme environment based on causal inference specifically includes the following steps:

[0035] Step 1. Collect historical environmental and operating condition data of the entire inverter degradation process in extreme climate areas, annotate the operating status and degradation type of the inverter, and construct a multi-label degradation dataset combining the environment and inverter working conditions, namely, the environment-operating condition multi-label degradation dataset.

[0036] Specifically, in extreme climate areas such as deserts, islands, high altitudes and extremely cold regions, the degradation process data of the communication system inverter is collected to construct an environment-operating condition multi-label degradation dataset, which includes environmental conditions, operating conditions and operating status data. In order to capture the degradation process, different types of degradation need to be labeled.

[0037] like Figures 3 to 5 As shown in the figure, the degradation types mainly include data fluctuation, data sudden change, and data drift. Among them, data fluctuation refers to abnormal changes in the law, period, amplitude, etc. of data fluctuation, which may be related to temperature fluctuation or humidity change. Data sudden change refers to sudden and drastic changes in data that exceed the normal threshold, which may be related to extreme weather such as strong winds and heavy rains or circuit short circuits caused by saline-alkali corrosion. Data drift refers to changes in the statistical characteristics of data such as mean and variance over time, which may be related to material aging caused by long-term high temperature or extreme cold environment.

[0038] Based on the historical environment and historical working condition data and the annotation of the operating status and degradation type of the inverter, a multi-label degradation dataset combining environment and working condition is constructed to obtain the environment-working condition multi-label degradation dataset :

[0039] .

[0040] in, Represents environmental factors, which include temperature, humidity, salinity, atmospheric pressure, wind speed, wind direction, and precipitation in extreme environments. , Represents the total number of environmental factors in the environment-operating condition multi-label degradation dataset; Indicates the working condition indicators, which include input voltage, output voltage, input frequency, output frequency, power, power consumption, and gain. , Represents the total number of working condition indicators in the environment-working condition multi-label degradation dataset; Indicates the running status label. , , , , The two represent normal operation, data fluctuation, data sudden change and data drift respectively.

[0041] Step 2. Build an inverter degradation prediction and identification model, including an environmental branch, a timing branch, a dual-branch feature fusion module, a Transformer module, and a multi-label classification module.

[0042] Among them, the environmental branch is used to extract the impact characteristics of the environment on the working conditions, namely the environmental characteristics, use causal inference technology to quantify the impact of various environmental factors on the working conditions, and integrate the impact of various environmental factors on the working conditions into a multi-dimensional environmental impact feature matrix.

[0043] The time series branch is used to extract the time series features of historical operating condition data.

[0044] A dual-branch feature fusion module is used to map environmental features into queries and temporal features into keys and values, then fuse local neighborhood features using a cross-attention mechanism and further enhance the fused features through convolution and feature interaction.

[0045] The Transformer module is used to predict future operating condition data based on the fusion features output by the dual-branch feature fusion module.

[0046] The multi-label classification module is used to determine whether the inverter is degraded and identify the degradation type based on historical operating condition data and future operating condition data.

[0047] Specifically, the process of using the environment branch to extract the characteristics of the impact of the environment on the working condition is as follows:

[0048] Based on the environment-operating condition multi-label degradation dataset constructed in step 1, causal inference techniques are used to quantify the impact of environmental factors on the operating conditions. By calculating the redundant, unique, and synergistic effects of each environmental factor on the operating conditions, the mechanism of environmental degradation on the operating conditions is deeply analyzed. Figure 6 The two environmental factors are shown , For a working condition Causal decomposition as an example.

[0049] First, define the variables and dataset:

[0050] Historical environmental data Defined as: .

[0051] in, Indicates Environmental factors such as temperature, humidity, etc. , Represents the total number of environmental factors in the environment-operating condition multi-label degradation dataset.

[0052] Historical operating data Defined as: .

[0053] in, Indicates working condition indicators, such as input voltage, output frequency, etc. , Represents the total number of operating condition indicators in the environment-operating condition multi-label degradation dataset.

[0054] Calculate mutual information:

[0055] Mutual information is an important tool to measure the interdependence between environmental factors and operating conditions. In order to quantify the impact of different environmental factors on the operating conditions, the mutual information between historical environmental data and historical operating conditions data is first calculated. , calculate the mutual information between historical environmental data and historical operating data :

[0056] .

[0057] in, is the joint probability distribution of historical environmental data and historical operating condition data, and They are the marginal probability distributions of historical environmental data and historical operating condition data respectively.

[0058] Decompose mutual information to determine different types of causal relationships between variables and calculate redundant, unique, and co-causal relationships:

[0059] In order to further analyze the redundant, unique and synergistic effects of environmental factors on working conditions, different types of interactions between environmental factors are identified by calculating redundant causality, unique causality and synergistic causality.

[0060] Redundant causality indicates the impact of multiple environmental factors on the working conditions. For some working conditions of the inverter, temperature and humidity may jointly affect the working stability of the equipment, so it is necessary to calculate the redundant impact. The calculation of redundant causality is expressed as:

[0061] .

[0062] in, Represents all historical environmental data Working condition indicators The redundant explanatory information is the redundant causal mutual information. All historical environmental data Working condition indicators The total mutual information between Each individual environmental factor And working condition indicators The sum of the mutual information between .

[0063] Unique causality indicates the unique impact of a single environmental factor on the working condition. For example, under certain extreme conditions, certain environmental factors such as temperature may be the only dominant factor. The calculation of unique causality is expressed as:

[0064] .

[0065] in, Represents each individual environmental factor Working condition indicators The unique explanatory information is the unique causal mutual information.

[0066] Synergistic causality indicates the synergistic effect of multiple environmental factors on the working conditions. This information cannot be obtained by considering each environmental factor separately or only the redundant part. The calculation of synergistic causality is expressed as:

[0067] .

[0068] in, Represents all historical environmental data Working condition indicators The collaborative explanation information is the collaborative causal mutual information. The collaborative explanation information is the result of the interaction between environmental factors. It can only be explained when multiple environmental factors are considered together. Produce additional explanations.

[0069] In extreme environments, the synergistic effects between environmental factors may be more significant. For example, in a high temperature and high humidity environment, the combined effect of temperature and humidity may be stronger than the effect of either factor alone.

[0070] Calculating causal leakage:

[0071] Calculates the error due to unobserved variables or incomplete models The loss of explanatory information is called causal leakage :

[0072] .

[0073] in, It is the working condition indicator The entropy of In extreme circumstances, causal leakage It may be larger because some factors in extreme environments, such as dust and salt spray, may be difficult to fully observe or quantify.

[0074] The environmental branch is used to extract the environmental impact characteristics on the working conditions, namely the environmental characteristics, and the redundant, unique and synergistic impacts of various environmental factors on the working conditions are integrated into a multi-dimensional environmental impact characteristic matrix:

[0075] When analyzing the impact of the environment on operating conditions, the causal relationship between environmental factors and operating condition indicators is quantified in detail and integrated through normalization and multi-dimensional feature matrices to further provide data support for modeling and prediction.

[0076] Normalized mutual information:

[0077] Under extreme environmental conditions, the operating conditions of the inverter may be affected by a combination of multiple environmental factors. The purpose of normalization is to standardize the influence range of each environmental factor to the same scale, which is between 0 and 1 in this embodiment, so that their sum is equal to 1. Normalizing the calculated redundant, unique and collaborative causal mutual information can help identify which environmental factors have the most prominent impact on the equipment operating conditions at a certain moment.

[0078] .

[0079] .

[0080] .

[0081] .

[0082] in, is the normalized redundant causal mutual information, is the normalized unique causal mutual information, is the normalized co-causal mutual information, is the normalized causal leakage, The value is between 0 and 1. When the value is 0, it means that all causal relationships are observed variables, that is, historical environmental data All environmental factors Explain, when When the value of is 1, it means that all causal relationships are not explained by the observed variables.

[0083] It is the causal leakage in causal inference, which is used to quantify the proportion of causal relationships that are not explained by observed variables, to judge the degree to which observed variables explain the causal relationship of target variables, and to serve as a reference when evaluating the causal relationship between environment and working conditions.

[0084] Construct a multidimensional environmental impact feature matrix:

[0085] Integrate normalized redundant, unique and co-causal mutual information into a multidimensional environmental impact feature matrix :

[0086] .

[0087] in, is the redundant, unique and synergistic causal influence feature vector of environmental factors on working condition indicators, The dimension is ,in represents the number of environmental factors, Indicates the number of working condition indicators, The 3 in the matrix represents the redundancy, uniqueness and collaborative causality, which represent the redundant, uniqueness and collaborative causal influence characteristics of environmental factors on operating indicators. This characteristic matrix helps to capture the complex interactive effects between environmental factors, and thus provide reliable data support for the maintenance, optimization and fault prediction of equipment in extreme environments.

[0088] Specifically, the process of extracting the time series features of historical operating condition data using time series branches is as follows:

[0089] In extreme environments, inverters usually work in harsh environments such as high temperature, low temperature, corrosion, and humidity changes. Their performance degradation will be significantly affected by these environmental factors. Therefore, extracting accurate timing features is crucial for equipment prediction and fault warning.

[0090] According to the historical working condition degradation data, i.e. the historical working condition data, the Transformer module is used in the time series branch to extract the time series features of the historical working condition data. The input of the Transformer module is the historical working condition data, which is input into the Transformer encoder after position encoding to form the representation of the time series data. After several layers of self-attention and feedforward networks, the time series feature representation is output.

[0091] Enter historical operating data:

[0092] In extreme environments, the working state of the inverter will be affected by factors such as temperature, humidity, and voltage fluctuations. The historical working condition data includes the impact of these environmental changes on the equipment status. First, obtain multi-dimensional historical working condition data from the sensor. , where each Indicates at a point in time The operating condition observations. Due to the particularity of the equipment operating environment, these data often contain long-term dependencies, so a model that can handle long-term dependencies is needed to extract time series features. Based on the historical operating condition degradation data, the Transformer module is applied to extract the time series features of the historical operating condition data. Due to its self-attention mechanism, the Transformer module is very suitable for processing sequence data and can capture the long-term dependencies between each time point in the sequence.

[0093] Transformer module architecture:

[0094] The self-attention mechanism in the Transformer module is used to capture the temporal characteristics of historical operating data. The input of the model is historical operating data The historical working condition data will be input into the Transformer encoder after position encoding to form a representation of time series data. After passing through several layers of self-attention and feedforward networks, the time series feature representation is output. The calculation is simplified to:

[0095] .

[0096] in, , , The historical working condition data The generated query, key, and value matrices are: is the dimension of the key, used to scale the calculation results to prevent the inner product value from being too large. is the normalized exponential function.

[0097] After position encoding, it is input into the encoder of Transformer. In order to strengthen the dependency between each time step in the time series data, a multi-head self-attention mechanism is used to calculate and fuse multiple groups of self-attention, and the final output is:

[0098] .

[0099] in, is a high-dimensional vector representation of the temporal characteristics of historical operating condition data. For the The self-attention output of each head, is the output weight matrix, Represents the connection operation between vectors. Through the multi-head self-attention mechanism, the model can learn multiple different dependencies in parallel, enhancing the ability to capture complex change patterns in historical working condition data.

[0100] Extracting time series features:

[0101] After several layers of self-attention mechanism and feedforward network, the time series features output by Transformer are still a high-dimensional vector representation, which needs to be further reduced and key information extracted through pooling operation. Pooling operation uses global average pooling or maximum pooling to obtain the final time series features. :

[0102] .

[0103] in, express , that is, using the multi-head self-attention mechanism to calculate and fuse multiple groups of self-attention to obtain the final output; represents the pooling operation, The feature vector representing each time point is the time series feature of the historical operating condition data. These time series features include the long-term dependency, trend changes and external environment influence of the historical operating condition data, providing a compact and effective feature representation for subsequent fault prediction, performance analysis and other tasks.

[0104] Dual-branch feature fusion and deep feature interaction:

[0105] The environmental features are mapped as queries and the temporal features are mapped as keys and values. The local neighborhood features are fused using the cross-attention mechanism, and the fused features are further enhanced through convolution and feature interaction.

[0106] In extreme environments, such as high temperature and strong electromagnetic interference, the fusion and interaction of environmental features and time series features need to consider the dynamic changes of the environment and the system. The convolution layer is used to map the environmental branch features to queries, and the time series branch features are mapped to keys and values ​​respectively through linear transformation. For each environmental query, the system extracts the local neighborhood key-value pairs centered on the query from the time series branch, and completes feature fusion with the help of the local neighborhood cross-attention mechanism. Finally, the fusion features are further enhanced with the help of convolution and feature interaction to improve the feature expression ability. The feature fusion flow chart of this process is shown below. Figure 7 shown.

[0107] Query, key and value generation:

[0108] Environmental impact characteristic matrix obtained for the environmental branch , generated using convolutional layers , , :

[0109] .

[0110] .

[0111] .

[0112] in, Queries that represent environmental characteristics, keys representing characteristics of the environment, represents the value of the environmental characteristic, , , is the convolution kernel, and its dimension is , represents the size of the convolution kernel, represents the depth of the convolution kernel, represents the convolution operation, Represents the ReLU activation function.

[0113] Timing characteristics obtained for timing branches , using linear transformation to generate the key of time series features and the values ​​of the time series features :

[0114] .

[0115] .

[0116] in, Represents a linear transformation operation.

[0117] Feature Fusion:

[0118] Before feature fusion, we first need to ensure that the dimensions of the environment branch and the timing branch are consistent. If the dimensions do not match, the features of the environment branch are adjusted through convolution. , making it consistent with the characteristics of sequential branching Dimension alignment. The adjusted environmental features and temporal features will be cross-attention calculated, and the expression is as follows:

[0119] .

[0120] in, It is the final output feature of the dual-branch feature fusion module and is used for subsequent prediction and classification tasks.

[0121] The processing of the dual-branch feature fusion module is described in detail below:

[0122] Feature fusion uses a cross-attention mechanism for local time windows. The calculation of cross-attention only considers the local time window in the time series. For each query in the environment branch , calculate its key with the time series feature The similarity of , get the cross attention of the environment branch and the timing branch :

[0123] .

[0124] in, It is the dimension of the key and plays a scaling role.

[0125] In order to enhance the feature representation of the environment branch, the self-attention inside the environment branch is calculated at the same time :

[0126] .

[0127] in, and are the keys and values ​​of the environment features. This process strengthens the influence of environment features on temporal features through the self-attention mechanism of environment queries.

[0128] Output the cross attention And the self-attention output After further transformation by linear layer and multi-layer perceptron MLP, the fusion feature is obtained. :

[0129] .

[0130] Feature Enhancements:

[0131] The fusion feature output of feature fusion Characteristics of sequential branches Splicing, time series fusion features under environmental feature query On this basis, we further strengthen the time series features to form a comprehensive fusion feature .

[0132] .

[0133] Respectively , and Perform convolution operation to generate intermediate features , , , and then the intermediate features are interactively processed through the convolution module to generate two output features , .

[0134] .

[0135] The two output features , and Concatenate and generate the final output features through the convolution layer .

[0136] .

[0137] In extreme environments, the system characteristics are highly dynamic. The feature fusion enhancement module effectively improves the robustness of the system in unstable environments through convolution and interactive processing, thereby ensuring that the inverter still maintains high prediction accuracy and responsiveness under extreme conditions.

[0138] Multi-label classification module:

[0139] In this embodiment, the multi-label classification module uses the XGBoost classifier to determine whether there is degradation and identify the degradation type, judges the operating status of the inverter based on historical and future operating data, and outputs a prediction result, that is, predicts the operating status of the inverter, including data fluctuations, data sudden changes, data drift, and normal operation.

[0140] Use including historical operating data And the corresponding running status label The degraded data is used to train the XGBoost classifier:

[0141] .

[0142] in, is the XGBoost classification algorithm, , , , , When represent normal operation, data fluctuation, data sudden change and data drift respectively, and M represents the multi-label classification module.

[0143] Step 3. Use the environment-operating condition multi-label degradation dataset to train the inverter degradation prediction and identification model.

[0144] Step 4. Use the historical operating condition data in the identification phase as the input of the time series branch and the future environmental data as the input of the environmental branch, and use the trained inverter degradation prediction and identification model to predict the inverter degradation.

[0145] In extreme environments, the system characteristics are highly dynamic, and the operating data of the inverter is strongly affected by environmental factors. Combining the historical operating data in the identification stage with the future environmental data can better capture the impact of the short-term environment on the inverter performance, thereby improving the accuracy of predicting future operating data.

[0146] Extract the long-term time series characteristics of historical operating condition data in the identification stage:

[0147] First, we identify the historical operating data of the stage Extract long-term time series features. Identify historical operating data of the stage Contains the operating status of the inverter in the past period of time, reflecting the long-term change trend of the equipment. As the input of the time series branch, the long-term time series features are extracted through the Transformer module :

[0148] .

[0149] in, Represents the Transformer module, long-term temporal features The dimension is , is the feature dimension.

[0150] Extract short-term environmental impact characteristics of future environmental data:

[0151] In order to accurately reflect the impact of short-term environmental factors on the inverter working conditions, the future environmental data As the input of the environment branch, the environment impact feature matrix Extract short-term environmental impact characteristics and convert future environmental data Mapped to the environmental impact feature space, the short-term environmental impact characteristics are obtained :

[0152] .

[0153] Among them, future environmental data The dimension is , is the time step, is the number of environmental factors. is the redundant, unique and synergistic causal influence feature vector of environmental factors on working condition indicators, The dimension is ,in Indicates the number of working condition indicators, The 3 in the figure represents the dimensions of redundancy, uniqueness, and co-causality. is the extracted future short-term environmental impact characteristics, The dimension is .

[0154] Extracting the future short-term environmental impact characteristics Integrate and reduce dimension to generate comprehensive impact characteristics of future short-term environment on working conditions :

[0155] .

[0156] in, is a dimensionality reduction operation. The dimension from Reduced to , thereby extracting local features in future environmental data and capturing the impact of short-term environmental changes on the inverter operating conditions, such as the impact of sudden temperature changes or humidity surges in a short period of time on the inverter operating conditions.

[0157] Dual-branch feature fusion and future working condition prediction:

[0158] Based on the dual-branch feature fusion module, the long-term time series features of the historical operating condition data in the recognition phase extracted by the time series branch are The comprehensive impact characteristics of the future short-term environment on the working conditions extracted by the environmental branch Fusion is performed to obtain fusion features :

[0159] .

[0160] in, Represents the dual-branch feature fusion module, the long-term time series features of historical operating data in the recognition phase of the time series branch extraction The dimension is , The comprehensive impact characteristics of the future short-term environment on the working conditions extracted by the environment branch The dimension is The fused features The dimension is , is the fusion feature dimension. and environmental characteristics Conduct in-depth interactions to capture the combined impact of environmental factors on the short-term impact of the inverter and the long-term degradation of the inverter.

[0161] The fusion features Input the Transformer module to predict future working condition data :

[0162] .

[0163] Predicted future operating data The dimension is , which can effectively reflect the working status of the inverter under future environmental changes and provide accurate prediction basis for equipment operation and maintenance.

[0164] Splicing historical working condition data and future working condition data in the identification stage:

[0165] Future working condition data Historical operating data during the identification phase Splicing is performed to provide more context information for degradation type identification and obtain degradation identification condition data :

[0166] .

[0167] Identification phase historical operating data The latest historical operating condition data is selected. Assuming that the degradation identification phase is 12 hours, the degradation identification operating condition data is obtained by splicing the latest 6 hours of historical operating condition data of the identification phase and the predicted 6 hours of future operating condition data. , the historical operating condition data Y refers to all historical data, which is used to build the inverter degradation prediction and identification model.

[0168] Degradation type identification:

[0169] Degradation identification condition data Input into the trained multi-label classification module for degradation pattern recognition:

[0170] .

[0171] The prediction result is Figure 3 The degradation prediction and identification results shown in the figure show that the predicted operating state of the inverter is:

[0172] when When , it indicates that the inverter is operating normally.

[0173] when , it indicates that the inverter has data fluctuation.

[0174] when , it indicates that the inverter data changes suddenly.

[0175] when , it indicates that the inverter has data drift.

[0176] Example 2

[0177] This embodiment 2 describes a computer device, which includes a memory and one or more processors.

[0178] An executable code is stored in the memory. When the processor executes the executable code, it is used to implement the steps of the inverter degradation prediction method under extreme environment based on causal inference in the above-mentioned embodiment 1.

[0179] In this embodiment, the computer device is any device or apparatus with data processing capability, which will not be described in detail here.

[0180] Example 3

[0181] This embodiment 3 describes a computer-readable storage medium on which a program is stored. When the program is executed by a processor, it is used to implement the steps of a method for predicting inverter degradation under extreme environments based on causal inference.

[0182] The computer-readable storage medium may be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc., equipped on the device.

[0183] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A method for predicting inverter degradation under extreme environment based on causal inference, characterized in that: The steps include: Step 1. Collect historical environmental and operating condition data of the inverter during the entire degradation process in extreme climate areas, annotate the operating status and degradation type of the inverter, and construct a multi-label degradation dataset combining the environment and inverter operating conditions; Step 2. Build a frequency converter degradation prediction and identification model, which includes the following structure: The environmental branch is used to extract the environmental impact characteristics on the working conditions, namely the environmental characteristics, quantify the impact of various environmental factors on the working conditions using causal inference technology, and integrate the impact of various environmental factors on the working conditions into a multi-dimensional environmental impact characteristic matrix; Time series branch, used to extract time series features of historical operating condition data; A two-branch feature fusion module that maps environmental features to queries and temporal features to keys and values, then fuses local neighborhood features using a cross-attention mechanism and further enhances the fused features through convolution and feature interaction; Transformer module, used to predict future operating condition data based on the fusion features output by the dual-branch feature fusion module; Multi-label classification module, used to determine whether the inverter is degraded and identify the degradation type based on historical operating data and future operating data; Step 3. Use the multi-label degradation data set to train the inverter degradation prediction and identification model; Step 4. Use the historical operating condition data in the identification phase as the input of the time series branch, and the future environmental data as the input of the environmental branch, and use the trained inverter degradation prediction and identification model to predict the inverter degradation; The process of using the environment branch to extract the characteristics of the impact of the environment on the working conditions is: Based on the environment-operating condition multi-label degradation dataset, historical environment data X and historical operating condition data Y are defined: X={x1,x2,…,x i ,…,x E },Y={y1,y2,…,y j ,…,y c }; Among them, x i represents environmental factors, i = 1,…,E, E represents the total number of environmental factors in the environment-operating condition multi-label degradation dataset; y j represents the working condition index, j = 1,…,C, C represents the total number of working condition indicators in the environment-working condition multi-label degradation dataset; Calculate the mutual information I(X;Y) between historical environmental data and historical operating condition data: Among them, p(x i ,y j ) is the joint probability distribution of historical environmental data and historical operating condition data, p(x i ) and p(y j ) represent the marginal probability distribution of historical environmental data and historical operating condition data respectively; Compute redundant causality: Among them, ΔI R Represents all historical environmental data X for working condition index y j The redundant explanatory information is the redundant causal mutual information, I(X; y j ) is all historical environmental data X and operating condition index y j The total mutual information between is each individual environmental factor x i And working condition index y j The sum of the mutual information between ; Compute unique causality: Among them, ΔI U Represents each individual environmental factor x i Working condition index y j The unique explanatory information is the unique causal mutual information; Compute co-causality: ΔI S =I(X;y j )-ΔI R -ΔI U ; Among them, ΔI S Represents all historical environmental data X for working condition index y j The collaborative explanation information is the collaborative causal mutual information; The redundant, unique and collaborative causal mutual information are normalized and integrated into a multidimensional environmental impact feature matrix F env .

2. The inverter degradation prediction method under extreme environment based on causal inference according to claim 1 is characterized in that: In step 1, the multi-label degradation dataset combining the environment and working conditions is constructed as follows: D={(x i ,y j ,z k )}; Where D is the environment-operating condition multi-label degradation dataset; x i represents environmental factors, i = 1,…,E, E represents the total number of environmental factors in the environment-operating condition multi-label degradation dataset; y j represents the working condition index, j = 1,…,C, C represents the total number of working condition indicators in the environment-working condition multi-label degradation dataset; z k Indicates the running status label.

3. The inverter degradation prediction method under extreme environment based on causal inference according to claim 1 is characterized in that: In step 2, a Transformer module is used in the time series branch to extract the time series features of the historical operating condition data; The input of the Transformer module is historical operating condition data. After position encoding, the historical operating condition data is input into the Transformer encoder to form a time series data representation. After passing through several layers of self-attention and feedforward networks, the time series feature representation is output.

4. The inverter degradation prediction method under extreme environment based on causal inference according to claim 3 is characterized in that: In step 2, the process of extracting the time series features of the historical operating condition data by using the time series branch is as follows: Use the self-attention mechanism in the Transformer module to capture the temporal characteristics of historical working condition data and calculate self-attention; The multi-head self-attention mechanism is used to calculate and fuse multiple groups of self-attention, output the high-dimensional vector representation of the time series features of the historical working condition data, and further obtain the final time series feature F through pooling operation. time .

5. The inverter degradation prediction method under extreme environment based on causal inference according to claim 1 is characterized in that: In step 2, the processing process of the dual-branch feature fusion module is: The environmental impact feature matrix F obtained for the environmental branch env ,Use convolutional layers to generate query ,keys and values ​​of environmental features; The timing feature F obtained for the timing branch time , using linear transformation to generate the keys and values ​​of time series features; If the dimensions of the environment branch and the timing branch do not match, the feature F of the environment branch is adjusted by convolution env , making it consistent with the characteristic F of the sequential branch time Dimension alignment; For each query in the environment branch, calculate its similarity with the key in the time series feature, and weighted sum the value of the time series feature to obtain the cross attention of the environment branch and the time series branch; Compute self-attention inside the environment branch; The cross attention output and self-attention output are further transformed and concatenated after passing through the linear layer and multi-layer perceptron MLP to obtain the fusion feature F s ′; The fusion feature F s ′ and the characteristic F of the sequential branch time Splicing to form a comprehensive fusion feature F c ; F c 、F s ′ and F time Perform convolution operation to generate intermediate features F c ′、F″ s , F′ time , and then the intermediate features are interactively processed through the convolution module to generate two output features F int1 、F int2 ; The two output features F int1 、F int2 With F c ′ concatenated, and the final output feature F is generated through the convolution layer o .

6. The inverter degradation prediction method under extreme environment based on causal inference according to claim 1 is characterized in that: In step 2, the multi-label classification module uses an XGBoost classifier; Use historical operating data Y and corresponding operating status label z k Train the XGBoost classifier to obtain a trained multi-label classification module.

7. The inverter degradation prediction method under extreme environment based on causal inference according to claim 1 is characterized in that: The step 4 is specifically as follows: The historical operating condition data of the inverter in the identification phase is used as the input of the time series branch, and the long-term time series features are extracted through the Transformer module; The future environmental data is used as the input of the environmental branch, and the future environmental data is mapped to the environmental impact feature space through the environmental impact feature matrix to obtain the short-term environmental impact features; Integrate and reduce the dimension of the extracted future short-term environmental impact features to generate comprehensive impact features of the future short-term environment on the working conditions; Based on the dual-branch feature fusion module, the long-term time series features of the historical working condition data in the identification phase extracted by the time series branch and the comprehensive impact features of the future short-term environment on the working condition extracted by the environment branch are fused to obtain the fusion features; Input the fused features into the Transformer module to predict future working condition data; The future working condition data is combined with the historical working condition data in the identification stage to obtain the degradation identification working condition data; The degradation identification condition data is input into the trained multi-label classification module to predict the inverter degradation and obtain the prediction result, that is, the predicted operating state of the inverter.

8. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, the steps of the method for predicting converter degradation under extreme environment based on causal inference are implemented as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for predicting converter degradation under extreme environment based on causal inference as described in any one of claims 1 to 7 are implemented.

Images