Method and device for predicting life of inverter, electronic equipment and storage medium

Abstract

The application is suitable for the technical field of inverter detection, and provides an inverter life prediction method and device, electronic equipment and storage medium. The inverter life prediction method comprises: inputting a plurality of sample data sets into an initial life prediction network, outputting a first life prediction value and a first prediction value corresponding to each physical feature; obtaining first training loss information based on the difference between the output of the initial life prediction network and the corresponding target life value and target value, adjusting the initial life prediction network to obtain a candidate life prediction model; outputting a second life prediction value and a second prediction value corresponding to each physical feature based on the candidate life prediction model, and adjusting the candidate life prediction model according to the second training loss information to obtain a target life prediction model; inputting the aging data of the target inverter into the target life prediction model to output the target life prediction value of the target inverter. The above scheme improves the accuracy and reliability of inverter life prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of inverter detection, and particularly relates to a method and device for predicting the service life of an inverter, an electronic device, and a storage medium. BACKGROUND

[0002] An inverter is a core component in energy devices such as photovoltaic systems, electric vehicles, and wind energy systems, and its service life is often significantly shorter than that of photovoltaic modules and other equipment, thus easily becoming a key factor limiting the stability and reliability of the system. In order to reduce operation and maintenance costs and prolong the overall service life of the equipment, it is of great significance to accurately predict the service life of the inverter.

[0003] Existing service life prediction methods are usually based on multi-dimensional sensor data during operation and use statistical analysis or machine learning models to establish a mapping relationship between the service life and the data. These methods can reflect the degradation trend of the inverter to some extent, but in complex working conditions or scenes with large environmental changes, they often show insufficient accuracy or unstable results. SUMMARY

[0004] Therefore, the embodiments of the present application provide a method and device for predicting the service life of an inverter, an electronic device, and a storage medium. By introducing a multi-stage loss function optimization mechanism in the training process, the service life prediction model has higher stability and robustness in complex operating environments and can efficiently converge, thereby improving the accuracy and reliability of the service life prediction of the inverter.

[0005] A first aspect of the embodiments of the present application provides a method for predicting the service life of an inverter, the method comprising:

[0006] inputting a plurality of sample data sets into an initial service life prediction network to obtain a first service life prediction value and a first prediction value corresponding to each physical feature; the sample data set is the aging data of a sample inverter, and each sample set corresponds to a first label, which is used to indicate the target service life value of the inverter and the target value corresponding to at least one physical feature;

[0007] obtaining first training loss information based on the difference between the first service life prediction value and the target service life value, and the difference between the first prediction value corresponding to each physical feature and the target value, adjusting the initial service life prediction network using the first training loss information to obtain a candidate service life prediction model;

[0008] inputting a plurality of sample data sets into the candidate service life prediction model to obtain a second service life prediction value and a second prediction value corresponding to each physical feature;

[0009] obtain second training loss information based on the difference between the second life prediction value and the target life value, the difference between the second prediction value of each physical feature and the target value, and a physical residual term, and adjust the candidate life prediction model using the second training loss information to obtain a target life prediction model; the physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent the deviation degree of the candidate life prediction model from the physical law;

[0010] input the aging data of the target inverter into the target life prediction model for life prediction, and output a target life prediction value of the target inverter.

[0011] The embodiments of the present application first input a plurality of sample data sets into an initial life prediction network to generate a first life prediction value and a first prediction value of each physical feature. The same batch of data provides a life label and a physical feature label at the same time, and the two are aligned at one time during training, so as to fully utilize the supervision information. Then, the difference between the first life prediction value and the target life value, and the difference between the first prediction value of each physical feature and the corresponding target value are used to form first training loss information for parameter updating, so as to quickly align the supervision and stabilize the early convergence, and obtain a candidate life prediction model. Continue to input the sample data set to obtain a second life prediction value and a second prediction value of the physical feature, which provides a comparable baseline output for the next step of constraint. On this basis, the difference between the second life prediction value and the target life value, the difference between the second prediction value of each physical feature and the target value, and a physical residual term are used to form second training loss information. The physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to measure and suppress the deviation from the physical law, thereby reducing overfitting and improving the stability across operating conditions, and further obtaining a target life prediction model. Finally, input the aging data of the target inverter into the target life prediction model to output a target life prediction value, and realize reliable life evaluation for real scenarios. It can be inferred that the present method takes into account both data supervision and physical consistency, so that the life prediction model has higher stability and robustness in complex operating environments, and can converge efficiently, thereby improving the accuracy and reliability of inverter life prediction.

[0012] In a possible implementation, the initial life prediction network at least includes a convolutional neural network layer, a multi-head self-attention layer, a feedforward neural network layer, and a fully connected layer; and the inputting of the plurality of sample data sets into the initial life prediction network to obtain the first life prediction value and the first prediction value corresponding to each physical feature includes:

[0013] input the plurality of sample data sets into the initial life prediction network, and process the input plurality of sample data sets in the convolutional neural network layer to extract local time sequence features and obtain first intermediate features;

[0014] inputting the first intermediate feature into the multi-head self-attention layer to capture the correlation of the first intermediate feature in the time sequence, to obtain a second intermediate feature;

[0015] inputting the second intermediate feature into the feed-forward neural network layer for nonlinear mapping, to obtain a third intermediate feature;

[0016] inputting the third intermediate feature into the fully connected layer, to obtain the first life prediction value and the first prediction value corresponding to each physical feature.

[0017] In a possible implementation, the first training loss information further includes an initial condition constraint term, the initial condition constraint term representing a deviation degree of the initial life prediction network from a known initial state at an initial time; and the adjusting the initial life prediction network by using the first training loss information to obtain a candidate life prediction model comprises:

[0018] combining a difference between the first life prediction value and the target life value, a difference between the first prediction value corresponding to each physical feature and a target value, and the initial condition constraint term into a first loss function according to weights;

[0019] performing back propagation and parameter updating on the initial life prediction network by using the first loss function, to obtain the candidate life prediction model.

[0020] In a possible implementation, the second training loss information further includes an initial condition constraint term, the initial condition constraint term further representing a deviation degree of the candidate life prediction model from a known initial state at an initial time; and the adjusting the candidate life prediction model by using the second training loss information to obtain a target life prediction model comprises:

[0021] combining a difference between the second life prediction value and the target life value, a difference between the second prediction value corresponding to each physical feature and a target value, a physical residual term, and the initial condition constraint term into a second loss function according to weights;

[0022] performing back propagation and parameter updating on the candidate life prediction model by using the second loss function, to obtain the target life prediction model.

[0023] In a possible implementation, the calculation of the physical residual term comprises the following steps:

[0024] calculating an inductance current residual and a capacitance voltage residual based on a BOOST circuit average model differential equation;

[0025] calculating a junction temperature residual based on an IGBT thermal model equation;

[0026] averaging the sample dimensions of the inductance current residual, the capacitance voltage residual, and the junction temperature residual to obtain the physical residual term.

[0027] In a possible implementation, the calculation of the initial condition constraint term comprises the following steps:

[0028] obtaining initial values of a BOOST inductance current, a BOOST capacitance voltage, and an IGBT junction temperature;

[0029] obtaining predicted values of each physical feature at an initial time by the initial life prediction network or the candidate life prediction model, and comparing the predicted values with corresponding initial values to determine deviations of the BOOST inductance current, the BOOST capacitance voltage, and the IGBT junction temperature at the initial time respectively;

[0030] calculating an initial condition constraint term corresponding to the initial life prediction network or the candidate life prediction model according to the deviations of the BOOST inductance current, the BOOST capacitance voltage, and the IGBT junction temperature respectively corresponding to the initial life prediction network or the candidate life prediction model.

[0031] In a possible implementation, the generation of the first label comprises the following steps:

[0032] constructing a target life value based on a ratio of a current remaining life to a rated life, the current remaining life being a remaining time length from a current sample collection time to a life termination;

[0033] For each physical feature, determining a target value corresponding to the physical feature based on a sample measurement value corresponding to the physical feature at a backward adjacent time of a current sample collection time.

[0034] In a possible implementation, the method further comprises:

[0035] saving a plurality of weights of the candidate life prediction model and the target life prediction model during the training process;

[0036] evaluating the models corresponding to the plurality of weights based on a root mean square error and a direction penalty score;

[0037] selecting a target life prediction model with an optimal evaluation result to output a target life prediction value of the inverter.

[0038] A second aspect of an embodiment of the present application provides an inverter life prediction device, the device comprising:

[0039] The first prediction module is configured to input a plurality of sample data sets into an initial life prediction network to obtain a first life prediction value and a first prediction value corresponding to each physical feature; the sample sets are aging data of sample inverters, and each sample set corresponds to a first label indicating a target life value of the inverter and a target value corresponding to at least one physical feature;

[0040] The first training module is configured to obtain a first training loss information based on a difference between the first life prediction value and the target life value, and a difference between the first prediction value corresponding to each physical feature and the target value, and adjust the initial life prediction network based on the first training loss information to obtain a candidate life prediction model.

[0041] The second prediction module is configured to input a plurality of sample data sets into the candidate life prediction model to obtain a second life prediction value and a second prediction value corresponding to each physical feature.

[0042] The second training module is configured to obtain a second training loss information based on a difference between the second life prediction value and the target life value, a difference between the second prediction value corresponding to each physical feature and the target value, and a physical residual term, and adjust the candidate life prediction model based on the second training loss information to obtain a target life prediction model; the physical residual term is configured to constrain a physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent a deviation degree of the candidate life prediction model from a physical law.

[0043] The target output module is configured to input aging data of a target inverter into the target life prediction model to perform life prediction, and output a target life prediction value of the target inverter.

[0044] The third aspect of the embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method according to the first aspect.

[0045] The fourth aspect of the embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the method according to the first aspect.

[0046] The fifth aspect of the embodiment of the present application provides a computer program product, which, when executed on an electronic device, causes the electronic device to perform the steps of the method according to the first aspect.

[0047] The beneficial effects of the second aspect to the fifth aspect can refer to the beneficial effects of the first aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0049] Figure 1 is a flow diagram of a method for predicting the life of an inverter provided by an embodiment of the present application;

[0050] Figure 2 is a flow diagram of another method for predicting the life of an inverter provided by an embodiment of the present application;

[0051] Figure 3 is a model training flow diagram;

[0052] Figure 4 is a structural diagram of a device for predicting the life of an inverter provided by an embodiment of the present application;

[0053] Figure 5 is a structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0054] In the following description, specific details such as specific system structures, techniques, etc. are presented in order to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to avoid unnecessary details that hinder the description of the present application.

[0055] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or sets thereof.

[0056] It should also be understood that the term "and / or" used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0057] As used in the specification and the appended claims herein, the term "if' can be interpreted as meaning "when" or "upon" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [the recited condition or event] is detected" can be interpreted as meaning "upon determining" or "in response to determining" or "upon detecting [the recited condition or event]" or "in response to detecting [the recited condition or event]" depending on the context.

[0058] In addition, in the description of the present application and the appended claims, the terms "first", "second", "third" and the like are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0059] It should be understood that the size of the serial number of each step in the embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0060] The inverter is a core component in energy devices such as photovoltaic systems, electric vehicles and wind power systems, and its service life is often significantly shorter than that of photovoltaic modules and other equipment, so it is easy to become a key factor limiting the stability and reliability of the system. In order to reduce operation and maintenance costs and prolong the overall life of the equipment, it is of great significance to accurately predict the service life of the inverter.

[0061] The existing life prediction method is usually based on multi-dimensional sensor data in the running process, and with the help of statistical analysis or machine learning model to establish the mapping relationship between life and data. These methods can reflect the degradation trend of the inverter to some extent, but in the scene of complex working conditions or large environmental changes, they often show the problems of insufficient accuracy or unstable results.

[0062] To solve the above problems, the application provides a method, device, electronic equipment and storage medium for inverter life prediction. The method for inverter life prediction in the application first inputs a plurality of sample data sets into an initial life prediction network to generate a first life prediction value and a first prediction value of each physical feature, and the same batch of data provides a life label and a physical feature label at the same time, and the two are aligned at the same time in one training, so as to fully utilize the supervision information; then the difference between the first life prediction value and the target life value, the difference between the first prediction value of each physical feature and the corresponding target value are used to form a first training loss information for parameter updating, to quickly align the supervision and stabilize the early convergence, and obtain a candidate life prediction model; continue to input the sample data set to obtain a second life prediction value and a second prediction value of the physical feature, to provide a comparable baseline output for the next step of constraint; on this basis, the difference between the second life prediction value and the target life value, the difference between the second prediction value of each physical feature and the target value and a physical residual term are used to form a second training loss information, wherein the physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, to measure and suppress the deviation from the physical law, so as to reduce the overfitting and improve the cross-condition stability, and then obtain a target life prediction model; finally, the target inverter aging data is input into the target life prediction model to output a target life prediction value, to realize reliable life evaluation for real scenes. Therefore, the method takes into account data supervision and physical consistency, so that the life prediction model has higher stability and robustness in a complex operating environment, and can converge efficiently, thereby improving the accuracy and reliability of inverter life prediction.

[0063] The method, device, electronic equipment, storage medium and computer program for inverter life prediction provided by the application are described in detail below with reference to the accompanying drawings.

[0064] Referring to Figure 1 , a flowchart of a method for inverter life prediction provided by the application is shown; as Figure 1 shown, the method can include the following steps:

[0065] Step 101, input a plurality of sample data sets into an initial life prediction network to obtain a first life prediction value and a first prediction value corresponding to each physical feature.

[0066] The sample data set is the aging data of a sample inverter, and each sample set corresponds to a first label, which is used to indicate the target life value of the inverter and the target value corresponding to at least one physical feature.

[0067] In this embodiment, the aging data of the sample inverters can be obtained based on aging experiments. Before conducting accelerated aging experiments, the normal operation of each sensor in the inverter device is first verified. The sensors are arranged around the DC / DC converter and DC / AC converter of the inverter, and the collected multi-dimensional sensor data includes, but is not limited to: temperature, power, DC voltage / current, three-phase AC voltage / current, switching frequency, etc. Subsequently, in the aging test laboratory, a batch of inverters of the same model (number P) are subjected to single-item aging experiments (such as high temperature, accelerated electrical stress, load stress, high humidity corrosion, accelerated switching frequency) and random phased multi-item comprehensive aging experiments; the experiment terminates when "the inverter cannot stably output the specified AC voltage", and P complete cycles of data are accumulated, with one complete cycle of data corresponding to one inverter.

[0068] For a complete cycle of data corresponding to a single sample inverter, the raw data from each sensor is recorded as follows: Q 0, its typical dimension is q 0, including: time t Input voltage (PV voltage) V in Boost circuit inductor current i L Boost circuit (BOOST) capacitor voltage v c IGBT output voltage V IGBT With output current I IGBT IGBT case temperature Tc Switching frequency f sw Duty cycle D Ambient temperature T a Etc. To avoid gradient vanishing or exploding during subsequent training, ... Q 0. Data cleaning: First, standardization is performed based on the reference device ratings or empirical maximum / minimum values ​​under extreme conditions; second, multicollinearity analysis is performed to remove features irrelevant to the target lifetime value, resulting in a cleaned dataset. Q (dimension) q Based on dataset A, a sliding window slice is made to generate a sample data set, so as to construct multiple sample data sets (each sample data set corresponds to a time segment, which is used to characterize the operating state of the inverter in that time segment).

[0069] In one possible implementation, each sample set corresponds to a first label, which indicates the target lifetime value of the inverter and the target value corresponding to at least one physical characteristic.

[0070] Specifically, the generation of the first label includes the following steps:

[0071] The target life value is constructed based on the ratio of the current remaining life to the rated life, the current remaining life referring to the remaining time length from the current sample collection time to the life termination;

[0072] For each physical feature, the target value corresponding to the physical feature is determined based on the sample measurement value corresponding to the backward adjacent time of the current sample collection time of the physical feature.

[0073] In the embodiments of the present application, it is assumed that the at least one physical feature includes: BOOST inductor current i L BOOST capacitor voltage v c IGBT junction temperature Tj Therefore, the first label at least includes the target life value and the target value of i L , v c Since the physical residual term is involved in the training process Tj , it is not necessary to set the corresponding target value in the first label.

[0074] Exemplarily, the target life value is constructed by the ratio of the current remaining life to the rated life:

[0075]

[0076] Wherein, t is the current sample collection time, and T is the rated life of the inverter from the initial operation to the life termination. Thus, the target life value sequence is obtained:

[0077]

[0078] Exemplarily, the construction of the target value corresponding to the physical feature value adopts the sample measurement value corresponding to the backward adjacent time of the current sample collection time as the supervision signal, that is, the measured value at the time t+1 is taken as the target value at the time t, and the corresponding initial value is preset at the initial time, that is, the target value set of the BOOST inductor current t , the target value set of the BOOST capacitor voltage t , as follows: I V

[0079]

[0080]

[0081] For the label set corresponding to the sample data set of the above P periods​​Y 、 I 、 V , the continuous data segments are obtained according to a sliding window, the window size is the batch_size during training, and the step is 1 / 2 of the batch_size; in the obtained data segment set, 1 / 10 is randomly extracted as a test set, and the rest is used as a training set, wherein the training set includes a plurality of sample data sets. Thus, each sample data set in the training data set has a first label, and the first label indicates the target value corresponding to the target life value and at least one physical feature, which meets the needs of subsequent input of the training data set into the initial life prediction network and calculation of the first training loss information.

[0082] wherein the target life value is calculated as (supervision for subsequent first training loss information).

[0083] In the embodiments of the present application, the initial life prediction network can be composed of a convolutional neural network layer, a multi-head self-attention layer, a feedforward neural network layer, and a fully connected layer, for mapping the hidden layer features to the percentage of the inverter remaining life cycle. For example, the initial life prediction network outputs a first life prediction value of 0.205 at a certain time, which corresponds to a "remaining life cycle of 20.5%".

[0084] In one possible implementation, the construction of the initial life prediction network includes the following steps:

[0085] extracting local time sequence features in the convolutional neural network layer;

[0086] introducing position encoding and using time sequence mask in the multi-head self-attention layer to capture time correlation;

[0087] performing nonlinear mapping in the feedforward neural network layer;

[0088] obtaining life prediction values and first prediction values corresponding to each physical feature in the fully connected layer.

[0089] Specifically, on this basis, in another possible implementation, a plurality of sample data sets are input into the initial life prediction network to obtain first life prediction values and first prediction values corresponding to each physical feature, including:

[0090] inputting a plurality of sample data sets into the initial life prediction network, and processing the input plurality of sample data sets in the convolutional neural network layer to extract local time sequence features and obtain first intermediate features;

[0091] inputting the first intermediate features into the multi-head self-attention layer to capture the correlation of the first intermediate features in the time sequence and obtain second intermediate features;

[0092] The second intermediate feature is input to a feedforward neural network layer for nonlinear mapping to obtain a third intermediate feature;

[0093] The third intermediate feature is input to a fully connected layer to obtain a first life prediction value and a first prediction value corresponding to each physical feature.

[0094] In the embodiments of the present application, based on the training data set obtained in the above steps, a plurality of sample data sets therein are input to an initial life prediction network in time sequence, and the forward calculation process thereof includes:

[0095] (1) Convolutional neural network layer: extract local time sequence features to obtain a first intermediate feature.

[0096] Let the input sequence corresponding to a single sample data set be (n is the length of the input sequence; is the dimension of the input feature vector). One-dimensional convolution is performed along the time dimension to obtain:

[0097]

[0098] H cnn is recorded as the first intermediate feature.

[0099] (2) Multi-head self-attention layer: introduce position encoding and use time mask to capture time correlation to obtain a second intermediate feature.

[0100] First linear mapping to obtain query / key / value:

[0101]

[0102] wherein, is the weight of QKV, is usually , is the number of attention heads; RoPE encodes position information in attention calculation, and rotates the Q and K vectors for each position p so that the position information is naturally included when dot product is performed. RoPE acts on the two-dimensional subspace of the vector through a rotation matrix to add position encoding to Q and K:

[0103]

[0104] QKV calculation formula after adding rotation position encoding:

[0105]

[0106] According to the QKV with rotation position encoding, the second intermediate feature can be obtained.

[0107] (3) A feedforward neural network layer: performing a nonlinear mapping to obtain third intermediate features.

[0108] Specifically, the second intermediate features can be mapped to the third intermediate features by using a nonlinear activation function.

[0109] (4) A fully connected layer: obtaining a first life prediction value.

[0110] The third intermediate features are linearly projected and mapped to the interval (0, 1) to obtain a life percentage sequence, i.e., the first life prediction value; the third intermediate features are linearly mapped to obtain the first prediction value corresponding to each physical feature. That is, the same layer shares features, and two fully connected heads are used to output the first life prediction value and the first prediction value corresponding to each physical feature, respectively.

[0111] In the embodiment of the present application, the corresponding first life prediction value is obtained for each sample data set in a batch, which is used to determine the first training loss information in the subsequent step and train the candidate life prediction model.

[0112] In step 102, the first training loss information is obtained based on the difference between the first life prediction value and the target life value, and the difference between the first prediction value corresponding to each physical feature and the target value. The initial life prediction network is adjusted using the first training loss information to obtain the candidate life prediction model.

[0113] The first training loss information is obtained based on the difference between the first life prediction value and the target life value, and the difference between the first prediction value corresponding to each physical feature and the target value.

[0114] In one possible implementation, the first training loss information is composed of two parts: one is the difference between the first life prediction value and the target life value, and the difference between the first prediction value corresponding to each physical feature and the target value; the other is an initial condition constraint term, wherein the initial condition constraint term represents the deviation degree of the initial life prediction network from the known initial state at the initial time.

[0115] Specifically, the initial life prediction network is adjusted using the first training loss information to obtain the candidate life prediction model, including:

[0116] The difference between the first life prediction value and the target life value, the difference between the first prediction value corresponding to each physical feature and the target value, and the initial condition constraint term are combined into a first loss function according to weights;

[0117] The initial life prediction network is back propagated and parameter updated using the first loss function to obtain the candidate life prediction model.

[0118] In the embodiments of the present application, the physical characteristics can include BOOST inductor current, BOOST capacitor voltage.

[0119] Specifically, according to the two-step loss design of the present application, only the life prediction difference, the physical characteristic difference and the initial condition constraint term are introduced in the first stage to form a first loss function L 1:

[0120]

[0121] wherein, is the combination of the life prediction difference and the physical characteristic difference, is the initial condition constraint term, is the weight coefficient of the initial condition constraint term. Assuming that the number of samples is N, The specific calculation formula of is as follows:

[0122]

[0123] In the above formula, , , respectively correspond to the three prediction values of the initial life prediction network: the first life prediction value, the first prediction value of the BOOST inductor current, and the first prediction value of the BOOST capacitor voltage. Y, I, and V are the target life value, the target value corresponding to the BOOST inductor current, and the target value corresponding to the BOOST capacitor voltage, respectively.

[0124] The initial condition constraint term is as follows:

[0125]

[0126] wherein, is the IGBT junction temperature prediction value corresponding to the initial life prediction network, , , are the initial value corresponding to the BOOST inductor current, the initial value corresponding to the BOOST capacitor voltage, and the IGBT case temperature, respectively.

[0127] Specifically, the calculation of the initial condition constraint term includes the following steps:

[0128] Obtain the initial values of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature;

[0129] Obtain the prediction values of the initial life prediction network for each physical characteristic at the initial time, and compare the prediction values with the corresponding initial values to determine the deviations of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature at the initial time, respectively;

[0130] Based on the deviations of the BOOST inductor current, BOOST capacitor voltage, and IGBT junction temperature in the initial lifetime prediction network, the initial condition constraints corresponding to the initial lifetime prediction network are calculated.

[0131] In the embodiments of this application, using L 1 is used as the first training loss information to perform forward computation on the initial lifetime prediction network, based on... L Backpropagation and parameter updates are performed on step 1 until the preset convergence / stopping conditions are met, resulting in a candidate lifetime prediction model. Throughout the training process, the supervision signal comes only from the target lifetime value, the target values ​​corresponding to each physical feature, and the initial condition constraints, thereby stabilizing the fitting of the primary lifetime task and constraining the initial state.

[0132] Step 103: Input multiple sample data sets into the candidate lifetime prediction model to obtain the second lifetime prediction value and the second prediction value corresponding to each physical feature.

[0133] In this embodiment of the application, the sample data set is the same as that in step 101, which is the aging data of the sample inverter. Each sample set has a first tag, which is used to indicate the target life value of the inverter and the target value corresponding to at least one physical feature.

[0134] The candidate lifetime prediction model is obtained by training the initial lifetime prediction network based on multiple sample data sets and the first training loss information.

[0135] Specifically, the multiple sample data sets obtained in step 101 are input into the candidate lifetime prediction model in chronological order. The candidate lifetime prediction model follows the forward computation process of the lifetime prediction network, sequentially extracting local temporal features through a convolutional neural network layer, capturing temporal correlations through a multi-head self-attention layer (including position encoding and temporal masking), performing nonlinear mapping through a feedforward neural network layer, and finally outputting simultaneously in a fully connected layer in a dual-branch manner. One branch obtains the second lifetime prediction value through normalization mapping, and the other branch obtains the second prediction value corresponding to each physical feature through linear mapping.

[0136] Step 104: Based on the difference between the second lifetime prediction value and the target lifetime value, the difference between the second prediction value and the target value corresponding to each physical feature, and the physical residual term, the second training loss information is obtained, and the candidate lifetime prediction model is adjusted using the second training loss information to obtain the target lifetime prediction model.

[0137] The second training loss information is based on the difference between the second lifetime prediction value and the target lifetime value, the difference between each physical feature and the target value, and the physical residual term.

[0138] The physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent the deviation degree of the candidate life prediction model from the physical law.

[0139] In a possible implementation, the second training loss information is composed of three parts, one of which is the difference between the second life prediction value and the target life value, and the difference between each physical feature and the target value; the second is the physical residual term; and the third is the initial condition constraint term. That is, the physical residual term is added on the basis of the first training loss information, and the same part as the first training loss information will not be described herein again. Please refer to the description in step 103, and the physical residual term and the initial condition constraint term in the present embodiment will be described below.

[0140] In the present embodiment, the physical residual term can consider the following two core contents, that is, BOOST circuit equation residual and IGBT thermal model residual.

[0141] The BOOST average model differential equation is:

[0142]

[0143]

[0144] The IGBT thermal model equation is:

[0145]

[0146] Suppose:

[0147]

[0148]

[0149]

[0150] Therefore, the physical residual term is designed as:

[0151]

[0152] In the above formula, , D, , , Five are sensor measurement values, which are input voltage value, duty cycle, environmental temperature, IGBT output current and IGBT output voltage; R, L, C, , , are resistance, inductance, capacitance, heat capacity, thermal resistance, power loss, and except for the power loss , the other five are constants, and the power loss is calculated by measuring values , , switching frequency , IGBT case temperature , and single switching energy loss , wherein The equation is fitted from datasheet curves.

[0153] In a possible implementation, the physical characteristics include: BOOST inductor current, BOOST capacitor voltage, and IGBT junction temperature;

[0154] The calculation of the physical residual term includes the following steps:

[0155] The inductor current residual and the capacitor voltage residual are calculated based on the BOOST circuit average model differential equation;

[0156] The junction temperature residual is calculated based on the IGBT thermal model equation;

[0157] The inductor current residual, the capacitor voltage residual, and the junction temperature residual are averaged in the sample dimension to obtain the physical residual term.

[0158] In the embodiments of the present application, the inductor current residual and the capacitor voltage residual are calculated based on the BOOST circuit average model differential equation, and the BOOST inductor current and the BOOST capacitor voltage are required; the junction temperature residual is calculated based on the IGBT thermal model equation, and the IGBT junction temperature is required.

[0159] Specifically, on this basis, in another implementation,

[0160] The second training loss information further includes an initial condition constraint term, and the initial condition constraint term is further used to represent the deviation degree of the candidate life prediction model from the known initial state at the initial moment; the candidate life prediction model is adjusted by using the second training loss information to obtain the target life prediction model, including:

[0161] The difference between the second life prediction value and the target life value, the difference between the second prediction value corresponding to each physical characteristic and the target value, the physical residual term, and the initial condition constraint term are combined according to weights to form a second loss function;

[0162] The candidate life prediction model is back propagated and parameter updated by using the second loss function to obtain the target life prediction model.

[0163] In the embodiments of the present application, the calculation of the initial condition constraint term includes the following steps:

[0164] The initial values of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature are obtained;

[0165] The candidate life prediction model obtains the predicted value of each physical feature at the initial moment, and compares the predicted value with the corresponding initial value to determine the deviation of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature at the initial moment, respectively.

[0166] According to the deviations of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature in the candidate life prediction model, the initial condition constraint term corresponding to the candidate life prediction model is calculated.

[0167] In the embodiment of the present application, the second loss function is: wherein, is a weight coefficient of the physical residual term.

[0168] In the embodiment of the present application, the forward calculation of the candidate life prediction model is performed with L 2 as the first training loss information, the back propagation and parameter update based on L 2 are performed until the preset convergence / stop condition is met, and the target life prediction model is obtained. In the entire training process, the supervision signal is derived from the target life value and the target value corresponding to each physical feature; at the same time, the physical residual term and the initial condition constraint term are used as constraint terms to participate in optimization, so as to improve the stability and reliability of the model under complex working conditions.

[0169] Step 105: input the aging data of the target inverter into the target life prediction model for life prediction, and output the target life prediction value of the target inverter.

[0170] In the embodiment of the present application, the aging data collected during the operation of the target inverter is preprocessed (including standard normalization and feature screening) in the same way as step 101, and a sample data set for inference is constructed in the same input format as the training phase; the sample data set is organized into a time series according to the same segmentation method (window size and step of the sliding window are consistent with the training) as the training phase, and is input into the target life prediction model for forward calculation to obtain a life percentage sequence; the position aligned with the current moment is output as the target life prediction value. In order to ensure the consistency of the training supervision, the order, dimension and dimension of the input features are consistent with the training phase, and the output range of the target life prediction model is limited to the interval (0, 1), corresponding to the proportion of the remaining life cycle of the inverter; when a single result is needed within a period of time, the end output or weighted fusion method can be used to determine the target life prediction value after aligning the outputs of the continuous windows by time.

[0171] In this embodiment, first, a plurality of sample data sets are input into an initial life prediction network to generate first life prediction values and first prediction values of each physical feature, and the same batch of data provides life labels and physical feature labels at the same time, and both are trained and aligned at the same time, so as to fully utilize the supervised information; then, the difference between the first life prediction values and the target life values, and the difference between the first prediction values of each physical feature and the corresponding target values are used to form first training loss information for parameter updating, to quickly align the supervision and stabilize the early convergence, and obtain a candidate life prediction model; continue to input the sample data sets to obtain second life prediction values and second prediction values of each physical feature, to provide comparable baseline outputs for the next step of constraint; on this basis, the difference between the second life prediction values and the target life values, the difference between the second prediction values of each physical feature and the target values, and a physical residual term are used to form second training loss information, wherein the physical residual term is used to constrain the physical quantities output by the candidate life prediction model to satisfy a preset physical equation, to measure and suppress the deviation from the physical law, thereby reducing overfitting and improving the cross-condition stability, and further obtaining a target life prediction model; finally, target inverter aging data are input into the target life prediction model to output target life prediction values, to realize reliable life evaluation facing real scenes. It can be inferred that the method takes into account data supervision and physical consistency, so that the life prediction model has higher stability and robustness under complex operating environments, and can efficiently converge, thereby improving the accuracy and reliability of inverter life prediction.

[0172] Referring to Figure 2 , a flowchart of another method for inverter life prediction provided by an embodiment of the present application is shown; as shown in Figure 2 , the method can include the following steps:

[0173] Step 201, a plurality of sample data sets are input into an initial life prediction network to obtain first life prediction values and first prediction values corresponding to each physical feature.

[0174] Step 202, first training loss information is obtained based on the difference between the first life prediction values and target life values, and the difference between the first prediction values corresponding to each physical feature and target values, and the initial life prediction network is adjusted using the first training loss information to obtain a candidate life prediction model.

[0175] Step 203, a plurality of sample data sets are input into the candidate life prediction model to obtain second life prediction values and second prediction values corresponding to each physical feature.

[0176] Step 204, second training loss information is obtained based on the difference between the second life prediction values and target life values, the difference between the second prediction values corresponding to each physical feature and target values, and a physical residual term, and the candidate life prediction model is adjusted using the second training loss information to obtain a target life prediction model.

[0177] The steps 201-204 of this embodiment are the same as the steps 101-104 of the foregoing embodiment, and can be mutually referred to. The steps 201-204 of this embodiment will not be described again here.

[0178] In step 205, multiple weights of the candidate life prediction model and the target life prediction model are saved during the training process.

[0179] In step 206, the models corresponding to the multiple weights are evaluated based on the root mean square error and the direction penalty score.

[0180] In step 207, the target life prediction model with the optimal evaluation result is selected to output the target life prediction value of the inverter.

[0181] In the embodiments of the present application, the multiple weights of the candidate life prediction model and the target life prediction model are saved at a preset frequency during the training process, and the models corresponding to the multiple weights are uniformly evaluated based on the test set divided in step 101 after the training is completed. In the evaluation, the target life value of each sample in the test set is recorded as , the life prediction value output by the corresponding model is recorded as , and the total number of samples is n.

[0182] To measure the consistency of the prediction and the reality, the root mean square error (RMSE) is calculated:

[0183]

[0184] The smaller the RMSE is, the smaller the overall error is. On the other hand, to reflect the influence of the error direction on the operation and maintenance risk, the direction penalty score (Effectiveness Score) is calculated, and different penalty weights are applied to overestimation and underestimation:

[0185]

[0186] The exponential form is used for nonlinear weighting, and the difference in the denominator is used to reflect the different intensity of the directional penalty, so as to preferentially reduce the potential risk caused by “prediction being too large” under the same RMSE.

[0187] After the above evaluation is completed, the RMSE and the direction penalty score of the models corresponding to the saved weights are comprehensively sorted, and the one with the best comprehensive performance is selected as the target life prediction model for deployment according to the criterion of “the smallest RMSE and the optimal direction penalty score”. Subsequently, according to step 105, the collected inverter aging data is input into the selected target life prediction model after being preprocessed in the same way as the training, and the target life prediction value at the corresponding time is calculated by forward calculation, so as to guide the repair and replacement strategy.

[0188] In a possible implementation, before the model is evaluated, the following model training can be performed:

[0189] The initial life prediction network is initialized with the Xavier algorithm, and the first loss function L 1 (consisting of the difference between the first life prediction value and the target life value, the difference between the first prediction value of each physical feature and the target value, and the initial condition constraint term) is trained until the preset convergence / stop condition is met, and the model M0 (i.e., the candidate life prediction model) is obtained.

[0190] With the parameters of M0 as the initial weights, the second loss function L 2 (including the difference between the second life prediction value and the target life value, the difference between the first prediction value of each physical feature and the target value, the physical residual term, and the initial condition constraint term) is used to continue training, and the model M1 (i.e., the target life prediction model) is obtained.

[0191] The initial life prediction network is reinitialized with the Xavier algorithm, and does not go through L 1 stage, directly uses the second loss function L 2 to train from scratch to obtain the model M2, which is used for comparison with M1 obtained by transfer learning.

[0192] The initial life prediction network is structurally simplified while keeping the rest of the structure and hyperparameters consistent: the position encoding is removed in the multi-head self-attention layer (for example, the rotational position encoding is removed) to investigate the influence of the position encoding on life prediction; the Xavier algorithm is used for initialization, and the first loss function L 1 is used for training to obtain the model M3.

[0193] The above training process can save multiple weights of each stage at a preset frequency to facilitate unified evaluation and optimal model selection based on the root mean square error and the direction penalty score.

[0194] Step 208: Input the aging data of the target inverter into the target life prediction model to perform life prediction, and output the target life prediction value of the target inverter.

[0195] Step 208 of this embodiment is the same as step 105 of the foregoing embodiments, and can be mutually referred to. This embodiment will not be described here again.

[0196] Compared with the previous embodiment, the application embodiment saves multiple sets of weights and evaluates and selects the optimal band based on the RMSE and the direction penalty score, which can bring higher landing reliability and safety. On the one hand, multiple weights are saved during the training process and evaluated offline, avoiding the contingency of "using the last training result as the deployment model", reducing the risk caused by overfitting and training fluctuations, and improving the generalization stability. On the other hand, in addition to RMSE, the direction penalty score is introduced to "punish more for overestimation", which can explicitly suppress optimistic bias, preferentially select a model that is more conservative and safe for operation and maintenance, and reduce the risk of missing the maintenance window due to overestimation of life. Finally, the combination of the two realizes the "multiple candidates-quantitative evaluation-optimal selection" closed loop, so that the selected target life prediction model is simultaneously constrained in accuracy (RMSE) and risk control (direction penalty), thereby improving the credibility and usability of the target life prediction value, and optimizing maintenance decisions and life cycle costs.

[0197] Referring to Figure 3 , a model training process schematic diagram is shown. As Figure 3 indicated, the model training process can be summarized as:

[0198] First, run the test on the same type of inverter in the accelerated aging experiment and collect the full cycle raw data. After completing the sensor self-check, the temperature, power, direct / alternating voltage and current, duty cycle, switching frequency and other multi-dimensional signals are merged into the raw data set. Then, data cleaning is performed, and each channel is normalized according to the rated value or extreme value, and multiple collinearity analysis is carried out to eliminate features unrelated to life, to obtain the cleaned "data".

[0199] On this basis, the complete cycle is divided into continuous data segments according to the sliding window, and a training data set composed of multiple sample data sets and their first labels is constructed, the first label is used to indicate the target life value of the inverter and the target value corresponding to at least one physical feature. In parallel, the "Loss set" is formed according to the loss design and splitting, which includes the first loss function used for the first stage training (composed of the difference between the first life prediction value and the target life value, the difference between the first prediction value corresponding to each physical feature and the target value, and the initial condition constraint term) and the second loss function used for the second stage training (composed of the difference between the second life prediction value and the target life value, the difference between the second prediction value corresponding to each physical feature and the target value, the physical residual term and the initial condition constraint term).

[0200] Then enter the "model training" node: the initial life prediction network is initialized with Xavier, the first loss function is used for back propagation and parameter update, and a candidate life prediction model is obtained; it is judged whether the last loss training is completed, if not, the second stage is entered, the second loss function is used for continuous training (transfer learning can be performed on the first model weight), until the preset convergence / stop condition is met, and the target life prediction model is output. After training, "model saving" is performed, and multiple weights can be saved at the frequency set during training for subsequent evaluation and optimization; thus the current training process ends.

[0201] Referring to Figure 4 , a structure diagram of an inverter life prediction device provided by an embodiment of the present application is shown; only parts related to the embodiments of the present application are shown for ease of description.

[0202] The inverter life prediction device 400 comprises:

[0203] The first prediction module 401 is configured to input a plurality of sample data sets into an initial life prediction network to obtain a first life prediction value and a first prediction value corresponding to each physical feature; the sample data set is the aging data of a sample inverter, and each sample set corresponds to a first label, which is used to indicate a target life value of the inverter and a target value corresponding to at least one physical feature;

[0204] The first training module 402 is configured to obtain first training loss information based on the difference between the first life prediction value and the target life value, and the difference between the first prediction value corresponding to each physical feature and the target value, and adjust the initial life prediction network by using the first training loss information to obtain a candidate life prediction model;

[0205] The second prediction module 403 is configured to input a plurality of sample data sets into the candidate life prediction model to obtain a second life prediction value and a second prediction value corresponding to each physical feature;

[0206] The second training module 404 is configured to obtain second training loss information based on the difference between the second life prediction value and the target life value, the difference between the second prediction value corresponding to each physical feature and the target value, and a physical residual term, and adjust the candidate life prediction model by using the second training loss information to obtain a target life prediction model; the physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent the deviation degree of the candidate life prediction model from the physical law;

[0207] The target output module 405 is configured to input the aging data of a target inverter into the target life prediction model for life prediction, and output a target life prediction value of the target inverter.

[0208] In the embodiment of the present application, the initial life prediction network at least includes a convolutional neural network layer, a multi-head self-attention layer, a feedforward neural network layer, and a fully connected layer; the first prediction module 401 can further include:

[0209] The first feature determination module is configured to input a plurality of sample data sets into the initial life prediction network, and process the plurality of input sample data sets in the convolutional neural network layer to extract local time sequence features and obtain first intermediate features.

[0210] The second feature determination module is configured to input the first intermediate features into the multi-head self-attention layer to capture the correlation of the first intermediate features in the time sequence and obtain second intermediate features.

[0211] The third feature determination module is configured to input the second intermediate features into the feedforward neural network layer for nonlinear mapping to obtain third intermediate features.

[0212] The first prediction value determination module is configured to input the third intermediate features into the fully connected layer to obtain the first life prediction value and the first prediction value corresponding to each physical feature.

[0213] In the embodiment of the present application, the first training loss information further includes an initial condition constraint term, and the initial condition constraint term represents the deviation degree of the initial life prediction network from the known initial state at the initial time; the first training module 402 can further include:

[0214] The first loss function generation module is configured to combine the difference between the first life prediction value and the target life value, the difference between the first prediction value corresponding to each physical feature and the target value, and the initial condition constraint term according to weights to generate a first loss function.

[0215] The first model determination module is configured to perform back propagation and parameter update on the initial life prediction network by using the first loss function to obtain a candidate life prediction model.

[0216] In the embodiment of the present application, the second training loss information further includes an initial condition constraint term, and the initial condition constraint term is further used to represent the deviation degree of the candidate life prediction model from the known initial state at the initial time; the second training module 404 can further include:

[0217] The second loss function generation module is configured to combine the difference between the second life prediction value and the target life value, the difference between the second prediction value corresponding to each physical feature and the target value, the physical residual term, and the initial condition constraint term according to weights to generate a second loss function.

[0218] The second model determination module is configured to perform back propagation and parameter update on the candidate life prediction model by using the second loss function to obtain a target life prediction model.

[0219] In the embodiments of the present application, the second training module 404 can further specifically include:

[0220] The first physical residual calculation module is configured to calculate the inductance current residual and the capacitance voltage residual based on the BOOST circuit average model differential equation;

[0221] The second physical residual calculation module is configured to calculate the junction temperature residual based on the IGBT thermal model equation;

[0222] The physical residual term determination module is configured to average the inductance current residual, the capacitance voltage residual, and the junction temperature residual in the sample dimension to obtain the physical residual term.

[0223] In the embodiments of the present application, the first training module and the second training module are further specifically configured to:

[0224] Obtain initial values of the BOOST inductance current, the BOOST capacitance voltage, and the IGBT junction temperature;

[0225] Obtain the predicted values of each physical feature at the initial time by the initial life prediction network or the candidate life prediction model, and compare the predicted values with the corresponding initial values to determine the deviations of the BOOST inductance current, the BOOST capacitance voltage, and the IGBT junction temperature at the initial time, respectively;

[0226] According to the deviations of the BOOST inductance current, the BOOST capacitance voltage, and the IGBT junction temperature corresponding to the initial life prediction network or the candidate life prediction model, respectively, calculate the initial condition constraint term corresponding to the initial life prediction network or the candidate life prediction model.

[0227] In the embodiments of the present application, the data acquisition module 401 can specifically include:

[0228] The target life value determination module is configured to construct a target life value based on the ratio of the current remaining life to the rated life of the inverter, and the current remaining life refers to the remaining time length from the current sample collection time to the end of life;

[0229] The target value determination module is configured to determine, for each physical feature, a target value corresponding to the physical feature based on the sample measurement value corresponding to the physical feature at the backward adjacent time of the current sample collection time.

[0230] In the embodiments of the present application, the inverter life prediction device 400 can further specifically include:

[0231] The weight saving module is configured to save a plurality of weights of the candidate life prediction model and the target life prediction model in the training process;

[0232] The model evaluation module is configured to evaluate the models corresponding to the plurality of weights based on the root mean square error and the direction penalty score.

[0233] The selection module is configured to select a target life prediction model with the optimal evaluation result as the target life prediction model of the inverter.

[0234] The inverter life prediction device 400 provided by the embodiments of the present application can be applied in the inverter life prediction method provided by the foregoing embodiments, and details are described in the description of the inverter life prediction method provided by the foregoing embodiments, which will not be repeated here.

[0235] Referring to Figure 5 , a structural schematic diagram of an electronic device provided by an embodiment of the present application is shown. As Figure 5 shown, the electronic device 500 of the embodiment includes at least one processor 510 (only one is shown in the figure), a memory 520, and a computer program 521 stored in the memory 520 and executable on the at least one processor 510, and the processor 510 implements the steps in the foregoing inverter life prediction method embodiments when executing the computer program 521. Figure 5

[0236] The electronic device 500 can be a server, a physical server, a computing device, etc. The electronic device can include, but is not limited to, the processor 510 and the memory 520. Those skilled in the art can understand that Figure 5 The electronic device 500 is only an example and does not constitute a limitation on the electronic device 500, and can include more or fewer components than shown, or combine certain components, or different components, for example, can also include an input / output device, a network access device, etc.

[0237] The processor 510 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or can also be any conventional processor.

[0238] ​The memory 520 can be an internal storage unit of the electronic device 500, such as a hard disk or a memory of the electronic device 500 in some embodiments. The memory 520 can also be an external storage device of the electronic device 500, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the electronic device 500 in other embodiments. Further, the memory 520 can include both an internal storage unit and an external storage device of the electronic device 500. The memory 520 is used to store an operating system, an application program, a Boot Loader, data, and other programs, such as program codes of computer programs, etc. The memory 520 can also be used to temporarily store data that has been output or will be output.

[0239] In a specific implementation, the processor 510, the memory 520, and the computer program 521 described in the embodiments of the present application can perform the embodiments of the method for predicting the life of the inverter of the present application, which will not be described here.

[0240] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual applications, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiments can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above integrated unit can be realized in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for easy distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0241] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in a certain embodiment can be referred to the related description of other embodiments.

[0242] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0243] In the embodiments provided by the present application, it should be understood that the disclosed apparatus / equipment and method can be implemented in other ways. For example, the apparatus / equipment embodiments described above are merely schematic. For example, the division of the modules or units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between each other can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or in other forms.

[0244] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place, or distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0245] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0246] The integrated modules / units, if implemented in the form of software functional units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content included in the computer readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electric carrier signals and telecommunication signals.

[0247] The above-mentioned embodiment methods can also be completed by a computer program product, when the computer program product runs on an electronic device, so that the electronic device can implement the steps in the above-mentioned various method embodiments.

[0248] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of inverter lifetime prediction, characterized by, The method comprises: inputting a plurality of sample data sets into an initial life prediction network to obtain first life prediction values and first prediction values corresponding to each physical feature; each sample data set is aging data of a sample inverter, and each sample data set corresponds to a first label indicating a target life value of the inverter and a target value corresponding to at least one physical feature; based on the difference between the first life prediction value and the target life value, the difference between the first prediction value corresponding to each physical feature and the target value, first training loss information is obtained, and the initial life prediction network is adjusted by using the first training loss information to obtain a candidate life prediction model; inputting a plurality of sample data sets into the candidate life prediction model to obtain second life prediction values and second prediction values corresponding to each physical feature; based on the difference between the second life prediction value and the target life value, the difference between the second prediction value corresponding to each physical feature and the target value, and a physical residual term, second training loss information is obtained, and the candidate life prediction model is adjusted by using the second training loss information to obtain a target life prediction model; the physical residual term is used to constrain the physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent the deviation degree of the candidate life prediction model from the physical law; inputting the aging data of a target inverter into the target life prediction model for life prediction, and outputting a target life prediction value of the target inverter; wherein the initial life prediction network comprises at least a convolutional neural network layer, a multi-head self-attention layer, a feedforward neural network layer, and a fully connected layer; inputting a plurality of sample data sets into the initial life prediction network to obtain first life prediction values and first prediction values corresponding to each physical feature comprises: inputting the plurality of sample data sets into the initial life prediction network, processing the plurality of input sample data sets in the convolutional neural network layer to extract local time sequence features, and obtaining first intermediate features; inputting the first intermediate features into the multi-head self-attention layer to capture the correlation of the first intermediate features in the time sequence, and obtaining second intermediate features; inputting the second intermediate features into the feedforward neural network layer for nonlinear mapping, and obtaining third intermediate features; inputting the third intermediate features into the fully connected layer to obtain the first life prediction values and the first prediction values corresponding to each physical feature; wherein the calculation of the physical residual term comprises the following steps: calculating inductance current residual and capacitance voltage residual based on a BOOST circuit average model differential equation; calculating junction temperature residual based on an IGBT thermal model equation; averaging the inductance current residual, the capacitance voltage residual, and the junction temperature residual in the sample dimension to obtain the physical residual term.

2. The method of claim 1, wherein, The first training loss information further comprises an initial condition constraint term, and the initial condition constraint term represents a deviation degree of the initial life prediction network from a known initial state at an initial time; and the adjusting the initial life prediction network by using the first training loss information to obtain a candidate life prediction model comprises: combining a difference between the first life prediction value and the target life value, a difference between a first prediction value corresponding to each physical feature and a target value, and the initial condition constraint term into a first loss function according to weights; performing back propagation and parameter updating on the initial life prediction network by using the first loss function to obtain the candidate life prediction model.

3. The method of claim 1, wherein, The second training loss information further comprises an initial condition constraint term, and the initial condition constraint term is further used to represent a deviation degree of the candidate life prediction model from a known initial state at an initial time; and the adjusting the candidate life prediction model by using the second training loss information to obtain a target life prediction model comprises: combining a difference between the second life prediction value and the target life value, a difference between a second prediction value corresponding to each physical feature and a target value, a physical residual error term, and the initial condition constraint term into a second loss function according to weights; performing back propagation and parameter updating on the candidate life prediction model by using the second loss function to obtain the target life prediction model.

4. The method according to any one of claims 2 or 3, characterized in that, The calculation of the initial condition constraint term comprises the following steps: obtaining initial values of a BOOST inductor current, a BOOST capacitor voltage, and an IGBT junction temperature; obtaining prediction values of each physical feature at an initial time by using the initial life prediction network or the candidate life prediction model, and comparing the prediction values with corresponding initial values to determine deviations of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature at the initial time, respectively; calculating an initial condition constraint term corresponding to the initial life prediction network or the candidate life prediction model according to deviations of the BOOST inductor current, the BOOST capacitor voltage, and the IGBT junction temperature in the initial life prediction network or the candidate life prediction model, respectively.

5. The method of claim 1, wherein, The generation of the first label comprises the following steps: constructing a target life value based on a ratio of a current remaining life to a rated life, wherein the current remaining life refers to a remaining time length from a current sample collection time to a life termination time; for each physical feature, determining a target value corresponding to the physical feature based on a sample measurement value corresponding to the physical feature at a backward adjacent time of a current sample collection time.

6. The method of claim 1, wherein, The method further comprises: saving a plurality of weights of the candidate life prediction model and the target life prediction model during a training process; evaluating models corresponding to the plurality of weights based on a root mean square error and a direction penalty score; selecting a target life prediction model with an optimal evaluation result to output a target life prediction value of the inverter.

7. An apparatus for inverter lifetime prediction, the apparatus comprising: The device comprises: The first prediction module is configured to input a plurality of sample data sets into an initial life prediction network to obtain a first life prediction value and a first prediction value corresponding to each physical feature; the sample data set is aging data of a sample inverter, and each sample data set corresponds to a first label indicating a target life value of the inverter and a target value corresponding to at least one physical feature; The first training module is configured to obtain a first training loss information based on a difference between the first life prediction value and the target life value, a difference between the first prediction value corresponding to each physical feature and the target value, adjust the initial life prediction network based on the first training loss information, and obtain a candidate life prediction model; The second prediction module is configured to input a plurality of sample data sets into the candidate life prediction model to obtain a second life prediction value and a second prediction value corresponding to each physical feature; The second training module is configured to obtain a second training loss information based on a difference between the second life prediction value and the target life value, a difference between the second prediction value corresponding to each physical feature and the target value, and a physical residual term, and adjust the candidate life prediction model based on the second training loss information to obtain a target life prediction model; the physical residual term is used to constrain a physical quantity output by the candidate life prediction model to satisfy a preset physical equation, so as to represent a deviation degree of the candidate life prediction model from a physical law; The target output module is configured to input aging data of a target inverter into the target life prediction model for life prediction, and output a target life prediction value of the target inverter; The initial life prediction network at least includes a convolutional neural network layer, a multi-head self-attention layer, a feedforward neural network layer, and a fully connected layer; the first prediction module includes: The first feature determination module is configured to input a plurality of sample data sets into an initial life prediction network, process the plurality of input sample data sets in the convolutional neural network layer to extract local time sequence features, and obtain first intermediate features; The second feature determination module is configured to input the first intermediate features into the multi-head self-attention layer to capture the correlation of the first intermediate features in the time sequence, and obtain second intermediate features; The third feature determination module is configured to input the second intermediate features into the feedforward neural network layer for nonlinear mapping to obtain third intermediate features; The first prediction value determination module is configured to input the third intermediate features into the fully connected layer to obtain the first life prediction value and the first prediction value corresponding to each physical feature; The second training module includes: The first physical residual calculation module is configured to calculate an inductor current residual and a capacitor voltage residual based on a BOOST circuit average model differential equation; The second physical residual calculation module is configured to calculate a junction temperature residual based on an IGBT thermal model equation; The physical residual term determination module is configured to average the inductor current residual, the capacitor voltage residual, and the junction temperature residual in the sample dimension to obtain the physical residual term.

8. The apparatus for predicting life of an inverter according to claim 7, wherein The first training loss information further comprises an initial condition constraint term, and the initial condition constraint term represents a deviation degree of the initial life prediction network from a known initial state at an initial time; The first training module further comprises: a first loss function generation module, configured to combine a difference between the first life prediction value and the target life value, a difference between the first prediction value corresponding to each physical feature and a target value, and the initial condition constraint term according to weights to generate a first loss function; a first model determination module, configured to perform back propagation and parameter update on the initial life prediction network by using the first loss function, to obtain a candidate life prediction model.

9. The apparatus for predicting life of an inverter according to claim 7, wherein The second training loss information further comprises an initial condition constraint term, and the initial condition constraint term is further used to represent a deviation degree of the candidate life prediction model from the known initial state at the initial time; The second training module further comprises: a second loss function generation module, configured to combine a difference between the second life prediction value and the target life value, a difference between the second prediction value corresponding to each physical feature and a target value, the physical residual term, and the initial condition constraint term according to weights to generate a second loss function; a second model determination module, configured to perform back propagation and parameter update on the candidate life prediction model by using the second loss function, to obtain a target life prediction model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the method in any one of claims 1 to 6.

11. A computer-readable storage medium, the computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 6.

Images