Substation equipment service life prediction method and device

Abstract

The application provides a substation equipment service life prediction method and device, by constructing a first set containing multi-dimensional features, the following steps are iteratively executed until the first set is empty: placing the feature with the minimum feature weight in the first set in a second set and setting the deletion probability of the feature with the minimum feature weight to 0; adding 1 to the deletion probability of each feature in the second set; deleting the feature with the minimum feature weight from the first set, updating the feature weight of each feature in the first set; after the iteration is completed, normalizing the deletion probability of each feature in the second set, and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determining the target feature based on the selection probability of each feature; determining the estimated service life of the to-be-predicted equipment based on the feature data of the target feature corresponding to the to-be-predicted equipment, and improving the accuracy of the substation equipment service life prediction task.

Description

Technical Field

[0001] This invention relates to the field of equipment evaluation technology, and in particular to a method and apparatus for predicting the service life of substation equipment. Background Technology

[0002] As the core hub of the ultra-high voltage (UHV) power transmission system, UHV substations undertake the critical tasks of transmitting, distributing, and regulating electricity. The stability and safety of equipment within these substations, such as transformers, reactors, and capacitors, are paramount. Therefore, predicting the remaining service life of substation equipment helps achieve precise operation and maintenance, avoids sudden equipment failures, and reduces maintenance costs. Furthermore, accurate service life prediction can optimize asset investment decisions, preventing losses from premature or delayed equipment replacement.

[0003] Current methods for predicting the lifespan of substation equipment mainly include physical model-based methods and machine learning / deep learning-based methods. Physical model-based methods establish lifespan prediction models based on the physical characteristics of equipment operation (such as material aging, fatigue curves, and thermal effects), while machine learning / deep learning-based methods train regression or classification models using large-scale monitoring data (such as temperature, humidity, current, voltage, and vibration signals) to predict the remaining lifespan of the equipment. Because physical models are complex to construct and have high requirements for internal parameters and physical characteristics of the equipment, they are difficult to adapt to different equipment and environments. Therefore, existing lifespan prediction methods mostly adopt machine learning / deep learning-based methods. However, many features are related to the lifespan of substation equipment. For example, environmental data, manufacturer data, maintenance data, and various operational data monitored during actual operation may all affect the lifespan of substation equipment to varying degrees. Therefore, machine learning / deep learning-based methods face the challenge of too many input features, and the model is prone to learning irrelevant or weakly relevant information, increasing computational complexity and reducing prediction accuracy. Summary of the Invention

[0004] This invention provides a method and apparatus for predicting the service life of substation equipment, which solves the defects of high computational complexity and low prediction accuracy in the prior art when faced with a large number of input features.

[0005] This invention provides a method for predicting the service life of substation equipment, comprising:

[0006] Data acquisition steps: Acquire the feature data and expected lifespan of the multi-dimensional features corresponding to the sample substation equipment, and construct a first set containing the multi-dimensional features;

[0007] Feature weight calculation steps: Place the feature with the smallest feature weight in the first set into the second set and set the deletion probability of the feature with the smallest feature weight to 0; increment the deletion probability of each feature in the second set by 1; delete the feature with the smallest feature weight from the first set and update the feature weight of each feature in the first set; wherein, the feature weight of each feature is determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and its expected lifespan;

[0008] Iterative steps: Repeat the feature weight calculation steps until the first set is empty; normalize the deletion probability of each feature in the second set, and determine the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determine the target feature based on the selection probability of each feature.

[0009] Lifetime prediction step: Based on the feature data of the target features corresponding to the device to be predicted, determine the estimated lifetime of the device to be predicted.

[0010] According to the substation equipment service life prediction method provided by the present invention, the feature weights of each feature are determined based on the following steps:

[0011] The λ parameter is updated based on the sum of the current adjustment factor and the current value of the λ parameter of the Lasso regression model;

[0012] Based on the feature data of multidimensional features corresponding to sample substation equipment and their expected lifespan, the feature weights of each feature are determined using the Lasso regression model.

[0013] If the number of features with a feature weight of 0 is greater than or equal to 2, then the current adjustment factor is reduced.

[0014] If the number of features with a feature weight of 0 is less than 2, then the current adjustment factor remains unchanged.

[0015] According to the present invention, a method for predicting the service life of substation equipment includes target features and non-sensor features; the step of determining the estimated service life of the equipment to be predicted based on feature data corresponding to the target features includes:

[0016] The embedding layer based on the lifetime prediction model obtains the embedding vector of the feature data corresponding to the target feature of the device to be predicted, and constructs an embedding matrix based on the embedding vector of the feature data corresponding to the sensor feature of the device to be predicted; wherein, one row of the embedding matrix is ​​the embedding vector of the feature data corresponding to any sensor feature of the device to be predicted.

[0017] The encoding and decoding layer based on the lifetime prediction model encodes and decodes the embedding matrix to obtain the feature vector of the embedding matrix;

[0018] The fusion layer based on the lifetime prediction model fuses the embedding vector of the feature data corresponding to the non-sensor features of the device to be predicted and the feature vector of the embedding matrix to obtain the fusion vector of the device to be predicted.

[0019] The prediction layer based on the lifetime prediction model performs lifetime prediction on the fusion vector of the device to be predicted, and obtains the estimated lifetime of the device to be predicted.

[0020] According to the present invention, a method for predicting the service life of substation equipment is provided, wherein the service life prediction model comprises multiple encoding and decoding layers; the encoding and decoding layers based on the service life prediction model encode and decode the embedding matrix to obtain the feature vector of the embedding matrix, including:

[0021] The encoding matrix output by the encoding module of the previous encoding / decoding layer is fused based on a sliding time window to obtain a fused encoding matrix;

[0022] The fused encoding matrix is ​​encoded based on the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer;

[0023] The decoding module of the current codec layer decodes the decoding matrix output by the decoding module of the previous codec layer and the encoding matrix output by the encoding module of the current codec layer to obtain the decoding matrix output by the decoding module of the current codec layer.

[0024] The decoding matrix output by the decoding module of the last encoding / decoding layer is vectorized to obtain the feature vector of the embedding matrix.

[0025] According to the present invention, a method for predicting the service life of substation equipment includes fusing each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer based on a sliding time window to obtain a fused encoding matrix, comprising:

[0026] For any row of the encoding matrix output by the encoding module of the previous encoding / decoding layer, the sub-vector within the current sliding time window is converted into a sub-vector of length k based on a preset transformation matrix and then concatenated to the fusion vector corresponding to the row. The initial value of the fusion vector corresponding to the row is empty, the length of any row of the encoding matrix is ​​P×k, the length of the sliding time window is 2k, and the sliding step size of the sliding time window is k. The fusion vector corresponding to each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer constitutes the fusion encoding matrix.

[0027] According to the present invention, a method for predicting the service life of substation equipment includes encoding the fused encoding matrix based on the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer, comprising:

[0028] Based on the first self-attention unit of the current encoding / decoding layer's encoding module, the first attention vector corresponding to each row of the fusion encoding matrix is ​​obtained, and a first attention matrix is ​​constructed based on the first attention vector corresponding to each row of the fusion encoding matrix.

[0029] Based on the second self-attention unit of the encoding module of the current encoding / decoding layer, the second attention vector corresponding to each column of the first attention matrix is ​​obtained, and the encoding matrix is ​​constructed based on the second attention vector corresponding to each column of the first attention matrix.

[0030] According to the substation equipment life prediction method provided by the present invention, the decoding matrix output by the decoding module of the previous encoding / decoding layer and the encoding matrix output by the encoding module of the current encoding / decoding layer are decoded by the decoding module of the current encoding / decoding layer to obtain the decoding matrix output by the decoding module of the current encoding / decoding layer, including:

[0031] Based on the third self-attention unit of the decoding module of the current codec layer, the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the previous codec layer is obtained respectively, and a second attention matrix is ​​constructed based on the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the previous codec layer.

[0032] Based on the fourth self-attention unit of the decoding module of the current encoding and decoding layer, the fourth attention vector corresponding to each column of the second attention matrix is ​​obtained respectively, and the basic decoding matrix is ​​constructed based on the fourth attention vector corresponding to each column of the second attention matrix.

[0033] Based on the cross-attention unit of the decoding module of the current encoding / decoding layer, attention transformation is performed on the basic decoding matrix and the encoding matrix output by the encoding module of the current encoding / decoding layer to obtain the decoding matrix.

[0034] The present invention also provides a substation equipment lifespan prediction device, comprising:

[0035] The data acquisition unit is used to acquire feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, and to construct a first set containing multi-dimensional features;

[0036] The target feature filtering unit is used to perform a feature weight calculation step and an iteration step. The feature weight calculation step includes: placing the feature with the smallest feature weight in the first set into the second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set and updating the feature weights of each feature in the first set; wherein the feature weights of each feature are determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and their expected lifespan. The iteration step includes: repeating the feature weight calculation step until the first set is empty; normalizing the deletion probability of each feature in the second set and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; and determining the target feature based on the selection probability of each feature.

[0037] The lifespan prediction unit is used to determine the estimated lifespan of the device to be predicted based on the feature data of the target features corresponding to the device to be predicted.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the substation equipment life prediction method as described above.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the substation equipment lifespan prediction method as described above.

[0040] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the substation equipment lifespan prediction method as described above.

[0041] The present invention provides a method and apparatus for predicting the service life of substation equipment. This method acquires feature data and expected service life of multi-dimensional features corresponding to sample substation equipment, constructs a first set containing these multi-dimensional features, and iteratively executes the following steps until the first set is empty: The feature with the smallest feature weight in the first set is placed in a second set, and the deletion probability of the feature with the smallest feature weight is set to 0; the deletion probability of each feature in the second set is incremented by 1; the feature with the smallest feature weight is deleted from the first set, and the feature weights of each feature in the first set are updated; after iteration, the deletion probability of each feature in the second set is normalized, and the selection probability of each feature is determined based on the normalized deletion probability; the target feature is determined based on the selection probability of each feature; thus, based on the feature data of the target feature corresponding to the equipment to be predicted, the estimated service life of the equipment to be predicted is determined. This method can screen out the key features that determine the service life of substation equipment, improving the accuracy of the substation equipment service life prediction task. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the substation equipment lifespan prediction method provided by the present invention.

[0044] Figure 2 This is a schematic diagram of the lifespan prediction model provided by the present invention;

[0045] Figure 3 This is a flowchart illustrating the encoding and decoding method provided by the present invention;

[0046] Figure 4 This is a schematic diagram of the substation equipment lifespan prediction device provided by the present invention;

[0047] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0049] Figure 1 This is a flowchart illustrating the substation equipment lifespan prediction method provided by the present invention, as shown below. Figure 1 As shown, the method includes:

[0050] Data acquisition step 110: Acquire the feature data and expected lifespan of the multi-dimensional features corresponding to the sample substation equipment, and construct a first set containing the multi-dimensional features;

[0051] Feature weight calculation step 120: Place the feature with the smallest feature weight in the first set into the second set and set the deletion probability of the feature with the smallest feature weight to 0; increment the deletion probability of each feature in the second set by 1; delete the feature with the smallest feature weight from the first set and update the feature weight of each feature in the first set; wherein, the feature weight of each feature is determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and its expected lifespan;

[0052] Iteration step 130: Repeat the feature weight calculation step until the first set is empty; normalize the deletion probability of each feature in the second set, and determine the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determine the target feature based on the selection probability of each feature.

[0053] Lifetime prediction step 140: Based on the feature data of the target features corresponding to the device to be predicted, determine the estimated lifetime of the device to be predicted.

[0054] Here, feature data corresponding to the multi-dimensional features of the sample substation equipment is acquired, including environmental data, maintenance data, and operational data monitored during actual equipment operation. Taking a transformer as an example, multi-dimensional features may include ambient temperature, ambient humidity, ambient pressure, maintenance object, maintenance time, transformer manufacturing year, voltage, current, load, and oil temperature. It should be noted that multi-dimensional features include sensor features, and the feature data of sensor features (such as ambient temperature, ambient humidity, ambient pressure, voltage, current, load, and oil temperature) are generally time-series data, that is, a data sequence monitored by the corresponding sensors over a period of time. The expected lifespan of the sample substation equipment can be obtained by labeling the actual decommissioning lifespan of the corresponding equipment, where the expected lifespan can be a time interval, such as 3.5-4 years. For example, based on the time difference between the actual decommissioning lifespan of the sample substation equipment and the acquisition time of the feature data, the time interval containing this time difference can be taken as the expected lifespan of the equipment according to a pre-defined time interval. Subsequently, a first set is constructed for the multi-dimensional features. The first set may contain only unique markers for each feature to indicate the corresponding feature.

[0055] To select features from multidimensional features that determine the remaining service life of substation equipment, the following feature weight calculation steps can be iteratively executed: The feature with the smallest feature weight in the first set is placed in the second set, and its deletion probability is set to 0; the deletion probability of each feature in the second set is incremented by 1; the feature with the smallest feature weight is deleted from the first set, and the feature weights of each feature in the first set are updated. It can be seen that after multiple iterations, features with lower feature weights are placed into the second set first, and the earlier a feature is placed into the second set, the greater its deletion probability. Specifically, determining the initial feature weights of each feature in the first set and updating the feature weights of each feature in the first set can be achieved by calling the feature weight calculation module. The feature data and expected service life of the features corresponding to the substation equipment whose feature weights need to be calculated / updated are input into the feature weight calculation module to realize the calculation of feature weights.

[0056] In some embodiments, each time the feature weight calculation module is invoked, the feature weight calculation module may perform the following steps:

[0057] The λ parameter is updated based on the sum of the current adjustment factor and the current value of the λ parameter of the Lasso regression model. Subsequently, based on the feature data and expected lifetime of the features for which feature weights need to be calculated / updated corresponding to the sample substation equipment input into the module, the feature weights of each feature are determined using the Lasso regression model. Specifically, features with a weight of 0 can be identified. If the number of features with a weight of 0 is greater than or equal to 2, the current adjustment factor is reduced; if the number of features with a weight of 0 is less than 2, the current adjustment factor remains unchanged. By setting the current adjustment factor each time the feature weight calculation module is called, thereby controlling the λ parameter of the Lasso regression model, the sparsity of the Lasso regression model can be flexibly controlled. This achieves the effect of compressing only one feature at a time to filter out the features with the lowest importance, which helps improve the accuracy of feature deletion probabilities during the above iteration process, thus improving the accuracy of feature selection.

[0058] The feature weight calculation steps described above are iteratively executed until the first set is empty, at which point the iteration ends. Subsequently, the deletion probability of each feature in the second set can be normalized, and the selection probability of each feature is determined based on the normalized deletion probability. Specifically, the selection probability of any feature can be obtained by subtracting the normalized deletion probability of any feature from the value 1. Based on the selection probability of each feature, several features with higher selection probabilities can be identified as target features. In some embodiments, the average selection probability of each feature in the second set can be calculated, and features with selection probabilities greater than this average can be identified as target features.

[0059] After selecting the target features, if remaining service life prediction is required, the estimated service life of the device to be predicted can be determined based on the feature data of the target features corresponding to the device to be predicted. In some embodiments, the target features include sensor features and non-sensor features, where the feature data of sensor features is time-series data and the non-sensor features is non-time-series data. During service life prediction, a linear regression model can be constructed based on the feature weights of each target feature, and this linear regression model can be used to predict the service life of the device to be predicted. However, there may be complex nonlinear relationships between the various features of substation equipment, and there may also be certain temporal correlations within sensor features, which are difficult to reflect by constructing a linear regression model. Therefore, a deep learning mechanism can be used to construct a service life prediction model based on the above target features, and the service life prediction model can be trained using the feature data of each target feature corresponding to the sample substation equipment and their expected service life, so that the trained service life prediction model can be used for service life prediction. It should be noted that the device to be predicted and the sample substation equipment are both of the same type.

[0060] In some embodiments, such as Figure 2 As shown, the lifetime prediction model includes an embedding layer, an encoding / decoding layer, a fusion layer, and a prediction layer. The embedding layer, which can be constructed based on the input embedding layer of the Transformer model, is used to obtain the embedding vectors of the feature data corresponding to the target features of the input device to be predicted. After obtaining the embedding vectors of the feature data corresponding to the target features of the device to be predicted, the embedding layer also combines the embedding vectors of the sensor feature data into an embedding matrix, where each row of the embedding matrix is ​​the embedding vector of the feature data corresponding to a certain sensor feature of the device to be predicted. The encoding / decoding layer of the lifetime prediction model is used to encode and decode the above embedding matrix to obtain the feature vector of the embedding matrix. In one embodiment, the lifetime prediction model may contain multiple sequentially connected encoding / decoding layers, each of which can be constructed based on an Encoder-Decoder model.

[0061] In some embodiments, when there are multiple codec layers, such as Figure 3 As shown, multiple encoding / decoding layers can perform encoding / decoding in the following manner to obtain the feature vector of the embedding matrix:

[0062] Step 310: Based on the sliding time window, each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer is fused to obtain the fused encoding matrix;

[0063] Step 320: Encode the fused coding matrix based on the coding module of the current coding layer to obtain the coding matrix output by the coding module of the current coding layer;

[0064] Step 330: Based on the decoding module of the current codec layer, decode the decoding matrix output by the decoding module of the previous codec layer and the encoding matrix output by the encoding module of the current codec layer to obtain the decoding matrix output by the decoding module of the current codec layer.

[0065] Step 340: The decoding matrix output by the decoding module of the last encoding / decoding layer is vectorized to obtain the feature vector of the embedding matrix.

[0066] Specifically, for any encoding / decoding layer, each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer can be fused based on a sliding time window to obtain a fused encoding matrix. Each row of the encoding matrix corresponds to a sensor feature. By fusing each row of the encoding matrix, temporal information of the same sensor feature at different time scales can be obtained, thereby enhancing the predictive ability of the lifetime prediction model. In some embodiments, when processing any row of the encoding matrix output by the encoding module of the previous encoding / decoding layer using a sliding time window, the sub-vector within the current sliding time window can be converted into a sub-vector of length k based on a preset transformation matrix (which can be learned during model training), and then concatenated to the corresponding fused vector of that row to form a new fused vector. Here, the initial value of the fused vector corresponding to any row is empty, the length of any row of the encoding matrix is ​​P×k, the length of the sliding time window is 2k, and the sliding step size of the sliding time window is k. If the sliding time window slides to the end of the row, it now contains only a sub-vector of length k, so this sub-vector can be directly concatenated to the corresponding fused vector of that row to form the final fused vector of that row. The fusion vectors corresponding to each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer constitute the fusion encoding matrix.

[0067] It should be noted that if the current encoding / decoding layer is the first encoding / decoding layer, the embedding vector is directly used as the above-mentioned fusion encoding matrix.

[0068] Subsequently, the fused encoding matrix is ​​encoded based on the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer. In some embodiments, considering that the encoder in the existing Encoder-Decoder model has difficulty in synchronously acquiring the relationship between features from different sensors, the encoder in the Encoder-Decoder model can be improved to form the encoding module in the embodiments of this application. Specifically, the encoding module may include a first self-attention unit and a second self-attention unit.

[0069] The first self-attention unit is used to obtain the first attention vector corresponding to each row of the fusion coding matrix, and construct a first attention matrix based on the first attention vectors corresponding to each row of the fusion coding matrix. Here, for any row of the fusion coding matrix, the first self-attention unit obtains the attention weight of each element in that row based on the self-attention mechanism, normalizes the attention weight of each element in that row, and then weights each element in that row to obtain the first attention vector corresponding to that row (i.e., the row vector composed of the weighted elements of that row). Through the first self-attention unit, the temporal relationship within a single sensor feature can be obtained. Subsequently, the second self-attention unit is used to obtain the second attention vector corresponding to each column of the first attention matrix, and construct a coding matrix based on the second attention vectors corresponding to each column of the first attention matrix. Here, for any column of the first attention matrix, the second self-attention unit obtains the attention weight of each element in that column based on the self-attention mechanism, normalizes the attention weight of each element in that column, and then weights each element in that column to obtain the second attention vector corresponding to that column (i.e., the column vector composed of the weighted elements of that column). Through the second self-attention unit, the correlation between sensor features can be obtained.

[0070] After obtaining the encoding matrix output by the encoding module of the current encoding / decoding layer, the decoding matrix output by the decoding module of the previous encoding / decoding layer and the encoding matrix output by the current encoding / decoding layer can be decoded based on the current encoding / decoding layer's decoding module to obtain the decoding matrix output by the current encoding / decoding layer's encoding module. It should be noted that if the current encoding / decoding layer is the first encoding / decoding layer, the input to the current encoding / decoding layer's decoding module is a preset initial matrix (which can be learned through the model training process) and the encoding matrix output by the current encoding / decoding layer's encoding module.

[0071] The decoding module includes a third self-attention unit and a fourth self-attention unit. The third self-attention unit obtains the third attention vector corresponding to each row of the decoding matrix (or a preset initial matrix) output by the decoding module of the previous encoding / decoding layer, and constructs a second attention matrix based on the third attention vectors corresponding to each row of the decoding matrix (or the preset initial matrix). Here, for any row of the decoding matrix (or the preset initial matrix), the third self-attention unit obtains the attention weight of each element in that row based on the self-attention mechanism, normalizes the attention weight of each element in that row, and then weights each element in that row to obtain the third attention vector corresponding to that row (i.e., the row vector composed of the weighted elements of that row). The fourth self-attention unit obtains the fourth attention vector corresponding to each column of the second attention matrix, and constructs a basic decoding matrix based on the fourth attention vectors corresponding to each column of the second attention matrix. Here, for any column of the second attention matrix, the fourth self-attention unit obtains the attention weight of each element in the column based on the self-attention mechanism, normalizes the attention weight of each element in the column, and then weights each element in the column to obtain the fourth attention vector corresponding to the column (i.e., the column vector composed of the weighted elements of the column).

[0072] The obtained base decoding matrix contains the correlations between sensor features and the temporal relationships within those features. Therefore, the decoding module of the current encoding / decoding layer can output this base decoding matrix as the decoding matrix. To further improve the accuracy of feature extraction, a cross-attention unit can be added to the decoding module. This cross-attention unit is used to perform attention transformation on the base decoding matrix and the encoding matrix output by the current encoding / decoding layer's encoding module based on the cross-attention mechanism, resulting in the decoding matrix. Specifically, a query vector can be constructed based on any row of the base decoding matrix (assuming it's the i-th row), and a key vector and value vector can be constructed based on the corresponding row of the encoding matrix output by the current encoding / decoding layer's encoding module (also the i-th row). This allows the attention weights of each element in the i-th row of the base decoding matrix to be obtained using the attention mechanism. The elements in the i-th row are then weighted based on these attention weights to obtain the fifth attention vector of the i-th row of the base decoding matrix. The fifth attention vectors of each row of the base decoding matrix can be combined to form the decoding matrix.

[0073] After obtaining the decoding matrices output by each codec layer in the above manner, the decoding matrix output by the decoding module of the last codec layer can be vectorized to obtain the feature vector of the embedding matrix.

[0074] Based on the eigenvectors of the embedding matrix, the fusion layer of the lifetime prediction model can fuse the embedding vectors of the non-sensor features corresponding to the device to be predicted with the eigenvectors of the embedding matrix to obtain the fused vector of the device to be predicted. Here, the fusion layer can achieve the above fusion by concatenation. Subsequently, the prediction layer of the lifetime prediction model can predict the lifetime of the device to be predicted using the fused vector, thus obtaining the estimated lifetime of the device. The prediction layer can be constructed based on fully connected layers and softmax layers. The prediction layer can output the probability of the device to be predicted for each predefined time interval, thereby determining the time interval with the highest probability as the estimated lifetime of the device to be predicted.

[0075] In summary, the method provided by this invention obtains feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, constructs a first set containing multi-dimensional features, and iteratively executes the following steps until the first set is empty: placing the feature with the smallest feature weight in the first set into a second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set and updating the feature weights of each feature in the first set; normalizing the deletion probability of each feature in the second set after iteration, and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determining the target feature based on the selection probability of each feature; thereby determining the estimated lifespan of the equipment to be predicted based on the feature data of the target feature corresponding to the equipment to be predicted, which can screen out the key features that determine the lifespan of substation equipment and improve the accuracy of the substation equipment lifespan prediction task.

[0076] The substation equipment lifespan prediction device provided by the present invention is described below. The substation equipment lifespan prediction device described below and the substation equipment lifespan prediction method described above can be referred to in correspondence.

[0077] Based on any of the above embodiments Figure 4 This is a schematic diagram of the substation equipment lifespan prediction device provided by the present invention, as shown below. Figure 4 As shown, the device includes:

[0078] The data acquisition unit 410 is used to acquire the feature data of the multi-dimensional features corresponding to the sample substation equipment and their expected lifespan, and to construct a first set containing the multi-dimensional features.

[0079] The target feature filtering unit 420 is used to perform a feature weight calculation step and an iteration step. The feature weight calculation step includes: placing the feature with the smallest feature weight in the first set into the second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set and updating the feature weights of each feature in the first set; wherein the feature weights of each feature are determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and their expected lifespan. The iteration step includes: repeating the feature weight calculation step until the first set is empty; normalizing the deletion probability of each feature in the second set and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; and determining the target feature based on the selection probability of each feature.

[0080] The lifespan prediction unit 430 is used to determine the estimated lifespan of the device to be predicted based on the feature data of the target features corresponding to the device to be predicted.

[0081] The apparatus provided in this invention acquires feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, constructs a first set containing multi-dimensional features, and iteratively executes the following steps until the first set is empty: placing the feature with the smallest feature weight in the first set into a second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set and updating the feature weights of each feature in the first set; normalizing the deletion probability of each feature in the second set after iteration, and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determining the target feature based on the selection probability of each feature; thereby determining the estimated lifespan of the equipment to be predicted based on the feature data of the target feature corresponding to the equipment to be predicted, which can screen out the key features that determine the lifespan of substation equipment and improve the accuracy of the substation equipment lifespan prediction task.

[0082] Based on any of the above embodiments, the feature weights of each feature are determined based on the following steps:

[0083] The λ parameter is updated based on the sum of the current adjustment factor and the current value of the λ parameter of the Lasso regression model;

[0084] Based on the feature data of multidimensional features corresponding to sample substation equipment and their expected lifespan, the feature weights of each feature are determined using the Lasso regression model.

[0085] If the number of features with a feature weight of 0 is greater than or equal to 2, then the current adjustment factor is reduced.

[0086] If the number of features with a feature weight of 0 is less than 2, then the current adjustment factor remains unchanged.

[0087] Based on any of the above embodiments, the target features include sensor features and non-sensor features; determining the estimated lifespan of the device to be predicted based on the feature data corresponding to the target features of the device to be predicted includes:

[0088] The embedding layer based on the lifetime prediction model obtains the embedding vector of the feature data corresponding to the target feature of the device to be predicted, and constructs an embedding matrix based on the embedding vector of the feature data corresponding to the sensor feature of the device to be predicted; wherein, one row of the embedding matrix is ​​the embedding vector of the feature data corresponding to any sensor feature of the device to be predicted.

[0089] The encoding and decoding layer based on the lifetime prediction model encodes and decodes the embedding matrix to obtain the feature vector of the embedding matrix;

[0090] The fusion layer based on the lifetime prediction model fuses the embedding vector of the feature data corresponding to the non-sensor features of the device to be predicted and the feature vector of the embedding matrix to obtain the fusion vector of the device to be predicted.

[0091] The prediction layer based on the lifetime prediction model performs lifetime prediction on the fusion vector of the device to be predicted, and obtains the estimated lifetime of the device to be predicted.

[0092] Based on any of the above embodiments, the lifetime prediction model includes multiple encoding and decoding layers; the encoding and decoding layers based on the lifetime prediction model encode and decode the embedding matrix to obtain the feature vector of the embedding matrix, including:

[0093] The encoding matrix output by the encoding module of the previous encoding / decoding layer is fused based on a sliding time window to obtain a fused encoding matrix;

[0094] The fused encoding matrix is ​​encoded based on the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer;

[0095] The decoding module of the current codec layer decodes the decoding matrix output by the decoding module of the previous codec layer and the encoding matrix output by the encoding module of the current codec layer to obtain the decoding matrix output by the decoding module of the current codec layer.

[0096] The decoding matrix output by the decoding module of the last encoding / decoding layer is vectorized to obtain the feature vector of the embedding matrix.

[0097] Based on any of the above embodiments, the step of fusing each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer based on a sliding time window to obtain a fused encoding matrix includes:

[0098] For any row of the encoding matrix output by the encoding module of the previous encoding / decoding layer, the sub-vector within the current sliding time window is converted into a sub-vector of length k based on a preset transformation matrix and then concatenated to the fusion vector corresponding to the row. The initial value of the fusion vector corresponding to the row is empty, the length of any row of the encoding matrix is ​​P×k, the length of the sliding time window is 2k, and the sliding step size of the sliding time window is k. The fusion vector corresponding to each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer constitutes the fusion encoding matrix.

[0099] Based on any of the above embodiments, the process of encoding the fused encoding matrix using the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer includes:

[0100] Based on the first self-attention unit of the current encoding / decoding layer's encoding module, the first attention vector corresponding to each row of the fusion encoding matrix is ​​obtained, and a first attention matrix is ​​constructed based on the first attention vector corresponding to each row of the fusion encoding matrix.

[0101] Based on the second self-attention unit of the encoding module of the current encoding / decoding layer, the second attention vector corresponding to each column of the first attention matrix is ​​obtained, and the encoding matrix is ​​constructed based on the second attention vector corresponding to each column of the first attention matrix.

[0102] Based on any of the above embodiments, the decoding module of the current codec layer decodes the decoding matrix output by the decoding module of the previous codec layer and the encoding matrix output by the encoding module of the current codec layer to obtain the decoding matrix output by the decoding module of the current codec layer, including:

[0103] Based on the third self-attention unit of the decoding module of the current codec layer, the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the previous codec layer is obtained respectively, and a second attention matrix is ​​constructed based on the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the previous codec layer.

[0104] Based on the fourth self-attention unit of the decoding module of the current encoding and decoding layer, the fourth attention vector corresponding to each column of the second attention matrix is ​​obtained respectively, and the basic decoding matrix is ​​constructed based on the fourth attention vector corresponding to each column of the second attention matrix.

[0105] Based on the cross-attention unit of the decoding module of the current encoding / decoding layer, attention transformation is performed on the basic decoding matrix and the encoding matrix output by the encoding module of the current encoding / decoding layer to obtain the decoding matrix.

[0106] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include: a processor 510, a memory 520, a communication interface 530, and a communication bus 540, wherein the processor 510, the memory 520, and the communication interface 530 communicate with each other through the communication bus 540. The processor 510 can call logic instructions in the memory 520 to execute a substation equipment lifespan prediction method. This method includes: a data acquisition step: acquiring feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, and constructing a first set containing multi-dimensional features; a feature weight calculation step: placing the feature with the smallest feature weight in the first set into a second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set, and updating the feature weight of each feature in the first set; wherein the feature weight of each feature is determined based on the feature data and expected lifespan of the multi-dimensional features corresponding to the sample substation equipment; an iteration step: repeating the feature weight calculation step until the first set is empty; normalizing the deletion probability of each feature in the second set, and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determining the target feature based on the selection probability of each feature; and a lifespan prediction step: determining the estimated lifespan of the equipment to be predicted based on the feature data of the target feature corresponding to the equipment to be predicted.

[0107] Furthermore, the logical instructions in the aforementioned memory 520 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the substation equipment lifespan prediction method provided by the above methods, the method comprising: a data acquisition step: acquiring feature data and expected lifespan of sample substation equipment corresponding to multi-dimensional features, and constructing a first set containing multi-dimensional features; a feature weight calculation step: placing the feature with the smallest feature weight in the first set into a second set and setting the deletion probability of the feature with the smallest feature weight to 0; and setting the deletion probability of each feature in the second set to 0. Increment the feature weight by 1; delete the feature with the smallest feature weight from the first set, and update the feature weight of each feature in the first set; wherein, the feature weight of each feature is determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and its expected lifespan; iterative steps: repeat the feature weight calculation steps until the first set is empty; normalize the deletion probability of each feature in the second set, and determine the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determine the target feature based on the selection probability of each feature; lifespan prediction steps: determine the estimated lifespan of the equipment to be predicted based on the feature data of the target feature corresponding to the equipment to be predicted.

[0109] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the aforementioned substation equipment lifespan prediction methods. The method includes: a data acquisition step: acquiring feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, and constructing a first set containing multi-dimensional features; a feature weight calculation step: placing the feature with the smallest feature weight in the first set into a second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set, and updating the feature weight of each feature in the first set; wherein the feature weight of each feature is determined based on the feature data and expected lifespan of the multi-dimensional features corresponding to the sample substation equipment; an iteration step: repeating the feature weight calculation step until the first set is empty; normalizing the deletion probability of each feature in the second set, and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determining a target feature based on the selection probability of each feature; and a lifespan prediction step: determining the estimated lifespan of the equipment to be predicted based on the feature data of the target feature corresponding to the equipment to be predicted.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the service life of substation equipment, characterized in that, include: Data acquisition steps: Acquire the feature data and expected lifespan of the multi-dimensional features corresponding to the sample substation equipment, and construct a first set containing the multi-dimensional features; Feature weight calculation steps: Place the feature with the smallest feature weight in the first set into the second set and set the deletion probability of the feature with the smallest feature weight to 0; increment the deletion probability of each feature in the second set by 1; delete the feature with the smallest feature weight from the first set and update the feature weight of each feature in the first set; wherein, the feature weight of each feature is determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and its expected lifespan; Iterative steps: Repeat the feature weight calculation steps until the first set is empty; normalize the deletion probability of each feature in the second set, and determine the selection probability of each feature based on the normalized deletion probability of each feature in the second set; determine the target feature based on the selection probability of each feature. Lifetime prediction step: Based on the feature data of the target features corresponding to the device to be predicted, determine the estimated lifetime of the device to be predicted; The feature weights of each feature are determined based on the following steps: updating the λ parameter based on the sum of the current adjustment factor and the current value of the λ parameter of the Lasso regression model; Based on the feature data of multidimensional features corresponding to sample substation equipment and their expected lifespan, the feature weights of each feature are determined using the Lasso regression model. If the number of features with a feature weight of 0 is greater than or equal to 2, then the current adjustment factor is reduced. If the number of features with a feature weight of 0 is less than 2, then the current adjustment factor remains unchanged.

2. The method for predicting the service life of substation equipment according to claim 1, characterized in that, The target features include sensor features and non-sensor features; The step of determining the estimated lifespan of the device to be predicted based on the feature data of the target features corresponding to the device to be predicted includes: obtaining the embedding vector of the feature data of the target features corresponding to the device to be predicted based on the embedding layer of the lifespan prediction model, and constructing an embedding matrix based on the embedding vector of the feature data of the sensor features corresponding to the device to be predicted; wherein, one row of the embedding matrix is ​​the embedding vector of the feature data of any sensor feature corresponding to the device to be predicted. The encoding and decoding layer based on the lifetime prediction model encodes and decodes the embedding matrix to obtain the feature vector of the embedding matrix; The fusion layer based on the lifetime prediction model fuses the embedding vector of the feature data corresponding to the non-sensor features of the device to be predicted and the feature vector of the embedding matrix to obtain the fusion vector of the device to be predicted. The prediction layer based on the lifetime prediction model performs lifetime prediction on the fusion vector of the device to be predicted, and obtains the estimated lifetime of the device to be predicted.

3. The method for predicting the service life of substation equipment according to claim 2, characterized in that, The lifetime prediction model includes multiple encoding and decoding layers; the encoding and decoding layers based on the lifetime prediction model encode and decode the embedding matrix to obtain the feature vector of the embedding matrix, including: fusing each row of the encoding matrix output by the encoding module of the previous encoding and decoding layer based on a sliding time window to obtain a fused encoding matrix; The fused encoding matrix is ​​encoded based on the encoding module of the current encoding / decoding layer to obtain the encoding matrix output by the encoding module of the current encoding / decoding layer; The decoding module of the current codec layer decodes the decoding matrix output by the decoding module of the previous codec layer and the encoding matrix output by the encoding module of the current codec layer to obtain the decoding matrix output by the decoding module of the current codec layer. The decoding matrix output by the decoding module of the last encoding / decoding layer is vectorized to obtain the feature vector of the embedding matrix.

4. The method for predicting the service life of substation equipment according to claim 3, characterized in that, The method of fusing each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer based on a sliding time window to obtain a fused encoding matrix includes: for any row of the encoding matrix output by the encoding module of the previous encoding / decoding layer, converting the sub-vector within the current sliding time window into a sub-vector of length k based on a preset transformation matrix and concatenating it to the fused vector corresponding to the row; wherein, the initial value of the fused vector corresponding to the row is empty, the length of any row of the encoding matrix is ​​P×k, the length of the sliding time window is 2k, and the sliding step size of the sliding time window is k; the fused vector corresponding to each row of the encoding matrix output by the encoding module of the previous encoding / decoding layer constitutes the fused encoding matrix.

5. The method for predicting the service life of substation equipment according to claim 3, characterized in that, The step of encoding the fused coding matrix based on the current coding layer to obtain the coding matrix output by the current coding layer includes: based on the first self-attention unit of the current coding layer, obtaining the first attention vector corresponding to each row of the fused coding matrix, and constructing a first attention matrix based on the first attention vector corresponding to each row of the fused coding matrix; Based on the second self-attention unit of the encoding module of the current encoding / decoding layer, the second attention vector corresponding to each column of the first attention matrix is ​​obtained, and the encoding matrix is ​​constructed based on the second attention vector corresponding to each column of the first attention matrix.

6. The method for predicting the service life of substation equipment according to claim 3, characterized in that, The decoding of the decoding matrix output by the decoding module of the previous encoding / decoding layer and the encoding matrix output by the encoding module of the current encoding / decoding layer by the decoding module of the current encoding / decoding layer to obtain the decoding matrix output by the decoding module of the current encoding / decoding layer includes: obtaining the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the current encoding / decoding layer based on the third self-attention unit of the decoding module of the current encoding / decoding layer; and constructing a second attention matrix based on the third attention vector corresponding to each row of the decoding matrix output by the decoding module of the previous encoding / decoding layer. Based on the fourth self-attention unit of the decoding module of the current encoding and decoding layer, the fourth attention vector corresponding to each column of the second attention matrix is ​​obtained respectively, and the basic decoding matrix is ​​constructed based on the fourth attention vector corresponding to each column of the second attention matrix. Based on the cross-attention unit of the decoding module of the current encoding / decoding layer, attention transformation is performed on the basic decoding matrix and the encoding matrix output by the encoding module of the current encoding / decoding layer to obtain the decoding matrix.

7. A device for predicting the service life of substation equipment, characterized in that, include: The data acquisition unit is used to acquire feature data and expected lifespan of multi-dimensional features corresponding to sample substation equipment, and to construct a first set containing multi-dimensional features; The target feature filtering unit is used to perform a feature weight calculation step and an iteration step. The feature weight calculation step includes: placing the feature with the smallest feature weight in the first set into the second set and setting the deletion probability of the feature with the smallest feature weight to 0; incrementing the deletion probability of each feature in the second set by 1; deleting the feature with the smallest feature weight from the first set and updating the feature weights of each feature in the first set; wherein the feature weights of each feature are determined based on the feature data of the multi-dimensional features corresponding to the sample substation equipment and their expected lifespan. The iteration step includes: repeating the feature weight calculation step until the first set is empty; normalizing the deletion probability of each feature in the second set and determining the selection probability of each feature based on the normalized deletion probability of each feature in the second set; and determining the target feature based on the selection probability of each feature. The lifespan prediction unit is used to determine the estimated lifespan of the device to be predicted based on the feature data of the target features corresponding to the device to be predicted. The feature weights of each feature are determined based on the following steps: updating the λ parameter based on the sum of the current adjustment factor and the current value of the λ parameter of the Lasso regression model; Based on the feature data of multidimensional features corresponding to sample substation equipment and their expected lifespan, the feature weights of each feature are determined using the Lasso regression model. If the number of features with a feature weight of 0 is greater than or equal to 2, then the current adjustment factor is reduced. If the number of features with a feature weight of 0 is less than 2, then the current adjustment factor remains unchanged.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the substation equipment lifespan prediction method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the substation equipment life prediction method as described in any one of claims 1 to 6.

Images