Temperature prediction method and device for vehicle motor

Abstract

The invention relates to a temperature prediction method and device for a vehicle motor, and the method comprises the steps: obtaining the temperature data of the vehicle motor in a current time period, dividing the temperature data, and obtaining at least two temperature subsequences; performing feature dimension raising processing on each temperature sub-sequence by adopting a first feature extraction network to obtain dimension raising features of each temperature sub-sequence in different network layers in the first feature extraction network; and adopting a second feature extraction network to perform feature dimension reduction processing on each temperature sub-sequence to obtain dimension reduction features of each temperature sub-sequence in different network layers in the second feature extraction network, performing fusion processing on the dimension raising features and the dimension reducing features of the temperature subsequences to obtain temperature features corresponding to the temperature subsequences; and predicting the motor temperature of the vehicle motor in the next time period according to the temperature characteristics corresponding to the temperature subsequences. By adopting the method, the accuracy and reliability of a subsequent motor temperature prediction result can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor performance monitoring, and in particular to a temperature prediction method and device for a vehicle motor. BACKGROUND

[0002] With the popularity and development of electric vehicles worldwide, their performance and safety are of great importance. The motor of an electric vehicle, as a core component, its temperature directly affects the operating efficiency, service life and safety of the motor. The motor is affected by driving conditions, external temperature, driving speed and other factors when working, and the temperature will change rapidly, therefore, accurately predicting and real-time monitoring the motor temperature is the key to ensuring the safe operation of the motor.

[0003] At present, although the traditional temperature prediction method for a vehicle motor can also mine the complex relationship between the motor temperature and various factors based on machine learning or deep learning, there are still limitations in the feature extraction process, resulting in low accuracy of the prediction result of the motor temperature. SUMMARY

[0004] Therefore, it is necessary to provide a temperature prediction method and device for a vehicle motor to improve the accuracy and reliability of motor temperature prediction.

[0005] In a first aspect, the present application provides a temperature prediction method for a vehicle motor, the method comprising:

[0006] obtaining temperature data of the vehicle motor in a current time period, and dividing the temperature data to obtain at least two temperature subsequences;

[0007] using a first feature extraction network to perform feature dimensionality increasing processing on each temperature subsequence to obtain dimensionality increased features of each temperature subsequence at different network layers in the first feature extraction network; and

[0008] using a second feature extraction network to perform feature dimensionality reducing processing on each temperature subsequence to obtain dimensionality reduced features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same;

[0009] respectively performing fusion processing on the dimensionality increased features and the dimensionality reduced features of each temperature subsequence to obtain temperature features corresponding to each temperature subsequence;

[0010] predicting the motor temperature of the vehicle motor in a next time period according to the temperature features corresponding to each temperature subsequence.

[0011] In a second aspect, the present application further provides a temperature prediction device for a vehicle motor, the device comprising:

[0012] The data acquisition module is configured to acquire temperature data of the vehicle motor in a current time period, and divide the temperature data to obtain at least two temperature subsequences.

[0013] The feature extraction module is configured to perform feature dimension increasing processing on each temperature subsequence by using a first feature extraction network to obtain dimension-increased features of each temperature subsequence at different network layers in the first feature extraction network, and perform feature dimension decreasing processing on each temperature subsequence by using a second feature extraction network to obtain dimension-decreased features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same.

[0014] The feature determination module is configured to perform fusion processing on the dimension-increased features and the dimension-decreased features of each temperature subsequence respectively to obtain temperature features corresponding to each temperature subsequence.

[0015] The temperature prediction module is configured to predict the motor temperature of the vehicle motor in a next time period according to the temperature features corresponding to each temperature subsequence.

[0016] In a third aspect, the present application further provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0017] The temperature data of the vehicle motor in a current time period is acquired, and the temperature data is divided to obtain at least two temperature subsequences.

[0018] The feature dimension increasing processing is performed on each temperature subsequence by using a first feature extraction network to obtain dimension-increased features of each temperature subsequence at different network layers in the first feature extraction network; and

[0019] The feature dimension decreasing processing is performed on each temperature subsequence by using a second feature extraction network to obtain dimension-decreased features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same.

[0020] The fusion processing is performed on the dimension-increased features and the dimension-decreased features of each temperature subsequence respectively to obtain temperature features corresponding to each temperature subsequence.

[0021] The motor temperature of the vehicle motor in a next time period is predicted according to the temperature features corresponding to each temperature subsequence.

[0022] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0023] obtain temperature data of the vehicle motor in a current time period, and divide the temperature data to obtain at least two temperature subsequences;

[0024] perform feature dimension increasing processing on each temperature subsequence by using a first feature extraction network to obtain dimension-increased features of each temperature subsequence at different network layers in the first feature extraction network; and

[0025] perform feature dimension decreasing processing on each temperature subsequence by using a second feature extraction network to obtain dimension-decreased features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same.

[0026] perform fusion processing on the dimension-increased features and the dimension-decreased features of each temperature subsequence respectively to obtain temperature features corresponding to each temperature subsequence.

[0027] predict the motor temperature of the vehicle motor in a next time period according to the temperature features corresponding to each temperature subsequence.

[0028] In a fifth aspect, the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the following steps:

[0029] obtain temperature data of the vehicle motor in a current time period, and divide the temperature data to obtain at least two temperature subsequences;

[0030] perform feature dimension increasing processing on each temperature subsequence by using a first feature extraction network to obtain dimension-increased features of each temperature subsequence at different network layers in the first feature extraction network; and

[0031] perform feature dimension decreasing processing on each temperature subsequence by using a second feature extraction network to obtain dimension-decreased features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same.

[0032] perform fusion processing on the dimension-increased features and the dimension-decreased features of each temperature subsequence respectively to obtain temperature features corresponding to each temperature subsequence.

[0033] predict the motor temperature of the vehicle motor in a next time period according to the temperature features corresponding to each temperature subsequence.

[0034] The temperature prediction method and device of the vehicle motor, by dividing the temperature data of the vehicle motor in the current time period, introducing at least two temperature subsequences, facilitate subsequent more detailed capture of feature information in the temperature data. The first feature extraction network is used to perform feature dimensionality processing on each temperature subsequence to obtain the dimensionality increased features of each temperature subsequence in different network layers of the first feature extraction network, and the second feature extraction network is used to perform feature dimensionality reduction processing on each temperature subsequence to obtain the dimensionality reduced features of each temperature subsequence in different network layers of the second feature extraction network, and the temperature features obtained by fusing the dimensionality increased features and the dimensionality reduced features are used for motor temperature prediction, that is, the dimensionality increased features and the dimensionality reduced features are introduced at the same time, which enriches the diversity of data feature extraction, so that the dimensionality increased features and the dimensionality reduced features are comprehensively utilized, which can avoid the limitations brought by only considering single feature dimension extraction in the traditional feature extraction process, and thus the accuracy and reliability of motor temperature prediction can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 An application environment diagram of the temperature prediction method of the vehicle motor provided by some embodiments of the present application is provided;

[0036] Figure 2 A flowchart of the temperature prediction method of the vehicle motor provided by some embodiments of the present application is provided;

[0037] Figure 3 A structure diagram of the network model with the bilinear feature enhancement mechanism provided by some embodiments of the present application is provided;

[0038] Figure 4 A flowchart of determining the temperature features provided by some embodiments of the present application is provided;

[0039] Figure 5 A flowchart of determining the target fusion features provided by some embodiments of the present application is provided;

[0040] Figure 6 A flowchart of determining the dimensionality increased enhanced features and the dimensionality reduced enhanced features provided by some embodiments of the present application is provided;

[0041] Figure 7 A flowchart of determining the semantic deviation degree of the dimensionality increased enhanced features and the dimensionality reduced enhanced features provided by some embodiments of the present application is provided;

[0042] Figure 8 A structure diagram of the network model with the bilinear feature enhancement mechanism provided by some other embodiments of the present application is provided;

[0043] Figure 9 A flowchart of determining the target fusion features provided by some other embodiments of the present application is provided;

[0044] Figure 10A flowchart for determining feature distances provided for some embodiments of the present application;

[0045] Figure 11 A flowchart for determining initial fused features provided for some embodiments of the present application;

[0046] Figure 12 A flowchart for determining abnormal feature elements provided for some embodiments of the present application;

[0047] Figure 13 A structural block diagram of a temperature prediction device for a vehicle motor provided for some embodiments of the present application;

[0048] Figure 14 An internal structural diagram of a computer device provided for some embodiments of the present application. DETAILED DESCRIPTION

[0049] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0050] It should be noted that the terms "first", "second", etc. used in the present application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "include" and "have" used in the present application and any variations thereof are intended to cover non-exclusive inclusion. The term "multiple" used in the present application refers to two and more than two. The term "and / or" used in the present application refers to one of the options, or any combination of multiple options.

[0051] The temperature prediction method for a vehicle motor provided by the embodiments of the present application can be applied to, for example, Figure 1The application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data required by the server 104 to process. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. Among them, the terminal 102 can be but not limited to an electric vehicle equipped with an electric motor such as an electric vehicle, an electric bus, or an on-board electronic device on an electric vehicle. In this embodiment, the terminal 102 takes an electric vehicle as an example, which has a connected vehicle controller and vehicle motor. The vehicle controller is used to obtain real-time running data of the vehicle motor, and the real-time running data of the vehicle motor includes but is not limited to temperature data. The vehicle controller can accurately predict the motor temperature of the vehicle motor in the next period according to the vehicle motor temperature prediction method of the application. The vehicle motor can be a permanent magnet synchronous motor, or other types of motors, such as an alternating current asynchronous motor. The vehicle controller can be but not limited to a battery management system (Battery Management System, BMS), an electronic control unit (Electronic Control Unit, ECU), etc.

[0052] Optionally, in another embodiment, the terminal 102 can also send the obtained real-time running data of the vehicle motor to the server 104, so that the server 104 accurately predicts the motor temperature of the vehicle motor in the next period according to the vehicle motor temperature prediction method of the application. Specifically, the server first obtains the temperature data of the vehicle motor in the current period, and divides the temperature data to obtain at least two temperature subsequences; using a first feature extraction network, each temperature subsequence is processed to increase the dimension of the feature to obtain the dimension-increased feature of each temperature subsequence at different network layers in the first feature extraction network, and using a second feature extraction network, each temperature subsequence is processed to reduce the dimension of the feature to obtain the dimension-reduced feature of each temperature subsequence at different network layers in the second feature extraction network; then the dimension-increased feature and the dimension-reduced feature of each temperature subsequence are fused to obtain the temperature feature corresponding to each temperature subsequence; then, according to the temperature feature corresponding to each temperature subsequence, the motor temperature of the vehicle motor in the next period can be predicted. In some embodiments, the electric vehicle is equipped with a temperature sensor to obtain and monitor the temperature data of the vehicle motor in real time. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0053] In one embodiment, as Figure 2 shown, the method is applied to the server in Figure 1 It can be understood that the method can also be applied to the terminal, and can also be applied to a system including the terminal and the server, and is realized through the interaction of the terminal and the server. In this embodiment, the method includes the following steps:

[0054] S202, acquire temperature data of the vehicle motor in a current time period, and divide the temperature data to obtain at least two temperature subsequences.

[0055] The temperature data in the current time period is a set of temperature values of the vehicle motor at each time point in a set time period including the current time. The current time period can be determined according to the prediction period or the prediction frequency of the model. The temperature data of the vehicle motor can be collected in real time by a temperature sensor installed on the motor. In order to facilitate subsequent more detailed capture of feature information in the temperature data and improve the accuracy of motor temperature prediction, the temperature data is divided in this embodiment to obtain a plurality of continuous temperature subsequences, and independent feature extraction and analysis are performed on these temperature subsequences respectively, so as to decompose the complex temperature change process into a plurality of relatively simple sub-processes. The temperature subsequences obtained after division contain part of the temperature values.

[0056] For example, the temperature data of the vehicle motor in 24 hours is first acquired, which covers the temperature change of the vehicle motor under different working conditions, such as city road driving, highway driving, climbing, etc., and the sampling data is collected by a fixed sampling frequency (for example, sampling once per second). In the 24 hours, a total of 86400 temperature values can be collected. After the temperature values are divided according to certain rules, smaller data sequences are obtained. The division rule can be equal interval division according to time sequence, or division according to other standards.

[0057] In a possible implementation, the collected temperature data can be divided according to a preset time window size, for example, the temperature data is divided into a plurality of temperature subsequences X i , i = 1, 2, …, n, that is, each temperature subsequence has 400 temperature data. The main purpose of the division of the time window size is to improve the efficiency of the model, capture local time sequence patterns, and reduce data redundancy and computational complexity. Considering that too large time window size will correspondingly lead to too large feature dimension in the dimension increasing and dimension reduction process, and too small time window size will correspondingly lead to too small feature dimension in the dimension increasing and dimension reduction process, that is, too large or too small time window size will affect feature extraction, therefore, based on data characteristics and empirical verification, the size of the time window is preferably 400. In addition, if the total number of temperature values cannot be divided by 400, for example, only 86390 data points are collected, then the last remaining temperature subsequence with less than 400 data points can be discarded.

[0058] S204, a first feature extraction network is used to perform feature dimensionality upscaling on each temperature subsequence to obtain the dimensionality upscaling features of each temperature subsequence in different network layers of the first feature extraction network, and a second feature extraction network is used to perform feature dimensionality reduction on each temperature subsequence to obtain the dimensionality reduction features of each temperature subsequence in different network layers of the second feature extraction network.

[0059] In this embodiment, the first and second feature extraction networks have the same number of network layers. The first feature extraction network is a dimensionality-upgrading convolutional neural network model, whose main function is to perform dimensionality-upgrading processing on the input temperature subsequence. Dimensionality-upgrading means mapping the temperature data to a higher-dimensional feature space, allowing the latent features in the temperature data to be fully mined and represented, such as temperature change trends and periodicity. That is, the dimensionality-upgrading features in this embodiment are feature representations after dimensionality-upgrading processing by the first feature extraction network, which can more comprehensively describe the features of the temperature data.

[0060] The second feature extraction network is a dimensionality-reducing convolutional neural network model, whose function is to perform feature dimensionality reduction processing on the input temperature subsequence. The main purpose of feature dimensionality reduction processing is to remove redundant information in the temperature data and extract the main features of the temperature data, thereby simplifying the representation of the temperature data. That is, the feature dimensionality reduction processing in this embodiment transforms the temperature subsequence from a high-dimensional feature space to a low-dimensional feature space. In other words, the dimensionality-reduced features in this embodiment are feature representations after dimensionality reduction processing by the second feature extraction network, and the dimensionality-reduced features can retain the main features of the temperature data. In other words, this embodiment mainly uses a bilinear one-dimensional convolutional neural network (1dCNN) to extract features from the data of each time window, i.e., the temperature subsequence, and uses the first feature extraction network and the second feature extraction network to mine the feature information of the temperature data from different perspectives, which helps to improve the accuracy of subsequent predictions.

[0061] In one possible implementation, the first and second feature extraction networks can have 3 to 5 layers. However, considering the large amount of temperature data and the efficiency requirements of model training, this embodiment preferably uses 3 layers. This also avoids overfitting. Specifically, as shown... Figure 3 As shown, Figure 3 This is a network model with a bilinear feature enhancement mechanism. In the up-dimensional convolutional neural network, the number of convolutional features is doubled after each convolution calculation; in the down-dimensional convolutional neural network, the number of convolutional features is divided by 2 after each convolution calculation. The up-dimensionality calculation process in this embodiment is as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] in, This represents a higher-dimensional convolution operation; Indicates the first One time window; Representing the A temperature subsequence; Representing temperature subsequence The increased-dimensional features obtained after the first increased-dimensional convolution, taking a time window size of 400 as an example, are the temperature subsequence. After the first dimensionality-increasing convolution, the size of the dimensionality-increasing feature becomes 800, which is twice the initial feature dimension. The increased-dimensional features obtained after the second increased-dimensional convolution are the input features. Twice as much; The increased-dimensional features obtained after the third increased-dimensional convolution are the input features. Twice as much.

[0066] The dimensionality reduction calculation process in this embodiment is as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] in, This represents a dimension reduction convolution operation; Representing temperature subsequence The dimensionality-reduced features obtained after the first dimensionality-reducing convolution, taking a time window size of 400 as an example, are the temperature subsequence. After the first dimensionality reduction convolution, the feature size becomes 200, which is half the initial feature dimension; The reduced-dimensionality features obtained after the second dimensionality reduction convolution are the input features. 1 / 2 times; The reduced-dimensionality features obtained after the third dimensionality reduction convolution are the input features. 1 / 2 times.

[0071] After the temperature subsequence passes the dimension-increasing convolution operation of the first feature extraction network and the dimension-decreasing convolution operation of the second feature extraction network, the dimension-increasing features of the temperature subsequence at the network layers of the first level become 800, and the dimension-decreasing features become 200; the dimension-increasing features of the temperature subsequence at the network layers of the second level become 1600, and the dimension-decreasing features become 100; the dimension-increasing features of the subsequence at the network layers of the third level become 3200, and the dimension-decreasing features become 50.

[0072] In S206, the dimension-increasing features and the dimension-decreasing features of each temperature subsequence are fused to obtain temperature features corresponding to each temperature subsequence.

[0073] The temperature features are features obtained by fusing the dimension-increasing features and the dimension-decreasing features of each temperature subsequence. The temperature features comprehensively utilize the feature information of each level and each dimension, and also comprehensively utilize the rich information of high-dimensional features and the simplicity of low-dimensional features, thereby better describing the change of the temperature of the vehicle motor and providing strong support for subsequent accurate prediction of the motor temperature.

[0074] Specifically, there are various ways to fuse the dimension-increasing features and the dimension-decreasing features of each temperature subsequence, for example, the fusion can be based on simple splicing, that is, the dimension-increasing features and the dimension-decreasing features are directly spliced together to form a longer feature vector, or the fusion can be based on linear combination, that is, different weights are assigned to the dimension-increasing features and the dimension-decreasing features, and then the weighted dimension-increasing features and the dimension-decreasing features are spliced together. In this embodiment, the dimension-increasing features and the dimension-decreasing features are directly spliced together as an example. When fusing the dimension-increasing features and the dimension-decreasing features of each temperature subsequence, the dimension-increasing features and the dimension-decreasing features of the network layers at the same level can be fused first to obtain fused features, and then the fused features are fused; or the dimension-increasing features and the dimension-decreasing features of the network layers at different levels can be fused first to obtain fused features, and then the fused features are fused to obtain the temperature features corresponding to each temperature subsequence.

[0075] In S208, the motor temperature of the vehicle motor in the next time period is predicted according to the temperature features corresponding to each temperature subsequence.

[0076] The next time period is a future time period immediately after the current time period, which can be a time period of a preset duration or the next time. There are various ways to predict the motor temperature of the vehicle motor in the next time period according to the temperature features, including but not limited to autoregressive integrated moving average model, exponential smoothing method, long short-term memory network, gated recurrent unit, etc., which can be selected according to actual use requirements.

[0077] Exemplarily, after obtaining the temperature features corresponding to each temperature subsequence, the temperature features are input into a Long Short-Term Memory (LSTM) network, time sequence information is extracted through the Long Short-Term Memory network, and finally temperature data of the vehicle motor within 1 min or at the next moment is obtained, so as to provide a reference for management and maintenance of the vehicle motor.

[0078] The temperature prediction method of the vehicle motor can improve the accuracy and reliability of the predicted temperature result of the motor.

[0079] In one embodiment, on the basis of the above-mentioned embodiment, the fusion processing of the dimension-increasing features and the dimension-reducing features of each temperature subsequence is further performed, and the process of obtaining the temperature features corresponding to each temperature subsequence is described in detail. As shown in Figure 4 The specific steps include the following steps:

[0080] S402, for each temperature subsequence, the dimension-increasing features and the dimension-reducing features of the temperature subsequence at the network layer of the same level are fused to obtain target fusion features of the temperature subsequence at different levels.

[0081] The target fusion features are features obtained by fusing the dimension-increasing features and the dimension-reducing features in the network layer of the same level. This fusion combines the advantages of the dimension-increasing features and the dimension-reducing features, retains the rich detailed information in the dimension-increasing features, and utilizes the key information after removing noise and redundancy in the dimension-reducing features.

[0082] Optionally, when the dimension-increasing features and the dimension-reducing features of the temperature subsequence at the network layer of the same level are fused, the feature elements in the dimension-increasing features and the feature elements in the dimension-reducing features can be directly spliced together to form a longer feature vector, or the fusion can be based on linear combination, that is, different weights are assigned to the dimension-increasing features and the dimension-reducing features, and then the weighted dimension-increasing features and the weighted dimension-reducing features are spliced together to obtain the target fusion features of the temperature subsequence at the corresponding level.

[0083] Exemplarily, the dimension-increasing features of the temperature subsequence X1 at the network layer of the first level of the first feature extraction network and the dimension-reducing features of the temperature subsequence X1 at the network layer of the first level of the second feature extraction network are fused to obtain the target fusion features of the temperature subsequence X1 at the first level. The dimension-increasing features of the temperature subsequence X1 at the network layer of the second level of the first feature extraction network and the dimension-reducing features of the temperature subsequence X1 at the network layer of the second level of the second feature extraction network are fused to obtain the target fusion features of the temperature subsequence X1 at the second level. The dimension-increasing features of the temperature subsequence X1 at the network layer of the third level of the first feature extraction network and the reduced dimension features of the network layer at the third level of the second feature extraction network After fusion, the target fusion feature of the temperature subsequence X1 at the third level can be obtained.

[0084] wherein, if the temperature subsequence X1 is in the network layer at the first level, the increased dimension feature The short-term fluctuation details of the temperature, such as the slight temperature change per minute, are extracted, while the reduced dimension feature The long-term trend of the temperature, such as the average temperature change per day, is captured, and after the fusion of the two features, the target fusion feature obtained contains both the short-term fluctuation information and the long-term trend information, that is, the target fusion feature can more accurately describe the characteristics of the temperature subsequence at this level.

[0085] Optionally, in another possible implementation, the increased dimension features and the reduced dimension features of the temperature subsequence at the network layers of different levels can be fused to obtain fusion features, and then the fusion features are fused to obtain the temperature feature corresponding to the temperature subsequence. In actual application, the fusion mode of the increased dimension features and the reduced dimension features of the temperature subsequence can be selected according to specific task requirements and data characteristics.

[0086] S404, fusing the target fusion features of the temperature subsequence at different levels to obtain the temperature feature corresponding to the temperature subsequence.

[0087] Specifically, after obtaining the target fusion features of the temperature subsequence at each level, the target fusion features are fused to obtain the final temperature feature. In this way, different dimension and level features can be mixed and processed to mine the deep relationship between different dimension features, so that the model can better utilize these different dimension features for prediction, and the limitations of single level features can be reduced to improve the robustness and stability of the features.

[0088] For example, the target fusion features of the above temperature subsequence X1 obtained at three levels are fused to obtain the final temperature feature of the temperature subsequence X1.

[0089] In this embodiment, by first fusing the increased dimension features and the reduced dimension features of the temperature subsequence at the network layers of the same level, and then fusing the target fusion features of the temperature subsequence at different levels, the consistency of the fused features in semantics can be ensured, and the complexity of calculation and resource consumption can be reduced, which is more convenient for the model to understand and process the fused features.

[0090] In one embodiment, on the basis of the above-mentioned embodiment, the process of further fusing the dimension-increased features and the dimension-reduced features of the temperature subsequence at the network layer of the same level to obtain the target fused features of the temperature subsequence at different levels is described in detail. As shown in Figure 5 The process specifically includes the following steps:

[0091] S502, for each level, performing semantic enhancement processing on the dimension-increased features and the dimension-reduced features of the temperature subsequence at the network layer of the level to obtain dimension-increased enhanced features and dimension-reduced enhanced features.

[0092] Since the original dimension-increased features and the dimension-reduced features may contain some unnecessary information or noise, the semantic enhancement processing can filter and strengthen these features, highlighting the information related to the key semantics such as temperature change patterns, trends, etc., to help the subsequent model better learn and identify the features of the temperature subsequence and improve the performance of the model. There are various ways of semantic enhancement processing, such as feature weighting, feature transformation, etc.

[0093] Optionally, for each level, the dimension-increased features and the dimension-reduced features of the temperature subsequence at the network layer of the level can be processed based on the feature transformation method to obtain the dimension-increased enhanced features and the dimension-reduced enhanced features. The dimension-increased enhanced features are the features obtained after the semantic enhancement processing of the dimension-increased features, containing more valuable semantic information. Similarly, the dimension-reduced enhanced features are the results after the semantic enhancement processing of the dimension-reduced features.

[0094] For example, assume that the dimension-increased features are the multi-scale features of the temperature subsequence obtained by wavelet transform, which contain components of different frequencies. After semantic enhancement processing, the specific frequency components related to motor failure can be weighted to make these key components more prominent in the features, thereby obtaining the dimension-increased enhanced features. For the dimension-reduced features, such as the principal component features obtained by the PCA (Principal Component Analysis) method, the semantic enhancement processing can remove some principal components generated by data noise and retain the principal components that truly reflect the temperature change trend, thereby forming the dimension-reduced enhanced features.

[0095] S504, fusing the dimension-increased enhanced features and the dimension-reduced enhanced features to obtain the target fused features of the temperature subsequence at the level.

[0096] In an alternative embodiment, the dimension-increasing enhanced features and the dimension-decreasing enhanced features can be directly spliced according to a preset fusion order to obtain the target fusion features of the temperature subsequence at the level. Since the enhanced features obtained through dimension-increasing and dimension-decreasing have different dimensions and characteristics, there may be some outliers that affect the performance of the model. Therefore, in another alternative embodiment, the outliers in the dimension-increasing enhanced features and the dimension-decreasing enhanced features can be removed first, and then the dimension-increasing enhanced features and the dimension-decreasing enhanced features with the outliers removed are fused to obtain the target fusion features of the temperature subsequence at the level.

[0097] In this embodiment, by first enhancing the dimension-increasing features and the dimension-decreasing features of the network layers of the temperature subsequence at the same level and then fusing them, the discrimination and expression ability of the features can be improved, the features are more representative, and thus the subsequent model can better learn and recognize the features of the temperature subsequence, improving the performance of the model.

[0098] In one embodiment, on the basis of the above-mentioned embodiment, the process of further performing semantic enhancement processing on the dimension-increasing features and the dimension-decreasing features of the network layers of the temperature subsequence at the level to obtain the dimension-increasing enhanced features and the dimension-decreasing enhanced features is described in detail. As shown in Figure 6 The specific steps include the following steps:

[0099] S602, determine the feature deviation between the dimension-increasing features and the dimension-decreasing features, which is a first index value for representing the fluctuation between different feature elements in the dimension-increasing features, and a second index value for representing the fluctuation between different feature elements in the dimension-decreasing features.

[0100] The feature deviation is also an inter-dimension average difference, which is used to represent the mean difference between the dimension-increasing features and the dimension-decreasing features at the same level in the feature space, and reflects the overall semantic deviation under different feature extraction modes. The first index value is used to represent the fluctuation between different feature elements in the dimension-increasing features, i.e., the dispersion degree in the feature space. The first index value can help determine the range and degree of semantic transformation of the dimension-increasing features, so that the enhanced dimension-increasing features can better retain semantic information while introducing appropriate changes. The second index value is used to represent the fluctuation between different feature elements in the dimension-decreasing features, i.e., the dispersion degree in the feature space. The second index value can help determine the range and degree of semantic transformation of the dimension-decreasing features, so that the enhanced dimension-decreasing features can better retain semantic characteristics while increasing a certain diversity.

[0101] Optionally, the average feature of the high-dimensional feature and the average feature of the low-dimensional feature can be obtained based on the calculation method of the feature mean value, and the feature deviation between the high-dimensional feature and the low-dimensional feature can be determined based on the average feature of the high-dimensional feature and the average feature of the low-dimensional feature. In addition, the first index value for representing the fluctuation between different feature elements in the high-dimensional feature and the second index value for representing the fluctuation between different feature elements in the low-dimensional feature can be obtained based on the calculation method of the feature standard deviation and in combination with the average feature of the high-dimensional feature and the average feature of the low-dimensional feature.

[0102] In S604, the distribution function composed of the feature deviation and the first index value is used to perform semantic enhancement processing on the high-dimensional feature to obtain a high-dimensional enhanced feature.

[0103] Optionally, the distribution function for performing semantic enhancement processing on the high-dimensional feature can be constructed based on the normal distribution law and the feature deviation and the first index value, and the corresponding high-dimensional enhanced feature can be obtained by inputting the high-dimensional feature into the distribution function.

[0104] It can be understood that, since the distribution function combines the feature deviation and the first index value, the semantic transformation range can be controlled while the stability of the high-dimensional feature is preserved. Furthermore, the high-dimensional enhanced feature obtained by performing semantic enhancement processing on the high-dimensional feature according to the distribution function will not deviate too far from the semantics of the original high-dimensional feature and still has the basic semantic structure and stability of the original high-dimensional feature.

[0105] In S606, the distribution function composed of the feature deviation and the second index value is used to perform semantic enhancement processing on the low-dimensional feature to obtain a low-dimensional enhanced feature.

[0106] Similarly, the distribution function for performing semantic enhancement processing on the low-dimensional feature can be constructed based on the normal distribution law and the feature deviation and the second index value, and the corresponding low-dimensional enhanced feature can be obtained by inputting the low-dimensional feature into the distribution function.

[0107] In this embodiment, the distribution function composed of the feature deviation and the first index value is used to perform semantic enhancement on the high-dimensional feature, and the distribution function composed of the feature deviation and the second index value is used to perform semantic enhancement on the low-dimensional feature, which can improve the expression ability and robustness of the features.

[0108] In one embodiment, the process of determining the feature deviation between the high-dimensional feature and the low-dimensional feature, the first index value for representing the fluctuation between different feature elements in the high-dimensional feature, and the second index value for representing the fluctuation between different feature elements in the low-dimensional feature is further described in detail based on the above-described embodiments. As shown in Figure 7 The specific steps include the following steps:

[0109] S702, perform feature averaging on each feature element in the dimensionality-upgrading feature to obtain the first average feature, and perform feature averaging on each feature element in the dimensionality-reducing feature to obtain the second average feature.

[0110] Optionally, based on the method of calculating the feature mean, the sum of each feature element in the increased-dimensional feature is first calculated, and then divided by the total number of feature elements in the increased-dimensional feature to obtain the first average feature. Similarly, based on the method of calculating the feature mean, the sum of each feature element in the decreased-dimensional feature is first calculated, and then divided by the total number of feature elements in the decreased-dimensional feature to obtain the second average feature.

[0111] Specifically, assuming dimensionality enhancement features It can be represented as { The dimensionality reduction feature can be represented as { },in Indicates the first In this embodiment, there are several network layers. =1,2,3; Indicates the first A temperature subsequence. The first average feature of the upgraded dimensionality feature can be obtained according to the following formula:

[0112] ;

[0113] in, It is in the The temperature subsequence at the ... The average feature of the upgraded features of each network layer, i.e., the first average feature; The number of feature elements in the upgraded feature; It is the first The temperature subsequence at the ... The first of the dimensionality-upgrading features of the network layer Each feature element.

[0114] In addition, the second average characteristic of the cooling feature can be calculated using the following formula:

[0115] ;

[0116] in, It is in the The temperature subsequence at the ... The average feature of the dimensionality reduction features of each network layer, i.e., the second average feature; The number of feature elements in the dimensionality reduction feature; It is the first The temperature subsequence at the ... The dimensionality reduction feature of the nth network layer Each feature element.

[0117] S704, determine the feature deviation between the dimensionality-upgrading feature and the dimensionality-reducing feature based on the difference between the first average feature and the second average feature.

[0118] Optionally, the difference between the first average feature and the second average feature can be defined as the feature deviation between the dimensionality-upgrading feature and the dimensionality-reducing feature.

[0119] Specifically, the characteristic deviation can be calculated using the following formula:

[0120] ;

[0121] in, It is the first A temperature subsequence in First average characteristic Second average characteristics The characteristic deviation between them.

[0122] S706. Based on the distance between each feature element in the upgraded feature and the first average feature, determine the standard deviation of the upgraded feature, and use it as the first index value to characterize the fluctuation between different feature elements in the upgraded feature.

[0123] Optionally, the difference between each feature element in the upgraded feature and the first average feature can be calculated first. Based on the standard deviation calculation method and the difference between each feature element and the first average feature, a first index value is obtained to characterize the fluctuation between different feature elements in the upgraded feature.

[0124] Specifically, the first indicator value can be calculated using the following formula:

[0125] = ;

[0126] in, For the first A temperature subsequence in The first indicator value.

[0127] S708. Based on the distance between each feature element in the dimensionality reduction feature and the second average feature, determine the standard deviation of the dimensionality reduction feature as a second index value used to characterize the fluctuation between different feature elements in the dimensionality reduction feature.

[0128] Specifically, the second indicator value can be calculated using the following formula:

[0129] = ;

[0130] in, For the first A temperature subsequence in The second indicator value.

[0131] For example, suppose that for =1, The temperature subsequence with value 2, the upgraded feature corresponding to the temperature subsequence of the vehicle motor. The sample set is {30, 32, 35, 33, 34}, that is... =30, =32, =35; =33; =34, then calculate the corresponding first average feature. = (30 + 32 + 35 + 33 + 34) / 5 = 32.8; Calculate the corresponding first indicator value. ≈1.72.

[0132] In this embodiment, by determining the feature deviation based on the first average feature of the dimensionality-upgrading feature and the second average feature of the dimensionality-reducing feature, and by determining the first index value and the second index value based on the feature deviation, it helps to ensure the accuracy and reliability of subsequent semantic enhancement processing of the dimensionality-upgrading feature and the dimensionality-reducing feature.

[0133] In one possible implementation, semantic enhancement processing can be performed on both dimensionality-upgrading and dimensionality-reducing features using a normal distribution. Specifically, for example... Figure 8 As shown, Figure 8 This is another network model with a bilinear feature enhancement mechanism. When performing semantic enhancement on the upgraded features, it first processes the upgraded features... Along the multivariate distribution N ( , The direction is randomly transformed N times to generate enhanced features. Then, the average value of the N enhanced features is calculated to obtain the final dimension-upgraded enhanced features. The calculation process is as follows:

[0134] ;

[0135] in, This is the final dimensionality enhancement feature of the temperature subsequence; As a feature of dimensional elevation The dimensionality-upgraded features after each semantic enhancement.

[0136] Similarly, when performing semantic enhancement on dimensionality-reduced features, first process the dimensionality-reduced features... Along the multivariate distribution N ( , The semantic transformation is performed N times randomly in the direction to generate the corresponding dimensionality reduction and enhancement features. wherein the number of semantic changes is set according to the requirement of model training, and it is necessary to consider avoiding generating too much redundant features and reducing the influence on feature extraction; then the average value of the N enhanced features is calculated to obtain the final reduced enhanced features, and the calculation process is as follows:

[0137]

[0138] wherein, is the final reduced enhanced feature of the temperature subsequence; is the reduced feature after each semantic enhancement.

[0139] For example, the number of times of semantic change processing of the dimensionality increased features and the dimensionality reduced features in the embodiment is preferably 5 times. Specifically, the dimensionality increased features of the temperature subsequence in the state of =3, =1 are subjected to N=5 times of semantic enhancement transformation. Taking the first semantic transformation as an example, the dimensionality increased features of the first temperature subsequence are subjected to multivariate normal distribution N( , ) to obtain the corresponding dimensionality increased enhanced features . Assuming that there is only one feature element =0.5, =0.1, =0.01 in the dimensionality increased features , according to the characteristics of the multivariate normal distribution, the feature element of the generated dimensionality increased enhanced features may be =0.5+0.1+ , wherein, is a value randomly generated according to the normal distribution of , , is a value randomly generated according to the normal distribution of =0.1, such as =0.05, then the generated dimensionality increased enhanced features is 0.65. According to this way, four times of random sampling are performed to obtain =0.62, =0.68, =0.63, =0.66, and the average value of the enhanced features after five times of semantic transformation is calculated as = (0.65+0.62+0.68+0.63+0.66) / 5=0.648, and then the final dimensionality increased enhanced features ​0.648. Similarly, if there are multiple feature elements in the dimensionality-increased feature, semantic enhancement is performed on each feature element in the above manner to obtain the final dimensionality-increased enhanced feature .

[0140] Similarly, the temperature subsequence at t = 3 is dimensionality-reduced to a dimensionality-reduced feature = 1. The dimensionality-reduced feature undergoes N = 5 times of semantic enhancement transformation, each time the dimensionality-reduced feature obeys a multivariate normal distribution N ( , ), and a dimensionality-reduced enhanced feature = 0.27, = 0.26, = 0.28, = 0.25, = 0.27 is obtained. The final dimensionality-reduced enhanced feature of the dimensionality-reduced feature is (0.27 + 0.26 + 0.28 + 0.25 + 0.27) / 5 = 0.266.

[0141] In this embodiment, the semantic change is performed by normal distribution, and the final enhanced feature can increase the diversity and richness of the data, so that the model can learn more different feature representations, thereby improving the generalization ability and robustness of the model.

[0142] In another possible implementation, when performing semantic enhancement processing on the dimensionality-increased feature and the dimensionality-reduced feature of the temperature subsequence, in order to improve the model training efficiency, semantic enhancement processing can be performed on each level of the dimensionality-increased feature and the dimensionality-reduced feature, or only on the dimensionality-increased feature of the network layer corresponding to the highest dimension and the dimensionality-reduced feature of the network layer corresponding to the lowest dimension. For example, when the number of network layers of the first feature extraction network and the second feature extraction network is three, only the dimensionality-increased feature of the third network layer and the dimensionality-reduced feature of the third network layer are subjected to semantic enhancement processing.

[0143] In one embodiment, on the basis of the above embodiment, the dimensionality-increased enhanced feature and the dimensionality-reduced enhanced feature are further fused to obtain the target fused feature of the temperature subsequence at the level. As shown in Figure 9 , the process specifically includes the following steps:

[0144] S902, determining the feature distance between the dimensionality-increased enhanced feature and the dimensionality-reduced enhanced feature.

[0145] Optionally, the feature distance between the enhanced high-dimensional feature and the enhanced low-dimensional feature can be determined based on a calculation formula of the Euclidean distance and each feature element in the enhanced high-dimensional feature and each feature element in the enhanced low-dimensional feature. In addition, the manner of determining the feature distance between the enhanced high-dimensional feature and the enhanced low-dimensional feature also includes but is not limited to Manhattan distance, cosine similarity, and the like. The specific calculation method can be determined according to the model training requirement or precision.

[0146] Since the enhanced high-dimensional feature contains more detailed information and local features, it is helpful to capture subtle changes in the data, while the enhanced low-dimensional feature may focus more on global information and key features, and can summarize the main trend of the data. Through the feature distance, the difference between the two features can be quantified, so that the features of different properties can be reasonably adjusted and integrated during feature fusion. The feature distance is used to represent the difference and relevance between the enhanced high-dimensional feature and the enhanced low-dimensional feature, and reflects the separation degree of the two features of different dimensions in the feature space. Moreover, the greater the feature distance, the greater the difference between the enhanced high-dimensional feature and the enhanced low-dimensional feature, and the weaker the relevance; on the contrary, the smaller the feature distance, the smaller the difference between the enhanced high-dimensional feature and the enhanced low-dimensional feature, and the stronger the relevance.

[0147] S904, the enhanced high-dimensional feature and the enhanced low-dimensional feature are spliced to obtain an initial fusion feature of the temperature subsequence at the level.

[0148] Optionally, when the enhanced high-dimensional feature and the enhanced low-dimensional feature are spliced, they can be spliced together in a certain order, for example, the feature elements of the enhanced high-dimensional feature are in front and the feature elements of the enhanced low-dimensional feature are in back. The enhanced feature obtained after splicing is the initial fusion feature, which can retain more information.

[0149] S906, the initial fusion feature of the temperature subsequence at the level is updated using the feature distance to obtain a target fusion feature of the temperature subsequence at the level.

[0150] Optionally, the feature weight can be determined according to the feature distance, and then the feature weight is multiplied by each feature element in the initial fusion feature to update the initial fusion feature, thereby obtaining the target fusion feature of the temperature subsequence at the corresponding level.

[0151] It can be understood that by updating the initial fusion feature through the feature distance, the weight of feature fusion can be automatically adjusted according to the distance and relevance between the features, so that the fused feature is more reasonable and effective, and the features with close distance can complement and strengthen each other during fusion, thereby improving the expression ability of temperature change.

[0152] The target fusion feature refers to a feature obtained by updating the initial fusion feature according to the feature distance. The target fusion feature comprehensively considers the information of the dimension-increased enhanced feature and the dimension-reduced enhanced feature, and can more comprehensively and accurately reflect the characteristics of the data.

[0153] In this embodiment, the initial fusion feature is updated according to the feature distance, so that the target fusion feature obtained finally can more comprehensively and accurately reflect the characteristics of the data, thereby helping to ensure the effect of model training and prediction.

[0154] In one embodiment, on the basis of the above embodiment, the process of determining the feature distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature is further described. As shown in the following figure, the process specifically includes the following steps: Figure 10

[0155] S1002, average features in the dimension-increased enhanced feature to obtain a third average feature, and average features in the dimension-reduced enhanced feature to obtain a fourth average feature.

[0156] Optionally, the features in the dimension-increased enhanced feature can be averaged to obtain the third average feature based on the calculation method of the feature mean. Similarly, the features in the dimension-reduced enhanced feature can be averaged to obtain the fourth average feature based on the calculation method of the feature mean.

[0157] S1004, determining the feature distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature according to the third average feature and the fourth average feature.

[0158] Optionally, the feature distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature can be obtained based on the calculation formula of the difference between the third average feature and the fourth average feature and the feature standard deviation.

[0159] It can be understood that the distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature can reflect the similarity or difference between the two features. If the distance between the two features is small, it means that the correlation between them is strong, and the fusion coefficient can be large; on the contrary, if the distance is large, the fusion coefficient can be small.

[0160] Specifically, the feature distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature is calculated according to the following calculation formula:

[0161] ;

[0162] wherein, represents the total number of feature elements of the dimension-reduced enhanced feature in the i th network layer of the i th temperature sub-sequence. It can be understood that the distance between the dimension-increased enhanced feature and the dimension-reduced enhanced feature can reflect the similarity or difference between the two features. If the distance between the two features is small, it means that the correlation between them is strong, and the fusion coefficient can be large; on the contrary, if the distance is large, the fusion coefficient can be small.​​​ is a dimension-reduced enhanced feature in the feature element; is a fourth average feature of the dimension-reduced enhanced feature of the jth temperature sub-sequence at the ith network layer; is a fourth average feature of the dimension-reduced enhanced feature of the jth temperature sub-sequence at the ith network layer; is a total number of feature elements in the dimension-increased enhanced feature of the jth temperature sub-sequence at the ith network layer; is a total number of feature elements in the dimension-increased enhanced feature of the jth temperature sub-sequence at the ith network layer; is a fourth average feature of the dimension-reduced enhanced feature in the feature element; is a third average feature of the dimension-increased enhanced feature of the jth temperature sub-sequence at the ith network layer; is a third average feature of the dimension-increased enhanced feature of the jth temperature sub-sequence at the ith network layer; represents a Hilbert space used for calculating the distribution distance.

[0163] For example, the fourth average feature of the dimension-reduced enhanced feature is 10.5, and the third average feature of the dimension-increased enhanced feature is 10.1, then In addition, in order to accelerate the convergence speed of the model, the calculated feature distance can be normalized to limit the value of the feature distance to (0, 1), and the normalized feature distance is the bilinear feature fusion coefficient. The existence of the bilinear feature fusion coefficient can facilitate subsequent feature splicing and product operations to realize the fusion of features of different dimensions and automatically adjust the weight of feature fusion, thereby improving the practicability and reliability of the model. Assuming that the maximum value of the distance is , and the minimum value is , then the normalized is still 0.16.

[0164] In one embodiment, on the basis of the above embodiment, the dimension-increased enhanced feature and the dimension-reduced enhanced feature are spliced to obtain the initial fusion feature of the temperature sub-sequence at the layer level. As shown in Figure 11 , the process specifically includes the following steps:

[0165] S1102, based on the perception network, perceiving the dimension-increased enhanced feature and the dimension-reduced enhanced feature to obtain abnormal feature elements in the dimension-increased enhanced feature and the dimension-reduced enhanced feature.

[0166] Optionally, the abnormal feature elements can be identified based on a perception network such as an autoencoder. Specifically, the dimension-increased enhanced feature and the dimension-decreased enhanced feature are input into the autoencoder for training. The goal of the autoencoder is to reconstruct the input data as accurately as possible, i.e., the input data is encoded into a low-dimensional representation by the encoder, and then reconstructed into data similar to the input by the decoder. After training, the dimension-increased enhanced feature and the dimension-decreased enhanced feature are input into the autoencoder again, and the reconstruction error of each feature element is calculated. The reconstruction error refers to the difference between the input data and the reconstructed data, which is usually measured by mean square error (MSE). Finally, a threshold of the reconstruction error is set, and when the reconstruction error of a certain feature element exceeds the threshold, it is determined as an abnormal feature element.

[0167] It can be understood that the perception network is a neural network model with feature recognition and analysis capabilities. The perception network can also be a convolutional neural network (CNN), a recurrent neural network (RNN), a long short-term memory network, etc. based on deep learning. The perception network can identify abnormal feature elements caused by noise in the data collection process, sensor failure, data transmission error, etc.

[0168] S1104, the abnormal feature elements in the dimension-increased enhanced feature and the dimension-decreased enhanced feature are removed to obtain the optimized dimension-increased enhanced feature and the optimized dimension-decreased enhanced feature.

[0169] Optionally, after the abnormal feature elements in the dimension-increased enhanced feature are deleted, the optimized dimension-increased enhanced feature is obtained. Similarly, after the abnormal feature elements in the dimension-decreased enhanced feature are deleted, the optimized dimension-decreased enhanced feature is obtained.

[0170] For example, the dimension-increased enhanced feature The original feature has 800 feature elements, and after deleting the 3 abnormal feature elements, there are still 797 feature elements left. The feature composed of the 797 feature elements is the optimized dimension-increased enhanced feature , which can be represented as .

[0171] S1106, the optimized dimension-increased enhanced feature and the optimized dimension-decreased enhanced feature are spliced to obtain the initial fusion feature of the temperature subsequence at the level.

[0172] Optionally, after the optimized dimension-increased enhanced feature and the optimized dimension-decreased enhanced feature are spliced in order, the initial fusion feature of the temperature subsequence at the corresponding level is obtained.

[0173] In this embodiment, by removing abnormal feature elements, noise and error information in the data can be removed, so that the optimized enhanced features in dimension increasing and dimension decreasing are more accurate and reliable. This helps to improve the training effect and prediction accuracy of the subsequent model, and avoids misleading the model by abnormal data.

[0174] In one embodiment, based on the above embodiment, the process of perceiving the enhanced features in dimension increasing and dimension decreasing based on the perception network is further described. As shown in Figure 12 , it specifically includes the following steps:

[0175] S1202, the enhanced features in dimension increasing and dimension decreasing are spliced to obtain spliced features.

[0176] Optionally, after splicing the feature elements of the enhanced features in dimension increasing and dimension decreasing in order, the spliced features can be obtained.

[0177] For example, the enhanced features in dimension increasing and the enhanced features in dimension decreasing are spliced to obtain a vector with a length of 3200+50=3250, which is the corresponding spliced features.

[0178] S1204, the spliced features are input into the perception network for perception to obtain the global semantic deviation of the spliced features and the global standard deviation for representing the fluctuation between different feature elements in the spliced features.

[0179] Optionally, in this embodiment, the perception network takes the multi-layer perception machine (Multi-Layer Perceptron, MLP) as an example. After inputting the spliced features into the multi-layer perception machine, the corresponding global semantic deviation and global standard deviation can be obtained.

[0180] It can be understood that the global semantic deviation is an index for measuring the deviation degree of the overall semantics, which is used to represent the difference between the average position of the overall semantic vector and the position of the pre-defined standard semantic vector. If the global semantic deviation is small, it means that the semantics of this group of data is overall close to the standard or expected semantics, and the semantic consistency of the data is high; on the contrary, if the global semantic deviation is large, it means that the semantics of the data deviates from the standard semantics more, and the semantic consistency is poor. The global standard deviation can reflect the degree of coordinated change between different semantic features, that is, how the changes of these features in the overall are related to each other. If the covariance of two semantic features is positive, it means that they tend to increase or decrease at the same time; if the covariance is negative, it means that when one feature increases, the other feature tends to decrease.

[0181] exemplarily, the enhanced features in the higher dimension and the enhanced features in the lower dimension The concatenated features are input into a multi-layer perceptron (MLP) to obtain corresponding global semantic bias = 0.5, and the global standard deviation = 0.1.

[0182] S1206, according to the distribution function composed of the global standard deviation and the global semantic bias, determine the semantic bias threshold.

[0183] Optionally, since the principle is based on the characteristics of the normal distribution, the data will be around the mean present a symmetrical distribution, the dispersion of the data is measured by the global standard deviation , and the calculation process does not involve complex iterations and optimization steps, and can quickly process a large amount of data, so in this embodiment, the global standard deviation and the global semantic bias obtained in the previous step can be combined to form a distribution function for determining the semantic bias threshold.

[0184] Specifically, after obtaining the global standard deviation and the global semantic bias, a distribution function can be formed according to the principle and , and the distribution function is used to determine the semantic bias threshold.

[0185] S1208, the feature elements deviating from the semantic bias threshold in the enhanced features in the higher dimension and the enhanced features in the lower dimension are determined as abnormal feature elements.

[0186] Specifically, when or , it is determined that is an abnormal feature element. Wherein, m in the enhanced features in the higher dimension is the mth element.

[0187] When or , it is determined that is an abnormal feature element. Wherein, m in the enhanced features in the lower dimension is the mth element. By confirming and deleting the abnormal feature elements in the above manner, it is helpful to reduce the interference of abnormal values on model training and prediction, thereby helping to improve the accuracy and stability of the model. In this way, after deleting the abnormal feature elements, the optimized enhanced features in the higher dimension are represented by , and the optimized enhanced features in the lower dimension are represented by​ is represented.

[0188] Exemplarily, the dimensionality-increasing enhanced feature is processed by the principle of outlier processing, for the dimensionality-increasing enhanced feature , if a certain feature element 0.5-0.3=0.2, or 0.5+0.3=0.8, it is deleted. Assuming that the 10th element of the dimensionality-increasing enhanced feature is 0.85, this element is deleted from the dimensionality-increasing enhanced feature , thus obtaining the optimized dimensionality-increasing enhanced feature .

[0189] In one possible implementation, after obtaining the optimized dimensionality-increasing enhanced feature and the optimized dimensionality-decreasing enhanced feature, the optimized dimensionality-increasing enhanced feature and the optimized dimensionality-decreasing enhanced feature are spliced to obtain the final target fusion feature.

[0190] Specifically, at this time, the feature distance between the optimized dimensionality-increasing enhanced feature and the optimized dimensionality-decreasing enhanced feature needs to be calculated according to the following formula:

[0191] ;

[0192] wherein, represents the total number of feature elements in the optimized dimensionality-decreasing enhanced feature ; is a feature element in the optimized dimensionality-decreasing enhanced feature ; is the fourth average feature of the optimized dimensionality-decreasing enhanced feature ; represents the total number of feature elements in the optimized dimensionality-increasing enhanced feature ; is a feature element in the optimized dimensionality-increasing enhanced feature ; is the third average feature of the optimized dimensionality-increasing enhanced feature .

[0193] After obtaining , the dimensionality-increasing enhanced feature and the dimensionality-decreasing enhanced feature are spliced and multiplied with the normalized , and then the features between different levels are spliced to obtain the final target fusion feature, and the splicing process is as follows:

[0194] ;

[0195] wherein || represents a concatenation operation; represents the final fusion feature output of the time window, i.e., the target fusion feature of the time window.

[0196] The embodiment adopts the above method, which can ensure the accuracy and reliability of the final target fusion feature, thereby helping to improve the training effect of the model and the reliability of the motor temperature prediction structure.

[0197] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps.

[0198] Based on the same inventive concept, the embodiment of the present application also provides a vehicle motor temperature prediction device for implementing the vehicle motor temperature prediction method described above. The problem-solving implementation scheme provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more vehicle motor temperature prediction device embodiments provided below can refer to the limitations of the vehicle motor temperature prediction method described above, and will not be repeated here.

[0199] In one embodiment, as shown in Figure 13 a vehicle motor temperature prediction device is provided, comprising: a data acquisition module 1302, a feature extraction module 1304, a feature determination module 1306, and a temperature prediction module 1308, wherein:

[0200] The data acquisition module 1302 is configured to acquire temperature data of the vehicle motor in a current time period, and divide the temperature data to obtain at least two temperature sub-sequences.

[0201] The feature extraction module 1304 is configured to perform feature dimension increasing processing on each temperature subsequence by using a first feature extraction network to obtain dimension-increased features of each temperature subsequence at different network layers in the first feature extraction network, and perform feature dimension decreasing processing on each temperature subsequence by using a second feature extraction network to obtain dimension-decreased features of each temperature subsequence at different network layers in the second feature extraction network, wherein the network layers with the same number of layers in the first feature extraction network and the second feature extraction network are the same.

[0202] The feature determination module 1306 is configured to perform fusion processing on the dimension-increased features and the dimension-decreased features of each temperature subsequence respectively to obtain temperature features corresponding to each temperature subsequence.

[0203] The temperature prediction module 1308 is configured to predict the motor temperature of the vehicle motor in a next time period according to the temperature features corresponding to each temperature subsequence.

[0204] In an embodiment, the feature determination module 1306 includes a feature fusion sub-module and a temperature feature generation sub-module, wherein the feature fusion sub-module is configured to perform fusion on the dimension-increased features and the dimension-decreased features of the temperature subsequence at the network layers of the same level for each temperature subsequence to obtain target fusion features of the temperature subsequence at different levels, and the temperature feature generation sub-module is configured to perform fusion on the target fusion features of the temperature subsequence at different levels to obtain the temperature features corresponding to the temperature subsequence.

[0205] In an embodiment, the feature fusion sub-module further includes a semantic enhancement sub-sub-module and a feature fusion sub-sub-module, wherein the semantic enhancement sub-sub-module is configured to perform semantic enhancement processing on the dimension-increased features and the dimension-decreased features of the temperature subsequence at the network layers of each level to obtain dimension-increased enhanced features and dimension-decreased enhanced features, and the feature fusion sub-sub-module is configured to perform fusion on the dimension-increased enhanced features and the dimension-decreased enhanced features to obtain target fusion features of the temperature subsequence at the level.

[0206] In an embodiment, the semantic enhancement sub-sub-module further includes a bias determination unit and a feature enhancement unit, wherein the bias determination unit is configured to determine a feature bias between the dimension-increased features and the dimension-decreased features, a first index value for characterizing fluctuation conditions between different feature elements in the dimension-increased features, and a second index value for characterizing fluctuation conditions between different feature elements in the dimension-decreased features, the feature enhancement unit is configured to perform semantic enhancement processing on the dimension-increased features according to a distribution function constituted by the feature bias and the first index value to obtain the dimension-increased enhanced features, and perform semantic enhancement processing on the dimension-decreased features according to a distribution function constituted by the feature bias and the second index value to obtain the dimension-decreased enhanced features.

[0207] In an embodiment, the bias determining unit further comprises an average feature determining subunit, a feature bias determining subunit, and an index value determining subunit, wherein the average feature determining subunit is configured to perform feature averaging on each feature element in the dimension-increasing feature to obtain a first average feature, and perform feature averaging on each feature element in the dimension-decreasing feature to obtain a second average feature; the feature bias determining subunit is configured to determine the feature bias between the dimension-increasing feature and the dimension-decreasing feature according to a difference between the first average feature and the second average feature; the index value determining subunit is configured to determine a standard deviation of the dimension-increasing feature as a first index value representing fluctuation among different feature elements in the dimension-increasing feature according to a distance between each feature element in the dimension-increasing feature and the first average feature, and determine a standard deviation of the dimension-decreasing feature as a second index value representing fluctuation among different feature elements in the dimension-decreasing feature according to a distance between each feature element in the dimension-decreasing feature and the second average feature.

[0208] In an embodiment, the feature fusion sub-module further comprises a feature distance determining unit, an initial fusion feature determining unit, and a target fusion feature generating unit, wherein the feature distance determining unit is configured to determine a feature distance between the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature; the initial fusion feature determining unit is configured to splice the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature to obtain the initial fusion feature of the temperature sub-sequence at the level; and the target fusion feature generating unit is configured to update the initial fusion feature of the temperature sub-sequence at the level by using the feature distance to obtain the target fusion feature of the temperature sub-sequence at the level.

[0209] In an embodiment, the feature distance determining unit further comprises an average feature calculating subunit and a feature distance calculating subunit, wherein the average feature calculating subunit is configured to perform average feature on each feature element in the dimension-increasing enhanced feature to obtain a third average feature, and perform average feature on each feature element in the dimension-decreasing enhanced feature to obtain a fourth average feature; and the feature distance calculating subunit is configured to determine the feature distance between the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature according to the third average feature and the fourth average feature.

[0210] In an embodiment, the initial fusion feature determining unit further comprises an abnormal feature element determining subunit, an abnormal feature element removing subunit, and a fusion feature determining subunit, wherein the abnormal feature element determining subunit is configured to perceive the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature based on the perception network to obtain abnormal feature elements in the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature; the abnormal feature element removing subunit is configured to remove the abnormal feature elements in the dimension-increasing enhanced feature and the dimension-decreasing enhanced feature to obtain an optimized dimension-increasing enhanced feature and an optimized dimension-decreasing enhanced feature; and the fusion feature determining subunit is configured to splice the optimized dimension-increasing enhanced feature and the optimized dimension-decreasing enhanced feature to obtain the initial fusion feature of the temperature sub-sequence at the level.

[0211] In one embodiment, the abnormal feature element determination subunit further includes a feature splicing subunit, an abnormal element perception subunit, a deviation threshold determination subunit, and an abnormal element determination subunit. The feature splicing subunit is used to splice the dimensionality-enhanced features and the dimensionality-reduced enhanced features to obtain spliced ​​features. The abnormal element perception subunit is used to input the spliced ​​features into a perceptual network for perception, obtaining the global semantic deviation of the spliced ​​features and the global standard deviation used to characterize the fluctuations between different feature elements in the spliced ​​features. The deviation threshold determination subunit is used to determine the semantic deviation threshold based on the distribution function formed by the global standard deviation and the global semantic deviation. The abnormal element determination subunit is used to determine the feature elements in the dimensionality-enhanced features and the dimensionality-reduced enhanced features that deviate from the semantic deviation threshold as abnormal feature elements.

[0212] The various modules in the aforementioned vehicle motor temperature prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0213] In one embodiment, a computer device is provided, which may be a terminal vehicle controller, and its internal structure diagram may be as follows: Figure 14 As shown. The computer device includes a processor, memory, communication interface, etc., connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for predicting the temperature of a vehicle motor. The display screen of the computer device can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0214] Those skilled in the art will understand that Figure 14 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0215] In an embodiment, a computer device is provided, comprising a memory and a processor, the memory has stored a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0216] In an embodiment, a computer readable storage medium is provided, having stored a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0217] In an embodiment, a computer program product is provided, comprising a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0218] It should be noted that the temperature data of the vehicle motor involved in the present application are all information and data authorized by the user or fully authorized by all parties.

[0219] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0220] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0221] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for predicting the temperature of a vehicle motor, characterized in that, The method includes: Obtain the temperature data of the vehicle motor within the current time period, and divide the temperature data to obtain at least two temperature subsequences; A first feature extraction network is used to perform feature upscaling on each of the temperature subsequences, obtaining the upscaled features of each temperature subsequence at different network layers in the first feature extraction network; and, A second feature extraction network is used to perform feature dimensionality reduction processing on each of the temperature subsequences to obtain the dimensionality reduction features of each of the temperature subsequences in different network layers of the second feature extraction network, wherein the first feature extraction network and the second feature extraction network have the same number of network layers. The dimensionality-upgrading and dimensionality-reducing features of each temperature subsequence are fused to obtain the temperature features corresponding to each temperature subsequence. Based on the temperature characteristics corresponding to each of the temperature subsequences, the motor temperature of the vehicle motor in the next time period is predicted.

2. The method according to claim 1, characterized in that, The process of fusing the dimensionality-upgrading and dimensionality-reducing features of each temperature subsequence to obtain the temperature features corresponding to each temperature subsequence includes: For each temperature subsequence, the dimensionality-upgrading and dimensionality-reducing features of the temperature subsequence at the same level of the network layer are fused to obtain the target fused features of the temperature subsequence at different levels. The temperature subsequence is fused with target fusion features at different levels to obtain the temperature features corresponding to the temperature subsequence.

3. The method according to claim 2, characterized in that, The process of fusing the dimensionality-upgrading and dimensionality-reducing features of the temperature subsequence at the same network layer to obtain the target fused features of the temperature subsequence at different layers includes: For each level, semantic enhancement processing is performed on the dimensionality-upgrading and dimensionality-reducing features of the temperature subsequence in the network layer of that level to obtain dimensionality-upgrading enhancement features and dimensionality-reducing enhancement features; The dimensionality-upgrading enhancement feature and the dimensionality-reducing enhancement feature are fused to obtain the target fused feature of the temperature subsequence at the level.

4. The method according to claim 3, characterized in that, The semantic enhancement processing of the dimensionality-upgrading and dimensionality-reducing features of the temperature subsequence at the network layer at the specified level yields dimensionality-upgrading enhanced features and dimensionality-reducing enhanced features, including: Determine the feature deviation between the dimensionality-upper feature and the dimensionality-reducing feature, a first index value used to characterize the fluctuation between different feature elements in the dimensionality-upper feature, and a second index value used to characterize the fluctuation between different feature elements in the dimensionality-reducing feature; Based on the distribution function formed by the feature deviation and the first index value, semantic enhancement processing is performed on the dimensionality-upgraded feature to obtain the dimensionality-upgraded enhanced feature; and, Based on the distribution function formed by the feature deviation and the second index value, the dimensionality-reduced features are semantically enhanced to obtain the dimensionality-reduced enhanced features.

5. The method according to claim 4, characterized in that, The determination of the feature deviation between the dimensionality-upgraded feature and the dimensionality-reduced feature, a first index value used to characterize the fluctuation between different feature elements in the dimensionality-upgraded feature, and a second index value used to characterize the fluctuation between different feature elements in the dimensionality-reduced feature, includes: The first average feature is obtained by averaging the features of each feature element in the dimensionality-upgrading feature, and the second average feature is obtained by averaging the features of each feature element in the dimensionality-reducing feature. The feature deviation between the dimensionality-upgrading feature and the dimensionality-reducing feature is determined based on the difference between the first average feature and the second average feature; Based on the distance between each feature element in the upgraded feature and the first average feature, the standard deviation of the upgraded feature is determined, serving as a first index value characterizing the fluctuation among different feature elements in the upgraded feature; and... Based on the distance between each feature element in the dimensionality reduction feature and the second average feature, the standard deviation of the dimensionality reduction feature is determined, which serves as a second index value to characterize the fluctuation between different feature elements in the dimensionality reduction feature.

6. The method according to claim 3, characterized in that, The process of fusing the dimensionality-upgrading enhancement features and the dimensionality-reducing enhancement features to obtain the target fused feature of the temperature subsequence at the level includes: Determine the feature distance between the dimension-up enhancement feature and the dimension-down enhancement feature; The dimensionality-upgrading enhancement feature and the dimensionality-reducing enhancement feature are concatenated to obtain the initial fusion feature of the temperature subsequence at the level. Using the aforementioned feature distance, the initial fusion features of the temperature subsequence at the specified level are updated to obtain the target fusion features of the temperature subsequence at the specified level.

7. The method according to claim 6, characterized in that, Determining the feature distance between the dimensionality-enhancing feature and the dimensionality-reducing feature includes: The third average feature is obtained by averaging the features of each feature element in the dimensionality-enhancing feature, and the fourth average feature is obtained by averaging the features of each feature element in the dimensionality-reducing feature. The feature distance between the dimension-up enhancement feature and the dimension-down enhancement feature is determined based on the third average feature and the fourth average feature.

8. The method according to claim 6, characterized in that, The concatenation of the dimensionality-upgrading enhancement features and the dimensionality-reducing enhancement features to obtain the initial fusion features of the temperature subsequence at the level includes: Based on a perceptual network, the dimensionality-upgrading enhancement features and the dimensionality-reducing enhancement features are perceived to obtain abnormal feature elements in the dimensionality-upgrading enhancement features and the dimensionality-reducing enhancement features; Abnormal feature elements are removed from the dimensionality-up enhancement features and the dimensionality-reduction enhancement features to obtain optimized dimensionality-up enhancement features and optimized dimensionality-reduction enhancement features; The optimized up-dimensional enhancement feature and the optimized down-dimensional enhancement feature are concatenated to obtain the initial fusion feature of the temperature subsequence at the level.

9. The method according to claim 8, characterized in that, The method based on a perceptual network perceives the dimensionality-enhancing features and the dimensionality-reducing features to obtain anomalous feature elements in the dimensionality-enhancing features and the dimensionality-reducing features, including: The dimensionality-up enhancement feature and the dimensionality-down enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are input into a perceptual network for perception, and the global semantic deviation of the spliced ​​features and the global standard deviation used to characterize the fluctuation between different feature elements in the spliced ​​features are obtained. The semantic deviation threshold is determined based on the distribution function formed by the global standard deviation and the global semantic deviation. The feature elements that deviate from the semantic deviation threshold in the dimensionality enhancement features and the dimensionality reduction enhancement features are identified as the anomalous feature elements.

10. A temperature prediction device for a vehicle motor, characterized in that, The device includes: The data acquisition module is used to acquire the temperature data of the vehicle motor in the current time period, and divide the temperature data to obtain at least two temperature subsequences; The feature extraction module is used to perform feature dimensionality upscaling on each of the temperature subsequences using a first feature extraction network to obtain the dimensionality upscaling features of each of the temperature subsequences in different network layers of the first feature extraction network, and to perform feature dimensionality reduction on each of the temperature subsequences using a second feature extraction network to obtain the dimensionality reduction features of each of the temperature subsequences in different network layers of the second feature extraction network, wherein the first feature extraction network and the second feature extraction network have the same number of network layers; The feature determination module is used to fuse the dimensionality-upgrading features and dimensionality-reducing features of each of the temperature subsequences to obtain the temperature features corresponding to each of the temperature subsequences. The temperature prediction module is used to predict the motor temperature of the vehicle motor in the next time period based on the temperature characteristics corresponding to each of the temperature sub-sequences.

Images