Efficient time series analysis method, device and medium based on lightweight convolutional neural network

Abstract

The present invention discloses an efficient time series analysis method, device and medium based on convolutional neural networks. The present invention obtains time series modeling and analysis results for different downstream tasks through the steps of raw data reading, sample standardization, adaptive time-frequency conversion, frequency domain component extraction and fitting, and time domain data reconstruction. In the time series analysis method, the time-frequency conversion and time domain data reconstruction steps are implemented by a dynamically adjustable wavelet analysis component, while the frequency domain data fitting step efficiently models the distribution of frequency components in the raw data through a convolutional neural network with low time and space complexity. Raw data reading and sample standardization provide a starting point for time series analysis for different tasks. The present invention is based on data-driven, has no special requirements for data, has strong universality, can maintain high modeling accuracy while significantly reducing time costs and hardware resource overhead, and has strong theoretical and practical properties.

Description

Technical Field

[0001] The present invention relates to a time series analysis method, and in particular to a time series modeling and analysis method, device and medium based on a lightweight convolutional neural network. Background Art

[0002] Time series analysis is a key branch of statistics and data science. It primarily studies chronologically ordered data points to identify trends, seasonality, cyclical patterns, and random fluctuations in the data. With the advent of the big data era, time series analysis is increasingly being used in fields such as finance, meteorology, healthcare, and economic forecasting. Existing time series analysis methods often employ statistical models and deep learning models, such as ARIMA and VAR. These models treat time series as variables and solve them based on statistical analysis. However, real-world data does not follow a linear distribution, meaning linear solutions are impossible. Furthermore, solving high-order nonlinear equations is a particularly challenging problem. Representative deep learning models include the sequence models LSTM and Transformer. These models are completely data-driven, using a loss function to iteratively optimize the original data to overcome the limitations of traditional methods in analyzing nonlinear data. However, both LSTM and Transformer require maintaining long-range historical information to establish long-term contextual relationships. This results in their time and space complexity growing quadratically with the length of the input data, making them unsuitable for large datasets. By transferring the time domain series to the frequency domain, the periodic composition of the time series can be efficiently modeled with a method with low computational time complexity and low hardware overhead, and long-distance modeling can be achieved without maintaining lengthy historical information. Summary of the Invention

[0003] This invention aims to improve and standardize the shortcomings of existing research and technologies, proposing an efficient time series analysis method based on a lightweight convolutional neural network. This method offers a new perspective on time series analysis, significantly reducing the time and space overhead required compared to existing research, while achieving higher modeling accuracy than existing technologies, thus offering greater practical value. Furthermore, as an efficient method, the proposed algorithm maintains a stable time and space complexity when the input data scale continues to expand, demonstrating high scalability.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] An efficient time series analysis method based on lightweight convolutional neural networks, including:

[0006] Step 1: Use an efficient convolutional neural network as a feature extractor and an adaptive wavelet analysis component as a time-frequency conversion component to build an efficient time series analysis model;

[0007] Step 2: Read the time series in the calibration unit of the corresponding time series dataset according to different types of time series downstream tasks;

[0008] Step 3: Use the time series data set in step 2 to eliminate the dimensions of different variables for each sequence variable and standardize the data;

[0009] Step 4: Input the data normalized in step 3 into the efficient time series analysis model constructed in step 1, perform adaptive wavelet decomposition on the normalized data through the wavelet analysis component, convert the time domain data into the frequency domain, and obtain the global frequency domain distribution;

[0010] Step 5: Separate the global frequency domain distribution into high-frequency components and low-frequency components, and use an efficient convolutional neural network to fit the component data distribution and generate high-frequency representations and low-frequency representations;

[0011] Step 6: The high-frequency representation and the low-frequency representation are passed to the wavelet analysis component in step 4 to implement adaptive wavelet reconstruction and restore them to time domain representation;

[0012] In step 7, the time domain representation restored in step 6 is mapped to an output matrix by the output layer, and the error is calculated together with the true value through the mean square error loss function. The error gradient is then derived and back-propagated to simultaneously optimize the efficient convolutional neural network and wavelet analysis components.

[0013] In step 8, according to different types of time series downstream tasks, the corresponding time series dataset is read into the efficient time series analysis model, and the modeling results are output after dimensionality upgrade and characterization operations on the read features, which are mapped to the final analysis results by the output layer.

[0014] Furthermore, step 2 includes:

[0015] Records the number of single-moment data records at a certain moment. The record entries at that moment contain timestamps and sequence variables, and combines multiple data records at different moments into a time series dataset.

[0016] Furthermore, step 3 includes:

[0017] Take a time series variable and calculate its mathematical expectation and standard deviation. After obtaining the mathematical expectation and standard deviation, subtract the mathematical expectation from the original sequence and divide it by the standard deviation to obtain the standardized sequence.

[0018] Furthermore, step 4 includes:

[0019] The wavelet analysis component receives the original data as a single parameter and performs frequency domain conversion. The conversion method is:

[0020] Initialize the high-pass decomposition filter function and the low-pass decomposition filter function;

[0021] Note a certain time series variable, shift the high-pass decomposition filter function backward by the sequence length time units and multiply it with the time series variable, shift the low-pass decomposition filter function backward by the sequence length units and also multiply it with the time series variable, after completing one operation, reset the high-pass decomposition filter function and the low-pass decomposition filter function to their original positions, shift them backward by the variable length minus one time unit and repeat the above operation until all the product results of the high-pass decomposition filter and the low-pass decomposition filter are obtained, and then accumulate the results respectively to obtain the high-frequency component spectrum and the low-frequency component spectrum.

[0022] Furthermore, step 5 includes:

[0023] In step 5.1, after obtaining the high-frequency component spectrum and the low-frequency component spectrum, they are characterized and modeled by two independent efficient convolutional neural networks. The convolution operation process is as follows:

[0024] Randomly initialize the high-frequency component convolution kernel and the low-frequency component convolution kernel;

[0025] The high-frequency component spectrum and the low-frequency component spectrum obtained in step 4 are shifted backward by the spectrum length time units and then multiplied with the high-frequency component spectrum. The low-frequency component convolution kernel is shifted backward by the variable length units and then multiplied with the low-frequency component spectrum. After completing one operation, the high-pass decomposition filter function and the low-pass decomposition filter function are reset to their original positions, and then the spectrum length minus one time unit is shifted backward and the above operation is repeated again until all the product results of the low-frequency component convolution kernel and the high-frequency component convolution kernel are obtained. The results are accumulated respectively to obtain the primary nonlinear high-frequency representation and the primary nonlinear low-frequency representation;

[0026] In step 5.2, after obtaining the preliminary representation, the representation pair is enhanced with the nonlinear expression ability through the GELU activation function to obtain the final representation pair. The calculation process is as follows: after integrating the exponential function from negative infinity to the representation length, the final nonlinear high-frequency representation is obtained by multiplying it with the primary nonlinear high-frequency representation; after integrating the exponential function from negative infinity to the representation length, the final nonlinear low-frequency representation is obtained by multiplying the primary nonlinear low-frequency representation.

[0027] Furthermore, step 6 includes:

[0028] The wavelet analysis component receives the nonlinear high-frequency representation and the nonlinear low-frequency representation obtained in step 5 as parameters to reconstruct the representation information in the time domain. The conversion method is:

[0029] Initialize the high-pass reconstruction filter function and the low-pass reconstruction filter function;

[0030] The restored time domain representation is recorded as follows: the high-pass reconstruction filter function is shifted backward by the representation length units and then multiplied by the high-frequency representation; the low-pass reconstruction filter function is shifted backward by the representation length units and then multiplied by the low-frequency representation. After completing one operation, the high-pass reconstruction filter and the low-pass reconstruction filter are reset to their original positions, and then the above operation is repeated again by shifting the representation length minus one unit until all the product results of the high-pass reconstruction filter and the low-pass reconstruction filter are obtained. All the results are accumulated and multiplied by 2 to obtain the restored time domain representation.

[0031] Furthermore, step 7 includes:

[0032] After obtaining the time domain representation restored in step 6, calculate the mean square error between it and the true value, and update all parameters of the wavelet analysis component and the efficient convolutional neural network based on the result of the mean square error. The mean square error calculation process is: for the predicted value of the time domain representation at each time point, calculate the difference between it and the true value at that point, obtain the difference of all time points, and then accumulate them, and divide the accumulated result by the total number of time points to obtain the final result.

[0033] The present invention also provides an efficient time series analysis device based on a lightweight convolutional neural network, comprising one or more processors for implementing the efficient time series analysis method based on a lightweight convolutional neural network as described above.

[0034] The present invention also provides a readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-mentioned efficient time series analysis method based on a lightweight convolutional neural network.

[0035] Compared with the prior art, the present invention has the following innovative advantages and significant effects:

[0036] 1) Using a new frequency-domain time series modeling paradigm, it can quickly and efficiently analyze the periodic composition of a given time series and map the modeling results in the frequency domain to the time domain to correspond to different types of downstream analysis tasks;

[0037] 2) For different types of time series data, the specific implementation methods of decomposition and reconstruction during time-frequency transformation can vary, which has strong flexibility and scalability;

[0038] 3) While improving the modeling accuracy of various time series tasks, it greatly reduces the time and hardware overhead required, and has good practicality and deployability in real-world applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1This is a flow chart of a time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0040] Figure 2 It is a high-frequency component amplitude diagram obtained after the adaptive time-frequency conversion in step 4 of the embodiment is applied to a time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0041] Figure 3 It is a low-frequency component amplitude diagram obtained after the adaptive time-frequency conversion in step 4 of the embodiment is applied to a time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0042] Figure 4 It is a high-frequency component representation diagram obtained by fitting the high-frequency component data distribution in step 5 of the time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0043] Figure 5 It is a low-frequency component representation diagram obtained by fitting the low-frequency component data distribution in step 5 of the time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0044] Figure 6 It is a prediction curve obtained by reconstructing the low-frequency representation and high-frequency representation obtained into time domain data in step 6 of the time series modeling and analysis method based on a lightweight convolutional neural network of the present invention.

[0045] Figure 7 This is a structural schematic diagram of an efficient time series analysis device based on a lightweight convolutional neural network of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figure 1 , an efficient time series analysis method based on lightweight convolutional neural network, including the following steps:

[0048] Step 1: Building an efficient time series analysis model: Based on the characteristics of time series data, we select an efficient convolutional neural network as the feature extractor and an adaptive wavelet analysis component as the time-frequency conversion component to build an efficient time series analysis model. Specifically, we include:

[0049] Convolutional neural networks (CNNs) are a type of feedforward neural network with a deep structure that incorporates convolutional computations. They are a representative algorithm for deep learning. CNNs possess the ability to learn representations and perform translation-invariant classification of input information based on their hierarchical structure. Efficient convolutional modeling is achieved by stacking small convolution kernels.

[0050] Wavelet analysis is a localized analysis method for time and frequency. It uses basis function scaling and translation operations to gradually refine numerical sequences at multiple scales, ultimately achieving time subdivision at high frequencies and frequency subdivision at low frequencies. It automatically adapts to the requirements of time-frequency signal analysis, allowing for the focus to be placed on any detail of the signal. Within the efficient time series analysis model, by setting learnable basis functions, gradient descent algorithms can be used to adaptively optimize these basis functions to the most suitable shape.

[0051] Step 2: Reading raw data: Based on different types of time series downstream tasks, read the time series in the calibration unit of the corresponding time series dataset, including:

[0052] Record The number of single moment data records at a moment is recorded as The record entry at that moment contains a timestamp , sequence variables , the time series dataset is recorded as ,in .

[0053] Step 3, Sample Standardization: Using the time series dataset in step 2 For each sequence variable Calculating the mathematical expectation and standard deviation Afterwards, through and Will Scaling is performed to eliminate the dimensions of different variables in the original time series, thus ensuring data comparability.

[0054] Among them, the single time series variable is recorded as , calculate the mathematical expectation and standard deviation , which is calculated as follows:

[0055]

[0056] Where k is the total number of samples and t is the time;

[0057] For sequence variables For standardization, the calculation formula is as follows:

[0058] .

[0059] Where, Represents the normalized sequence.

[0060] Step 4, adaptive time-frequency conversion: Input the standardized data in step 3 into the efficient time series analysis model constructed in step 1, and perform adaptive wavelet decomposition on the standardized data through the wavelet analysis component to convert the time domain data into the frequency domain to obtain the global frequency domain distribution. , specifically including:

[0061] In the adaptive wavelet decomposition stage, the wavelet analysis component receives the original data as a single parameter for frequency domain conversion. The conversion method is:

[0062] Initialize the high-pass decomposition filter function With the low-pass decomposition filter function ;

[0063] Let the single time series variable be , calculate the high frequency component spectrum and low-frequency component spectrum , which is calculated as follows:

[0064]

[0065] Where, represents the discrete convolution operator, Indicates the translation step size of the filter function.

[0066] Step 5, frequency domain distribution pattern fitting: First, the global frequency domain distribution Perform separation operation to obtain high frequency component and low frequency component ,Will Use efficient convolutional neural networks to fit data distribution and enhance nonlinear expression capabilities through activation functions to obtain spectral representation after modeling. ,in represents high-frequency representation, Indicates low-frequency characterization. Specifically including:

[0067] Step 5.1, in the distribution pattern fitting stage, the spectrum pair is obtained. Then, the representation model is performed through an efficient convolutional neural network. The convolution operation process is as follows:

[0068] Randomly initialize the high-frequency component convolution kernel Convolution kernel with low frequency component ;

[0069] The high-frequency component spectrum obtained from step 4 and low-frequency component spectrum , and its characterization calculation formula is as follows:

[0070]

[0071] Where, represents the primary nonlinear high-frequency characterization, Represents the primary nonlinear low-frequency characterization.

[0072] Step 5.2: After obtaining the primary nonlinear representation pair, the representation pair is activated by the GELU function to enhance the nonlinear expression ability and obtain the final representation pair. , the calculation process is as follows:

[0073]

[0074]

[0075] Where, Represents the GELU activation function.

[0076] Step 6, adaptive time domain feature reconstruction: Use the frequency component representation obtained in step 5 to As the initial value, it is passed as a parameter to the wavelet analysis component in step 4 to achieve adaptive wavelet reconstruction and restore it to the time domain representation, specifically including:

[0077] In the stage of time domain feature reconstruction, the wavelet analysis component receives the two spectral representation pairs obtained in step 5. As a parameter to reconstruct the representation information in the time domain, the conversion method is:

[0078] Initialize the high-pass reconstruction filter function with the low-pass reconstruction filter function ;

[0079] The restored time domain representation is , which is calculated as follows:

[0080] .

[0081] Step 7, Error Backpropagation and Optimization: After obtaining the restored time domain representation through step 6, the time domain representation is mapped into different shapes through the output layer according to different downstream analysis tasks, and the error is calculated through the mean square error loss function after being compared with the true value. After the error is obtained, it is propagated back through the derivative gradient to simultaneously optimize the efficient convolutional neural network and wavelet analysis components. Specifically including:

[0082] Get characterization results Then, calculate the mean square error between it and the true value , and according to the mean square error The results update the wavelet analysis component and all parameters of the convolutional network, the mean square error The calculation method is:

[0083]

[0084] Where k is the total number of samples, is the true value of the i-th sample, is the predicted value of the i-th sample.

[0085] Step 8, time series analysis: Based on different types of time series downstream tasks, the trained efficient time series analysis model obtained in step 7 is fed with the corresponding time series dataset. The efficient time series analysis model performs dimension upgrade and characterization operations on the read features to produce modeling results, which are then mapped to the final analysis results by the output layer.

[0086] See also Figure 7 An embodiment of the present invention provides an efficient time series analysis device based on a lightweight convolutional neural network, including one or more processors for implementing an efficient time series analysis method based on a convolutional neural network in the above embodiment.

[0087] An embodiment of the present invention's efficient time series analysis device based on a lightweight convolutional neural network can be applied to any device with data processing capabilities, which can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, if Figure 7 As shown in the figure, it is a hardware structure diagram of any device with data processing capability where an efficient time series analysis device based on a lightweight convolutional neural network of the present invention is located. Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in the embodiment may also include other hardware according to the actual function of the device with data processing capabilities, which will not be described in detail.

[0088] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0089] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] An embodiment of the present invention also provides a readable storage medium having a program stored thereon. When the program is executed by a processor, an efficient time series analysis method based on a convolutional neural network in the above embodiment is implemented.

[0091] The readable storage medium may be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or memory. The readable storage medium may also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and may also be used to temporarily store data that has been output or is to be output.

[0092] Example

[0093] This example processes the oil temperature time series of transformers at a substation in a certain location, collected from July 1, 2016, to June 26, 2018. The specific variables and related data included in the dataset are shown in Table 1:

[0094] Table 1 Partial time series data of transformer oil temperature in a certain place

[0095]

[0096] The implementation data set of the multivariate time series analysis method in this embodiment is the transformer oil temperature time series data of a certain location mentioned above. The method result is the future prediction curve of the oil temperature variable obtained. The detailed implementation steps are as follows:

[0097] Step 1: Build an efficient time series analysis model: Based on the characteristics of time series data, we select an efficient convolutional neural network as the feature extractor and an adaptive wavelet analysis component as the time-frequency conversion component to build an efficient time series analysis model. The adaptive wavelet analysis component is responsible for converting the time series into a frequency spectrum, while the efficient convolutional neural network characterizes the distribution pattern of the data in the frequency spectrum.

[0098] Step 2: Obtain demand based on the epidemic trend curve and read historical transformer tank temperature data in hours. The number of single-hour data records at time is recorded as This information includes the record timestamp , high useful load variable {v1}, high useless load variable {v2}, medium useless load variable {v3}, low useful load variable {v4}, low useless load variable {v5} and oil temperature variable {v6}, and record the time series data set as ,in ; According to the data information provided in Table 1, it includes all the necessary information required in this step.

[0099] Step 3: Using the time series dataset from step 2 , first calculate the mathematical expectation for each time series and standard deviation , and on this basis calculate Value, get the standardized result of each record, the results are shown in Table 2:

[0100] Table 2 Time series data after standardization (no unit after standardization)

[0101]

[0102] Step 4: Input the standardized data in step 3 into the efficient time series analysis model constructed in step 1, and perform adaptive wavelet decomposition on the standardized data through the wavelet analysis component to convert the time domain data into the frequency domain to obtain the global frequency domain distribution, which is recorded as The calculation method is as follows:

[0103]

[0104] In this embodiment, the high-frequency spectrum and low-frequency spectrum calculated from the original data are as follows: Figure 2 and Figure 3 shown.

[0105] In step 5, the frequency domain distribution pattern fitting step, the specific fitting method is as follows:

[0106] a) Preset two independent convolutional neural networks and And randomly initialize the weights of the convolution kernel of the neural network.

[0107] b) Based on the frequency components calculated in step 4 , and pass them as parameters into the corresponding feature extraction network for data fitting and modeling. The convolutional neural network used in this embodiment is a one-dimensional convolution. The convolution layer performs a convolution operation on the input data by sliding a fixed-size window, extracting the features within the window, and then mapping these features to the next layer after enhancing the nonlinear expression. It consists of multiple convolutional layers and pooling layers alternating, and finally uses a fully connected layer to map the extracted features to the output. During the training process, the neural network uses the backpropagation algorithm to update the model parameters to minimize the loss function. This step is carried out in two stages. The initial feature extraction calculation process in the first stage is as follows:

[0108]

[0109] Get the primary nonlinear characterization pair After that, the nonlinear expression ability is enhanced by the GELU operator. The calculation method is as follows:

[0110]

[0111]

[0112] After the entire representation modeling phase is completed, the fitted frequency component representation pair with a fixed prediction time T of 96 can be obtained. , respectively as Figure 4 、 Figure 5 shown.

[0113] Step 6, in the adaptive time domain feature reconstruction step, use the , and pass it as a parameter to the wavelet analysis component to reconstruct the frequency domain periodic information into a time domain signal The calculation method is as follows:

[0114]

[0115] After the reconstruction is completed, use The fitted features are output as prediction curves for evaluation and analysis of different downstream tasks. Figure 6 shown.

[0116] The present invention provides an efficient time series analysis method based on a lightweight convolutional neural network, which mainly obtains the final time series prediction curve through steps such as sample standardization, adaptive time-frequency conversion, frequency domain distribution pattern fitting, and adaptive time domain feature reconstruction. Figure 1 This is the specific process of the time series efficient modeling method. The entire embodiment is based on Figure 1 According to the process shown in , the transformer oil temperature dataset is processed and the temperature prediction curve is finally obtained. Figure 2 、 Figure 3To calculate the high-frequency spectrum and low-frequency spectrum of the original data using the present invention, Figure 4 、 Figure 5 In order to characterize the distribution pattern of frequency information in the spectrum fitted by the present invention, Figure 6 The final temperature prediction curve is obtained. Time-frequency transformation and frequency domain fitting of the oil temperature data can produce a prediction curve that is more in line with the actual situation than the traditional time series analysis model. While meeting the higher degree of data authenticity, it also has the characteristics of high real-time performance, low resource overhead, and strong scalability. It provides a more practical solution for subsequent related research on time series modeling tasks with high real-time requirements.

[0117] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An efficient time series analysis method based on lightweight convolutional neural network, characterized by: include: Step 1: Use an efficient convolutional neural network as a feature extractor and an adaptive wavelet analysis component as a time-frequency conversion component to build an efficient time series analysis model; Step 2: Read the time series in the calibration unit of the transformer oil temperature time series dataset, including: recording the number of single-moment data records at a certain moment, where the record entry at that moment includes a record timestamp, a high useful load variable, a high useless load variable, a medium useless load variable, a low useful load variable, a low useless load variable, and an oil temperature variable, and combining multiple data records at different moments into a time series dataset; Step 3: Using the transformer oil temperature time series data set in step 2, eliminate the dimension of different variables for each sequence variable to standardize the data; Step 4: Input the standardized data from step 3 into the efficient time series analysis model constructed in step 1. Perform adaptive wavelet decomposition on the standardized data through the wavelet analysis component to convert the time domain data into the frequency domain to obtain the global frequency domain distribution, including: The wavelet analysis component receives the original data as a single parameter and performs frequency domain conversion. The conversion method is: Initialize the high-pass decomposition filter function and the low-pass decomposition filter function; Take a certain time series variable, shift the high-pass decomposition filter function backward by the sequence length time units and multiply it with the time series variable, shift the low-pass decomposition filter function backward by the sequence length units and multiply it with the time series variable, after completing one operation, reset the high-pass decomposition filter function and the low-pass decomposition filter function to their original positions, shift the variable length minus one time unit and repeat the above operation again, until all the product results of the high-pass decomposition filter and the low-pass decomposition filter are obtained, and then accumulate the results respectively to obtain the high-frequency component spectrum and the low-frequency component spectrum; Step 5: Separate the global frequency domain distribution into high-frequency components and low-frequency components, and use an efficient convolutional neural network to fit the component data distribution and generate high-frequency representations and low-frequency representations; Step 6: The high-frequency representation and the low-frequency representation are passed to the wavelet analysis component in step 4 to implement adaptive wavelet reconstruction and restore them to time domain representation; In step 7, the time domain representation restored in step 6 is mapped to an output matrix by the output layer, and the error is calculated together with the true value through the mean square error loss function. The error gradient is then derived and back-propagated to simultaneously optimize the efficient convolutional neural network and wavelet analysis components. In step 8, the transformer oil temperature time series dataset is read into the efficient time series analysis model, and the modeling results are output after dimensionality upgrade and characterization operations on the read features, which are mapped into the final future prediction curve of the oil temperature variable by the output layer.

2. The efficient time series analysis method based on a lightweight convolutional neural network according to claim 1, characterized in that: Step 3 includes: Take a time series variable and calculate its mathematical expectation and standard deviation. After obtaining the mathematical expectation and standard deviation, subtract the mathematical expectation from the original sequence and divide it by the standard deviation to obtain the standardized sequence.

3. The efficient time series analysis method based on lightweight convolutional neural network according to claim 2 is characterized in that: Step 5 includes: In step 5.1, after obtaining the high-frequency component spectrum and the low-frequency component spectrum, they are characterized and modeled by two independent efficient convolutional neural networks. The convolution operation process is as follows: Randomly initialize the high-frequency component convolution kernel and the low-frequency component convolution kernel; The high-frequency component spectrum and the low-frequency component spectrum obtained in step 4 are shifted backward by the spectrum length time units and then multiplied with the high-frequency component spectrum. The low-frequency component convolution kernel is shifted backward by the variable length units and then multiplied with the low-frequency component spectrum. After completing one operation, the high-pass decomposition filter function and the low-pass decomposition filter function are reset to their original positions, and then the spectrum length minus one time unit is shifted backward and the above operation is repeated again until all the product results of the low-frequency component convolution kernel and the high-frequency component convolution kernel are obtained. The results are accumulated respectively to obtain the primary nonlinear high-frequency representation and the primary nonlinear low-frequency representation; In step 5.2, after obtaining the preliminary representation, the representation pair is enhanced with the nonlinear expression ability through the GELU activation function to obtain the final representation pair. The calculation process is as follows: after integrating the exponential function from negative infinity to the representation length, the final nonlinear high-frequency representation is obtained by multiplying it with the primary nonlinear high-frequency representation; after integrating the exponential function from negative infinity to the representation length, the final nonlinear low-frequency representation is obtained by multiplying the primary nonlinear low-frequency representation.

4. The efficient time series analysis method based on a lightweight convolutional neural network according to claim 3 is characterized in that: Step 6 includes: The wavelet analysis component receives the nonlinear high-frequency representation and the nonlinear low-frequency representation obtained in step 5 as parameters to reconstruct the representation information in the time domain. The conversion method is: Initialize the high-pass reconstruction filter function and the low-pass reconstruction filter function; The restored time domain representation is recorded as follows: the high-pass reconstruction filter function is shifted backward by the representation length units and then multiplied by the high-frequency representation; the low-pass reconstruction filter function is shifted backward by the representation length units and then multiplied by the low-frequency representation. After completing one operation, the high-pass reconstruction filter and the low-pass reconstruction filter are reset to their original positions, and then the above operation is repeated again by shifting the representation length minus one unit until all the product results of the high-pass reconstruction filter and the low-pass reconstruction filter are obtained. All the results are accumulated and multiplied by 2 to obtain the restored time domain representation.

5. The efficient time series analysis method based on lightweight convolutional neural network according to claim 1, characterized in that: Step 7 includes: After obtaining the time domain representation restored in step 6, calculate the mean square error between it and the true value, and update all parameters of the wavelet analysis component and the efficient convolutional neural network based on the result of the mean square error. The mean square error calculation process is: for the predicted value of the time domain representation at each time point, calculate the difference between it and the true value at that point, obtain the difference of all time points, and then accumulate them, and divide the accumulated result by the total number of time points to obtain the final result.

6. An efficient time series analysis device based on a lightweight convolutional neural network, characterized in that: The invention comprises one or more processors for implementing an efficient time series analysis method based on a lightweight convolutional neural network according to any one of claims 1 to 5.

7. A readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, an efficient time series analysis method based on a lightweight convolutional neural network as described in any one of claims 1 to 5 is implemented.

Images