Method and device for constructing signal prediction model and method and device for signal prediction

By preprocessing the sensor signal and training the multi-neural network fusion model, the combined loss function is constructed based on the signal amplitude and frequency loss function, the cost and complexity problems caused by the increase in sensors in the vehicle are solved, and the accuracy of the signal prediction model is achieved. Cost reduction.

CN119066342APending Publication Date: 2024-12-03NIO TECH ANHUI CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411197827.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-28
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The increase in sensors in vehicles leads to increased production costs and design complexity, and the prior art is difficult to achieve accurate signal prediction without adding sensors.

Method used

By obtaining the preprocessed sensor signal, using the variational modal decomposition algorithm for noise reduction processing, and inputting the signal to the multi-neural network fusion model, building a combined loss function based on the signal amplitude and frequency loss function to train the model, and building a signal prediction model .

Benefits of technology

It realizes accurate fitting of the new sensor output signal without adding sensors, reducing hardware cost and design complexity, and improving the prediction accuracy of the signal prediction model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066342A_ABST
    Figure CN119066342A_ABST
Patent Text Reader

Abstract

The invention relates to the field of signal processing, in particular to a method for constructing a signal prediction model, computer equipment, a computer readable storage medium and a signal prediction method, computer equipment and a computer readable storage medium. The method for constructing the signal prediction model in one aspect of the invention comprises the following steps: acquiring a training sample of a sensor signal, wherein the training sample of the sensor signal comprises a preprocessed sensor signal; inputting the training sample of the sensor signal into a multi-neural network fusion model to output a prediction signal corresponding to the training sample of the sensor signal; and constructing a signal amplitude loss function and a signal frequency loss function at least based on the prediction signal, and training the multi-neural network fusion model based on the constructed combined loss function of the signal amplitude loss function and the signal frequency loss function to construct the signal prediction model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of signal processing, and more specifically to a method, a computer device, a computer-readable storage medium for constructing a signal prediction model, and a method, a computer device, and a computer-readable storage medium for signal prediction. Background Art

[0002] In order to improve the safety and intelligence level of vehicles, the types and number of sensors used in vehicles are increasing. For example, various sensors are set in the vehicle's electric drive system, safety system, intelligent driving assistance system, battery management system, etc. However, the setting of various sensors increases the production cost of the vehicle and increases the complexity of vehicle design. Summary of the invention

[0003] In order to solve or at least alleviate one or more of the above problems, the following technical solutions are provided.

[0004] According to a first aspect of the present application, a method for constructing a signal prediction model is provided, the method comprising the following steps: obtaining training samples of sensor signals, the training samples of sensor signals comprising preprocessed sensor signals; inputting the training samples of the sensor signals into a multi-neural network fusion model to output a prediction signal corresponding to the training samples of the sensor signals; and constructing a signal amplitude loss function and a signal frequency loss function based on at least the prediction signal, and training the multi-neural network fusion model based on a combined loss function of the constructed signal amplitude loss function and the signal frequency loss function to construct the signal prediction model.

[0005] According to the method for constructing a signal prediction model described in an embodiment of the present application, the preprocessed sensor signal includes one or more of the following: a preprocessed vibration signal, a preprocessed current signal, and a preprocessed resolver signal.

[0006] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, the preprocessed sensor signal is obtained in the following manner: obtaining the sensor signal; performing denoising on the sensor signal using a variational mode decomposition algorithm; and performing one or more of time domain processing operations, frequency domain processing operations, and smoothing processing operations on the denoised sensor signal to obtain the preprocessed sensor signal.

[0007] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, performing noise reduction processing on the sensor signal using a variational mode decomposition algorithm includes: decomposing the sensor signal into a set of intrinsic mode functions with different center frequencies and finite bandwidths using a variational mode decomposition algorithm; identifying noise components in the sensor signal based on the set of intrinsic mode functions; and removing the identified noise components in the sensor signal from the sensor signal to perform noise reduction processing on the sensor signal.

[0008] According to the method for constructing a signal prediction model described in one embodiment or any one of the above embodiments of the present application, the time domain processing operation, the frequency domain processing operation and the smoothing processing operation are respectively performed in the following ways: obtaining the sensor signal values ​​of the denoised processing at multiple moments and performing the time domain processing operation on the sensor signal of the denoised processing based on the correlation between the multiple moments; obtaining the frequency domain information of the sensor signal values ​​of the denoised processing at multiple moments and performing the frequency domain processing operation on the sensor signal of the denoised processing based on the inverse transform of the main frequency component of the frequency domain information; and obtaining the sensor signal values ​​of the denoised processing at multiple moments and smoothing the sensor signal values ​​at the multiple moments to perform the smoothing processing operation on the sensor signal of the denoised processing.

[0009] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, the multi-neural network fusion model includes at least two of the following neural network models: a neural network model based on a gradient boosting decision tree, a neural network model for tabular data, and a time-recurrent neural network model.

[0010] According to the method for constructing a signal prediction model according to one embodiment of the present application or any one of the above embodiments, the training samples of the sensor signal are input into a multi-neural network fusion model to output a prediction signal corresponding to the training samples of the sensor signal, including: inputting the training samples of the sensor signal into each neural network model of the multi-neural network fusion model respectively to obtain the output results of each neural network model respectively; and performing weighted averaging processing on the output results of each neural network model to output a prediction signal corresponding to the training samples of the sensor signal.

[0011] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, constructing a signal amplitude loss function based on at least the prediction signal includes: using a dynamic time warping algorithm to determine the difference between the prediction signal and the sample label corresponding to the training sample; and constructing the signal amplitude loss function based on the difference.

[0012] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, constructing a signal frequency loss function based at least on the prediction signal includes: performing Fourier transform on the prediction signal and the sample label corresponding to the training sample to respectively obtain the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample; determining the difference information between the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample; and constructing the signal frequency loss function based on the difference information.

[0013] According to the method for constructing a signal prediction model according to one embodiment of the present application or any of the above embodiments, the combined loss function is obtained in the following manner: weighted summation of the constructed signal amplitude loss function and the signal frequency loss function is performed to obtain the combined loss function.

[0014] According to the method for constructing a signal prediction model described in one embodiment or any of the above embodiments of the present application, during the weighted summation, the weight corresponding to the constructed signal amplitude loss function and the weight corresponding to the signal frequency loss function are dynamically adjusted in the following manner: in response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being greater than the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function is adjusted to be greater than the weight corresponding to the signal frequency loss function; in response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being less than the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function is adjusted to be less than the weight corresponding to the signal frequency loss function; and in response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being equal to the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function and the weight corresponding to the signal frequency loss function are respectively adjusted to preset weights.

[0015] According to the second aspect of the present application, a computer device for constructing a signal prediction model is provided, the computer device comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, the execution of the computer program results in the execution of the steps of the method for constructing a signal prediction model according to the first aspect of the present application.

[0016] According to a third aspect of the present application, a computer storage medium is provided, which includes instructions, and the instructions, when run, execute the steps of the method for constructing a signal prediction model according to the first aspect of the present application.

[0017] According to the fourth aspect of the present application, a method for signal prediction is provided, the method comprising the following steps: acquiring a first sensor signal and a second sensor signal; inputting the first sensor signal and the second sensor signal into a signal prediction model constructed according to the method for constructing a signal prediction model described in the first aspect of the present application to predict a third sensor signal.

[0018] According to the fifth aspect of the present application, a computer device for signal prediction is provided, the computer device comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, the running of the computer program resulting in the execution of the steps of the signal prediction method according to the fourth aspect of the present application.

[0019] According to a sixth aspect of the present application, a computer storage medium is provided, which includes instructions, and the instructions, when executed, execute the steps of the signal prediction method according to the fourth aspect of the present application.

[0020] According to one or more embodiments of the present application, the scheme for constructing a signal prediction model can train a multi-neural network fusion model through a combined loss function of a signal amplitude loss function and a signal frequency loss function to construct a signal prediction model, thereby being able to fully consider the amplitude information in the time domain and the periodic information in the frequency domain of the sensor signal to accurately restore the time series shape of the signal to be predicted, thereby improving the prediction accuracy of the signal prediction model. According to one or more embodiments of the present application, the signal prediction scheme can achieve accurate fitting of the output signal of the sensor to be added without adding the sensor to be added, thereby reducing hardware costs and design complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or other aspects and advantages of the present application will become clearer and easier to understand through the following description of various aspects in conjunction with the accompanying drawings, in which the same or similar units are represented by the same reference numerals. The accompanying drawings include:

[0022] Figure 1 A flowchart of a method for constructing a signal prediction model according to one or more embodiments of the present application is shown.

[0023] Figure 2 A flow chart of a signal prediction method according to one or more embodiments of the present application is shown.

[0024] Figure 3 A schematic block diagram of a computer device for constructing a signal prediction model according to one or more embodiments of the present application is shown.

[0025] Figure 4 A schematic block diagram of a computer device for signal prediction according to one or more embodiments of the present application is shown. DETAILED DESCRIPTION

[0026] The present application is described more fully below with reference to the accompanying drawings in which illustrative embodiments of the present application are illustrated. However, the present application may be implemented in different forms and should not be interpreted as being limited to the embodiments given herein. The above embodiments are given to make the disclosure herein comprehensive and complete, so as to more fully convey the scope of protection of the present application to those skilled in the art.

[0027] In this specification, terms such as "comprise" and "include" indicate that in addition to the units and steps directly and explicitly stated in the specification and claims, the technical solution of the present application does not exclude the situation where there are other units and steps that are not directly or explicitly stated.

[0028] Unless otherwise specified, terms such as "first" and "second" do not indicate the order of the units in terms of time, space, size, etc., but are merely used to distinguish the units.

[0029] Hereinafter, various exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings.

[0030] Figure 1 A flowchart of a method for constructing a signal prediction model according to one or more embodiments of the present application is shown.

[0031] like Figure 1 As shown in FIG. 1 , in step S101 , a training sample of a sensor signal is acquired, and the training sample of the sensor signal includes a preprocessed sensor signal.

[0032] Optionally, in step S101, the training samples of the sensor signals may include vibration signals, current signals and resolver signals output by various sensors of the vehicle electric drive system. In one embodiment, the preprocessed sensor signals may include one or more of the following: preprocessed vibration signals, preprocessed current signals and preprocessed resolver signals. Exemplarily, the vibration signals, current signals and resolver signals output by various sensors of the vehicle electric drive system may be collected on a test bench. For example, the current signals and resolver signals are used as output signals of existing sensors, and the vibration signals are used as output signals of sensors to be added, so that the vibration signals, current signals and resolver signals are used together as training samples of sensor signals.

[0033] In one or more embodiments, the preprocessing operation of the sensor signal may include noise reduction processing and one or more of time domain processing operations, frequency domain processing operations, and smoothing processing operations performed on the noise-reduced sensor signal. The noise reduction processing of the sensor signal, the time domain processing operations, frequency domain processing operations, and smoothing processing operations performed on the noise-reduced sensor signal will be further described below according to one or more embodiments. In one embodiment, the vibration signal, the current signal, and the resolver signal may be preprocessed separately to use the preprocessed vibration signal, current signal, and resolver signal together as training samples of the sensor signal.

[0034] In one embodiment, a sensor signal may be acquired, and a variational mode decomposition algorithm may be used to perform noise reduction processing on the acquired sensor signal. In one embodiment, one or more of a time domain processing operation, a frequency domain processing operation, and a smoothing processing operation may be performed on the noise-reduced sensor signal to obtain a preprocessed sensor signal.

[0035] In one embodiment, a variational mode decomposition algorithm can be used to decompose a sensor signal into a set of intrinsic mode functions with different center frequencies and limited bandwidths, and noise components in the sensor signal can be identified based on the set of intrinsic mode functions, and the noise components in the identified sensor signal can be removed from the sensor signal to perform noise reduction on the sensor signal. In one embodiment, a permutation entropy method can be used to remove components with permutation entropy greater than a predetermined threshold as high-noise components, and the remaining components can be summed to generate a sensor signal time series sequence after noise reduction to perform noise reduction on the sensor signal. Exemplarily, the predetermined threshold can be determined based on the training effect of the constructed signal prediction model. For example, when the training effect indicates that the background noise component of the training sample of the sensor signal is high, the size of the predetermined threshold can be increased.

[0036] In one embodiment, the denoised sensor signal values ​​at multiple moments may be obtained, and the time domain processing operation may be performed on the denoised sensor signal based on the correlation between the multiple moments. In one embodiment, the frequency domain information of the denoised sensor signal values ​​at multiple moments may be obtained, and the frequency domain processing operation may be performed on the denoised sensor signal based on the inverse transformation of the main frequency component of the frequency domain information. In one embodiment, the denoised sensor signal values ​​at multiple moments may be obtained, and the sensor signal values ​​at multiple moments may be smoothed to perform the smoothing operation on the denoised sensor signal.

[0037] As an example, the process of time domain processing operation is described by taking the current sensor signal as an example. Assuming that the values ​​of the current sensor signal after noise reduction processing obtained at moments t1, t2, t3, t4, and t5 are 5, 7, 9, 10, and 11 (the unit is assumed to be ampere), based on the correlation between moments t1, t2, t3, t4, and t5 (for example, the correlation between moments t1, t2, and t3, the correlation between moments t2, t3, and t4, and the correlation between moments t3, t4, and t5), the time domain processing operation can be performed on the values ​​5, 7, 9, 10, and 11 of the current sensor signal after noise reduction processing to obtain training samples of the current sensor signal as [5, 7, 9], [7, 9, 10], and [9, 10, 11], so that the current values ​​of the training samples have correlations between multiple moments, so as to improve the training effect of the subsequent signal prediction model. It should be noted that the above-mentioned time domain processing operation can be applied to preprocessed vibration signals, preprocessed current signals, and preprocessed resolver signals.

[0038] As an example, the process of frequency domain processing operation is described by taking the current sensor signal as an example. Assuming that the discrete time domain information of the current sensor signal after noise reduction processing is obtained at multiple moments (for example, the values ​​of the current sensor signal after noise reduction processing obtained at moments t1, t2, t3, t4, and t5 are 5, 7, 9, 10, and 11 (the unit is assumed to be ampere)), the discrete time domain information can be converted into continuous frequency domain information through Fourier transform (for example, fast Fourier transform) and the main frequency component of the continuous frequency domain information (for example, the frequency component corresponding to the maximum amplitude) can be extracted through a filter. Then, the main frequency component of the extracted continuous frequency domain information is subjected to inverse Fourier transform to introduce part of the frequency domain information into the noise reduction processed current sensor signal, so that there is frequency correlation between the training samples to improve the training effect of the subsequent signal prediction model. It should be noted that the above frequency domain processing operation can be applied to preprocessed vibration signals, preprocessed current signals, and preprocessed resolver signals.

[0039] As an example, the process of the smoothing operation is described by taking the current sensor signal as an example. Assuming that the time series waveform of the current sensor signal after noise reduction is obtained at multiple moments, the time series waveform can be smoothed by multi-level exponential smoothing to accurately capture the trend and seasonal information of the time series waveform, thereby improving the training effect of the subsequent signal prediction model. For example, triple exponential smoothing (TES) can be used to construct a level equation, a trend equation and a seasonal equation based on the time series waveform of the current sensor signal, and the level, trend and seasonality of the time series waveform are weighted averaged by introducing smoothing coefficients corresponding to the level, trend and seasonality in the equation, thereby smoothing the time series waveform. It should be noted that the above-mentioned smoothing operation can be applied to preprocessed vibration signals, preprocessed current signals and preprocessed rotary transformer signals.

[0040] In step S103, the training samples of the sensor signals are input into the multi-neural network fusion model to output prediction signals corresponding to the training samples of the sensor signals.

[0041] Optionally, in step S103, the multi-neural network fusion model may include a fusion of at least two of the following neural network models: a neural network model based on a gradient boosting decision tree (e.g., LightGBM (Light GradientBoosting Machine)), a neural network model for tabular data (e.g., TabNet model), and a time recurrent neural network model (e.g., LSTM (Long Short-Term Memory)).

[0042] Optionally, in step S103, the training samples of the sensor signal can be input into each neural network model of the multi-neural network fusion model respectively to obtain the output results of each neural network model respectively, and the output results of each neural network model can be weighted averaged to output a prediction signal corresponding to the training samples of the sensor signal.

[0043] In step S105, a signal amplitude loss function and a signal frequency loss function are constructed based on at least the predicted signal, and a multi-neural network fusion model is trained based on a combined loss function of the constructed signal amplitude loss function and the signal frequency loss function to construct a signal prediction model.

[0044] Optionally, in step S105, a dynamic time warping algorithm may be used to determine the difference between the predicted signal and the sample label corresponding to the training sample, and a signal amplitude loss function may be constructed based on the difference.

[0045] Exemplarily, the DTW (Dynamic Time Warping) loss function can be used to construct a signal amplitude loss function, that is, a distance matrix can be constructed for two time series, in which each element represents the distance between two sequence points at the corresponding moment (for example, using Euclidean distance); the minimum cumulative distance from the starting point to each point is calculated by dynamic programming, and these cumulative distances are used to construct a cost matrix; the minimum path in the cost matrix is ​​found, and this path represents the best alignment between the two sequences; the cumulative distance of the best alignment path is calculated, and this distance is the DTW distance, which is a measure of the difference between the two sequences. Exemplarily, the above two time series can be respectively implemented as a predicted signal and a sample label corresponding to a training sample to determine the degree of difference between the predicted signal and the sample label corresponding to the training sample.

[0046] Optionally, in step S105, the prediction signal and the sample label corresponding to the training sample can be Fourier transformed to obtain the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample, respectively, determine the difference information between the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample, and construct a signal frequency loss function based on the difference information. Exemplarily, a norm (e.g., a first-order norm) can be used to determine the difference information between the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample, and construct a signal frequency loss function based on the difference information.

[0047] Optionally, in step S105, a weighted summation may be performed on the constructed signal amplitude loss function and the signal frequency loss function to obtain a combined loss function.

[0048] In one embodiment, the weight corresponding to the constructed signal amplitude loss function and the weight corresponding to the signal frequency loss function can be dynamically adjusted by the output results of the constructed signal amplitude loss function and the signal frequency loss function, so as to perform weighted summation on the constructed signal amplitude loss function and the signal frequency loss function. In one embodiment, when the amplitude difference between the predicted signal and the sample label corresponding to the training sample is greater than the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function can be adjusted to be greater than the weight corresponding to the signal frequency loss function; when the amplitude difference between the predicted signal and the sample label corresponding to the training sample is less than the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function can be adjusted to be less than the weight corresponding to the signal frequency loss function; and when the amplitude difference between the predicted signal and the sample label corresponding to the training sample is equal to the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function can be adjusted to the preset weights respectively.

[0049] Exemplarily, the combined loss function L ts (Y,f(x)) can be implemented as:

[0050] L ts (Y,f(x))=αL amp (Y,f(x))+βL fre (Y,f(x))

[0051] Among them, Y represents the sample label corresponding to the training sample, f(x) represents the prediction signal, and L amp (Y,f(x)) represents the signal amplitude loss function, L fre (Y, f(x)) represents the signal frequency loss function, α represents the weight corresponding to the constructed signal amplitude loss function, and β represents the weight corresponding to the signal frequency loss function. Exemplarily, α and β can be dynamically adjusted based on the output results of the signal amplitude loss function and the signal frequency loss function. For example, when the output result of the signal amplitude loss function indicates that the amplitude of the predicted signal deviates greatly from the sample label, α can be increased; when the output result of the signal frequency loss function indicates that the frequency of the predicted signal deviates greatly from the sample label, β can be increased. By constructing a combined loss function, the multi-neural network fusion model can fully learn the temporal shape and frequency information of the training samples of the sensor signal, thereby improving the prediction accuracy of the signal prediction model.

[0052] Optionally, in step S105, in the process of training the multi-neural network fusion model based on the constructed combined loss function of the signal amplitude loss function and the signal frequency loss function, the various parameters of the multi-neural network fusion model can be adjusted by minimizing the output result of the combined loss function to construct a signal prediction model based on the adjusted parameters.

[0053] According to the method for constructing a signal prediction model in accordance with one or more embodiments of the present application, the multi-neural network fusion model can be trained through a combined loss function of a signal amplitude loss function and a signal frequency loss function to construct a signal prediction model, thereby being able to fully consider the amplitude information in the time domain and the periodic information in the frequency domain of the sensor signal to accurately restore the time series shape of the signal to be predicted, thereby improving the prediction accuracy of the signal prediction model. According to the signal prediction scheme in accordance with one or more embodiments of the present application, the output signal of the sensor to be added can be accurately fitted without adding the sensor to be added, thereby being able to reduce the hardware cost of the vehicle.

[0054] Figure 2 A flow chart of a signal prediction method according to one or more embodiments of the present application is shown.

[0055] like Figure 2 As shown in FIG. 1 , in step S201 , a first sensor signal and a second sensor signal are acquired.

[0056] Optionally, in step S201, the acquired first sensor signal and second sensor signal may be a current signal and a resolver signal of a vehicle electric drive system, respectively.

[0057] In step S203 , the first sensor signal and the second sensor signal are input into a signal prediction model constructed according to the method for constructing a signal prediction model according to one or more embodiments of the present application to predict a third sensor signal.

[0058] Optionally, in step S203, when the first sensor signal and the second sensor signal are respectively the current signal and the resolver signal of the vehicle's electric drive system, the third sensor signal obtained by the signal prediction model can be implemented as a vibration signal of the vehicle's electric drive system, thereby eliminating the need to install a vibration sensor and reducing the hardware cost of the vehicle.

[0059] It should be noted that the above description of the current signal, resolver signal and vibration signal with the aid of the vehicle electric drive system is only exemplary. Without departing from the spirit and scope of the present application, when the first sensor signal, the second sensor signal and the third sensor signal are correlated with each other, the signal prediction method according to one or more embodiments of the present application can predict the third sensor signal based on the first sensor signal and the second sensor signal, thereby eliminating the need for setting up the third sensor and saving hardware costs.

[0060] According to one or more embodiments of the present application, the signal prediction method can achieve accurate fitting of the output signal of the sensor to be added without adding the sensor to be added, thereby reducing hardware costs and design complexity.

[0061] Figure 3 A schematic block diagram of a computer device for constructing a signal prediction model according to one or more embodiments of the present application is shown.

[0062] like Figure 3 As shown in FIG, a computer device 30 for constructing a signal prediction model includes a memory 310, a processor 320, and a computer program 330 stored in the memory 310 and executable on the processor 320. The processor 320 executes the computer program 330 to implement a method for constructing a signal prediction model according to one aspect of the present application.

[0063] Figure 4 A schematic block diagram of a computer device for signal prediction according to one or more embodiments of the present application is shown.

[0064] like Figure 4 As shown in FIG. 4 , the computer device 40 for signal prediction includes a memory 410, a processor 420, and a computer program 430 stored in the memory 410 and executable on the processor 420. The processor 420 executes the computer program 430 to implement a signal prediction method according to one aspect of the present application.

[0065] In addition, the present application may also be implemented as a computer storage medium, in which a program is stored for causing a computer to execute a method for constructing a signal prediction model according to one aspect of the present application. In addition, the present application may also be implemented as a computer storage medium, in which a program is stored for causing a computer to execute a method for signal prediction according to one aspect of the present application.

[0066] Computer storage media can be implemented as disks (e.g., magnetic disks, optical disks, etc.), cards (e.g., memory cards, optical cards, etc.), semiconductor memories (e.g., ROM, non-volatile memory, etc.), tapes (e.g., magnetic tapes, cassettes, etc.), etc.

[0067] In the applicable situation, hardware, software or a combination of hardware and software can be used to realize the various embodiments provided by the application. Moreover, in the applicable situation, without departing from the scope of the application, the various hardware components and / or software components set forth herein can be combined into a composite component comprising software, hardware and / or both. In the applicable situation, without departing from the scope of the application, the various hardware components and / or software components set forth herein can be divided into subcomponents comprising software, hardware or both. In addition, in the applicable situation, it is contemplated that the software component can be implemented as a hardware component, and vice versa.

[0068] Software (such as program code and / or data) according to the present application can be stored on one or more computer storage media. It is also contemplated that the software identified herein can be implemented using one or more general or special-purpose computers and / or computer systems, networked and / or otherwise. Where applicable, the order of the various steps described herein can be changed, combined into composite steps, and / or divided into sub-steps to provide the features described herein.

[0069] The embodiments and examples set forth herein are provided to best illustrate embodiments according to the present application and its specific applications, and thereby enable those skilled in the art to implement and use the present application. However, those skilled in the art will appreciate that the above description and examples are provided only for ease of illustration and example. The description set forth is not intended to cover all aspects of the present application or to limit the present application to the precise form disclosed.

Claims

1. A method for constructing a signal prediction model, characterized in that: The method comprises the following steps: Acquire a training sample of a sensor signal, wherein the training sample of the sensor signal includes a preprocessed sensor signal; Inputting the training samples of the sensor signal into a multi-neural network fusion model to output a prediction signal corresponding to the training samples of the sensor signal; as well as A signal amplitude loss function and a signal frequency loss function are constructed based on at least the predicted signal, and the multi-neural network fusion model is trained based on a combined loss function of the constructed signal amplitude loss function and the signal frequency loss function to construct the signal prediction model. 2 . The method of claim 1 , wherein the preprocessed sensor signal comprises one or more of: a preprocessed vibration signal, a preprocessed current signal, and a preprocessed resolver signal.

3. The method according to claim 1, wherein the preprocessed sensor signal is obtained by: Get sensor signals; Performing noise reduction processing on the sensor signal using a variational mode decomposition algorithm; and performing one or more of a time domain processing operation, a frequency domain processing operation and a smoothing processing operation on the noise reduction processed sensor signal to obtain the preprocessed sensor signal, The method of performing noise reduction processing on the sensor signal by using a variational mode decomposition algorithm includes: Decomposing the sensor signal into a set of intrinsic mode functions with different center frequencies and finite bandwidths using a variational mode decomposition algorithm; identifying a noise component in the sensor signal based on the set of intrinsic mode functions; as well as removing the identified noise component in the sensor signal from the sensor signal to perform noise reduction processing on the sensor signal, The time domain processing operation, the frequency domain processing operation and the smoothing processing operation are respectively performed in the following manners: Acquire the denoised sensor signal values ​​at a plurality of moments and perform the time domain processing operation on the denoised sensor signal based on the correlation between the plurality of moments; Acquire frequency domain information of the noise reduction processed sensor signal values ​​at multiple moments and perform the frequency domain processing operation on the noise reduction processed sensor signal based on an inverse transform of a main frequency component of the frequency domain information; as well as The smoothing operation is performed on the noise-reduced sensor signal by acquiring the noise-reduced sensor signal values ​​at a plurality of time points and smoothing the sensor signal values ​​at the plurality of time points.

4. The method according to claim 1, wherein the multi-neural network fusion model comprises at least two of the following neural network models: a neural network model based on a gradient boosting decision tree, a neural network model for tabular data, and a time-recurrent neural network model, The step of inputting the training samples of the sensor signals into a multi-neural network fusion model to output a prediction signal corresponding to the training samples of the sensor signals comprises: Inputting the training samples of the sensor signal into each neural network model of the multi-neural network fusion model respectively to obtain output results of each neural network model respectively; as well as The output results of each neural network model are processed by weighted average to output a prediction signal corresponding to the training sample of the sensor signal.

5. The method according to claim 1, wherein constructing a signal amplitude loss function based at least on the predicted signal comprises: Determining the difference between the prediction signal and the sample label corresponding to the training sample using a dynamic time warping algorithm; as well as Constructing the signal amplitude loss function based on the difference, Wherein constructing a signal frequency loss function based at least on the prediction signal comprises: Performing Fourier transform on the prediction signal and the sample label corresponding to the training sample to respectively obtain the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample; Determining difference information between the main frequency component of the prediction signal and the main frequency component of the sample label corresponding to the training sample; and Constructing the signal frequency loss function based on the difference information, The combined loss function is obtained in the following way: The constructed signal amplitude loss function and signal frequency loss function are weightedly summed to obtain the combined loss function. During the weighted summation, the weight corresponding to the constructed signal amplitude loss function and the weight corresponding to the signal frequency loss function are dynamically adjusted in the following manner: In response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being greater than the frequency difference between the predicted signal and the sample label corresponding to the training sample, adjusting the weight corresponding to the constructed signal amplitude loss function to be greater than the weight corresponding to the signal frequency loss function; In response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being smaller than the frequency difference between the predicted signal and the sample label corresponding to the training sample, adjusting the weight corresponding to the constructed signal amplitude loss function to be smaller than the weight corresponding to the signal frequency loss function; as well as In response to the amplitude difference between the predicted signal and the sample label corresponding to the training sample being equal to the frequency difference between the predicted signal and the sample label corresponding to the training sample, the weight corresponding to the constructed signal amplitude loss function and the weight corresponding to the signal frequency loss function are respectively adjusted to preset weights.

6. A computer device for constructing a signal prediction model, characterized in that: The computer device comprises: Memory; a processor coupled to the memory; and A computer program stored on the memory and running on the processor, wherein the running of the computer program results in the execution of the method for constructing a signal prediction model according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium comprises instructions, which, when executed, execute the method for constructing a signal prediction model according to any one of claims 1 to 5.

8. A signal prediction method, characterized in that: The method comprises the following steps: Acquire a first sensor signal and a second sensor signal; The first sensor signal and the second sensor signal are input into a signal prediction model constructed according to the method for constructing a signal prediction model according to any one of claims 1 to 5 to predict a third sensor signal.

9. A computer device for signal prediction, characterized in that: The computer device comprises: Memory; a processor coupled to the memory; and A computer program stored on the memory and running on the processor, the execution of the computer program causing the method for signal prediction according to claim 8 to be performed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium comprises instructions, which when executed, perform the method for signal prediction according to claim 8.

Citation Information

Cited By

  • Sensor signal prediction method, system and equipment for humanoid robot

    CN120995103A