A linear prediction method based on orthogonal compensation

Abstract

The application discloses a linear prediction method based on orthogonal compensation, relates to the technical field of data processing and machine learning, and calculates pixel offset between two adjacent images after initial linear compensation and corresponding offset data through an image matching algorithm; offset data is used as a difference value between an observation value and a prediction value, and a reference value of reference data is used as a state quantity; if the offset is normal, an interactive multiple model based on Bayesian theory is used to fuse real-time data, a prediction value of the interactive multiple model based on Bayesian theory is used as input, linear data error of a system linear data error model is used as output, and the system linear data error model is subjected to model training; if the offset is abnormal, a compensation data information is used to calculate a reference value, so that modified data information is obtained, and when an abnormal type that has not been sufficiently learned is encountered, the recognition capability can be greatly reduced, and even a missed detection or a false detection can occur.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and machine learning, in particular to a linear prediction method based on orthogonal compensation. BACKGROUND

[0002] With the continuous progress of science and technology, data processing and machine learning as an important training model data source of modern electronic equipment, its performance and security detection becomes crucial. In the performance analysis of machines and devices, accurately identifying energy loss anomalies is a key factor in evaluating their quality and service life. However, traditional data processing and machine learning analysis methods often rely on a large number of learning examples, and in the face of complex and variable abnormal situations, the recognition accuracy and efficiency are limited to a certain extent. Existing data processing and machine learning analysis techniques, although they can identify anomalies by comparing and analyzing data processing and machine learning data, in actual operation, due to the complexity and diversity of machine learning performance data, these techniques often have difficulty accurately capturing all types of abnormal losses. In addition, when encountering an abnormal type that has not been fully learned, the recognition ability of existing technologies will be greatly reduced, and even may miss or misjudge the situation. Based on this, a linear prediction method based on orthogonal compensation is proposed to solve the above problems. SUMMARY

[0003] The purpose of the present application is to provide a linear prediction method based on orthogonal compensation to solve the problems raised in the background art.

[0004] By adopting the above technical solution, a linear prediction method based on orthogonal compensation includes:

[0005] S1: Collect real-time data and set the reference value of the reference data;

[0006] Based on the reference value of the reference data;

[0007] S2: Fill in the data of the prediction data until the data length of the to-be-predicted time domain data reaches the expected target length;

[0008] S3: Calculate the compensation value of the initial inertia data by a linear compensation algorithm and perform initial linear compensation;

[0009] S4: Calculate the pixel offset between the adjacent two frames of images after initial linear compensation and the corresponding offset data by an image matching algorithm;

[0010] The offset data is used as the difference between the observation value and the prediction value, and the reference value of the reference data is used as the state quantity; if the offset quantity is normal, the real-time data is fused by using the interactive multiple model based on the Bayesian theory, and the prediction value of the interactive multiple model based on the Bayesian theory is used as the input, and the linear data error of the system linear data error model is used as the output, so as to train the model of the system linear data error model;

[0011] If the offset quantity is abnormal, the recursive orthogonal search algorithm is used to model the system linear data error model online, the linear data error of the system linear data error model is estimated in combination with the prediction value of the interactive multiple model, so as to compensate the output data information, and the compensated data information is used to calculate the reference value, so as to obtain the corrected data information.

[0012] According to the above technical scheme, the interactive multiple model is obtained by the following method:

[0013] A plurality of first time domain data are obtained;

[0014] Part of the data in each first time domain data is extracted, and a plurality of second time domain data corresponding to the plurality of first time domain data are constructed respectively, and the data length of each second time domain data is the same as that of the corresponding first time domain data;

[0015] The plurality of first time domain data are subjected to Fourier transform processing to obtain a plurality of first frequency domain data, and the plurality of second time domain data are subjected to Fourier transform processing to obtain a plurality of second frequency domain data;

[0016] The plurality of second frequency domain data and the plurality of first frequency domain data are used as the training set of the preset convolutional neural network model, the preset convolutional neural network model is trained, when the training times of the preset convolutional neural network model exceed the preset number of times, the training is stopped, and the second frequency domain data in the training set is input into the convolutional neural network model after the training is stopped, and the output third frequency domain data is obtained;

[0017] The third frequency domain data is subjected to inverse Fourier transform processing to obtain third time domain data.

[0018] According to the above technical scheme, the first effective data of the third time domain data is compared with the second effective data of the corresponding second time domain data, when the data length of the first effective data is greater than that of the second effective data, it is determined that the data length of the two satisfies the preset verification condition;

[0019] When the data length of the first valid data is less than or equal to the data length of the second valid data, it is determined that the data length of the two does not satisfy the preset check condition, wherein the valid data is data in the part of data extracted from the corresponding first time domain data; after the check succeeds, it is determined that the trained convolutional neural network model converges, and an orthogonal compensation data linear prediction model is obtained.

[0020] According to the above technical solution, the preset convolutional neural network model comprises K convolutional layers, each convolutional layer is provided with P pooling layers, the input of each convolutional layer is the output of all other layers in front of the convolutional layer, each convolution kernel size is R, and the input-output relationship of the convolutional neural network model can be defined as:

[0021] y=f(x, alpha);

[0022] Wherein, x is the input layer of the convolutional neural network, y is the output layer of the convolutional neural network, x and y are both frequency domain data, alpha is the parameter of the convolutional neural network model, and f(x, alpha) represents the nonlinear mapping relationship of the convolutional neural network.

[0023] According to the above technical solution, a large amount of frequency domain data easy to analyze is used as the input and output of the preset convolutional neural network model, the preset convolutional neural network model is trained, the training rate of the model is improved, the training is stopped when the model reaches the convergence condition, and a trained nuclear magnetic resonance data linear prediction model is obtained.

[0024] According to the above technical solution, the convergence condition of the convolutional neural network model further comprises that the number of model training reaches a preset number threshold, and the check condition is satisfied, a plurality of second frequency domain data and a plurality of first frequency domain data are used as a training set of the preset convolutional neural network model, the preset convolutional neural network model is trained, the training is stopped when the number of training of the preset convolutional neural network model exceeds the preset number value, the trained convolutional neural network model is checked through the training set, and it is determined that the trained convolutional neural network model converges when the check succeeds.

[0025] According to the above technical solution, the algorithm of the interactive multi-model tracking mainly comprises the following four steps, i.e., input model interaction, filter filtering, probability model updating and new data fusion.

[0026] Compared with the prior art, the beneficial effects of the present application are:

[0027] 1. A linear prediction method based on orthogonal compensation, pixel displacement between two adjacent frames of images after initial linear compensation and corresponding displacement data are calculated by an image matching algorithm; the difference between the displacement data and the prediction value is taken as the observation value, and the reference value of the reference data is taken as the state quantity; if the displacement is normal, the real-time data is fused using an interactive multiple model based on Bayesian theory, and the prediction value of the interactive multiple model based on Bayesian theory is taken as the input, and the linear data error of the system linear data error model is taken as the output, and the system linear data error model is trained; if the displacement is abnormal, the recursive orthogonal search algorithm is used to model the system linear data error model online, and the linear data error of the system linear data error model is estimated in combination with the prediction value of the interactive multiple model, so as to compensate the output data information, and the calibrated value is calculated by using the compensated data information, so as to obtain the corrected data information, and when an abnormal type that is not fully learned is encountered, the recognition ability is greatly reduced, and even the situation of missed detection or false detection may occur.

[0028] 2. A linear prediction method based on orthogonal compensation, the algorithm mainly includes the following four steps, that is, input model interaction, filter filtering, probability model updating and new data fusion, the filtering recursive process of the interactive multiple model Kalman filter tracking algorithm, the interactive output of each time is taken as the interactive input value of the next time, so as to complete the cycle of the whole filtering tracking process, solve the problem that the linear motion compensation is invalid due to inaccurate recognition ability estimation or estimation means cannot be used in the previous method, and the data processing and the machine can be automatically and quickly corrected in the passive state, so as to perform accurate linear motion compensation. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The figure is a schematic diagram of the interactive multiple model algorithm flow structure of the application;

[0030] Figure 2 The figure is a CV and CA simulation running diagram of the application;

[0031] Figure 3 The figure is a model weight change diagram of the application;

[0032] Figure 4 The figure is a CV and CA probability change diagram of the application;

[0033] Figure 5 The figure is a comparison 1 diagram of the MM algorithm and the Kalman filter algorithm of the application;

[0034] Figure 6 The figure is a comparison 2 diagram of the MM algorithm and the Kalman filter algorithm of the application. DETAILED DESCRIPTION

[0035] Clearly, the embodiments described are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort are within the protection scope of the present application.

[0036] Embodiment one:

[0037] A preferred embodiment of the linear prediction method based on orthogonal compensation provided by the present application is shown in the following figure: Figures 1 to 6 The linear prediction method based on orthogonal compensation comprises the following steps:

[0038] S1: Collect real-time data and set a reference value of reference data;

[0039] based on the reference value of the reference data;

[0040] S2: Fill data to the predicted data until the data length of the to-be-predicted time domain data reaches the expected target length;

[0041] S3: Obtain a compensation value of initial inertia data by a linear compensation algorithm and perform initial linear compensation;

[0042] S4: Calculate the pixel offset between the adjacent two images after initial linear compensation and the corresponding offset data by an image matching algorithm;

[0043] Take the difference between the offset data and the predicted value as the state quantity, and if the offset is normal, use the interactive multiple model based on the Bayesian theory to fuse the real-time data, and take the predicted value of the interactive multiple model based on the Bayesian theory as the input and the linear data error of the system linear data error model as the output to train the system linear data error model;

[0044] If the offset is abnormal, use the recursive orthogonal search algorithm to perform online modeling on the system linear data error model, estimate the linear data error of the system linear data error model in combination with the predicted value of the interactive multiple model, compensate the output data information, calculate the reference value by using the compensated data information, and obtain the corrected data information.

[0045] In the above technical solution, specifically,

[0046] Further,

[0047] wherein,

[0048] The data length of the plurality of first time domain data is the same and is L, and the first time domain data is defined as F1, and the expression of F1 is as follows:

[0049]

[0050] wherein N represents the number of peaks of the spectrum, A n represents the amplitude of the nth peak, Tn represents the relaxation time of the nth peak, t represents the sampling time, w n represents the frequency of the nth peak.

[0051] Extracting partial data in each first time domain data, and constructing a plurality of second time domain data corresponding to the plurality of first time domain data respectively.

[0052] The data length of each second time domain data is the same as that of the corresponding first time domain data, the extracted partial data is continuous data, and the continuous data is extracted from the first data of the first time domain data, and the data length of the continuous data is less than that of the corresponding first time domain data. The data length of the second time domain data is the same as that of the first time domain data, and the partial data in the first time domain data is half of the data of the first time domain data, i.e. the second time domain data only contains the first half of the data of the first time domain data.

[0053] According to the preset threshold range, a plurality of extracted data are obtained, and data filling is performed on the plurality of extracted data, so that the data length of each extracted data reaches the data length of the corresponding first time domain data, and a plurality of second time domain data are obtained; wherein the preset threshold range is a range starting from the first data of the time domain data, and the data filling process is a process of increasing the value of 0.

[0054] The first half of the data in the second time domain data is extracted from the first time domain data, and the amplitude of the first half of the data is the same as that of the corresponding first time domain data. In order to control the range of linear prediction and improve the accuracy of model training, the data length of the second time domain data is increased to the same as that of the corresponding first time domain data, and the data value of the increased part is set to 0, as shown in the input second time domain data.

[0055] The data length of the first time domain data is 10, the first time domain data includes (1, 2, 3, 4, 5, 6, 7, 8, 9, 10), and the preset threshold range is 1-5. The extracted data is (1, 2, 3, 4, 5), and then the extracted data is filled with data with a value of 0, so that the data length of the extracted data reaches the data length 10 of the corresponding first time domain data, and the second time domain data is obtained. The second time domain data is (1, 2, 3, 4, 5, 0, 0, 0, 0, 0).

[0056] By setting the value of the filled data to 0, the model can mainly predict the data of the part with the value of 0 during training, so that the model has the ability to convert the data of the part with the value of 0 into predicted data.

[0057] The manner of constructing the second time domain data further includes: keeping the first m points of the first time domain data F1 unchanged, setting the points after the m points to 0, and marking them as F2, and the expression of F2 is as follows:

[0058] F2=F1;

[0059] ID2(m+1:1)=0;

[0060] Where 1 represents the index of the last element of F2.

[0061] The plurality of first time domain data is subjected to Fourier transform processing to obtain a plurality of first frequency domain data, and the plurality of second time domain data is subjected to Fourier transform processing to obtain a plurality of second frequency domain data.

[0062] By performing Fourier transform processing on the plurality of first time domain data and the plurality of second time domain data, time domain data that is difficult to process can be converted into frequency domain data that is easy to analyze, so that the frequency domain data that is easy to analyze can be directly processed or processed, thereby speeding up the prediction efficiency of the data. The plurality of second frequency domain data is used as input data, the plurality of first frequency domain data is used as output data, a preset convolutional neural network model is trained until convergence, and a data linear prediction model is obtained.

[0063] The preset convolutional neural network model includes K convolutional layers, each convolutional layer has P pooling layers, the input of each convolutional layer is the output of all other layers in front of the convolutional layer, each convolution kernel size is R, and the input-output relationship of the convolutional neural network model can be defined as:

[0064] y=f(x, a);

[0065] Where x is the input layer of the convolutional neural network, y is the output layer of the convolutional neural network, x and y are frequency domain data, a is a parameter of the convolutional neural network model, and y=f(x, a) represents a nonlinear mapping relationship of the convolutional neural network.

[0066] A large amount of frequency domain data that is easy to analyze is used as the input and output of the preset convolutional neural network model, and the preset convolutional neural network model is trained, which can improve the training rate of the model. Training is stopped when the model reaches the convergence condition, and a trained nuclear magnetic resonance data linear prediction model is obtained.

[0067] The convergence condition of the convolutional neural network model includes that the accuracy of the model reaches a certain value, for example, when the output accuracy of the model reaches 98%, it is determined that the model has reached the convergence condition, that is, the training is stopped, and the model with an output accuracy of 98% is taken as the trained nuclear magnetic resonance data linear prediction model.

[0068] The convergence condition of the convolutional neural network model also includes that the number of times of training of the model reaches a preset number threshold, and a verification condition is met, a plurality of second frequency domain data and a plurality of first frequency domain data are taken as a training set of the preset convolutional neural network model, the preset convolutional neural network model is trained, when the number of times of training of the preset convolutional neural network model exceeds the preset number value, the training is stopped, the convolutional neural network model after stopping training is verified through the training set, when the verification is successful, it is determined that the convolutional neural network model after stopping training converges. When the number of times of training of the model reaches 50000 times, the training is stopped, and the model is verified by using the training set, when the model passes the verification, it is determined that the model meets the convergence condition.

[0069] Embodiment two:

[0070] On the basis of embodiment one, a preferred implementation of a linear prediction method based on orthogonal compensation provided by the application is shown in the figure: a CV model is established, and the state space of the CV model can be represented as:

[0071]

[0072] Then the transfer function is:

[0073]

[0074] Alternatively, the complete state space of the CV model can also be represented as:

[0075]

[0076] The matrix is represented as:

[0077]

[0078] Further, the main idea of the interactive multiple model (IMM) control algorithm is the automatic identification and switching between models based on the Bayesian theory: at any tracking time, real-time maneuvering model detection is performed by setting model filters corresponding to the number of possible target models, a weight coefficient and a model update probability are set for each filter, and finally the current optimal estimation state is calculated by weighting, so that the purpose of model adaptive tracking is achieved.

[0079] The following is the IMM algorithm flow Figure 1 :

[0080] As can be seen from the figure, the algorithm of the interactive multi-model tracking mainly contains the following four steps, namely, input model interaction, filter filtering, probability model updating and new data fusion.

[0081] Suppose that r motion models are set in the interactive multi-model algorithm, and the state equation is as follows:

[0082] X(k+1)=A j X(k)+W j (k);j=1,2,3...

[0083] Where X(k) is the state vector of the system, A j is the state transition matrix of the target, W j (k) is a Gaussian white noise with mean 0 and covariance matrix Q j .

[0084] The observation equation of any model j in the model set is:

[0085] Z(k)=H j X(k)+V j (k);

[0086] Where Z(k) is the measurement vector, H j is the observation matrix of model j, V j (k) is a white noise matrix with mean 0 and covariance R j .

[0087] The model transition matrix can be expressed as:

[0088]

[0089] Where P ij represents the transition matrix from model i to model j;

[0090] 1. Input interaction;

[0091] According to the system state estimation and covariance estimation at the previous time, the current measurement value Z(k) is predicted, and then the model is reinitialized and calculated, wherein the new initial value is obtained through the Markov operation matrix between different models, and the state optimal estimation value and estimation covariance matrix of model k-1 time are set as follows:

[0092] X j (k-1|k-1) and p j (k-1|k-1);

[0093] Then the initial conditions of the model after interaction are further derived:

[0094]

[0095]

[0096] wherein:

[0097]

[0098]

[0099] 2. filtering of the filter;

[0100] Since the Kalman filtering algorithm has excellent tracking effect, the algorithm is usually used in multi-model interactive tracking to filter and predict the target, and the input of the j model at k-1 time is the optimal estimation X oj (k-1|k-1) and the update covariance p 0j (k-1|k-1), the tracking process is similar to the single model Kalman filtering algorithm, and the Kalman filtering process of the IMM algorithm is as follows:

[0101] The one-time prediction equation and the prediction covariance matrix of the Kalman filter are:

[0102] X I (k-1|k-1) = A j X oj (k-1|k-1);

[0103] p j (k|k-1) = A j p oj (k-1|k-1)A T j + Q j ;

[0104] The Kalman filter gain is:

[0105] K j (k) = p j (k|k-1)H T [Hp j (k|k-1)H T + Q j ] -1 ;

[0106] The system state estimation and covariance estimation corresponding to the output of a single filter at k time are respectively:

[0107] X j (k|k) = X j (k|k-1) + K j (k)[Z(k) - HX j (k|k-1)];

[0108] p j(k|k) = [I - K j (k)H]p j (k|k-1);

[0109] 3. Update of model probability;

[0110] In IMM algorithm, the update of model directly affects the efficiency of the algorithm. The algorithm uses maximum likelihood function method to realize the update of model: the weight of the model which is most suitable for tracking is given by calculating the similarity between the current model and the state of the current moving target.

[0111] The maximum likelihood function which is most matched with model j at time k is:

[0112]

[0113] where S j (k) = Hp j (k|k-1)H T + R, d j (k) = Z j (k) - H j X j (k|k-1);

[0114] Then the update probability of model j can be obtained as:

[0115]

[0116] where c is a normalization constant, i.e.:

[0117]

[0118] 4. New fused data;

[0119] The final output result of the interaction data at time k is given according to the tracking result calculated by each model and the weight of model matching, i.e.:

[0120] The final output result of the interaction data at time k is given according to the tracking result calculated by each model and the weight of model matching, i.e. the overall estimation value:

[0121]

[0122] The overall covariance:

[0123]

[0124] The above is one filtering recursion process of the interactive multiple model Kalman filtering tracking algorithm, each time the interaction output of the last time is taken as the interaction input value of the next time, so as to complete the cycle of the whole filtering tracking process.

[0125] It is to be noted that, in the present text, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Finally, it should be noted that the above-mentioned preferred embodiments of the present application are merely intended to illustrate the present application and are not intended to limit the present application, and that the technical solutions described in the foregoing embodiments can be modified or equivalent replacements can be made to some of the technical features by those skilled in the art, without departing from the spirit and principle of the present application. Any modifications, equivalent replacements, improvements, and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A linear prediction method based on orthogonal compensation, characterized in that, The method comprises the following steps: S1: collecting real-time data and setting a reference value of reference data; based on the reference value of the reference data; S2: filling data to the predicted data until the data length of the to-be-predicted time domain data reaches the expected target length; S3: obtaining a compensation value of initial inertia data through a linear compensation algorithm and performing initial linear compensation; S4: calculating a pixel offset between adjacent two images after initial linear compensation and corresponding offset data through an image matching algorithm; taking the difference between the offset data as the observation value and the prediction value, and taking the reference value of the reference data as the state quantity; if the offset is normal, using an interactive multiple model based on Bayesian theory to fuse the real-time data, and taking the prediction value of the interactive multiple model based on the Bayesian theory as the input and the linear data error of the system linear data error model as the output to train the system linear data error model; if the offset is abnormal, using a recursive orthogonal search algorithm to perform online modeling on the system linear data error model, combining the prediction value of the interactive multiple model to estimate the linear data error of the system linear data error model, thereby compensating the output data information, and calculating the reference value by using the compensated data information, thereby obtaining the corrected data information.

2. The linear prediction method based on orthogonal compensation according to claim 1, characterized in that: The interactive multiple model is obtained by the following method: obtaining a plurality of first time domain data; extracting part of the data in each first time domain data to construct a plurality of second time domain data corresponding to the plurality of first time domain data, and the data length of each second time domain data is the same as that of the corresponding first time domain data; performing Fourier transform processing on the plurality of first time domain data to obtain a plurality of first frequency domain data, and performing Fourier transform processing on the plurality of second time domain data to obtain a plurality of second frequency domain data; taking the plurality of second frequency domain data and the plurality of first frequency domain data as a training set of a preset convolutional neural network model, training the preset convolutional neural network model; when the training times of the preset convolutional neural network model exceed a preset number of times, stop training; input the second frequency domain data in the training set into the convolutional neural network model after stopping training to obtain output third frequency domain data; performing inverse Fourier transform processing on the third frequency domain data to obtain third time domain data.

3. The linear prediction method based on orthogonal compensation according to claim 2, characterized in that: comparing the first effective data of the third time domain data with the second effective data of the corresponding second time domain data; when the data length of the first effective data is greater than the data length of the second effective data, it is determined that the data lengths of the two meet the preset verification condition; when the data length of the first effective data is less than or equal to the data length of the second effective data, it is determined that the data lengths of the two do not meet the preset verification condition; wherein the effective data is the data in the part of the data extracted from the corresponding first time domain data; when the verification is successful, it is determined that the convolutional neural network model after stopping training converges, and an orthogonal compensation data linear prediction model is obtained.

4. The linear prediction method based on orthogonal compensation according to claim 2, characterized in that: The preset convolutional neural network model comprises K convolutional layers, each convolutional layer is provided with P pooling layers, the input of each convolutional layer is the output of all other layers in front of the convolutional layer, each convolutional kernel size is R, and an input-output relationship of the convolutional neural network model can be defined as y=f(x, a). y=f(x, a); Wherein, x is an input layer of the convolutional neural network, y is an output layer of the convolutional neural network, x and y are both frequency domain data, a is a parameter of the convolutional neural network model, and f(x, a) represents a nonlinear mapping relationship of the convolutional neural network.

5. The linear prediction method based on orthogonal compensation according to claim 4, characterized in that: A large amount of frequency domain data easy to analyze is used as input and output of the preset convolutional neural network model, the preset convolutional neural network model is trained, the rate of model training can be improved, the model is stopped when the model reaches a convergence condition, and a trained nuclear magnetic resonance data linear prediction model is obtained.

6. The linear prediction method based on orthogonal compensation according to claim 2, characterized in that: The convergence condition of the convolutional neural network model also comprises that the number of model training reaches a preset number threshold, and a verification condition is met, a plurality of second frequency domain data and a plurality of first frequency domain data are used as a training set of the preset convolutional neural network model, the preset convolutional neural network model is trained, the training is stopped when the number of training of the preset convolutional neural network model exceeds the preset number value, the convolutional neural network model after stopping training is verified through the training set, and the convolutional neural network model after stopping training is determined to be converged when the verification is successful.

7. The linear prediction method based on orthogonal compensation according to claim 1, characterized in that: The algorithm of the interactive multi-model tracking mainly comprises the following four steps, namely input model interaction, filter filtering, probability model updating and new data fusion.

Images