CN110309491A - A Transient Phase Division Method and System Based on Local Gaussian Mixture Model - Google Patents

Abstract

The invention discloses a phase division method and system based on a local Gaussian mixture model, comprising: S1, collecting samples and creating a historical training data set; S2, selecting some samples from the historical training data set to create a first Gaussian distribution model to Determine the first steady-state phase; S3. Based on the previously determined Gaussian model, create a Gaussian mixture model containing two Gaussian components to determine the next steady-state phase; S4. Based on the determined two adjacent steady-state phases The state phase model determines the possible transient phase between the two steady state phases; S5, repeating S3 and S4 to complete the phase division of all sample data. The invention greatly reduces redundant calculation, improves calculation efficiency, adopts a step-by-step update strategy, and determines the steady-state phase and the transient phase step by step according to the sampling time sequence, and the number of phase divisions does not need to be pre-specified and divided. As a result, there is no need for subsequent processing and other advantages.

Description

A Transient Phase Division Method and System Based on Local Gaussian Mixture Model

technical field

The invention relates to the technical field of batch process statistical modeling, in particular to a method and system for transient phase division in a multiphase batch process.

Background technique

Batch process is a very common production method in modern industry and is widely used in fine chemical, pharmaceutical, metallurgy and semiconductor industries. With the development of technology and the diversification of requirements, the batch process has become more and more complex. The direct manifestation is that a batch process contains multiple different operation stages, or multiple different reaction/change stages, so that The process is called a multiphase batch process, and each such phase is called a phase. For example, in the penicillin fermentation process, suppose the time of a penicillin fermentation process is 400h, the first 45h is the pre-cultivation stage, and the last 355h is the feed-feeding stage, that is, the raw materials are added to the reaction kettle from the 45th hour. In terms of reaction mechanism, a typical penicillin fermentation process can be divided into four phases (stages), including lag stage, exponential growth stage, stable stage and autolysis stage. The penicillin fermentation process is a typical slow time-varying process. The transition between different phases is not a mutation process, but a slowly changing process. Therefore, the transition between different phases is not so obvious, and the following one will occur. In this case, the samples between two steady-state phases not only retain the characteristics of the first steady-state phase, but also contain the characteristics of the next new steady-state phase. How to accurately and reasonably divide a multi-phase batch process into different phases is beneficial to further understanding of the process mechanism and improving the accuracy of process modeling.

At present, there have been some research results for multiphase division, including the multiphase principal component analysis method similar to the exhaustive method, using the repetition factor index for batch division, but for the process with no obvious change inflection point, this is method is no longer applicable. The clustering method is widely used in the phase division of the batch process. However, in addition to pre-specifying the number of division categories based on the K-means algorithm, the temporal relationship between samples is not considered, resulting in confusion in the division results and the need for further follow-up. Handling, not easy to explain and other issues. Therefore, in the phase, the timing relationship between samples is an important factor that cannot be ignored, and it is necessary not only to divide the multi-stable phase effectively, but also to accurately divide the transient phase. That is to say, the above two methods do not consider the problems of timing and transient phase division.

SUMMARY OF THE INVENTION

Based on this, a new phase division method based on local Gaussian mixture model is proposed in view of the shortcomings of the existing phase division methods that do not consider timing and transient phase division.

A phase division method based on a local Gaussian mixture model, comprising the following steps:

S1. Collect samples and create a historical training data set;

S2. Select some samples from the historical training data set to create the first Gaussian distribution model according to the sampling time sequence to determine the first steady-state phase;

S3. Based on the previously determined Gaussian model, create a Gaussian mixture model containing two Gaussian components to determine the next steady state phase;

S4. Determine a possible transient phase between the two steady-state phases based on the determined two adjacent steady-state phase models;

S5. Repeat S3 and S4 to complete the phase division of all sample data.

Optionally, in one embodiment, the selecting partial samples from the historical training data set to create the first Gaussian distribution model to determine the first steady-state phase data includes:

S21. Select the first N ₁ samples from the historical training data set in turn and calculate their mean and variance to obtain a corresponding Gaussian distribution model p(x|1), where p(x|1) represents the first Gaussian distribution The probability density function of the model, x represents the sample data collected;

S22, extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that satisfy the first verification condition and the corresponding sample points that satisfy the first verification condition serial number marked as The verification condition is

where ρ is a pre-specified threshold;

S23. Determine whether N ₁ is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let And return to step S21 to iterate until N ₁ is equal to

Optionally, in one embodiment, creating a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model to determine the next steady state phase includes:

S31. Based on the previously determined Gaussian model, create and train a mixture model containing two Gaussian distribution functions, without loss of generality, assuming that the first c-1 steady-state phases have been determined, and c is an integer greater than or equal to 2, The formula corresponding to the mixed model is

p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c)

Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm;

S32, extracting sample points to verify the steady-state phase of the hybrid model,

That is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that meet the second verification condition and mark the sequence number corresponding to the sample point that meets the second verification condition as The verification condition is

where ρ is a pre-specified threshold;

S33. Judgment Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let And return to step S31 to iterate until is equal to N _c +N _c-1 .

Optionally, in one of the embodiments, the determining, based on the determined two adjacent steady-state phase models, a transient phase that may exist between the two steady-state phases includes:

From the two determined two stable phases X _c-1 and X _c , start the test from the first sample point of the c-th stable phase and find out the continuous sample points that satisfy p(x _n |c)<ρ , denoted as the transient phase X _c-1,c .

In addition, in order to solve the shortcomings of traditional techniques, a phase division system based on local Gaussian mixture model is also proposed.

A phase division system based on a local Gaussian mixture model, including:

a collection unit for collecting samples and creating historical training data sets;

a first Gaussian distribution creation unit, which is configured to select partial samples from the historical training data set to create a first Gaussian distribution model in the order of sampling time to determine the first steady-state phase data;

The Gaussian mixture model creation unit is used to create a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model to determine the next steady state phase, and cooperate with the transient phase acquisition unit to complete the analysis of all sample data. phase division;

A transient phase acquisition unit, configured to determine corresponding transient phase data based on two adjacent steady-state phase data.

Optionally, in one embodiment, the first Gaussian distribution creation unit includes:

A first data acquisition module, which is used to sequentially select the first N ₁ samples from the historical training data set and calculate their mean and variance to obtain a corresponding Gaussian distribution function p(x|1), where p(x|1 ) represents the probability density function of the first Gaussian distribution model, and x represents the collected sample data;

The first steady-state phase verification module is used for extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that satisfy the first verification condition and will satisfy the first The serial number corresponding to the sample point of the verification condition is marked as The verification condition is

where ρ is a pre-specified threshold;

The _first steady-state phase determination module, which is used to determine whether N1 is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let and iterates again by the _first steady-state phase verification module until N1 equals

Optionally, in one embodiment, the Gaussian mixture model creation unit includes:

The second data acquisition module is used to create and train a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model, assuming that the first c-1 steady-state phases have been determined, and c is greater than or equal to 2 Integer, the formula corresponding to the mixed model is

p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c)

Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm;

The above-mentioned mixed model is trained by using the maximum expectation algorithm, namely the EM algorithm, wherein, denoting X _m ={X _c-1 ,X _c }, the corresponding training difference θ _c ={α _c-1 ,α _c ,μ _c ,Σ _c };

The second steady-state phase verification module is used to extract sample points to perform steady-state phase verification on the mixed model, that is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that satisfy the second verification condition and mark the serial number corresponding to the sample point that satisfies the second verification condition as The verification condition is

where ρ is a pre-specified threshold;

The second steady-state phase determination module, which is used to determine Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let and iterates again by the second steady-state phase verification module until is equal to N _c +N _c-1 .

Optionally, in one of the embodiments, the processing procedure of the transient phase acquisition unit includes: from the two determined two steady-state phases X _c-1 and X _c , from the c-th steady-state phase A sample point starts to test and find out the continuous sample points that satisfy p(x _n |c)<ρ, which is recorded as the transient phase X _c-1,c .

In addition, in order to solve the shortcomings of the traditional technology, a computer-readable storage medium is also proposed, which includes computer instructions, when the computer instructions are executed on the computer, the computer can execute the method.

In the implementation of the embodiments of the present invention, in addition to solving the deficiencies in the existing phase division methods that do not consider timing and transient phase division, the present invention also has the following beneficial effects: namely (1) the present invention uses a Gaussian distribution from the perspective of a An independent Gaussian distribution describes a steady-state phase, and a mixture model of two adjacent Gaussian distributions is used to describe the transient phase, so that the phase division method can effectively divide the steady-state phase and determine the transient phase at the same time; (2 ) Each iteration of the present invention only uses local data for modeling verification, which greatly reduces redundant calculations and improves computational efficiency; (3) the present invention adopts a step-by-step update strategy, and according to the sampling time sequence, step by step determines The steady-state phase and the transient phase have the advantages that the number of phase divisions does not need to be pre-specified and the division results do not require subsequent processing.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

in:

Fig. 1a is a schematic diagram of initial model update in one embodiment;

Fig. 1b is a schematic diagram of hybrid model updating in one embodiment;

2 is a schematic diagram of phase division of a local Gaussian mixture model in one embodiment;

Fig. 3 is a schematic diagram of penicillin fermentation process in one embodiment;

FIG. 4 is a flow chart of core steps in one embodiment.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terms used herein in the description of the present invention are for the purpose of describing specific embodiments only, and are not intended to limit the present invention. It will be understood that the terms "first", "second", etc., as used herein, may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, a first element could be termed a second element, and, similarly, a second element could be termed a first element, without departing from the scope of the present application. Both the first element and the second element are elements, but they are not the same element.

In view of the lack of consideration of timing and transient phase division in the existing phase division methods, in this embodiment, a phase division method based on a local Gaussian mixture model is proposed, which uses an improved local Gaussian mixture model. The method uses step-by-step probability modeling to perform phase division on the multiphase batch process; that is, each time a local Gaussian mixture model is established, that is, an initial model with only one Gaussian component and a mixture model containing two Gaussian components, The steady-state phase and the transient phase in the process can be simultaneously determined by an iterative method. Specifically, as shown in Figure 4, the method includes the following steps:

S1, collect samples and create a historical training data set; in some specific embodiments, by collecting batch process data and expand into Get historical training datasets;

S2. Select some samples from the historical training data set according to the sampling time sequence to create the first Gaussian distribution model to determine the first steady-state phase data. The purpose of this step is to establish an initial model containing only one Gaussian component. A schematic diagram of model update is shown in Figure 1(a); in some specific embodiments, selecting partial samples from the historical training data set to create the first Gaussian distribution model to determine the first steady-state phase data includes:

S21. Select the first N ₁ samples from the historical training data set in turn and calculate their mean and variance to obtain a corresponding Gaussian distribution model p(x|1), where p(x|1) represents the first Gaussian distribution The probability density function of the model, x represents the sample data collected;

S22, extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that meet the first verification condition serial number marked as The verification condition is

Among them, ρ is a pre-specified threshold, for example ρ=0.001, if N ₁ /2 is a non-integer, it can be rounded forward or backward, that is, if N ₁ /2 is 16.5, it can be 16 or 17 can be;

S23. Determine whether N ₁ is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let And return to step S21 to iterate until N ₁ is equal to

S3. Based on the previously determined Gaussian model, create a Gaussian mixture model containing two Gaussian components to determine the next steady state phase. The purpose of this step is to establish a mixture model containing two Gaussian components without loss of generality. Since the first c-1 steady-state phases have been determined, the c-th steady-state phase can now be determined, and the schematic diagram of the mixture model update is shown in Figure 1(b); in some specific embodiments, the creation of the Gaussian mixture The model to determine the next steady state phase data includes:

S31. Create and train a mixture model including two Gaussian distribution functions. The formula corresponding to the mixture model is:

p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c)

Among them, it is assumed that the first c-1 stable phases have been determined, and the c-1 stable phase containing N _c-1 samples is denoted as X _c-1 , and the c stable phase containing N _c samples is X _c , c is an integer greater than or equal to 2;

The above-mentioned mixed model is trained by using the maximum expectation algorithm, namely the EM algorithm, wherein, denoting X _m ={X _c-1 ,X _c }, the corresponding training difference θ _c ={α _c-1 ,α _c ,μ _c ,Σ _c }; Since the first c-1 steady-state phases have been determined, such as the mean and variance of the first Gaussian component have been determined by S2, this step only calculates the training difference θ _c ={α _c-1 ,α _c , μ _c , Σ _c } can be;

S32, extracting sample points to perform steady-state phase verification on the mixed model, that is, starting from the N _c-1 /2th sample point for verification to find three consecutive sample points that meet the second verification condition The serial number corresponding to the sample point is marked as The verification condition is

where ρ is a pre-specified threshold, such as ρ=0.001;

S33. Judgment Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let And return to step S31 to iterate until is equal to N _c +N _c-1 .

Based on the previously determined Gaussian model, creating a Gaussian mixture model containing two Gaussian components to determine the next steady state phase includes:

S31. Based on the previously determined Gaussian model, create and train a mixture model containing two Gaussian distribution functions, without loss of generality, assuming that the first c-1 steady-state phases have been determined, and c is an integer greater than or equal to 2, The formula corresponding to the mixed model is

p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c)

Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm;

S32, extracting sample points to verify the steady-state phase of the hybrid model,

That is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that meet the second verification condition and mark the sequence number corresponding to the sample point that meets the second verification condition as The verification condition is

where ρ is a pre-specified threshold;

S33. Judgment Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let And return to step S31 to iterate until is equal to N _c +N _c-1 .

S4. The transient phase that may exist between the two steady-state phases is determined based on the determined two adjacent steady-state phase models. The schematic diagram of the phase division of the present invention is shown in FIG. 2 . In some specific embodiments, the determining, based on the determined two adjacent steady-state phase models, a transient phase that may exist between the two steady-state phases includes: from the two already determined two steady-state phases The state phase X _c-1 and X _c , starting from the first sample point of the c-th steady state phase, and find out the continuous sample points that satisfy p(x _n |c)<ρ, recorded as the transient phase X _{c- 1,c} .

S5. Repeat S3 and S4 to complete the phase division of all sample data.

In addition, in order to solve the shortcomings of the traditional technology, a phase division system based on the local Gaussian mixture model is also proposed, which includes:

a collection unit for collecting samples and creating historical training data sets; in some specific embodiments, by collecting batch process data and expand into Get historical training datasets;

a first Gaussian distribution creation unit, which is configured to select partial samples from the historical training data set to create a first Gaussian distribution model in the order of sampling time to determine the first steady-state phase data; in some specific embodiments, The first Gaussian distribution creation unit includes:

A first data acquisition module, which is used to sequentially select the first N ₁ samples from the historical training data set and calculate their mean and variance to obtain a corresponding Gaussian distribution model p(x|1), where p(x|1 ) represents the probability density function of the first Gaussian distribution model, and x represents the collected sample data;

The first steady-state phase verification module is used for extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that satisfy the first verification condition and will satisfy the first The serial number corresponding to the sample point of the verification condition is marked as The verification condition is

where ρ is a pre-specified threshold, such as ρ=0.001;

The _first steady-state phase determination module, which is used to determine whether N1 is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let and iterates again by the _first steady-state phase verification module until N1 equals

The Gaussian mixture model creation unit is used to create a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model to determine the next steady state phase, and cooperate with the transient phase acquisition unit to complete the analysis of all sample data. The phase division is that the steady-state phase is determined by the Gaussian mixture model creation unit, and the transient phase is determined by the transient phase acquisition unit; in some specific embodiments, the Gaussian mixture model creation unit includes:

The second data acquisition module is used to create and train a mixture model containing two Gaussian distribution functions based on the previously determined Gaussian model, without loss of generality, assuming that the first c-1 steady-state phases have been determined, c is an integer greater than or equal to 2, the formula corresponding to the mixed model is

p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c)

Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm;

The second steady-state phase verification module is used to extract sample points to perform steady-state phase verification on the mixed model, that is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that satisfy the second verification condition and mark the serial number corresponding to the sample point that satisfies the second verification condition as The verification condition is

where ρ is a pre-specified threshold, such as ρ=0.001;

The second steady-state phase determination module, which is used to determine Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let and iterates again by the second steady-state phase verification module until is equal to N _c +N _c-1 .

A transient phase acquisition unit, configured to determine corresponding transient phase data based on two adjacent steady-state phase data to complete the phase division of all sample data. In some specific embodiments, the processing procedure of the transient phase acquisition unit includes: from the two determined two steady-state phases X _c-1 and X _c , from the first sample of the c-th steady-state phase The point starts to check and find out the continuous sample points that satisfy p(x _n |c)<ρ, which is recorded as the transient phase X _c-1,c .

Based on the same inventive concept, the present invention also provides a computer-readable storage medium, comprising computer instructions, which, when the computer instructions are executed on a computer, cause the computer to execute the method.

Based on the above-mentioned technical solution, the following takes a specific experimental example—the penicillin fermentation process as an example to verify the effectiveness of the present invention. The schematic diagram of the penicillin fermentation process is shown in FIG. 3 .

specific:

In the stage of collecting samples and creating historical training data sets: a total of 20 batches of normal data are generated for phase division, and white noise of size N(0, 0.04) is added to each batch of data; The reaction time is 400h, and sampling is performed every 1h, so each batch contains 400 sample points, each sample point contains 11 variables, see Table 1.

Table 1

In the stage of determining the first steady-state phase data and the next steady-state phase data in the order of sampling time and dividing the phases of all sample data: the number of sample points of the first phase is set to three times the number of variables as the initial construction A sample point, ie N ₁ =33. When ρ=0.001, the division results are shown in Table 2. It can be seen from the table that the whole process is roughly divided into 10 steady-state phases and three transient phases. There are three small phases between the first and fifth steady state phases. In the actual fermentation reaction process, the initial stage is the pre-culture stage which is relatively stable, corresponding to the first steady-state phase. Then enter the violent reaction phase, corresponding to the following three small steady-state phases. Then the process enters the feed-feeding stage, and after a period of conversion, the process enters the stable fermentation stage, and finally the autolysis stage. It can be seen that the results divided by this method can correspond well with the real process stages.

Table 2

Implementing the embodiment of the present invention will have the following beneficial effects:

In addition to solving the shortcomings of the existing phase division methods that do not consider timing and transient phase division, the present invention also has the following beneficial effects: (1) the present invention describes a Gaussian distribution with an independent Gaussian distribution. Steady-state phase, the transient phase is described by two adjacent Gaussian distribution mixture models, so that the phase division method can not only effectively divide the steady-state phase, but also determine the transient phase at the same time; (2) each iteration of the present invention Only local data is used for modeling verification, which greatly reduces redundant calculation and improves calculation efficiency; (3) The present invention adopts a step-by-step update strategy, and determines the steady-state phase and transient state step by step according to the sampling time sequence. The phase has the advantages that the number of phase divisions does not need to be pre-specified and the division result does not need subsequent processing.

The above-mentioned embodiments only represent several embodiments of the present application, and the descriptions thereof are relatively specific and detailed, but should not be construed as a limitation on the scope of the patent of the present application. It should be pointed out that for those skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the scope of protection of the patent of the present application shall be subject to the appended claims.

Claims (8)

1. A phase division method based on a local Gaussian mixture model, comprising the steps of: S1. Collect samples and create a historical training data set; S2. Select some samples from the historical training data set to create the first Gaussian distribution model according to the sampling time sequence to determine the first steady-state phase; S3. Based on the previously determined Gaussian model, create a Gaussian mixture model containing two Gaussian components to determine the next steady state phase; S4. Determine a possible transient phase between the two steady-state phases based on the determined two adjacent steady-state phase models; S5. Repeat S3 and S4 to complete the phase division of all sample data. 2. The phase division method according to claim 1, wherein selecting partial samples from the historical training data set to create the first Gaussian distribution model to determine the first steady-state phase data comprises: S21. Select the first N ₁ samples from the historical training data set in turn and calculate their mean and variance to obtain a corresponding Gaussian distribution model p(x|1), where p(x|1) represents the first Gaussian distribution The probability density function of the model, x represents the sample data collected; S22, extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that satisfy the first verification condition and the corresponding sample points that satisfy the first verification condition serial number marked as The verification condition is where ρ is a pre-specified threshold; S23. Determine whether N ₁ is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let And return to step S21 to iterate until N ₁ is equal to 3. phase division method according to claim 2, is characterized in that, Based on the previously determined Gaussian model, creating a Gaussian mixture model containing two Gaussian components to determine the next steady state phase includes: S31. Based on the previously determined Gaussian model, create and train a mixture model containing two Gaussian distribution functions, assuming that the first c-1 steady-state phases have been determined, c is an integer greater than or equal to 2, and the mixture model corresponds to The formula is p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c) Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm; S32, extracting sample points to verify the steady-state phase of the hybrid model, That is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that meet the second verification condition and mark the sequence number corresponding to the sample point that meets the second verification condition as The verification condition is where ρ is a pre-specified threshold; S33. Judgment Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let And return to step S31 to iterate until is equal to N _c +N _c-1 . 4 . The phase division method according to claim 3 , wherein the determining, based on the determined two adjacent steady-state phase models, a transient phase that may exist between the two steady-state phases comprises: Two determined steady-state phases X _c-1 and X _c , start from the first sample point of the c-th steady-state phase and find out the continuous sample points that satisfy p(x _n |c)<ρ, record is the transient phase X _c-1,c . 5. A phase division system based on a local Gaussian mixture model, characterized in that, comprising: a collection unit for collecting samples and creating historical training data sets; a first Gaussian distribution creation unit, which is configured to select partial samples from the historical training data set to create a first Gaussian distribution model in the order of sampling time to determine the first steady-state phase data; The Gaussian mixture model creation unit is used to create a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model to determine the next steady state phase, and cooperate with the transient phase acquisition unit to complete the analysis of all sample data. phase division; A transient phase acquisition unit, which is used for determining a possible transient phase between the two steady-state phases based on the determined two adjacent steady-state phase models. 6. The system according to claim 5, wherein the first Gaussian distribution creation unit comprises: A first data acquisition module, which is used to sequentially select the first N ₁ samples from the historical training data set and calculate their mean and variance to obtain a corresponding Gaussian distribution function p(x|1), where p(x|1 ) represents the probability density function of the first Gaussian distribution model, and x represents the collected sample data; The first steady-state phase verification module is used for extracting sample points for steady-state phase verification, that is, starting from the N ₁ /2th sample point for verification to find three consecutive sample points that satisfy the first verification condition and will satisfy the first The serial number corresponding to the sample point of the verification condition is marked as The verification condition is where ρ is a pre-specified threshold; The _first steady-state phase determination module, which is used to determine whether N1 is equal to Yes, it means that the result is converged, that is, the first steady-state phase has been determined and the next step is performed; otherwise, let and iterates again by the _first steady-state phase verification module until N1 equals 7. The system according to claim 6, wherein the Gaussian mixture model creation unit comprises: The second data acquisition module is used to create and train a Gaussian mixture model containing two Gaussian components based on the previously determined Gaussian model, assuming that the first c-1 steady-state phases have been determined, and c is greater than or equal to 2 Integer, the formula corresponding to the mixed model is p(x|θ _c )=α _c-1 p(x|c-1)+α _c p(x|c) Among them, the probability density function p(x|c-1) of the c-1 th Gaussian model has been determined, and the c-1 th steady-state phase including N _c-1 samples is X _c-1 , and the c th The probability density function p(x|c) of the Gaussian model is to be determined. Assume that the c-th steady-state phase containing N _c samples is X _c , and denote X _m ={X _c-1 , X _c } for this Gaussian mixture model. Training data, the corresponding training parameters θ _c = {α _c-1 , α _c , μ _c , Σ _c }, α _c-1 and α _c are the c-1th and cth Gaussian in the mixed Gaussian model, respectively Combination coefficients of components, μ _c and Σ _c are the mean vector and variance matrix in the c-th Gaussian probability density function p(x|c) respectively, and the above-mentioned mixed model is trained by using the maximum expectation algorithm, ie the EM algorithm; The second steady-state phase verification module is used to extract sample points to perform steady-state phase verification on the mixed model, that is, start the verification from the N _c-1 /2th sample point to find three consecutive sample points that satisfy the second verification condition and mark the serial number corresponding to the sample point that satisfies the second verification condition as The verification condition is where ρ is a pre-specified threshold; The second steady-state phase determination module, which is used to determine Whether it is equal to N _c +N _c-1 , if it is, it means that the result is converged, that is, the c-th steady-state phase has been determined; otherwise, let and iterates again by the second steady-state phase verification module until is equal to N _c +N _c-1 . 8 . The system according to claim 7 , wherein the processing procedure of the transient phase acquisition unit comprises: from the two determined two steady-state phases X _c-1 and X _c , from the c-th The first sample point of the steady-state phase starts to test and finds out the continuous sample points satisfying p(x _n |c)<ρ, which is recorded as the transient phase X _c-1,c .