CN105809203A - Hierarchical clustering-based system steady state detection algorithm - Google Patents

Abstract

The invention discloses a system steady-state detection algorithm based on hierarchical clustering. For industrial data in a continuous time interval, calculate the distance between all two classes, get the matrix, find the minimum distance value in the matrix and merge it into a new class, and delete it, and then process to obtain the updated matrix, iterative calculation Obtain the expression matrix of the hierarchical clustering tree; for the expression matrix, calculate the rationality value of each clustering result from the last iteration value to the first iteration value in sequence, and the clustering result rationality obtained by each calculation Calculate the value to obtain the final threshold value, use it as the final threshold value to perform clustering calculation on industrial data, obtain the final clustering result sequence, and obtain the steady state of the system through joint timing judgment. The invention has strong applicability, and avoids the phenomenon of sharp increase of calculation amount caused by data dimension in processing data.

Description

A System Steady State Detection Algorithm Based on Hierarchical Clustering

technical field

The invention relates to a system detection method, in particular to a system steady-state detection algorithm based on hierarchical clustering.

Background technique

Steady state is the most important and common assumption in the study of a process system. Whether the system is in a steady state is directly related to the subsequent modeling, control and optimization methods of the system. The internal principle and structure of the process system are complicated, which is manifested by the strong coupling of physical quantities and the strong nonlinearity of the system. When the system is in an unstable state, the data characteristics of each variable of the system change drastically, and the numerical value is unstable or abnormal, and there is a large deviation from the input-output relationship of the real system. Only when the system is in a stable working condition can each parameter and variable have a strong state consistency. Based on this situation, the evaluation of equipment operation performance, the analysis of object characteristics and controller effects, all need to obtain the stable operation of the system as the premise.

With the development of the process industry, the system and production methods in the actual production process tend to be more complex, and the process system objects involved are often multi-variable, high-dimensional and strongly coupled, and the overall system is nonlinear and time-varying. , uncertainty and incompleteness. Although the complex system has caused great difficulties to mechanism research and modeling, more and more process data have been recorded due to the application of DCS control system and smart instruments. The strong and complex coupling relationship between variables in the actual process system makes it possible to study the stability of the entire system through measurement data. With the continuous development and improvement of data mining and statistical learning theory, the field of process industry It is also gradually using related algorithms to solve practical problems, resulting in fields such as statistical process control. In dealing with the problem of steady-state detection, the ideas and methods of big data are also different from the traditional process control research methods. The latter judges whether the system is stable or not, and often uses several key variables to formulate steady-state standards for decision-making, while the former makes decisions. It is required to analyze the system starting from all the data. Theoretically speaking, the results based on all the data will reflect the actual situation of the system more comprehensively and truly, so more accurate and reliable results can be obtained.

Since the steady-state detection problem was proposed in the 1980s, many scholars at home and abroad have proposed different steady-state detection methods. However, due to the complexity of field data, many existing steady-state detection methods have been verified in simulation tests The detection results obtained by state detection methods in practical applications are not very accurate. Moreover, various methods are limited by their own influence and specific requirements in practical applications, and the application occasions are also different.

Clustering is a very important technique in the field of unsupervised learning. It is used to divide individuals in data samples into different categories, so that individuals within a class have as high a similarity as possible, and individuals between classes have as high a difference as possible. The idea of hierarchical clustering was first proposed by JohnsonSC. in 1967. Unlike other clusters such as EM and K-means, hierarchical clustering does not need to specify the number of categories in advance, and does not require an iterative optimization process. Through a given The clustering result can be obtained by using the similarity function and the threshold. Since it does not involve optimization and other problems, it can avoid the "curse of dimensionality", but because hierarchical clustering needs to calculate the distance between two samples, the complexity will increase quadratically with the increase of sample size.

The flow chart of the traditional computing method structure is shown in Figure 1. Assuming that there are N sample individuals in the sample data set to be clustered, the hierarchical clustering algorithm specifically has the following steps:

1) Define the similarity function and the algorithm termination condition threshold (generally the maximum intra-class distance or the minimum inter-class distance);

2) Cluster each individual in the data set into one category, a total of N categories;

3) The current data set is divided into k categories, and the similarity between each category is calculated. d(i,j) indicates the similarity between i and j categories. If d(i,j) is the current The minimum value in the similarity set, then merge the i-th and j-th classes. At this time, the number of clusters changes from k to k-1;

4) Calculate the termination condition value at this time, and end the algorithm if the threshold is met;

5) Back to 3), when k=1, all sampling points are clustered into the same cluster, and the clustering algorithm cannot continue, and the algorithm stops.

Therefore, the main disadvantage of the traditional method is the strong dependence on the process object, poor universality, and mainly for single-variable steady-state detection, so the applicability is not strong, and the calculation due to the data dimension in processing data cannot be avoided. problem of increasing volume.

Contents of the invention

In order to solve the problems existing in the background technology, the present invention proposes a steady-state detection algorithm based on hierarchical clustering.

The technical scheme adopted in the present invention is:

STEP1 generates a clustering tree:

1.1) For the industrial data {d _i } containing N sampling points in a continuous time interval, i=1, 2, 3..., N, the sampling points in the set are used as the class, and d _i represents the industrial data of the class;

1.2) Calculate the distance between all two classes, and get the matrix A of N×N, the elements of row i and column j in matrix A are marked as a _ij , a _ii =0, and a _ij represents the difference between class d _i and class d _j The distance between the obtained matrix A is as follows:

1.3) Find the minimum distance value a _mn in the matrix A, where a _mn is its class distance. a _mn means that the mth class and the nth class are the closest classes, the mth class and the nth class are merged into a new class, the m rows, n rows, m columns and n columns in the matrix A are deleted, and the matrix A The remaining classes in and the new class obtained by merging are processed in the same way as step 1.2) to obtain the updated matrix A;

1.4) Repeat steps 1.2) to 1.3) to iterate until the matrix A becomes a 1×1 matrix, record m, n and a _mn in each iterative calculation process, and form an N×3 expression matrix of the hierarchical clustering tree Z;

$\mathrm{<span\; class="notranslate">\; Z</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{ccc}{\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>& <span\; class="notranslate">\; .</span>\\ {\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}& {\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ...</span>& {\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}\mathrm{<span\; class="notranslate">\; N</span>}}\end{array}\right]$

STEP2 threshold selection:

2.1) For the expression matrix Z, the sequence from the value of the last iteration to the value of the first iteration is calculated in the following manner:

The industrial data {d _i } is clustered with the current minimum distance value a _mn as the threshold, and the obtained clustering result sequence is T _k , k represents the iteration number, and the clustering result sequence T _k is an integer sequence of 1×N , wherein, T _k (i) represents the i-th clustering result of the clustering result sequence T _k , T _k (i)=p; calculate and obtain the differential sequence D of the clustering result sequence T _k , that is, in the sequence Each adjacent element of is subtracted to obtain the difference, and the number of zeros in the difference sequence D is calculated as the clustering result rationality value D_zero(k).

2.2) The clustering result rationality value sequence D_zero is formed from the clustering result rationality value D_zero(k) obtained each time, and the difference sequence of the clustering result rationality value sequence D_zero is calculated, that is, for each sequence in the sequence The difference is obtained by subtracting two adjacent elements, from which the largest difference value and its sequence number k in the difference sequence are found, and Z _3k is used as the final threshold.

In the step STEP2, the sequential calculation from the last iteration value to the first iteration value refers to calculation until the first iteration value of k=1 or the intermediate iteration value of k=N/2.

When k is too small, the clustering result is meaningless for the final steady-state identification. In order to reduce the amount of calculation, k=N/2 is generally selected as the stop condition of the algorithm.

STEP3 joint timing judgment steady state:

If the element Ti in the final clustering result sequence T satisfies the following conditions, the system between the mth sampling point and the m+ _k -1th sampling point is considered to be in a steady state:

T _i = c, i = m, m+1, m+2,..., m+k-1

Among them, T _i is the clustering result corresponding to the i-th sampling point in the final clustering result sequence T, c represents the result constant, k=τ/T _s , τ is the above-mentioned time length threshold, and T _s is the sampling time interval of the data .

The size of τ is determined according to the time characteristics of the researched object. Generally, τ=3t ^* is taken, and t ^* is the unit step response adjustment time of the system.

The new class obtained by merging in step 1.3) is placed at the end of the updated matrix A, and a new row/column is added at the end, which can be written as [a _1(N+1) a _2(N+1) ...] ^T.

The row and column numbers of the remaining classes in the matrix A remain unchanged, and the row and column numbers of the new class obtained by merging adopt new row and column numbers and are different from the row and column numbers of all previous classes, so that the entire system process can be processed every time Classes have a unique row and column numeric sequence number.

The distance between the classes that are merged in the step 1.3) and the remaining classes in the matrix A is calculated by the following formula:

a _si =αa _mi +(1-α)a _n*

Among them, α is the weight parameter, 0≤α≤1, a _si represents the distance between class s and class i, a _ni represents the distance between class n and class i, and a _mi represents the distance between class m and class i spacing.

Each sampling point of the system is regarded as a "feature vector" (or feature point) of a state space. When the system is in a steady state, the feature point fluctuates within a certain interval, presenting as a high point centered on a specific point. Gaussian distribution. When the system is in a transition state or an unsteady state, the state points will deviate from the original Gaussian distribution, and then present another scattered distribution state. Based on the fact that when the system is in an unsteady state, there is a large difference between the states before and after the time, the steady state of the process system can be detected by comparing the degree of difference between the system state eigenvectors. The clustering algorithm is introduced here. In the steady-state detection, the steady-state data has a high similarity, and will be clustered into one class in the clustering. At the same time, there is a big difference between the dynamic data points and the steady-state data. , will be divided into different categories.

Due to the characteristics of the system, it can only be guaranteed that the state distribution is relatively concentrated when the system is in a steady state, but if the system is in a fluctuating state, the fluctuation of its distribution in the state space is very large. That is to say, while there is similarity between steady-state sampling points, the similarity between dynamic sampling points is very small. In the clustering results, dynamic points in continuous time may be clustered in many different clusters. in class. Therefore, it is impossible to determine the number of the last existing classes before clustering, and the hierarchical clustering algorithm of the present invention is adopted based on these.

The beneficial effects of the present invention are:

The present invention introduces the idea and technology of "big data" into the steady-state detection, performs steady-state detection by comparing the similarity between each sampling point in the data set, and combines the time series characteristics of the data, and proposes how to determine the clustering threshold method.

The present invention is characterized by strong applicability, and avoids the phenomenon of sharp increase in calculation amount caused by data dimension in processing data.

Description of drawings

Figure 1 is a flowchart of the hierarchical clustering algorithm.

Fig. 2 is a flowchart of the threshold selection process of the method of the present invention.

Fig. 3 is a flow chart of the method of the present invention for judging a steady state in combination with time series.

FIG. 4 is an input-output curve diagram of the second-order simulation system in Embodiment 1. FIG.

Fig. 5 is a diagram of the clustering results of the second-order system state points in Embodiment 1.

FIG. 6 is a diagram of the clustering detection results in Example 1. FIG.

Fig. 7 is an input-output curve diagram of the second-order simulation system in Embodiment 2.

Fig. 8 is a comparison diagram of clustering results obtained when different thresholds are used in Example 2.

FIG. 9 is a graph of data to be processed inputted in Embodiment 3.

FIG. 10 is the clustering result in the 50th iteration of Embodiment 3.

FIG. 11 is a distribution diagram of the number of 0s in D and its variation in all 50 iterations of Embodiment 3.

FIG. 12 is a representative partial clustering result of Example 3.

FIG. 13 is a diagram of the steady-state detection results of Example 3. FIG.

detailed description

The present invention will be further described below in conjunction with drawings and embodiments.

The method of the present invention is specifically applied to the detection of the steady state of the process system. As shown in Figure 2, there is the following process: for the data point {d _i } to be classified, first define the distance between the points in its state space and the measurement of the distance between classes (Since the state points of the process system are randomly distributed in the corresponding state space according to a certain probability and degree of aggregation, we choose to use Euclidean distance for measurement in the method of steady-state detection). When measuring the distance between classes, the steady state we care about is a class in the clustering results. We will try to increase the degree of aggregation of each class, and choose the maximum distance to meet the above requirements.

The input of the algorithm: system historical data status points {d _i } sorted by time, where {d _i } is p-dimensional vector data, i=1, 2, 3..., N, N is the number of sampling points.

The output of the algorithm: the category T corresponding to each sampling point, T is a one-dimensional sequence of length N, that is, the sequence of sampling point category numbers. The category numbers here are only used to distinguish categories without any physical or numerical meaning.

According to the steps of hierarchical clustering, the algorithm then iterates by calculating the distance between points in the data set, checking whether the termination condition of the algorithm is reached at each step, and output the clustering result T when the algorithm ends.

Since the analyzed data objects are time series arranged according to time, when the system is in a steady state for a period of time, its spatial distribution is relatively concentrated. Therefore, using the timing of the data, it can be considered as a steady state if the following two conditions are met: 1) The data sampling points are continuous in time; 2) The similarity of the data sampling points is very high (that is, the clustering algorithm gathers the data into one class).

To obtain a steady state through clustering results, it is necessary to determine the threshold of clustering and the minimum time interval T for judging stability. Under the selected threshold δ, when the length of the system is not less than T, the state points of the system are clustered in the same class , which means that the system is in a steady state during this period. The determination of the minimum time interval T can be based on the time characteristics of the system object. The following mainly discusses the selection of the clustering threshold δ and the influence of the threshold on the algorithm.

In the process of clustering calculation, there are N samples in the data set to be clustered, and the process of gradually aggregating the data from N classes to 1 class is recorded in the N*3 clustering tree matrix Z, where Z _i1 records the i-th Merge the serial numbers of the first class in the two classes during the first step of clustering, Z _i2 records the serial numbers of the second class in the merged two classes in the ith step of clustering, and Z _i3 records the distance between the merged two classes in the ith step of clustering. Therefore, to obtain the optimal clustering threshold is to find a row in Z as the end row of the algorithm output, where Z _i3 is the threshold. Here we enumerate the results of each row to obtain the optimal threshold.

In addition, the innovation of the present invention is to use specified reasonable quantitative indicators to find a reasonable threshold for calculation. The specific principle is: T _k is the clustering result obtained by the kth enumeration, and its difference item D=diff(T _k ) is used to calculate To measure the overall clustering effect, as the threshold shrinks, the number of 0 elements in D gradually decreases, generally as in (1) to (9), the number of 0 in D decreases slowly, but when a When the class steady state is split, the result will appear that the system points frequently switch between the two classes for a period of time, which will make the number of 0s in D decrease sharply. Therefore, it is necessary to enumerate the clustering results until the enumeration Stop when the threshold is too small for steady-state identification. Find the step where the number of 0s in D changes the most, that is, the termination step where the clustering threshold is reduced, and the corresponding threshold is the appropriate steady-state threshold required by the current data clustering. The flow chart for determining the threshold is shown in 3 .

Embodiments of the present invention are as follows:

Example 1

Embodiment 1 is aimed at a second-order single-input-single-out system, adopting the input and output after simulation processing of the method of the present invention as shown in Figure 4, the simulation time length is 100s, and the sampling interval is 0.1s, and Figure 4 (a) represents the input variable Curve, Figure 4(b) shows the output variable curve.

Noise is not added in the variable of embodiment 1, it can be seen from the curve of Fig. 4 that the system is in the steady state at the initial stage (t=0～300), and a ramp adjustment is generated at t=300 place input, and the system enters the transition state, When t=600, the ramp signal ends, and the output also tends to be stable, and the final system enters another steady state. Since no noise is added, the degree of aggregation of the system state is very high in the steady state. Here, instead of using the threshold selection method of the above algorithm, a small threshold t=0.01 is directly manually selected, and the data is clustered. The obtained results are as follows Figure 5 shows.

From the curve in Figure 4, it can be seen that the system has two different stable states in the whole process. In the clustering results in Figure 5, the labels of the data points in the same steady state are consistent. It can be seen that there are The two horizontal sections correspond in time to the two steady states of the system.

The result of clustering represents the distance of the state points in the space, and judging whether the system is in a steady state needs to be combined with the timing of the data. When the system is in a steady state for a period of time, the data in this period of time will be clustered into the same category because of their high similarity. Conversely, this can be used as a steady state criterion, when all the data within a continuous time τ are clustered into one category by the clustering algorithm, the system is considered to be in a steady state. The length τ of the time period is determined according to the time characteristics of the researched object. Here, τ=2t* is taken, and t* is the unit step response adjustment time of the system.

Therefore, the obtained system stability results are shown in Figure 6. Figure 6(a) shows the clustering results of the second-order system state data points, and Figure 6(b) shows the final system steady-state detection results. In the steady-state results, 1 means stable and 0 Indicates instability. It can be seen from the results that the system breaks away from the first steady state around t=300, enters a transition state, and enters a steady state again when it is close to t=700. Comparing the input and output, although the input is at t= At 600, the ramp signal ends and becomes stable, but the output y is stabilized after a period of adjustment.

Example 2

Embodiment 2 added noise for simulation, and further explained the threshold selection process based on the single-input-single-out system process used in Embodiment 1. The input and output of the simulation system with noise added are shown in Figure 7, and Figure 7(a) shows the input variables curve, Figure 7(b) shows the output variable curve:

It can be seen from Figure 7 that the system is in a steady state within the interval of time t=0~300, and then input u produces a ramp change, and after a period of transition state, the system enters another steady state. In the state space of the whole process, the state points of the two steady states are highly aggregated, while the transition state points are scattered. The selection of the threshold is enumerated from large to small. As the clustering threshold shrinks, the degree of concentration within each class becomes higher and higher. When the jth step is reached, if the threshold is reduced to a certain extent, the points belonging to the same stable state will be divided into multiple categories. We think that the threshold at this time is too small to exceed our expectations, and it is unnecessary to continue to reduce the threshold. . At this time, the threshold reached in the previous step, that is, the j-1th step, is the current appropriate value.

The first 12 steps of the enumeration and clustering of the single-input-single-out second-order system of Embodiment 2 are shown in Figure 8: It can be seen from Figure 8 that the threshold in Figure 8 (1) is relatively large, and the algorithm only divides the data into In Figure 8(2) to Figure 8(9), the data in the intermediate state is continuously separated under the drive of the algorithm. As the clustering threshold shrinks, the degree of aggregation within each class becomes higher and higher, and the results of the present invention become more accurate. However, in Figure 8(10), when t is 700-1000, the state points of the system originally belong to the same class, but after the threshold is narrowed in this step, many points are separated from it, and the two classes separated out in time sequentially interleaved with each other. If these two categories are regarded as the steady state of the system, Figure 8(10) shows that the system frequently switches between the two steady states during this period, and there is no transition state—this is in the natural process system does not exist in .

From this, it is considered that the state in Figure 8(10) has reached the limit of threshold shrinkage, and the reasonable distance threshold should take the threshold corresponding to the result in Figure 8(9) as the final threshold.

Example 3:

Embodiment 3 is applied to the historical data generated in the actual process, and the boiler data of a 60MW unit of a certain power plant is used for data experimentation. The data mainly includes: boiler load command, power generation, coal intake, air intake, steam water temperature and pressure at each measurement point, etc., a total of 180 dimensions. Select 10,000 points of data and use the above clustering algorithm for steady-state detection. The changes in the main variables during this period are shown in Figure 9:

It can be seen from Figure 9 that during the time period studied, there were two major adjustments in the system load, roughly producing three periods of steady-state data within the interval. Both the main steam temperature and the main steam pressure fluctuate greatly, and the fluctuation of the system cannot even be seen from the steam temperature curve. Next, the state points composed of all the variables of the system are clustered. According to the method of selecting the threshold value above, the total number of iterations is selected as 50 times. The data test proves that the result of the 50th time is shown in Figure 10, and it is no longer possible to obtain useful results from the results. The news, so stop to continue iteration here, the number of 0 in D and its changes finally obtained are shown in Figure 11, Figure 11(a) shows the distribution of the number of 0 in D in 50 iterations, and Figure 11(b) shows Changes in the number of 0s in D in 50 iterations.

Embodiment In step 16, the number of 0 elements in D changes suddenly. In order to see the change of clustering more intuitively, some representative iteration results are selected as shown in FIG. 12 . It can be seen from Figure 12 that when iterating from step 16 to step 17, the points in the time period from 0 to 2000 are split into two categories, and the picture on the clustering result is consistent with the sudden change in the number of 0 elements in D. It also illustrates the rationality of the method of the present invention.

According to the results obtained by clustering, an appropriate time length is selected for steady-state detection. According to the characteristics of industrial objects, when the system remains stable within 500 sampling points, the system can be considered to be in a steady state. In the clustering results of the embodiment, if the clustering numbers of all the points in the interval of not less than 500 points in the clustering results are the same, the system is considered to be in a stable state, and the steady state detection output of the system is obtained accordingly, as shown in Figure 13 13(c) where 1 means stable and 0 means unstable.

Finally, it can be seen that the present invention has its remarkable technical effect, strong applicability of steady-state detection, and avoids the phenomenon of sharp increase in calculation amount caused by data dimension in processing data.

The above-mentioned embodiments do not limit the present invention, and the present invention is not limited to the above-mentioned embodiments, as long as the requirements of the present invention are met, they all belong to the protection scope of the present invention.

Claims (6)

1. the systematic steady state detection algorithm based on hierarchical clustering, it is characterised in that:

STEP1 generates clustering tree:

1.1) comprised to the industrial data { d of N number of sampled point one period of continuous time in interval_i, i=1,2,3 ..., N, in set, sampled point is as class, d_iRepresent the industrial data of class；

1.2) calculate all distances between class between two, obtain the matrix of N × NIn matrix A, the element of p row q row is designated as a_pq, a_pqRepresent class d_pWith class d_qBetween distance, a_pp=0；

1.3) the class spacing minima a in matrix A is found_mnNamely m class and the n-th class are two classes that current distance is nearest, m class and the n-th class are merged into a new class, m row in puncture table A, n row, m arrange and n row, the class of current residual is adopted step 1.2 again with merging the new class obtained) identical mode calculates class spacing, it is thus achieved that the matrix A after renewal；

1.4) step 1.2 is repeated)～1.3) be iterated, until matrix A becomes the matrix of 1 × 1, record m, n and a in each iterative process_mn, constitute expression matrix Z, the m of N × 3 of hierarchical clustering tree, n and a_mnThe first, second, third columns value respectively as matrix Z；

$Z=\left[\begin{array}{ccc}{Z}_{11}& Z_{21}& Z_{31}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ Z_{N1}& Z_{N1}...& Z_{3N}\end{array}\right]$

STEP2 threshold value is chosen:

For expression matrix Z, calculate cluster result reasonability value D_zero (k) obtaining each time successively to the order of first time iterative numerical by last iterative numerical, asked for calculate by each calculated cluster result reasonability value D_zero (k) and obtain final threshold value, with z_3kAs final threshold value to industrial data { d_iCarry out cluster calculation, it is thus achieved that final cluster result sequence T；

STEP3 combines sequential and judges stable state:

Element T in final cluster result sequence T_iIf meeting the following conditions, then it is assumed that the system between m-th sampled point to the m+l-1 sampled point is in stable state:

T_i=c, i=m, m+1, m+2 ..., m+l-1

Wherein, T_iFor the cluster result that the i-th sampled point in final cluster result sequence T is corresponding, c represents result constant, l=τ/T_s, τ is above-mentioned time span threshold value, T_sSampling time interval for data.

2. a kind of systematic steady state detection algorithm based on hierarchical clustering according to claim 1, it is characterized in that: described step 1.3) the middle end merged in the matrix A after the new class obtained is placed in renewal, in described matrix A, the ranks digital number of remaining class remains unchanged, and the ranks rank-numeral merging the new class obtained adopts new ranks digital number and the ranks digital number with all classes before all to differ.

3. a kind of systematic steady state detection algorithm based on hierarchical clustering according to claim 1, it is characterized in that: described step 1.3) in merge in the class that obtains and matrix A between each class remaining distance calculation as follows, class after note m class and the merging of n class is numbered s, i class is any sort except s class, then the distance calculating s class and i class has:

a_si=α a_mi+(1-α)α_ni

Wherein, α is weight parameter, 0≤α≤1, a_siRepresent the spacing between s class and i class, a_niRepresent the spacing between n class and i class, a_miRepresent the spacing between m class and i class.

4. a kind of systematic steady state detection algorithm based on hierarchical clustering according to claim 1, it is characterised in that: described step STEP2 calculates every time and calculates all in the following ways: with current lowest distance value a_mnFor threshold value to industrial data { d_iCarrying out cluster calculation, the cluster result sequence obtained is T_k, k represents iteration ordinal number, cluster result sequence T_kIt is the integer sequence of 1 × N, calculates and ask for cluster result sequence T_kDifference sequence D, calculate in difference sequence D the number of zero as cluster result reasonability value D_zero (k).

5. a kind of systematic steady state detection algorithm based on hierarchical clustering according to claim 1, it is characterized in that: described step STEP2 is asked for by each calculated cluster result reasonability value D_zero (k) calculating and obtains final threshold value particularly as follows: be made up of cluster result reasonability value sequence D_zero each calculated cluster result reasonability value D_zero (k), calculate the difference sequence asking for cluster result reasonability value sequence D_zero, therefrom find the sequence number j at difference value maximum in difference sequence and place thereof, and with Z_3jAs final threshold value.

6. a kind of systematic steady state detection algorithm based on hierarchical clustering according to claim 1, it is characterised in that: described step STEP2 is calculated, to the order of first time iterative numerical, the centre time iterative numerical referring to first time iterative numerical or the k=N/2 calculating k=1 by last iterative numerical successively.