Remote measurement parameter change real-time detection method based on representation learning

Abstract

The invention discloses a remote measurement parameter change real-time detection method based on representation learning. According to the remote measurement parameter classification method, the effects that the parameter change trend in classification is similar, and the parameter change difference between classification is obvious are achieved; the characterization sequence of each class can accurately describe the variation trend of the class of parameters; in real-time detection, when the telemetering parameters have changes represented by the representation sequences of the telemetering parameters, the changes can be correctly detected, the occurrence time of the parameter changes can be accurately obtained, and a feasible new thought is provided for real-time detection, automatic interpretation and correlation analysis of the changes of the rocket telemetering parameters.

Description

Technical Field

[0001] The invention belongs to the field of rocket launching technology, and in particular relates to a real-time detection method for telemetry parameter changes based on representation learning. Background Art

[0002] The ignition, shutdown, and stage separation of each stage of the launch vehicle are the key control points of the launch process. In these links, the onboard control system issues a series of instructions, triggers the actuator to move, and produces corresponding state changes. The state changes of various systems of the rocket occur intensively and rapidly. Whether these changes occur correctly and strictly according to the designed time is directly related to the success or failure of the rocket launch. Telemetry parameters are the main source of information reflecting the operating status of various systems of the rocket. The time when telemetry parameter changes occur and their correctness are important bases for the evaluation of key control points of rocket launches.

[0003] Telemetry parameters include command parameters such as switch quantity, control word, pulse, and state parameters such as voltage, pressure, temperature, speed, liquid level, overload, and vibration [1]. There is an inevitable logical connection between the changes in command parameters and state parameters. However, the change patterns of various parameters are numerous and difficult to detect automatically. Currently, only the change time of command parameters can be automatically detected in real-time tasks. Whether the remaining large number of state parameters change correctly and whether the change time is consistent with the design still depends on the system experts to manually interpret it afterwards.

[0004] In order to improve the level of real-time change detection and automatic interpretation of telemetry parameters, many scholars have conducted in-depth research. From the perspective of system architecture, the design of the automatic interpretation process of telemetry parameters and the establishment of a hierarchical evaluation model are discussed. For time instruction parameters, an improved method for timing parameter interpretation is proposed. Focusing on the time series characteristics of telemetry parameters, from the perspective of time series similarity, the change detection and interpretation methods of slowly varying telemetry parameters are studied. An automatic interpretation method of telemetry parameters based on the statistical characteristics of historical data is proposed, and the parameter robust estimation is used to effectively improve the interpretation efficiency and accuracy. In the above studies, there are more post-event overall interpretations of certain types of telemetry parameters with clear change characteristics, but there is a lack of real-time detection methods for local changes and change time series that are applicable to most telemetry parameters. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a real-time detection method for telemetry parameter changes based on representation learning.

[0006] A real-time detection method for telemetry parameter changes based on representation learning of the present invention comprises the following steps:

[0007] S1. According to the actual characteristics of telemetry parameter time series and key timing windows, complete data preprocessing and select sliding correlation distance as the similarity measure of telemetry parameter time series;

[0008] S2. Using the density clustering algorithm under the selected distance metric, the remote sensing parameters are classified according to the similarity of their change trends;

[0009] S3. For each type of time series, select or define a characterization sequence to summarize its change pattern, and calculate the detection threshold of the characterization parameter change;

[0010] S4. Calculate the matching degree between the measured sequence of each telemetry parameter and its corresponding characterization sequence, and compare it with the detection threshold to realize the real-time detection of telemetry parameter changes.

[0011] The data preprocessing in S1 refers to intercepting the telemetry parameters from 30 seconds before ignition to the moment of separation of the booster, fairing, first stage or second stage as valid data segments, and performing normalization and time alignment processing.

[0012] The formula for the normalization process is:

[0013]

[0014] Where x is the measured time series, max(x) and min(x) are the maximum and minimum values ​​of the series in the valid data segment, respectively. t and x t norm are the original value and normalized value of the sequence at time t respectively.

[0015] The time alignment process is to perform time alignment on the normalized time series. In the valid data segment, linear interpolation is performed at a frequency of 10 Hz. The interpolation process is shown in the following formula:

[0016]

[0017] Among them, t i is the measured time, i=1,2,...,m,t s is the time after interpolation, s=1,2,...,n, Δt s is the interpolation interval, which is a constant.

[0018] The calculation formula of the sliding related distance in S1 is:

[0019]

[0020] Among them, x and y are two time series, and their time series are taken in a given time window [t a ,t b ] as the distance metric dist(x,y) between them.

[0021] The classification of S2 includes the following steps:

[0022] S201. Randomly select a reference sample and calculate the distance between each sample and the reference sample;

[0023] S202. All samples are sorted according to the above distances, and then density clustering is performed according to the distance values ​​to obtain a number of clusters; the distances between samples within a cluster are dense, that is, the similarity is high; the distances between samples between clusters are sparse, that is, the similarity is low.

[0024] The density clustering algorithm process of S2 is as follows:

[0025] T1. Find the ∈-neighborhood of each sample in the dataset D and determine the core object set Ω;

[0026] T2. Randomly select a core object x from Ω i , find out the cluster C1 composed of samples whose density can be reached;

[0027] T3. Remove C1 from Ω and repeat step T2 until Ω is empty. The output cluster partition C = {C1, C2, ..., C k};

[0028] The clusters C1 to C1 output by the algorithm k , which is the result of time series division with highly similar sample change trends within the cluster.

[0029] The specific method of S3 is as follows: calculate the sliding correlation distance between the samples in the cluster, sort them from near to far, take the first 50% of the sample pairs, and count the samples with the highest frequency; at this time, the average distance between this sample and other samples is the shortest, and under the current distance measurement, it can best represent the characteristics of the samples in the cluster, so it is selected as the central sample; at the same time, take the maximum distance d between the central sample and other samples in the cluster max As the detection threshold for characterizing the sequence.

[0030] The real-time detection process of S4 includes the following steps:

[0031] S401. For the telemetry parameter K, obtain its characterization sequence p K The measured value sequence x in the time window of equal length K ;

[0032] S402.x K After the interpolation is completed and the time scale is aligned, the sliding related distance dist(x K ,p K );

[0033] S403. When dist(x K ,p K )<ρ K , that is, the characterization sequence is detected, indicating that the parameter K has undergone p Kthe changes represented;

[0034] S404. Take each dist(x K ,p K )<ρ K Until the next time dist(x K ,p K )>ρ K Within the time range, dist(x K ,p K )The time T corresponding to the minimum value K , which is the time when the parameter changes.

[0035] The detection time window in S401 is slid as needed for the historical mission data of the rocket flight, so as to further mine the time series of parameter changes; for the real-time flight data of the rocket, the detection time window is taken from the current moment and p K After a period of equal length has passed, real-time detection of changes in telemetry parameters is performed.

[0036] The beneficial effect of the present invention is that the telemetry parameter classification method of the present invention achieves the effect that the parameter change trends within the classification are similar and the parameter change differences between the classifications are obvious; the characterization sequence of each classification can accurately characterize the change trend of the parameters of this type; in real-time detection, when the telemetry parameter undergoes the change represented by its characterization sequence, it can be correctly detected, and the time when the parameter change occurs can be accurately obtained.

[0037] The proposed method provides a feasible new idea for real-time detection, automatic interpretation and correlation analysis of rocket telemetry parameter changes. In the future, the quantitative characteristics of parameter characterization sequences can be further mined, and historical data can be fully utilized to carry out real-time data evaluation, fault mode identification and flight status prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the time series density clustering process of the present invention.

[0039] Figure 2 It is a flow chart of the DBSCAN density clustering algorithm of the present invention.

[0040] Figure 3 It is a flow chart of parameter changes in characterization sequence detection of the present invention.

[0041] Figure 4 It is the main classification result of the telemetry parameter time series of the present invention.

[0042] Figure 5 This is the real-time detection effect of the characterization sequence of the present invention. DETAILED DESCRIPTION

[0043] Embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0044] A real-time detection method for telemetry parameter changes based on representation learning of the present invention comprises the following steps:

[0045] S1. According to the actual characteristics of telemetry parameter time series and key timing windows, complete data preprocessing and select sliding correlation distance as the similarity measure of telemetry parameter time series;

[0046] S101. Preprocessing of time series

[0047] In order to unify the dimensions of the time series of each telemetry parameter, the valid data segment of each parameter is first intercepted and normalized. For example, for the telemetry parameters of the booster, fairing, first stage and second stage, the valid data segment is intercepted from ignition-30 seconds to the separation time of the stage, and normalized as follows:

[0048]

[0049] Where x is the measured time series, max(x) and min(x) are the maximum and minimum values ​​of the series in the valid data segment, respectively. t and x t_norm are the original value and normalized value of the sequence at time t respectively.

[0050] S102. Perform time alignment processing on the normalized time series

[0051] In the valid data segment, linear interpolation is performed at a frequency of 10 Hz. Assume that the actual measurement time is t i ,i=1,2,...,m, the time after interpolation is t s ,s=1,2,...,n,Δt s is the interpolation interval, is a constant, and the interpolation process is as follows:

[0052]

[0053] After completing the time alignment, a time window covering as many parameter changes as possible is selected according to the time when the key event occurs, and the time series of all parameters within the window are intercepted.

[0054] After the above processing, a telemetry parameter time series with consistent dimension, consistent time length and uniform scale can be obtained.

[0055] S103. Sliding related distance measurement

[0056] The Pearson correlation coefficient is insensitive to differences in the magnitude of data changes, requires less calculation, and can better measure the changing trend of time series.

[0057] Considering the need to eliminate the influence of different change moments on distance calculation, for two time series x and y, we take their distances in a given time window [t a ,t b ] as the distance metric dist(x,y) between them. The calculation formula is as follows:

[0058]

[0059] Under this distance metric, if the change trends of two time series are completely consistent or completely opposite, the distance is 0; the greater the difference in the change trends, the greater the distance value, and the closer it is to 1. After determining the sliding correlation distance metric, the "distance" in the following text refers to the sliding correlation distance.

[0060] S2. Using the density clustering algorithm under the selected distance metric, the remote sensing parameters are classified according to the similarity of their change trends;

[0061] After completing the preprocessing of the telemetry parameter time series and determining the sliding correlation distance as the similarity metric, the telemetry parameter time series is classified under this distance metric. The problem is described as follows: For the time series x of i telemetry parameters i [t],t∈[t a ,t b ], we need to find k categories so that the time series samples in each category have similar trends and the differences between categories are as large as possible. Because the telemetry parameters have many different change patterns, it is impossible to determine the value of k in advance, so it should be treated as a clustering problem. The process is as follows Figure 1 As shown:

[0062] S201. Randomly select a reference sample and calculate the distance between each sample and the reference sample;

[0063] S202. All samples are sorted according to the above distances, and then density clustering is performed according to the distance values ​​to obtain a number of clusters; the distances between samples within a cluster are dense, that is, the similarity is high; the distances between samples between clusters are sparse, that is, the similarity is low.

[0064] DBSCAN is a commonly used density clustering algorithm that divides samples with sufficiently high density into a cluster. For the data set D(x1,x2,...,x n ), has the following definition:

[0065] 1)∈-neighborhood: for x j ∈D, distance dist(x i ,x j )≤∈i A collection of components;

[0066] 2) Core object: If x j The ∈-neighborhood of contains at least MinPts samples, then x j It is a core object;

[0067] 3) Density direct access: If x i Located at x j ∈-neighborhood, and x j is a core object, then x i By x j Density direct access;

[0068] 4) Density can be reached: If there exists a sample sequence p∈[x i ,x j ], p i+1 By p i The density is directly proportional to x i By x j Density can reach;

[0069] 5) Density connected: If x k So that x i and x j All by x k The density can reach, then x i and x j Density connected.

[0070] Based on the above definition, the DBSCAN algorithm defines a cluster as: a set of density-connected samples derived from a density-reachable relationship.

[0071] Input parameters (∈, MinPts), where ∈ is the minimum distance between samples of any two clusters, and MinPts is the minimum number of samples in a cluster. The algorithm flow is as follows Figure 2 As shown:

[0072] T1. Find the ∈-neighborhood of each sample in the dataset D and determine the core object set Ω;

[0073] T2. Randomly select a core object x from Ω i , find out the cluster C1 composed of samples whose density can be reached;

[0074] T3. Remove C1 from Ω and repeat step T2 until Ω is empty. The output cluster partition C = {C1, C2, ..., C k};

[0075] The clusters C1 to C1 output by the algorithm k , which is the result of time series division with highly similar sample change trends within the cluster.

[0076] S3. For each type of time series, select or define a characterization sequence to summarize its change pattern, and calculate the detection threshold of the characterization parameter change;

[0077] In the above time series division results, there is a "central sample" in each type of time series sample, which has the shortest average distance to all other samples of the same type and best represents the change characteristics of the time series of this type. This "central sample" can be selected as the characterization sequence of the telemetry parameter of this type to characterize the change trend of this type of parameter.

[0078] This application selects the central sample according to the following method: calculate the sliding correlation distance between the samples in the cluster, sort them from near to far, take the first 50% of the sample pairs, and count the samples with the highest frequency. At this time, the average distance between this sample and other samples is the closest. Under the current distance measurement, it can best represent the characteristics of the samples in the cluster, so it is selected as the central sample. At the same time, take the maximum distance d between the central sample and other samples in the cluster max As the detection threshold for characterizing the sequence.

[0079] S4. Calculate the matching degree between the measured sequence of each telemetry parameter and its corresponding characterization sequence, and compare it with the detection threshold to realize the real-time detection of telemetry parameter changes.

[0080] Through the above method, for each telemetry parameter K that changes within the event range of interest, a characterization sequence p can be obtained. K and the detection threshold ρ K .according to Figure 3 The steps shown are to detect whether parameter changes occur:

[0081] S401. For the telemetry parameter K, obtain its characterization sequence p K The measured value sequence x in the time window of equal length K ; The detection time window, for the historical mission data of the rocket flight, slides as needed to further explore the parameter change sequence; for the real-time flight data of the rocket, it starts from the current moment and takes the value corresponding to p K After a period of equal length has passed, real-time detection of changes in telemetry parameters is performed.

[0082] S402.x K After the interpolation is completed and the time scale is aligned, the sliding related distance dist(x K ,p K );

[0083] S403. When dist(x K ,p K )<ρ K , that is, the characterization sequence is detected, indicating that the parameter K has undergone p K the changes represented;

[0084] S404. Take each dist(x K ,p K )<ρ K Until the next time dist(x K ,p K )>ρ K Within the time range, dist(x K ,p K )The time T corresponding to the minimum value K , which is the time when the parameter changes.

[0085] Example

[0086] Taking the telemetry parameters of a certain type of launch vehicle as an example, the classification effect of the telemetry parameter time series and the detection effect of the characterization sequence are analyzed in detail.

[0087] The shutdown time of a certain engine [t-4s, t+2s] is selected as the time window, the number of time series samples is 1002, the frequency of the preprocessed data is 10Hz, and the density clustering input parameters are (0.1, 3). Figure 4 The main results of time series division based on density clustering are shown in the figure. The four classification results with the highest proportion of the total number of samples are given in the figure. The thick solid line in each figure is the characterization sequence of each classification, and the other curves are the three sample sequences listed in each category. The subtitles of each figure are the number of samples in each category and the proportion of the total number. It can be seen from the figure that the change trend of each classification sample and the characterization sequence of this category is highly consistent.

[0088] Figure 4 The distances between the representation sequences of the four categories and the detection distance thresholds of the representation sequences of each category are shown in Table 1. The bold part in the table is the detection distance threshold of category M, and the rest is the distance between the representation sequences of category M and category N. As can be seen from the table, the detection distance thresholds of each category are smaller than the distances between the representation sequences of each category. On the one hand, it shows that the classification effect has been achieved with similar internal change trends and obvious changes between categories; on the other hand, it shows that the representation sequences of each category can reflect the change characteristics of the category and the differences between categories, and can be correctly distinguished according to the distance threshold in actual detection.

[0089] Table 1 Characterization sequence distance and detection threshold of each category

[0090]

[0091] Based on the above classification results, the telemetry data of rockets of the same model but different batches are used to test the detection effect of each classification characterization sequence. For each telemetry parameter, the length of its characterization sequence is used as a sliding window, and the S4 method is used to perform sliding detection within the effective time range of the parameter. Two parameters are selected as examples for analysis in the detection results, such as Figure 5 As shown, parameter 1 corresponds to category 1, and parameter 2 corresponds to category 3. In (b), the distance between the measured sequence and the characterization sequence of parameter 1 is less than the detection threshold between 173.77s and 174.22s, the peak time is 174.0s, and the peak distance is 0.00028, indicating that parameter 1 has changed to category 1 at 174.0s, which is consistent with (a); in (d), the distance between the measured sequence and the characterization sequence of parameter 2 is less than the detection threshold between 172.02s and 172.35s, the peak time is 172.2s, and the peak distance is 0.09377, indicating that parameter 2 has changed to category 3 at 172.2s, which is consistent with (c). The above results show that the characterization sequence of each telemetry parameter can be correctly detected, and the time of occurrence of each parameter change can be accurately obtained.

[0092] Parameter change timing detection

[0093] Real-time detection of telemetry parameter changes, and obtaining the “parameter change-time” set {K,T K}, so that the timing of parameter changes can be further detected and evaluated.

[0094] Detection ideas:

[0095] 1) Take the detected instruction parameter changes as the starting point of each timing detection;

[0096] 2) Compare the instruction parameter change timing with the theoretical timing sequence to judge the correctness of the instruction timing;

[0097] 3) After the instruction parameter changes, the state parameter changes detected within a period of time (such as 2s) are classified according to the physical meaning and sorted by time as the aftereffect sequence of the instruction;

[0098] 4) Evaluate the execution effect of instructions based on the aftereffect sequence.

[0099] The results show (as shown in Table 2):

[0100] 1) It can correctly detect the timing of all instruction changes contained in the theoretical timing sequence, and the measured timing sequence is consistent with the theoretical timing sequence, but there is a difference in time;

[0101] 2) By detecting the changes in state parameters, the aftereffect parameters of each instruction can be summarized, which can be used as a verification of whether the instruction has occurred and the execution effect.

[0102] Table 2 Parameter change timing detection results

[0103]

[0104] Aiming at the problem of real-time and automatic detection of the time series of telemetry parameter changes during rocket flight, this paper proposes a detection method based on representation learning. The paper studies the problems of time series similarity measurement, time series clustering, parameter change detection and time series detection, and verifies the feasibility of the proposed method through experiments.

[0105] Although the above embodiments have been shown and described, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. Changes, modifications, substitutions and variations of the above embodiments by those of ordinary skill in the art are all within the scope of protection of the present invention.

Claims

1. A real-time detection method for telemetry parameter changes based on representation learning, characterized in that: The following steps are involved: S1. According to the actual characteristics of telemetry parameter time series and key timing windows, complete data preprocessing and select sliding correlation distance as the similarity measure of telemetry parameter time series; S2. Using the density clustering algorithm under the selected distance metric, the remote sensing parameters are classified according to the similarity of their change trends; S3. For each type of time series, select or define a characterization sequence to summarize its change pattern, and calculate the detection threshold of the characterization parameter change; S4. Calculate the matching degree between the measured sequence of each telemetry parameter and its corresponding characterization sequence, and compare it with the detection threshold to realize the real-time detection of telemetry parameter changes.

2. The real-time detection method for telemetry parameter changes based on representation learning according to claim 1 is characterized in that: The data preprocessing in S1 refers to intercepting the telemetry parameters from 30 seconds before ignition to the moment of separation of the booster, fairing, first stage or second stage as valid data segments, and performing normalization and time alignment processing.

3. The real-time detection method for telemetry parameter changes based on representation learning according to claim 2 is characterized in that: The formula for the normalization process is: Where x is the measured time series, max(x) and min(x) are the maximum and minimum values ​​of the series in the valid data segment, respectively. t and x t_norm are the original value and normalized value of the sequence at time t respectively.

4. The real-time detection method for telemetry parameter changes based on representation learning according to claim 2 is characterized in that: The time alignment process is to perform time alignment on the normalized time series. In the valid data segment, linear interpolation is performed at a frequency of 10 Hz. The interpolation process is shown in the following formula: Among them, t i is the measured time, i=1,2,...,m,t s is the time after interpolation, s=1,2,...,n, Δt s is the interpolation interval.

5. The real-time detection method for telemetry parameter changes based on representation learning according to claim 1 is characterized in that: The calculation formula of the sliding related distance in S1 is: Among them, x and y are two time series, and their time series are taken in a given time window [t a ,t b ] as the distance metric dist(x,y) between them.

6. The real-time detection method for telemetry parameter changes based on representation learning according to claim 1 is characterized in that: The classification of S2 includes the following steps: S201. Randomly select a reference sample and calculate the distance between each sample and the reference sample; S202. All samples are sorted according to the above distances, and then density clustering is performed according to the distance values ​​to obtain a number of clusters; the distances between samples within a cluster are dense, that is, the similarity is high; the distances between samples between clusters are sparse, that is, the similarity is low.

7. The real-time detection method for telemetry parameter changes based on representation learning according to claim 1 is characterized in that: The density clustering algorithm process of S2 is as follows: T1. Find the ∈-neighborhood of each sample in the dataset D and determine the core object set Ω; T2. Randomly select a core object x from Ω i , find out the cluster C1 composed of samples whose density can be reached; T3. Remove C1 from Ω and repeat step T2 until Ω is empty. The output cluster partition C = {C1, C2, ..., C k }; The clusters C1 to C1 output by the algorithm k , which is the result of time series division with highly similar sample change trends within the cluster.

8. The real-time detection method for telemetry parameter changes based on representation learning according to claim 7 is characterized in that: The specific method of S3 is as follows: calculate the sliding correlation distance between the samples in the cluster, sort them from near to far, take the first 50% of the sample pairs, and count the samples with the highest frequency; at this time, the average distance between this sample and other samples is the shortest, and under the current distance measurement, it can best represent the characteristics of the samples in the cluster, so it is selected as the central sample; at the same time, take the maximum distance d between the central sample and other samples in the cluster max As the detection threshold for characterizing the sequence.

9. The real-time detection method for telemetry parameter changes based on representation learning according to claim 1 is characterized in that: The real-time detection process of S4 includes the following steps: S401. For the telemetry parameter K, obtain its characterization sequence p K The measured value sequence x in the time window of equal length K ; S402.x K After the interpolation is completed and the time scale is aligned, the sliding related distance dist(x K ,p K ); S403. When dist(x K ,p K )<ρ K , that is, the characterization sequence is detected, indicating that the parameter K has undergone p K the changes represented; S404. Take each dist(x K ,p K )<ρ K Until the next time dist(x K ,p K )>ρ K Within the time range, dist(x K ,p K )The time T corresponding to the minimum value K , which is the time when the parameter changes.

10. The real-time detection method for telemetry parameter changes based on representation learning according to claim 9 is characterized in that: The detection time window in S401 is slid as needed for the historical mission data of the rocket flight, so as to further mine the time series of parameter changes; for the real-time flight data of the rocket, the detection time window is taken from the current moment and p K After a period of equal length has passed, real-time detection of changes in telemetry parameters is performed.

Images