A method for detecting anomalies in non-standard time-series spacecraft parameters by combining coarse and fine granularity

Abstract

To address the challenges of balancing accuracy and computational cost in spacecraft parameter detection, the difficulty of applying time-series processing methods such as Long Short-Term Memory (LSTM) neural networks to non-normalized time series, and the introduction of significant redundant data when using interpolation to normalize time series, this invention proposes a coarse- and fine-grained method for detecting anomalies in non-normalized time-series spacecraft parameters. The method includes: reading historical telemetry data files to obtain a sequence of historical telemetry engineering values ​​for the spacecraft; generating a coarse-grained normalized time series based on the historical telemetry engineering value sequence using an adaptive dynamic time window; normalizing the normalized time series to generate subsequences; training an LSTM neural network based on the set of subsequences; generating adaptive thresholds under strong and weak constraints based on the LSTM neural network; using the strong-constraint adaptive thresholds to detect anomalies in coarse-grained spacecraft parameters, and using the weak-constraint adaptive thresholds to detect anomalies in fine-grained spacecraft parameters.

Description

Technical Field

[0001] This invention belongs to the field of spacecraft testing and health assessment technology, and in particular relates to a method for detecting anomalies in non-standard time-series spacecraft parameters that combines coarse and fine granularity. Background Technology

[0002] Anomaly detection of spacecraft parameters is an important technology for assessing the health of spacecraft. Existing spacecraft parameters are diverse and involve large amounts of data. Utilizing automated detection and interpretation technologies can effectively improve detection efficiency.

[0003] Traditional automated interpretation uses fixed threshold alarms. Long-term practice has verified the stability and practicality of this method. However, for parameters of periodically changing modes, this method has a wide threshold, low sensitivity, poor adaptability, and cannot determine whether the spacecraft has performance degradation or unexpected logical changes in operating mode.

[0004] Deep learning technology, which has emerged in recent years, has received widespread attention in the field of anomaly detection. It utilizes models to learn the characteristics and distribution of data, deriving adaptive detection thresholds that dynamically adjust with data input, effectively solving the problems of wide fixed thresholds and low sensitivity. Due to the high safety requirements of spacecraft, the number of anomaly samples is extremely small. Therefore, their health assessment can use unsupervised learning-based time-series anomaly detection methods, such as Long Short-Term Memory (LSTM) neural networks. However, for spacecraft that are in a normal state most of the time, using fine-grained, high-precision neural network detection for parameters at all times is cumbersome and unnecessary.

[0005] Furthermore, during the actual sampling and transmission process in spacecraft, the time intervals of parameters vary, failing to meet the normalization requirements of such methods for the input time series. A common solution is to use interpolation to normalize the time series, but interpolation introduces a large amount of redundant data, increasing the computational burden on the model, reducing detection performance, and making it even more difficult to meet real-time detection needs. Therefore, it is difficult to apply time series processing methods based on long short-term memory neural networks for equipment health diagnosis during the spacecraft integrated testing phase. Summary of the Invention

[0006] To address the challenges of balancing accuracy and computational cost in spacecraft parameter detection, the difficulty of applying time series processing methods based on long short-term memory neural networks to non-standard time series, and the introduction of a large amount of redundant data when using interpolation to standardize time series, this invention proposes a coarse- and fine-grained method for detecting anomalies in non-standard time series spacecraft parameters.

[0007] The present invention is achieved through the following technical solution.

[0008] A method for detecting anomalies in non-standard time-series spacecraft parameters by combining coarse and fine granular approaches includes the following steps:

[0009] Step 1: Read the historical telemetry data file to obtain the sequence of historical telemetry engineering values ​​for the spacecraft;

[0010] Step 2: Generate a coarse-grained normalized time series based on the historical telemetry engineering value sequence using an adaptive dynamic time window;

[0011] Step 3: Normalize the normalized time series to generate subsequences;

[0012] Step 4: Train a long short-term memory neural network based on the set of subsequences;

[0013] Step 5: Based on the Long Short-Term Memory Neural Network, generate adaptive thresholds under strong and weak constraints;

[0014] Step 6: Use the adaptive threshold to perform coarse-grained anomaly detection on the spacecraft parameters. If the coarse-grained parameters of the spacecraft are normal, continue to acquire the data transmitted down from the spacecraft; if the coarse-grained parameters of the spacecraft are abnormal, proceed to step 7.

[0015] Step 7: Utilize the adaptive threshold to perform fine-grained anomaly detection on spacecraft parameters.

[0016] The beneficial effects of this invention are:

[0017] 1. This invention utilizes coarse-grained detection to first exclude most of the normal parameters of the spacecraft, and only performs fine-grained detection on the corresponding specific moments for a small portion of the time. This saves computing costs while ensuring the coverage and accuracy of spacecraft telemetry parameter detection, reducing detection time and improving interpretation efficiency.

[0018] 2. This invention adopts a dynamic timing normalization method, which can greatly reduce data redundancy compared with interpolation methods, and provides normalized input for long short-term memory neural networks or other timing processing methods;

[0019] 3. This invention proposes an anomaly detection method that utilizes dynamic timing specifications and strong and weak constraint thresholds to achieve full coverage and high-precision testing of spacecraft telemetry parameters with non-equal interval timing specifications.

[0020] 4. Compared with traditional anomaly detection methods based on interpolation to normalize time series, the present invention can significantly reduce data redundancy, improve the performance of long short-term memory neural networks, and also provide normalized input for other time series processing methods;

[0021] 5. This invention fills the gap in the field of non-standard time-series spacecraft parameter anomaly detection using long short-term memory neural network methods;

[0022] 6. The sequence of the present invention has been normalized, and the time interval of the telemetry parameter sequence is consistent. The amount of redundancy is significantly reduced compared with the original data, avoiding the huge redundancy introduced by the difference filling method, and can make the time interval consistent. Attached Figure Description

[0023] Figure 1 This is a flowchart of the non-standard timing spacecraft parameter anomaly detection method of the present invention;

[0024] Figure 2 This is a schematic diagram of the time series standardization method based on adaptive dynamic time window of the present invention;

[0025] Figure 3 This is a schematic diagram of telemetry parameters K58F sampled and transmitted by the spacecraft during a certain time period of the present invention.

[0026] Figure 4 This is a schematic diagram of the time series of some telemetry parameters K58F after standardization according to the present invention;

[0027] Figure 5 This is the result of the spacecraft parameter anomaly detection method combining coarse and fine granularity of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the present invention, and are not intended to limit the scope of the present invention.

[0029] like Figure 1 As shown, the present invention provides a method for detecting anomalies in non-standard time-series spacecraft parameters by combining coarse and fine granularity, specifically including the following steps:

[0030] Step 1: Read the historical telemetry data file to obtain the sequence of historical telemetry engineering values ​​for the spacecraft;

[0031] In this embodiment, the historical telemetry data uses telemetry parameters of a healthy spacecraft within a specified time interval, which are then converted to obtain structured time data and telemetry data; the structure of the data at each moment is shown in the table below:

[0032] item content Parameter name String Parameter code (parameter number) String time Numerical parameters Engineering value Numerical parameters Source code String

[0033] In practice, considering that the downlink interval of digital data of spacecraft parameters is between several milliseconds and several seconds, while analog data is generally located in the slow telemetry packets of spacecraft, with a downlink interval between several seconds and tens of seconds; therefore, in order to ensure that the data information is sufficient, the time span of the historical telemetry data in this embodiment is selected to be greater than four hours.

[0034] Step 2: Generate a coarse-grained normalized time series based on the historical telemetry engineering value sequence using an adaptive dynamic time window;

[0035] The idea behind this step is that the sampling intervals for digital and analog data of spacecraft parameters are different, and the intervals within spacecraft parameter classes are also different, making it difficult to directly apply time series processing algorithms. Therefore, this embodiment uses an adaptive dynamic time window to generate a normalized time series.

[0036] In this embodiment, the length of the time window is 1×n, where 1 represents the spacecraft parameter class and n represents the length of the window, i.e., the engineering values ​​of the telemetry parameters at n consecutive moments within the window. In specific implementation, the window length n is not fixed, but the time interval k seconds represented by the window is fixed, i.e., the number of engineering values ​​within k seconds is n, where the integer k is defined as follows:

[0037]

[0038] Where, γ∈N + The window scaling factor is represented by t, the average time interval is represented by t, the total time length is represented by T, N represents the number of parameters within that time interval, and α and β represent... and The weighting coefficients between them are represented by the int(·) function, which rounds down the time. In practice, if the window time interval k seconds is too large, the changes in spacecraft parameter points will be difficult to detect; if the interval is too small, the spacecraft parameter points will not be included in the time window. Therefore, the time window k is set to 60 seconds for spacecraft analog parameters and 1 second for spacecraft digital parameters.

[0039] like Figure 2 As shown in this embodiment, the generation of coarse-grained normalized time series is specifically as follows:

[0040] A dynamic time window is used to perform sliding window sampling on the spacecraft's historical telemetry engineering value sequence x. Coarse-grained normalized spacecraft parameter time series is generated using the in-window parameter statistics. Each value in the coarse-grained normalized spacecraft parameter time series can be traced back to the telemetry engineering value of the original sequence through a corresponding mapping relationship for further fine-grained detection.

[0041] Example 1: Traverse the historical telemetry engineering value sequence x, with a dynamic window size of 1×n (not fixed), a window representing a time length of k seconds, and a sliding window step size of dynamic value n, to obtain a set of window sequences.

[0042]

[0043] n iLet X represent the number of engineering values ​​in the i-th time window. The set X contains m vector sequences y, where... Let x represent the engineering values ​​of the spacecraft parameter sequence x within k seconds starting from time j+1. Based on the statistics of each vector element in the window sequence set X, including the mean, maximum, and minimum values, three normalized time series vectors x are generated. MEAN x MAX and x MIN ,in:

[0044]

[0045]

[0046]

[0047] The functions mean(·), max(·), and min(·) represent the mean, maximum, and minimum values ​​of a vector element, respectively. (Time series x) MEAN Used for training Long Short-Term Memory neural networks, time series x MAX and x MIN Used for coarse-grained temporal anomaly detection, x MEAN The correspondence f between the original sequence X and the window sequence set is:

[0048]

[0049] This correspondence allows us to map coarse-grained time series points back to obtain the telemetry engineering values ​​of the original sequence, facilitating fine-grained detection.

[0050] Step 3: Normalize the normalized time series to generate subsequences; specifically:

[0051] The normalized time series vector x obtained in step two MEAN Perform maximum and minimum value normalization on it, where the maximum value ma = max(x) MEAN The minimum value mi = min(x) MEAN After normalization, the time series vector is obtained. The calculation method is as follows:

[0052]

[0053] When the maximum value ma and the minimum value mi are the same, then It is a zero vector.

[0054] Meanwhile, for the time series x MAX and x MIN Normalization was also performed, resulting in the following time series vectors: and The calculation method is as follows:

[0055]

[0056] In this embodiment, the vector Generate subsequences to obtain the training set for training the Long Short-Term Neural Network, specifically as follows: Utilize a window of size 1×w, in... Perform sliding window segmentation with a step size of 1 to obtain the set of segmented subsequences. Each subsequence Represents the i-th window subsequence. The last l parameters of the segmented subsequence are used as the network's prediction output, where l can be set to 1. Therefore, the corresponding label set...

[0057] Step 4: Train a long short-term memory neural network based on the set of subsequences;

[0058] Based on the subsequence set obtained in step three Train a long short-term memory neural network and use the backpropagation algorithm to iteratively update the hyperparameters of the neural network.

[0059] Example 2: The pseudocode for training the Long Short-Term Memory Neural Network is shown in Table 1 below:

[0060] Table 1. Pseudocode for training a Long Short-Term Memory Neural Network

[0061]

[0062] Step 5: Based on the Long Short-Term Memory Neural Network, generate adaptive thresholds under strong and weak constraints;

[0063] The idea behind this step is that the adaptive threshold generation method under strong and weak constraints can be used for anomaly detection and health diagnosis in coarse-grained anomaly detection and fine-grained anomaly detection, respectively. The strong constraint means tightening the threshold range, which can make ambiguous points that are difficult to distinguish fall into the anomaly range. Then, according to the correspondence f formed in step two, the ambiguous points are traced back to obtain the telemetry engineering value of the original sequence. Finally, the weak constraint with a larger threshold range is used for anomaly detection and localization.

[0064] In this embodiment, the generation of adaptive thresholds under strong and weak constraints is specifically as follows:

[0065] Set training set The Long Short-Term Memory Neural Network obtained after training in step four is f θ The model prediction output is obtained as follows The actual output data is Calculate the mean square error (MSE) between the predicted output and the actual output for any subsequence to be predicted. Its predicted output is For both strong and weak constraint thresholds, a general adaptive upper bound is used. and lower limit The definition is as follows:

[0066]

[0067]

[0068] Where m represents the constraint coefficient, and m>0;

[0069] The constraint coefficients for strong and weak constraints are different. In the case of strong constraint adaptive threshold, m = m strong In the case of weakly constrained adaptive threshold, m = m weak And satisfy m strong <m weak .

[0070] Step 6: Use strong constraint adaptive thresholds to perform anomaly detection on coarse-grained spacecraft parameters. If the coarse-grained spacecraft parameters are normal, continue to acquire downlink data from the spacecraft; if the coarse-grained spacecraft parameters are abnormal, proceed to Step 7.

[0071] In this embodiment, a strong constraint adaptive threshold is used to detect anomalies in coarse-grained spacecraft parameters. The specific steps are as follows:

[0072] Based on the upper and lower thresholds of the strong constraints obtained in step five, x reMAX With strong constraint upper limit Compare, and put x reMIN With strong constraint lower bound Compare, among which:

[0073]

[0074]

[0075] When the detection result of the spacecraft's coarse-grained parameter anomaly is normal, then:

[0076] and

[0077] At this point, the spacecraft's coarse-grained parameters are considered normal, and no further fine-grained detection is required. The spacecraft continues to acquire downlink data to obtain a new sequence to be detected and continues to perform coarse-grained detection.

[0078] Otherwise, the coarse-grained parameters of the spacecraft are considered to be ambiguous time points where anomalies may exist.

[0079] Step 7: Utilize weakly constrained adaptive thresholds to perform anomaly detection on fine-grained spacecraft parameters; the specific steps are as follows:

[0080] First, regarding the prediction Spacecraft coarse-grained parameters at any fuzzy time point T0 Calculate the spacecraft coarse-grained parameters at time points T0-1 and T0+1 respectively. and The slopes ks1 and ks2 between them are given by the following formulas:

[0081]

[0082]

[0083] Where k represents the time interval of the dynamic time window as k seconds;

[0084] Then, according to the correspondence f, the fine-grained set of the original sequence X corresponding to the fuzzy time point T0 is determined. The dynamic window time period is T1 to T2 = T1 + k. If T0 is considered the midpoint of this time period, then the adaptive upper and lower thresholds from T1 to T0 are considered... and They are as follows:

[0085]

[0086]

[0087] Adaptive upper and lower thresholds from T0 to T2 and They are as follows:

[0088]

[0089]

[0090] As can be seen, the above processing enables the upper and lower thresholds to adapt to the spacecraft parameter change patterns within a short time window during fine-grained detection.

[0091] Finally, the original sequence Each telemetry value is compared with its upper and lower limits for the corresponding time period. If each value is less than the upper threshold and greater than the lower threshold, the original sequence is considered normal, and data transmitted from the spacecraft continues to be acquired to obtain a new sequence to be tested. Otherwise, telemetry values ​​that do not meet the above rules are considered anomalies, and the location of real anomalies and anomaly alarms are triggered in real time.

[0092] Example 3: The following is an example of anomaly detection of a telemetry parameter of a certain sensor on a certain spacecraft. The parameter settings are shown in Table 2:

[0093] Table 2 Examples of Spacecraft Parameter Anomaly Detection Parameters

[0094]

[0095]

[0096] Figure 3 This is a schematic diagram of the telemetry parameters K58F sampled and transmitted by the spacecraft over a certain period. Table 3 shows some of the engineering values ​​of K58F, and it can be seen that the sampling time of K58F is non-uniform.

[0097] Table 3. Engineering values ​​of some K58F telemetry parameters.

[0098] time K58F …… …… 2021-11-10 12:35:19.418000 2.54 2021-11-10 12:35:20.415000 2.54 2021-11-10 12:35:21.400000 2.54 2021-11-10 12:35:22.410000 2.54 2021-11-10 12:35:22.896000 2.54 2021-11-10 12:35:23.393000 2.54 2021-11-10 12:35:23.896000 2.54 2021-11-10 12:35:24.394000 2.54 …… ……

[0099] After normalizing the parameter sequence using a dynamic time window of k=60s, the results of the telemetry parameter time series are shown in Table 4 below. It can be seen that after normalization, the time intervals of the telemetry parameter sequence are consistent, the amount of redundancy is significantly reduced compared to the original data, and the massive redundancy introduced by the interpolation filling method is avoided. Furthermore, the time intervals are consistent. Therefore, time series processing methods, such as long short-term memory neural networks, can be used to process it. Figure 3 The normalized results of spacecraft sampling data over a medium time period are shown in the attached figure. Figure 4 As shown.

[0100] Table 4. Engineering values ​​of K58F telemetry parameters after dynamic time window standardization.

[0101]

[0102]

[0103] Using telemetry parameter data up to November 2021, after processing according to steps two through four, a trained Long Short-Term Memory (LSTM) neural network was obtained. Steps five through seven were then used to perform anomaly detection on parameters for a specific day in November 2021. Anomalies were identified in the parameter sequence through fault injection. The output results of coarse-grained and fine-grained detection are as follows: Figure 5As shown in the diagram. In the coarse-grained detection results, the sequence of maximum telemetry parameters is compared with the upper envelope under strong constraints, and the sequence of minimum values ​​is compared with the lower envelope under strong constraints, resulting in three fuzzy points: point 1, point 2, and point 3. Then, fine-grained detection is performed on the time period data corresponding to each fuzzy point. Following the method in step seven, fine-grained anomaly detection upper and lower envelopes under weak constraints are generated. By comparing the mapped real telemetry parameter points with the upper and lower envelopes, the anomaly points can be identified and their occurrence time located. Simultaneously, the inconsistent sampling intervals are clearly visible in the parameter sequences corresponding to points 2 and 3, indicating that this scheme can effectively achieve high-precision detection of anomalies in spacecraft parameters within non-standard time series.

[0104] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A coarse and fine-grained combined non-normal timing spacecraft parameter anomaly detection method, characterized in that, The method comprises the following steps: Step one, reading a historical telemetry data file to obtain a spacecraft historical telemetry engineering value sequence; Step two, generating a coarse-grained normalized time sequence according to the historical telemetry engineering value sequence using an adaptive dynamic time window; Step three, performing normalization processing on the normalized time sequence to generate a subsequence; Step four, training a long short-term memory neural network according to the subsequence set; Step five, generating an adaptive threshold under strong and weak constraints based on the long short-term memory neural network; Step six, performing abnormality detection on coarse-grained spacecraft parameters using the strong constraint adaptive threshold, if the coarse-grained spacecraft parameters are normal, continue to obtain spacecraft downlink data; if the coarse-grained spacecraft parameters are abnormal, go to step seven; Step seven, performing abnormality detection on fine-grained spacecraft parameters using the weak constraint adaptive threshold; The generation of the coarse-grained normalized time sequence is specifically as follows: Adopting dynamic time window in the spacecraft history telemetry engineering value sequence Sliding window sampling is performed on the spacecraft history telemetry engineering value sequence, and a coarse-grained normalized spacecraft parameter time sequence is generated by using the window parameter statistics, wherein each value of the coarse-grained normalized spacecraft parameter time sequence can be traced back to the telemetry engineering value of the original sequence through a corresponding mapping relationship, for further fine-grained detection.

2. The method of claim 1, wherein, The historical telemetry data uses healthy spacecraft telemetry parameters in a specified time interval, and after format conversion, structured time data and telemetry data are obtained.

3. The method of claim 2, wherein, The time span of the historical telemetry data is greater than four hours.

4. The method of claim 1 or 2 or 3, characterized in that, When training the long short-term memory neural network, the error back propagation algorithm is used to perform neural network hyperparameter iterative updating.

5. The method of claim 1 or 2 or 3, wherein, The generation of the adaptive threshold under strong and weak constraints is specifically as follows: Set training set wherein a represents the number of vector sequences of engineering values of the spacecraft parameter sequences contained in the window sequence set, w represents the sliding window width; the long short-term memory neural network obtained after step four is and the model prediction output is , the real output data is , the mean square error of the prediction output and the real output is calculated , for any to-be-predicted sub-sequence , the prediction output is , for the strong constraint and the weak constraint threshold, the general adaptive upper limit and the lower limit are defined as follows: wherein represents a constraint coefficient, .

6. The method of claim 5, wherein, The constraint coefficients of the strong constraint and the weak constraint are different, at the strong constraint adaptive threshold, , at the weak constraint adaptive threshold, , and satisfy .

7. The method of claim 1 or 2 or 3, wherein, The abnormality detection on the coarse-grained spacecraft parameters using the strong constraint adaptive threshold is specifically as follows: According to the upper and lower thresholds of the strong constraint obtained in step five, compare with the upper threshold of the strong constraint and compare with the lower threshold of the strong constraint wherein: When the coarse-grained spacecraft parameter abnormality detection result is normal, then: and At this time, it is considered that the coarse-grained spacecraft parameters are normal, and there is no need to further perform fine-grained detection, and the spacecraft downlink data is continuously obtained to obtain a new to-be-detected sequence and continuously perform coarse-grained detection; Otherwise, it is considered that the coarse-grained spacecraft parameters are a fuzzy time point that may exist abnormality.

8. The method of claim 1 or 2 or 3, wherein, The abnormality detection on the fine-grained spacecraft parameters using the weak constraint adaptive threshold is specifically as follows: First, regarding the prediction Any fuzzy time point within Spacecraft coarse-grained parameters Calculate its time before and after the time intervals. and Spacecraft coarse-grained parameters and slope between and ; Then, according to the correspondence relationship Fuzzy time point Corresponding set original sequence Fine granularity Dynamic window time period To , , think The middle time point of the time period, think To Adaptive upper and lower thresholds And ; Adaptive upper and lower thresholds And ;​ Finally, compare each telemetry engineering value in the original sequence with the upper and lower limits in the corresponding period, if each engineering value is less than the upper limit and greater than the lower limit, the original sequence is normal, and the spacecraft downlink data is continuously acquired to obtain a new sequence to be detected, otherwise, the telemetry engineering value that does not satisfy the above rule is considered as an abnormal point.

9. The method of claim 8, wherein, When it is considered that the telemetry engineering value that does not satisfy the above rule is an abnormal point, real-time positioning and abnormality alarm of the real abnormal point are performed.

Images