A large-scale low-orbit satellite network congestion awareness and failure prediction method

Abstract

The application discloses a large-scale low-orbit satellite network congestion awareness and fault prediction method, and belongs to the field of satellite fault positioning.The application performs space-time fusion coding on a link state sequence, and reduces the sparsity of link state information. By selecting an optimal matrix as the input of a prediction model, the data training and prediction efficiency is improved. The Informer model is used to train and predict the link state sequence, and a long sequence is extracted through a multi-head ProbSparse self-attention layer and a self-attention distillation layer. The decoder generates predictive prediction through a masked multi-head ProbSparse self-attention layer and a multi-head attention layer, and outputs a multi-step prediction matrix, thereby improving the speed of fault prediction. The fault detection interval based on the difference probability distribution is calculated, and the detection message overhead is reduced. Through dynamic updating of the fault detection parameter and the cycle of prediction to detection to prediction, the accuracy of satellite node congestion awareness and fault prediction is continuously improved while the detection overhead is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of satellite fault positioning, and particularly relates to a large-scale low-orbit satellite network congestion awareness and fault prediction method BACKGROUND

[0002] With the rapid growth of mobile devices, pure ground communication networks and expensive geostationary satellites have been insufficient to meet the future network capacity requirements, and the construction of low-orbit satellite networks has become imminent. However, the complex space environment brings great challenges to maintaining the stability of satellite communication. A large number of high-energy particles and cosmic radiation can damage the electronic equipment on the satellite, and then cause long-time link interruption failure; and the very limited computing and storage resources of the satellite can cause frequent congestion failure of the node. Therefore, the low-orbit satellite network urgently needs an accurate and efficient fault positioning method.

[0003] Satellite network devices differ from ground network devices in design and performance, and are subject to strict power, weight and size constraints, which limits their computing and storage capabilities to some extent. The energy required for on-board communication, storage and calculation is all converted from the power of solar panels. Therefore, compared with ground networks, satellite systems will more frequently experience congestion when processing data streams. However, satellite congestion has certain regularity: 1: As most Internet users are located in the northern hemisphere, the task request volume to the satellite is larger, and congestion is more likely to occur in the northern latitude area; 2: As the ground station connects the satellite network and the ground network, congestion is more likely to occur near the satellite near the ground station. In addition, the high dynamics of low-orbit satellite networks easily cause inter-satellite communication link failures, causing business data transmission to be interrupted and unable to guarantee the quality of service of end users. However, the spatiotemporal periodicity of the satellite's movement around the earth also makes the link failure have certain regularity, for example: due to the limitations of antenna tracking, the inter-plane link of the satellite after entering the polar region will be cut off and need to be rebuilt, the signal acquisition and tracking process is very complex and consists of six main steps, namely open-loop pointing, active end scanning, passive end acquisition, active end acquisition, double-end acquisition and stable tracking. The satellite may cause calculation errors or delays in some steps due to software bugs or computing resource overload, and then cause long-time link interruption. The regularity of these link interruptions makes it possible to predict satellite failures. In addition, due to the difference in time scale between congestion and link failure, it is easy to distinguish whether the predicted outage result is node congestion or link failure.

[0004] With the development of artificial intelligence, the method of time series prediction is also evolving. Traditional time series prediction methods, such as moving average model (MA), integrated moving average autoregressive model (ARIMA), etc. have been widely used for a long time. In recent years, deep learning models have become the new darling of time series prediction due to their powerful functions and excellent performance. Deep learning models such as recurrent neural network (RNN), long short-term memory network (LSTM) and gated recurrent unit (GRU) are applied to capture the complex features of time series data. The Informer model is the latest deep learning model based on Transformer. It introduces the ProbSpare self-attention mechanism to reduce the computational complexity, and efficiently processes extremely long input sequences through self-attention distillation technology. These improvements make the Informer model have broad development prospects in the satellite fault prediction scenario. SUMMARY

[0005] The purpose of the present application is to provide a large-scale low-orbit satellite network congestion awareness and fault prediction method, which can realize large-scale low-orbit satellite network congestion awareness and fault prediction, has the advantages of high prediction efficiency, high precision and low detection overhead.

[0006] The large-scale in the large-scale low-orbit satellite network refers to the number of low-orbit satellites being greater than 4000.

[0007] The purpose of the present application is achieved by the following technical solutions.

[0008] The application discloses a large-scale low-orbit satellite network congestion awareness and fault prediction method, which is based on BFD protocol to collect link information between two satellites; data is regularized on the link information to obtain a standard link state sequence; the standard link state sequence is spatio-temporally fused and coded according to a circuit breaking starting position and a duration to obtain a link state matrix, and the link state matrix under different combinations is compared to select an optimal matrix as an input of a prediction model. The link state matrix is input into an Informer model, the Informer model performs embedding operation on the link state matrix through dimension lifting, position coding and time feature coding to obtain an embedded vector; the embedded vector is input into an encoding layer, a decoding layer and a full connection layer to obtain a predicted link state matrix. The encoder extracts robust long-range dependencies of a long sequence input through a multi-head ProbSparse self-attention layer and a self-attention distillation layer; the decoder performs generative prediction through a masked multi-head ProbSparse self-attention layer and a multi-head attention layer to directly output a multi-step prediction matrix, thereby avoiding accumulated error problems in a propagation process and improving the prediction speed. Fault detection parameters are initialized, the fault type of a circuit breaking is judged according to the predicted link state matrix, and the fault type is divided into node congestion and link fault; a fault detection interval based on a difference probability distribution is calculated according to the difference between a predicted circuit breaking starting value and an actual circuit breaking starting value in a test set, so that the detection accuracy of the circuit breaking is ensured while the overhead of a detection packet is reduced; and a reaction time of judging the fault type after the circuit breaking is calculated according to the circuit breaking duration in the test set. After the circuit breaking is predicted each time, the node congestion and the link fault are evaluated according to the parameters set in step three. A plurality of parameters are updated: a detection point {t1,...,t i ,...,t w} after the circuit breaking is predicted next time is updated based on a circuit breaking starting point; a node congestion duration threshold is recorded to update ΔT, and ΔT' and ΔT'' are updated based on the circuit breaking duration. The link state data is added to the original link state sequence after data preprocessing, and then input into the Informer model for training and prediction of the next circuit breaking. Through dynamic updating of the fault detection parameters and the cycle of prediction, detection and prediction, the detection overhead is reduced while the accuracy of the satellite node congestion awareness and fault prediction is continuously improved.

[0009] The application discloses a large-scale low-orbit satellite network congestion awareness and fault prediction method, which comprises the following steps:

[0010] Step one, through BFD protocol in the collection of two satellite link information; link information data regularity obtained, obtained standard link state sequence; according to the starting position and duration of the circuit, the space-time fusion coding of standard link state sequence is carried out, and the link state matrix is obtained. The link state matrix under different combinations is compared, and the optimal matrix is selected as the input of the prediction model.

[0011] Step 1.1: After the establishment of BFD session between adjacent two satellite nodes, BFD packets are periodically sent on the channel between the two satellites. When one party does not receive BFD packet within the negotiated detection time, it is determined that the circuit is broken. The BFD protocol detection interval is set to B milliseconds, the channel is recorded as 1, the circuit is recorded as 0, and the original link state sequence is recorded as X={x1,...,x f ,...,x N}, that is, the link state data is obtained by using bidirectional forwarding detection BFD protocol. The obtained link state data is divided into training set and test set.

[0012] Step 1.2: In view of the difference in time scale between node congestion and link failure, according to formula (1), the circuit with duration of 1s-10s and the circuit with duration of 1min or more in the original link state sequence are reserved:

[0013]

[0014] The standard link state sequence is obtained, recorded as X'={x1',...,x f ′,...,x N ′}

[0015] Step 1.3: According to the redundancy and continuity of link state sequence, the standard link state sequence is compressed by space-time fusion coding, and the sparseness of link state sequence is reduced.

[0016] Redundancy refers to that the link state sequence is composed of numerical circuit 0 and channel 1, and since the link circuit state will appear as an unusual exception, the link state sequence is mostly channel 1, which has high redundancy.

[0017] Continuity refers to that the link channel and circuit state will last for a period of time, so the circuit 0 and channel 1 of the sequence have continuity and will not appear alternately.

[0018] According to formula (2), the starting value Z1:{z1,...z i …,z u} and time interval Z2:{z'1,...z'\ i …,z' u} of standard link state sequence statistics are set:

[0019]

[0020] Among them, I min I is the minimum interval between the starting values ​​of adjacent consecutive disconnections in the standard link state sequence. max Let be the maximum interval between the starting values ​​of adjacent consecutive disconnections in the standard link-state sequence, O be the duration of the most frequent congestion, and L be the length of the standard link-state sequence. At this point, for Z1 and Z2, u... 2 Combinations (z) i ,z′ i Based on Z2, the standard link state sequence is divided into s = [L / Z2] subsequences, and the number c of persistent disconnections in the j-th subsequence is recorded. j Starting position of circuit breaker and duration of circuit breaker j∈[1,[L / Z2]], the starting position of the circuit break is x jcj As the first value of the column-number pair, the duration of the circuit break is... As the second value, the column-directed pair form is: By vertically arranging the column pairs contained in each subsequence, the standard link-state sequence is increased in dimension to obtain the link-state matrix, which takes the following form:

[0021]

[0022] The link state matrix has a dimension of 2max{c j Given that the number of consecutive faults within each subsequence is unequal, for c j <c max The time interval column, missing parts are used To complete it in the form of, but The lack of fault information leads to wasted resources. The goal is to minimize the number of faults in the sequence while maximizing the time interval Z2. Useless information is used to calculate the link-state matrix after spatiotemporal fusion coding for each combination, and the total number of column pairs is recorded. Number of column pairs. First, select the combination with the smallest total number of column pairs, then select from these combinations... The combination with the smallest number of column pairs is used as the input to the Informer prediction model, and the output link state matrix under this combination is used as the input to the Informer prediction model.

[0023] Step two, input the link state matrix into the Informer model, the Informer model embeds the link state matrix through dimensionality, position encoding and time feature encoding to obtain the embedded vector; the embedded vector is input into the encoder layer, the decoder layer and the full connection layer to obtain the predicted link state matrix. The encoder extracts the robust long-range dependencies of the long sequence input through the multi-head ProbSparse self-attention layer and the self-attention distillation layer; the decoder generates the predicted matrix through the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, directly outputs the multi-step predicted matrix, avoids the cumulative error problem caused by the propagation process, and improves the prediction speed.

[0024] Step 2.1: The dimensionality of the Informer model is converted from a 2max{c j}dimensional vector to a 512-dimensional vector through one-dimensional convolution, and the position encoding method in the Transformer model is used to retain the time sequence information in the time sequence data by using fixed displacement embedding:

[0025]

[0026]

[0027] wherein, d model is the feature dimension of the input representation, In order to enable the similarity calculation of self-attention to access the global context, a 1-D convolution filter is used to project the scalar context to a d model -dimensional vector After superimposing the dimensionality, position encoding operation, the final input vector is obtained:

[0028]

[0029] Step 2.2: The encoder extracts the long sequence input through the multi-head ProbSparse self-attention layer and the self-attention distillation layer, and distinguishes the few point products that contribute to the main attention in the multi-head ProbSparse self-attention layer; according to the self-attention mechanism, the attention of the a th query is defined as a kernel smoother in the form of probability:

[0030]

[0031] Based on equation (6), the attention of the a th query to all keys is defined as the probability p(k e | q a ), the dominant point product will make the attention probability distribution of the corresponding query far away from the uniform distribution; according to the KL divergence formula (7), the distance between the two distributions is measured:

[0032]

[0033] Neglecting the constant term, the sparsity index of the ath query is defined as:

[0034]

[0035] According to formula (6) (7) (8), the multi-head ProbSparse self-attention mechanism allows each key to focus only on part of the dominant query:

[0036]

[0037] Based on the multi-head ProbSparse self-attention mechanism, different sparse queries are generated for each head, thereby avoiding severe information loss; according to the distillation of the output of the multi-head ProbSparse self-attention layer through one-dimensional convolution, the long-range dependency coupling between the output and the input is efficiently captured, and the specific form is as follows:

[0038]

[0039] wherein, denotes an attention block, which contains a multi-head ProbSparse self-attention matrix and basic operations, ELU is an activation function, and MaxPool is a maximum pooling.

[0040] Step 2.3: The decoder performs generative prediction and directly outputs a multi-step prediction matrix: the input form of the decoder is a sequence of length N token +N output , that is, the splicing sequence of the starting token of the target sequence and the placeholder of the target sequence and the timestamp, and the placeholder is set to 0:

[0041]

[0042] After sequentially passing through the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, an output matrix of (N token +N output xd is obtained, d is the dimension of the embedding. The matrix is then passed through a fully connected layer to obtain an output matrix, which is the predicted link state matrix.

[0043] The error function of the test set uses the indicators of normalized root mean square error nRMSE and relative root mean square error rRMSE to measure the prediction performance:

[0044]

[0045]

[0046] wherein, y iis the true value of the test set part of the Informer model, is the predicted value of the test set part.

[0047] Step three, initialize fault detection parameters, judge the fault type of the circuit break according to the predicted link state matrix, the fault type is divided into node congestion and link fault; Calculate the fault detection interval based on the difference probability distribution according to the difference between the predicted circuit break start value and the actual circuit break start value in the test set, reduce the detection message overhead while ensuring the accuracy of circuit break detection; Calculate the reaction time of judging the fault type after detecting the circuit break according to the circuit break duration in the test set.

[0048] Step 3.1: Initialize fault detection parameters, judge the fault type of the circuit break according to the predicted link state matrix, the fault type is divided into node congestion and link fault; Because of the obvious difference in nature between congestion and link fault, the predicted result of the circuit break in seconds is defined as congestion, and the predicted result in minutes is defined as link fault.

[0049] Step 3.2: Calculate the fault detection interval based on the difference probability distribution according to the data in the test set. After predicting the general trend of the time series based on the Informer model, in order to accurately locate the actual circuit break start point, it is necessary to continuously detect the link state around the predicted start point. The fault detection interval based on the difference probability distribution reduces the detection message overhead while ensuring the accuracy of circuit break detection. The predicted circuit break start point in the test set is recorded as The actual circuit break start point is recorded as Then T' c and T c The interval sequence ΔT1=T c -T' c .

[0050] Calculate the fault detection interval according to the probability distribution of ΔT1, the specific steps are as follows:

[0051] Use the DBSCAN algorithm based on density to eliminate outliers in ΔT1. DBSCAN algorithm uses two parameters of minimum point number minPts and neighborhood radius Eps for clustering. Arbitrarily select a point in the data set for clustering. If there are at least minPts points within the radius of Eps, mark the point as a core point and expand a cluster with the point as the center; If there are less than minPts points within the radius of Eps, but there are core points, then the point is marked as a boundary point and added to the cluster of core points; If the point is neither a core point nor a boundary point, it is marked as an abnormal point. Iterate minPts and Eps to find outliers and eliminate them; Until all points are visited to complete clustering.

[0052] After removing outliers, all intervals are arranged in ascending order and duplicate intervals are removed to obtain the interval sequence {Δt1,...,Δt}. g ,...,Δt n}, calculate the probability of each interval appearing in the interval sequence in ΔT1, and obtain the probability distribution P:{P1,...,P} n}; When w detection points {t1,...,t} are set h ,...,t w The predicted starting point of the node disconnection is t0, the first detection point is t1 = t0 + Δt1, and the last detection point is t. w =t0+Δt n The interval between adjacent detection points when the probability distributions are different is determined by the following formula:

[0053] if Then the h-th detection point

[0054] Step 3.3: Initialize the three reaction times for determining the fault type after detecting a circuit breaker based on the circuit breaker duration in the standard link state sequence. Upon detecting a circuit breaker predicted to be caused by node congestion, a load balancing strategy is immediately issued. This strategy involves the controller selecting half of all flow tables with high traffic demand passing through congested nodes. After deleting the congested node from the network topology, the shortest path algorithm is recalculated for the source nodes of the corresponding flow tables and the path is reissued. If the circuit breaker is caused by node congestion, the node congestion recovers quickly under the influence of the load balancing strategy; otherwise, the load balancing strategy will not promote link recovery. Upon detecting a circuit breaker predicted to be caused by link failure, a new routing strategy is immediately issued. This new routing strategy involves the controller updating the original routing table based on the shortest path algorithm in the network topology after removing the link. However, if the actual circuit breaker is caused by node congestion, the link recovers quickly. To further evaluate the accuracy of the predicted fault types, the congestion duration and link fault duration in the standard link state sequence were recorded. The DBSCAN algorithm was used to remove outliers. The initial response time ΔT for determining whether a circuit breaker is a node congestion after the load balancing policy is issued and the response time ΔT′ for determining whether a circuit breaker is a node congestion after the new routing policy is issued were initialized to the minimum congestion duration after removing outliers. The response time ΔT″ for determining whether a circuit breaker is a link fault after the new routing policy is issued was initialized to the maximum link fault duration after removing outliers.

[0055] Step 4: After each predicted circuit breaker, evaluate both node congestion and link failure scenarios based on the parameters set in Step 3. Update various parameters from Step 3: Update the detection points {t1,...,t} after the next predicted circuit breaker based on the circuit breaker start point. w};Based on a satellite memory threshold record node congestion duration update ΔT, based on the duration of the circuit breaker update ΔT' and ΔT''. The link state data is added to the original link state sequence after data preprocessing input Informer model training and prediction of the next circuit breaker. Through the dynamic update of the fault detection parameters and the prediction of the detection to the prediction cycle, while reducing the detection overhead, the accuracy of satellite node congestion awareness and fault prediction is continuously improved.

[0056] Step 4.1: After data preprocessing of the collected original link state sequence, the Informer model is trained and the link state matrix is predicted. After predicting the circuit breaker, the fault detection point {t1,...,t w} are sent in turn, if no circuit breaker is detected, the original link state sequence end to t w The time of the path duration is added to the original link state sequence to retrain and predict. If a circuit breaker is detected, the fault type of the circuit breaker is determined according to the predicted circuit breaker duration, if it is determined that the circuit breaker is node congestion, step 4.2 is performed, if it is determined that the circuit breaker is link fault, step 4.2 is performed. In order to adapt to the change of link characteristics, the difference between the predicted circuit breaker starting point and the detected circuit breaker starting point is added to the ΔT1 sequence in step 3.2 to recalculate the fault detection interval, and update the detection point {t1,...,t w} after the next prediction of the circuit breaker.

[0057] Step 4.2: If it is determined that the circuit breaker is node congestion, the load balancing strategy is issued. If the reason for the circuit breaker is node congestion, under the influence of the load balancing strategy, the node receives a large amount of traffic, the buffer queue is continuously reduced, and the node congestion is quickly alleviated.

[0058] The node monitors the queue occupancy rate at any time, and after the node memory is reduced to a suitable threshold, it is indicated that the node enters a state that is not easy to congest at this time, and the time is recorded and sent to the controller. In order to obtain the expected effect of the above mechanism, the setting of the threshold is very important, which is denoted as T1, in order to calculate the value, a value representing the busy state of the node buffer is added, denoted as T2.

[0059] T1 and T2 should be able to cope with the following extreme cases: when the satellite checks the buffer queue, the queue occupancy rate is close to but still lower than T1, after checking, the queue occupancy rate just exceeds T1, but the satellite needs to wait for the check interval Δt Q In order to ensure that the queue occupancy rate does not exceed T2 during Δt Q , T1 and T2 should satisfy:

[0060] (T2-T1)·L·S≥(I-O) max ·ΔtQ (14)

[0061] where L denotes the size of the buffer queue, S denotes the average packet size, (I-O) max denotes the maximum difference between the input and output traffic of the node during this period.

[0062] If the queue occupancy just exceeds T2 after a check, the satellite also needs to wait Δt Q to realize the change in queue occupancy. In order to avoid discarding packets due to buffer overflow, T2 should satisfy:

[0063] (1-T2)·L·S≥(I'-O') max ·Δt Q (15)

[0064] According to formula (15), we get:

[0065]

[0066]

[0067] Considering the extreme case, T1 and T2 are set as:

[0068]

[0069]

[0070] If the queue occupancy monitored by the congested node drops below T1 within ΔT time, it indicates that the prediction of the outage duration is correct, and the node congestion has recovered. The original congested node needs to feed back the time point when the queue occupancy drops to T1 to the controller to update the fault detection parameter ΔT. The controller subtracts the time point when the queue occupancy drops to T1 from the detected outage starting point to obtain the congestion duration, and ΔT takes the maximum value of the duration of the current node congestion and the duration of the node congestion after the last load balancing strategy is issued. The congestion duration is added as incremental data to the original link state sequence, preprocessed, and then trained and predicted again in the Informer model.

[0071] If the controller does not receive a node passage notification within ΔT time, it indicates that the prediction of the outage duration is incorrect. The error and the reason for the outage are link failure, and a new routing strategy is issued. The adjacent nodes continuously monitor the link state, record the fault duration, and return to the Informer model for retraining and prediction.

[0072] Step 4.3: If the circuit breaker is judged as link failure, a new routing strategy is issued. If the controller receives a node passage notification within ΔT', it indicates that the prediction of the circuit breaker duration is wrong, the cause of the circuit breaker is node congestion, the new routing table removes the circuit breaker link, and the traffic entering the node through the link is reduced to 0, and the node congestion is quickly alleviated. ΔT' takes the maximum of the duration of the current node congestion and the duration of the node congestion after the last time the new routing strategy is issued. The congestion duration is added to the original link state sequence, which is preprocessed and input into the Informer model for retraining and prediction.

[0073] If the controller receives a node passage notification within the time range of ΔT'', it indicates that the prediction result is correct, and the node has a link failure. Update ΔT'' based on the duration of the link failure, and ΔT'' takes the maximum of the duration of the current link failure and the duration of the last link failure. The duration of the link failure is added to the original link state sequence, which is input into the Informer model for retraining and prediction.

[0074] If the controller does not receive a node passage notification within the time range of ΔT'', it is determined that the node has a permanent and unrecoverable hardware failure. At this time, the control center collects the alarm information of the failed node and delivers it to manual processing.

[0075] Step four: Through the dynamic update of the fault detection parameters and the prediction from detection to prediction, the accuracy of the satellite node congestion awareness and fault prediction is continuously improved while reducing the detection overhead.

[0076] Advantages:

[0077] 1. The large-scale low-orbit satellite network congestion awareness and fault prediction method disclosed in the application encodes the link state sequence in space-time fusion to obtain a link state matrix, thereby reducing the sparsity of the link state information. The link state matrices under different combinations are compared, and the optimal matrix is selected as the input of the prediction model, thereby improving the data training and prediction efficiency.

[0078] 2. The large-scale low-orbit satellite network congestion awareness and fault prediction method disclosed in the application uses the Informer model to train and predict the link state sequence, extracts the robust long-range dependencies of the long sequence input through the multi-head ProbSparse self-attention layer and the self-attention distillation layer, and uses the decoder to perform generative prediction through the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, thereby directly outputting a multi-step prediction matrix, avoiding the cumulative error problem that may exist in the propagation process, and improving the speed of fault prediction.

[0079] 3, The large-scale low-orbit satellite network congestion awareness and fault prediction method disclosed by the application is based on a difference probability distribution fault detection interval calculation method, and the difference probability distribution fault detection interval is calculated according to the difference between the predicted circuit breaking starting value and the actual circuit breaking starting value, so that the detection message overhead is reduced while the circuit breaking detection accuracy is ensured.

[0080] 4, The large-scale low-orbit satellite network congestion awareness and fault prediction method disclosed by the application evaluates the node congestion and link fault in two cases respectively according to the parameters set in step three after each prediction of the circuit breaking. The link state data is added to the original link state sequence after data preprocessing, and the data is input into the Informer model for training and prediction of the next circuit breaking. Through the dynamic updating of the fault detection parameters and the cycle from prediction to detection to prediction, the accuracy of the satellite node congestion awareness and fault prediction is continuously improved while the detection overhead is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 The large-scale low-orbit satellite network congestion awareness and fault prediction method of the application.

[0082] Figure 2 Structure diagram of the Informer prediction model.

[0083] Figure 3 Distribution diagram of the fault detection interval based on the difference probability distribution.

[0084] Figure 4 Flowchart of the prediction result evaluation. DETAILED DESCRIPTION

[0085] In order to better illustrate the purpose and advantages of the application, the content of the application is further described below in combination with the drawings and examples.

[0086] Example 1

[0087] As Figure 1 shown, the large-scale low-orbit satellite network congestion awareness and fault prediction method disclosed by the application specifically realizes the following steps:

[0088] Step one, collect the link information between two satellites through the BFD protocol; regularize the link information to obtain a standard link state sequence; perform space-time fusion coding on the standard link state sequence according to the circuit breaking starting position and the duration to obtain a link state matrix; compare the link state matrices under different combinations, and select the optimal matrix as the input of the prediction model.

[0089] Step 1.1: After two adjacent satellite nodes establish a BFD session, the BFD packets are periodically sent on the channel between the two satellites. When one party does not receive the BFD packet within the negotiated detection time, it is determined that a circuit is occurring on this channel. The BFD protocol detection interval is set to 200 milliseconds, the channel is recorded as 1, the circuit is recorded as 0, and the original link state sequence is recorded as X = {x1,...,x f ,...,x N}, that is, the link state data is obtained by using the Bidirectional Forwarding Detection BFD protocol. The obtained link state data is divided into a training set and a test set.

[0090] Step 1.2: In view of the difference in time scale between node congestion and link failure, the circuit with a duration of 1s-10s and the circuit with a duration of 1min or more in the original link state sequence are reserved according to formula (1):

[0091]

[0092] The standard link state sequence is obtained, denoted as X' = {x1',...,x f ′,...,x N ′}

[0093] Step 1.3: According to the redundancy and continuity of the link state sequence, the standard link state sequence is compressed by spatiotemporal fusion coding, and the sparseness of the link state sequence is reduced.

[0094] The redundancy refers to the fact that the link state sequence is composed of only numerical circuits 0 and channels 1, and since the link circuit state will appear as an unusual exception, the link state sequence is mostly a channel 1, which has a high degree of redundancy.

[0095] The continuity refers to the fact that the link channel and circuit state will last for a period of time, so the circuit 0 and channel 1 of the sequence have continuity and do not appear in an interleaved manner.

[0096] According to formula (2), the starting value Z1 of the standard link state sequence statistics is set to {z1,...z i …,z u} and the time interval Z2 is set to {z'1,...z' i …,z' u}:

[0097]

[0098] Wherein, I min is the minimum interval of the starting value of the adjacent continuous circuit in the standard link state sequence, I maxLet be the maximum interval between the starting values ​​of adjacent consecutive disconnections in the standard link-state sequence, O be the duration of the most frequent congestion, and L be the length of the standard link-state sequence. At this point, for Z1 and Z2, u... 2 Combinations (z) i ,z′ i Based on Z2, the standard link state sequence is divided into s = [L / Z2] subsequences, and the number c of persistent disconnections in the j-th subsequence is recorded. j Starting position of circuit breaker and duration of circuit breaker j∈[1,[L / Z2]], the starting position of the circuit breaker As the first value of the column-number pair, the duration of the circuit break is... As the second value, the column-directed pair form is: By vertically arranging the column pairs contained in each subsequence, the standard link-state sequence is increased in dimension to obtain the link-state matrix, which takes the following form:

[0099]

[0100] For example, in the standard link state sequence, there exists a sequence {1111000011111110000011111111}, which is then spatiotemporally fused and encoded as follows:

[0101] If the initial value is 1 and the time interval is 12, then the link state matrix is:

[0102] If the initial value is the second 1 and the time interval is 12, then the link state matrix is:

[0103] If the initial value is 1 and the time interval is 24, then the link state matrix is:

[0104] The link state matrix has a dimension of 2max{c j Given that the number of consecutive faults within each subsequence is unequal, for c j <c max The time interval column, missing parts are used To complete it in the form of, but The lack of fault information leads to wasted resources. The goal is to minimize the number of faults in the sequence while maximizing the time interval Z2. Useless information is used to calculate the link-state matrix after spatiotemporal fusion coding for each combination, and the total number of column pairs is recorded. Number of column pairs. First, select the combination with the smallest total number of column pairs, then select from these combinations... The combination of the column direction number and the minimum, under which the output link state matrix is input into the Informer prediction model.

[0105] Step two, the Informer model structure is as shown in Figure 2 The link state matrix is input into the Informer model, and the Informer model performs embedding operation on the link state matrix through dimensionality increasing, position encoding and time feature encoding to obtain an embedded vector; the embedded vector is input into an encoder, a decoder and a full connection layer to obtain a predicted link state matrix. The encoder extracts robust long-range dependencies of long sequence input through a multi-head ProbSparse self-attention layer and a self-attention distillation layer; the decoder generates a predicted matrix through a masked multi-head ProbSparse self-attention layer and a multi-head attention layer, directly outputs a multi-step predicted matrix, avoids cumulative error problems caused by a propagation process, and improves the prediction speed.

[0106] Step 2.1: The dimensionality increasing of the Informer model is to convert the link state matrix 2max{c j}dimensional vector into a 512-dimensional vector through one-dimensional convolution, and use the position encoding method in the Transformer model to retain the time sequence information in the time sequence data by using fixed displacement embedding:

[0107]

[0108]

[0109] wherein, d model is the feature dimension of the input representation, In order to enable the similarity calculation of self-attention to access global context, a 1-D convolution filter is used to project the scalar context into a d model -dimensional vector After superimposing the dimensionality increasing and position encoding operations, the final input vector is obtained:

[0110]

[0111] The specific changes of the encoder input embedding are as follows: the dimension of the link state matrix is converted into 32x96x2max{c j}, 32 represents that one batch contains 32 samples, 96 represents that one sample represents 96 time point data, and 2max{c j}is the number of data dimensions contained for each time point. Assuming the time interval of the link state matrix is 15 min and the time length is 1 week, the global timestamp dimension is 32x96x4, 4 represents the time dimension, including minutes, hours, days, and weeks. The encoder input embedding converts the 32x96x2max{c j}matrix into a 32x96x512 matrix input into the encoder.

[0112] Step 2.2: The encoder extracts long sequence input through two parts of the multi-head ProbSparse self-attention layer and the self-attention distillation layer. The multi-head ProbSparse self-attention layer distinguishes a few point products that contribute to the main attention. Due to the multi-head attribute, 512 dimensions are divided into eight parts, i.e., the matrix dimension input into the multi-head ProbSparse self-attention layer is 32x8x96x64. According to the self-attention mechanism, the attention of the athquery is defined as a kernel smoother in the form of probability:

[0113]

[0114] Based on equation (6), the attention of the athquery to all keys is defined as the probability p(k e |q a ), the dominant point product will cause the attention probability distribution of the corresponding query to deviate from the uniform distribution; according to the KL divergence formula (7), the distance between the two distributions is measured:

[0115]

[0116] Neglecting the constant term, the sparsity index of the athquery is defined as:

[0117]

[0118] According to equations (6), (7), and (8), the multi-head ProbSparse self-attention mechanism allows each key to focus on only a part of the dominant query:

[0119]

[0120] Based on the multi-head ProbSparse self-attention mechanism, different sparse query pairs are generated for each head, thereby avoiding severe information loss; according to the distillation of the output of the multi-head ProbSparse self-attention layer through one-dimensional convolution, the long-range dependency coupling between the output and the input is efficiently captured, and the specific form is as follows:

[0121]

[0122] wherein, Attention block, which contains multi-head probability sparse self-attention matrix and basic operations, ELU as activation function, MaxPool as maximum pooling.

[0123] Step 2.3: The decoder performs generative prediction, directly outputting a multi-step prediction matrix: the input form of the decoder is a sequence of length N token +N output , that is, the splicing sequence of the starting token of the target sequence and the placeholder of the target sequence and the timestamp, and the placeholder is set to 0:

[0124]

[0125] After sequentially passing through the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, an output matrix of (N token +N output )xd is obtained, where d is the dimension of the embedding. The matrix is then passed through a fully connected layer to obtain an output matrix, which is the predicted link state matrix.

[0126] The error function of the test set uses the indicators of normalized root mean square error nRMSE and relative root mean square error rRMSE to measure the prediction performance:

[0127]

[0128]

[0129] Where y i is the true value of the test set part in the Informer model, the predicted value of the test set part.

[0130] Step three, initialize the fault detection parameters, judge the fault type of the circuit break according to the predicted link state matrix, and the fault type is divided into node congestion and link fault; calculate the fault detection interval based on the difference probability distribution according to the difference between the predicted circuit break starting value and the actual circuit break starting value in the test set, which reduces the detection message overhead while ensuring the accuracy of circuit break detection; calculate the reaction time of judging the fault type after detecting the circuit break according to the circuit break duration in the test set.

[0131] Step 3.1: Initialize the fault detection parameters, judge the fault type of the circuit break according to the predicted link state matrix, and the fault type is divided into node congestion and link fault; due to the obvious difference in nature between congestion and link fault, the prediction result in the order of seconds is defined as congestion, and the prediction result in the order of minutes is defined as link fault.

[0132] Step 3.2: According to the data in the test set, the fault detection interval based on the difference probability distribution is calculated. After predicting the approximate trend of the time series based on the Informer model, in order to accurately locate the actual outage starting point, it is necessary to continuously detect the link state around the predicted starting point. The fault detection interval based on the difference probability distribution reduces the detection message overhead while ensuring the accuracy of outage detection. The outage starting point predicted by the data in the test set is denoted as The actual outage starting point is denoted as Then T' c and T c The interval sequence of T c -T' c .

[0133] According to the probability distribution of ΔT1, the fault detection interval is calculated, and the specific steps are as follows:

[0134] The DBSCAN algorithm based on density is used to remove outliers in ΔT1. The DBSCAN algorithm uses two parameters, minPts and Eps, to cluster. Select a point in the data set to cluster. If there are at least minPts points within the radius of Eps, mark the point as a core point and expand a cluster centered on the point. If there are less than minPts points within the radius of Eps, but there are core points, mark the point as a boundary point and add it to the cluster of core points. If the point is neither a core point nor a boundary point, it is marked as an outlier. Iterate minPts and Eps to find outliers and remove them; until all points are visited to complete clustering.

[0135] The selection steps of minPts and Eps are as follows: calculate the distance between each point in the interval sequence and the second nearest point, and sort the calculation results from large to small, draw a k-distance graph, and find the distance of the inflection point position, which is the value of the minimum Eps. The selection of minPts is at least 3. Then, by changing the Eps value to the distance after the inflection point in the k-distance graph and increasing minPts by interval 1, the model parameters are iterated until the outliers account for 1% of the interval sequence.

[0136] After removing outliers, arrange all intervals from small to large and remove duplicate intervals to obtain the interval sequence {Δt1,...,Δt g ,...,Δt n}, and calculate the probability of each interval in the interval sequence ΔT1, to obtain the probability distribution P: {P1,...,P n}; when setting w detection points {t1,...,t h ,...,t w}, the predicted node outage starting point is t0, the first detection point is t1 = t0 + At1, and the last detection point is t w = t0 + At n The interval between adjacent detection points at different time is determined according to the following formula:

[0137] If The hth detection point is

[0138] Assuming w = 21, the interval distribution of 21 detection points is shown in Figure 3 , which presents the characteristics that the higher the probability density, the denser the fault detection points, so as to reduce the packet overhead while ensuring detection accuracy.

[0139] Step 3.3: Initialize three reaction times for judging the fault type after detecting the outage according to the outage duration in the standard link state sequence. After detecting the predicted node congestion outage, immediately issue a load balancing strategy. The load balancing strategy is to select half of all flow tables with high traffic demand through the congested node for the controller, and after deleting the congested node in the network topology, recalculate the path for the corresponding flow table source node using the shortest path algorithm and issue it. If the outage is caused by node congestion, the node congestion will quickly recover under the influence of the load balancing strategy, otherwise the load balancing strategy will not promote the link recovery. After detecting the predicted link fault outage, immediately issue a new routing strategy. The new routing strategy is to update the original routing table according to the shortest path algorithm based on the network topology after removing the link. However, if the actual outage is caused by node congestion, the link will quickly recover. To further evaluate the accuracy of the predicted fault type, record the congestion duration and link fault duration in the standard link state sequence, and use DBSCAN algorithm to remove outliers, respectively. The initial reaction time At of determining whether the outage is node congestion after issuing the load balancing strategy and the reaction time At' of determining whether the outage is node congestion after issuing the new routing strategy are initialized to the minimum value of the congestion duration after removing outliers, and the reaction time At" of determining whether the outage is link fault after issuing the new routing strategy is initialized to the maximum duration of the link fault after removing outliers.

[0140] Step four: After predicting the outage each time, evaluate the node congestion and link fault for two cases respectively according to the parameters set in step three. The flow of evaluating the accuracy of the Informer model prediction is shown in Figure 4 , and the parameters in step three are updated: based on the outage starting point, update the detection points {t1,...,t w};Based on a satellite memory threshold record node congestion duration update ΔT, based on the duration of the circuit breaker update ΔT' and ΔT''. The link state data is added to the original link state sequence after data preprocessing input Informer model training and prediction of the next circuit breaker. Through the dynamic update of the fault detection parameters and the prediction of the detection to the prediction cycle, while reducing the detection overhead, the accuracy of satellite node congestion awareness and fault prediction is continuously improved.

[0141] Step 4.1: After data preprocessing of the collected original link state sequence, the Informer model is trained and the link state matrix is predicted. After predicting the circuit breaker, the fault detection point {t1,...,t w} is sent in turn, if no circuit breaker is detected, the original link state sequence end to t w The duration of the path at the moment is added to the original link state sequence to retrain and predict. If a circuit breaker is detected, the fault type of the circuit breaker is determined according to the predicted duration of the circuit breaker. If it is determined that the circuit breaker is node congestion, step 4.2 is performed, and if it is determined that the circuit breaker is a link fault, step 4.2 is performed. To adapt to the changes of the link characteristics, the difference between the predicted circuit breaker starting point and the detected circuit breaker starting point is added to the ΔT1 sequence in step 3.2 to recalculate the fault detection interval and update the detection point {t1,...,t i ,...,t w} after the next prediction of the circuit breaker.

[0142] Step 4.2: If it is determined that the circuit breaker is node congestion, the load balancing strategy is issued. If the reason for the circuit breaker is node congestion, under the influence of the load balancing strategy, the node receives a large amount of traffic, the buffer queue is continuously reduced, and the node congestion is quickly alleviated.

[0143] The node monitors the queue occupancy rate at all times. When the node memory is reduced to a suitable threshold, it indicates that the node is in a state that is not prone to congestion at this time, and the time is recorded and sent to the controller. In order to obtain the expected effect of the above mechanism, the setting of the threshold is very important, which is denoted as T1. In order to calculate this value, a value representing the busy state of the node buffer is added, denoted as T2.

[0144] T1 and T2 should be able to cope with the following extreme cases: when the satellite checks the buffer queue, the queue occupancy rate is close to but still below T1, after checking, the queue occupancy rate just exceeds T1, but the satellite needs to wait for the check interval Δt Q to detect the change of the queue occupancy rate. In order to ensure that the queue occupancy rate does not exceed T2 during Δt Q , T1 and T2 should satisfy:

[0145] (T2-T1)·L·S≥(I-O)max • Δt Q (14)

[0146] where L denotes the size of the buffer queue, S denotes the average packet size, (I-O) max denotes the maximum difference between the input and output traffic of the node during this period.

[0147] If the queue occupancy just exceeds T2 after one check, the satellite also needs to wait Δt Q to realize the change of queue occupancy. In order to avoid discarding packets due to buffer overflow, T2 should satisfy:

[0148] (1-T2)·L·S≥(I'-O') max • Δt Q (15)

[0149] According to equation (15), we get:

[0150]

[0151]

[0152] Considering the extreme case, T1 and T2 are set as:

[0153]

[0154]

[0155] If the queue occupancy drops below T1 within ΔT time, the congestion node monitors that the prediction of the outage duration is correct, and the node congestion has recovered. The original congestion node needs to feed back the time point when the queue occupancy drops to T1 to the controller to update the fault detection parameter ΔT. The controller subtracts the time point when the queue occupancy drops to T1 from the detected outage starting point to obtain the congestion duration, and ΔT takes the maximum value of the duration of the current node congestion and the duration of the node congestion after the last load balancing strategy is issued. The congestion duration is added as incremental data to the original link state sequence, preprocessed, and then trained and predicted again in the Informer model.

[0156] If the controller does not receive a node passage notification within ΔT time, it indicates that the prediction of the outage duration is wrong. The error and the reason for the outage are link failure, and a new routing strategy is issued. The adjacent nodes continuously monitor the link state, record the fault duration, and return to the Informer model for retraining and prediction.

[0157] Step 4.3: If the circuit break is judged as link failure, a new routing strategy is issued. If the controller receives a node passage notification within ΔT', it indicates that the prediction of the circuit break duration is wrong, the cause of the circuit break is node congestion, the new routing table removes the circuit break link, and the traffic entering the node through the link is reduced to 0, and the node congestion is quickly alleviated. ΔT' takes the maximum of the duration of the current node congestion and the duration of the node congestion after the last time the new routing strategy is issued. The congestion duration is added to the original link state sequence, preprocessed, and input into the Informer model for retraining and prediction.

[0158] If the controller receives a node passage notification within the time range of ΔT", it indicates that the prediction result is correct, and the node has a link failure. Update ΔT" based on the duration of the link failure, and ΔT" takes the maximum of the duration of the current link failure and the duration of the last link failure. Add the duration of the link failure to the original link state sequence, input into the Informer model for retraining and prediction.

[0159] If the controller does not receive a node passage notification within the time range of ΔT", it is determined that the node has a permanent and unrecoverable hardware failure. At this time, the control center collects the alarm information of the failed node and delivers it to manual processing.

[0160] Step four: Through the dynamic update of the fault detection parameters and the cycle from detection to prediction, the accuracy of satellite node congestion awareness and fault prediction is continuously improved while reducing the detection overhead.

[0161] The above specific description further details the purpose, technical solutions, and benefits of the invention. It should be understood that the above description is only a specific embodiment of the invention and does not limit the protection scope of the invention. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the invention should be included in the protection scope of the invention.

Claims

1. A method for congestion awareness and failure prediction in large scale low earth orbit satellite networks, characterized by: Comprising the following steps, Step one, collecting the link information between two satellites through the BFD protocol; data regularization is performed on the link information to obtain a standard link state sequence; the standard link state sequence is spatio-temporally fused and coded according to the starting position and duration of the outage to obtain a link state matrix, and the link state matrices under different combinations are compared to select the optimal matrix as the input of the prediction model; Step two, input the link state matrix into the Informer model, and the Informer model embeds the link state matrix through dimensionality lifting, position coding and time feature coding to obtain an embedded vector; the embedded vector is input into the coding layer, decoding layer and full connection layer to obtain the predicted link state matrix; the encoder extracts the robust long-range dependencies of the long sequence input through the multi-head ProbSparse self-attention layer and the self-attention distillation layer; The decoder generates a prediction matrix by the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, and directly outputs the multi-step prediction matrix to avoid the cumulative error problem caused by the propagation process; Step three, initialize the fault detection parameters, judge the fault type according to the predicted link state matrix, and the fault type is divided into node congestion and link fault; calculate the fault detection interval based on the difference probability distribution according to the difference between the predicted outage starting value and the actual outage starting value in the test set, which can reduce the detection message overhead while ensuring the accuracy of the outage detection; calculate the reaction time of the fault type after detecting the outage according to the outage duration in the test set; Step four: After each predicted outage, evaluate the node congestion and link failure respectively according to the parameters set in step three; update the parameters in step three: update the detection points {t1,...,t w} based on the outage starting point, update ΔT based on the node congestion duration threshold, update ΔT' and ΔT'' based on the outage duration, add the link status data to the original link status sequence after data preprocessing to input the Informer model for training and predicting the next outage; through the dynamic update of the fault detection parameters and the cycle from prediction to detection to prediction, the accuracy of satellite node congestion awareness and fault prediction is continuously improved while reducing the detection overhead.

2. The method of claim 1, wherein: The implementation method of step one is, Step 1.1: After two adjacent satellite nodes establish a BFD session, BFD packets are periodically sent on the channel between the two satellites. If one party does not receive the BFD packet within the negotiated detection time, it is determined that a circuit is occurring on the channel; the BFD protocol detection interval is set to B milliseconds, the channel is recorded as 1, the circuit is recorded as 0, and the original link state sequence is recorded as X={x1,...,x f ,...,x N} that is, the link state data is obtained by using the Bidirectional Forwarding Detection BFD protocol; the obtained link state data is divided into a training set and a test set; Step 1.2: Due to the difference in time scale between node congestion and link fault, according to formula (1), the original link state sequence is reserved for the outage with a duration of 1s-10s and the outage with a duration of 1min or more: The standard link state sequence is obtained, denoted as X' = {x1',...,x f ',...,x N ',...,x Step 1.3: According to the redundancy and continuity of the link state sequence, the standard link state sequence is compressed by spatio-temporal fusion coding to reduce the sparsity of the link state sequence; The redundancy refers to the fact that the link state sequence is composed of only numerical outage 0 and pass 1, and since the link outage state will appear as an unusual exception, the link state sequence is mostly pass 1, which has a high degree of redundancy; The continuity refers to the fact that the pass and outage states of the link will last for a period of time, so the outage 0 and pass 1 of the sequence have continuity and do not appear alternately; According to equation (2), set the starting value Z1 for standard link state sequence statistics: {z1,…z1,…z2,…z3,…z4,…z5,…z6,…z7,…z8,…z9,…z1 ... i …,z u } and time interval Z2:{z′1,…z′ i …,z′ u }: where I min is the minimum interval of the starting value of adjacent continuous outages in the standard link state sequence, I max is the maximum interval of the starting value of adjacent continuous outages in the standard link state sequence, O is the most frequent congestion duration, and L s is the length of the standard link state sequence; at this time, for Z1, Z2, there are u 2 combinations (z i , z′ i ), according to Z2, the standard link state sequence is divided into s = [L s / Z2] sub-sequences, the number of continuous outages c j in the jth sub-sequence is recorded, the starting position X j of the outage is recorded, and the duration of the outage is recorded. j ∈ [1, [L s / Z2]], the starting position X of the outage is taken as the first value of the column vector pair, and the duration of the outage is taken as the second value, and the column vector pair is in the form of By vertically arranging the column vector pairs contained in each sub-sequence, the standard link state sequence is upgraded to obtain a link state matrix, which is in the following form: The link state matrix has a dimension of 2max{c j Given that the number of consecutive faults within each subsequence is unequal, for c j <c max The time interval column, missing parts are used The sequence is padded in the form of [missing information] to ensure that the sequence contains as few elements as possible while maximizing the time interval Z2. Useless information is used to calculate the link-state matrix after spatiotemporal fusion coding for each combination, and the total number of column pairs is recorded. Count the number of column pairs; first select the combination with the smallest total number of column pairs, then select from these combinations... The combination with the smallest number of column pairs is used as the input to the Informer prediction model, and the output link state matrix under this combination is used as the input to the Informer prediction model.

3. The method of claim 2, wherein: The implementation method of step two is, Step 2.1: The dimensionality of the Informer model is increased from 2max{c j} dimensional vectors to 512 dimensional vectors by one-dimensional convolution, and the position encoding method in the Transformer model is used to retain the time order information in the time series data by using fixed displacement embedding: where d model is the dimension of the input representation, To enable the similarity computation of self-attention to access global context, a 1-D convolution filter is used to project the scalar context into a d model dimensional vector which is stacked with the dimensionality-raising, position encoding operation to get the final input vector: Step 2.2: The encoder extracts the long sequence input through the multi-head ProbSparse self-attention layer and the self-attention distillation layer, and distinguishes the few point products that contribute to the main attention in the multi-head ProbSparse self-attention layer; according to the self-attention mechanism, the attention of the a th query is defined as a kernel smoother in the form of probability: Based on equation (6), the attention degree of the ath query to all keys is defined as the probability p(k e | q a ), the dominant dot product pair will make the attention probability distribution of the corresponding query far away from the uniform distribution; the distance between the two distributions is measured according to the KL divergence formula (7): Neglecting the constant term, the sparsity index of the a th query is defined as: According to formulas (6), (7) and (8), the multi-head ProbSparse self-attention mechanism allows each key to focus on only part of the dominant queries: Different sparse queries are generated for each head based on the multi-head ProbSparse self-attention mechanism, thereby avoiding serious information loss; according to distillation of the output of the multi-head ProbSparse self-attention layer through one-dimensional convolution, long-range dependency coupling between the output and the input is efficiently captured, and the specific form is as follows: wherein, denotes the attention block, which contains multi-head probability sparse self-attention matrix and basic operations, ELU is the activation function, and MaxPool is the max-pooling. Step 2.3: Decoder performs generative prediction, directly outputs multi-step prediction matrix: Decoder input is in the form of a sequence of length N, i.e., a concatenation of the target sequence start token and the target sequence placeholder set to 0, and a timestamp: token +N output Step 2.3: Decoder performs generative prediction, directly outputs multi-step prediction matrix: Decoder input is in the form of a sequence of length N, i.e., a concatenation of the target sequence start token and the target sequence placeholder set to 0, and a timestamp: After sequentially passing through the masked multi-head ProbSparse self-attention layer and the multi-head attention layer, an output matrix of (N token +N output )xd is obtained, where d is the dimension of embedding; the matrix passes through a fully connected layer to obtain an output matrix, which is a predicted link state matrix; The error function of the test set adopts the indicators of normalized root mean square error nRMSE and relative root mean square error rRMSE to measure the prediction performance: where y i is the true value of the test set portion in the Informer model, the predicted value of the test set portion.

4. The method of claim 3, wherein: The implementation method of step three is, Step 3.1: Initialize the fault detection parameters, and judge the fault type of the circuit break according to the predicted link state matrix. The fault type is divided into node congestion and link fault. Due to the obvious difference in the nature of congestion and link fault, the prediction result in the order of seconds is defined as congestion, and the prediction result in the order of minutes is defined as link fault; Step 3.2: According to the data in the test set, calculate the fault detection interval based on the difference probability distribution; after predicting the approximate trend of the time series based on the Informer model, in order to accurately locate the actual outage starting point, continuously detect the link state around the predicted starting point, and use the fault detection interval based on the difference probability distribution to reduce the detection message overhead while ensuring the accuracy of the outage detection; the outage starting point predicted by the data in the test set is denoted as The actual outage starting point is denoted as Then T' c and T c The interval sequence ΔT1=T c -T' c ; According to the probability distribution of ΔT1, the fault detection interval is calculated, and the specific steps are as follows: The DBSCAN algorithm based on density is used to eliminate outliers in ΔT1. The DBSCAN algorithm uses two parameters of minimum point number minPts and neighborhood radius Eps for clustering. An arbitrary point in the data set is selected for clustering. If there are at least minPts points within the radius of Eps, the point is recorded as a core point and a cluster is expanded with the point as the center. If there are less than minPts points within the radius of Eps, but there are core points, the point is recorded as a boundary point and added to the cluster of core points. If the point is neither a core point nor a boundary point, it is recorded as an abnormal point. minPts and Eps are iterated to find and eliminate outliers. Clustering is completed until all points are visited; After removing outliers, all intervals are arranged from small to large and duplicate intervals are removed to obtain interval sequence {Δt1,...,Δt g ,...,Δt n}, the probability of each interval appearing in the interval sequence in ΔT1 is counted to obtain the probability distribution P:{P1,...,P n}; when setting w detection points {t1,...,t h ,...,t w}, the predicted node breakdown starting point is t0, the first detection point t1=t0+Δt1, and the last detection point is recorded as t w =t0+Δt n , the intervals between adjacent detection points with different probability distributions are determined according to the following formula: If then the hth detection point Step 3.3: Three reaction times of judging the fault type after detecting the circuit according to the circuit duration in the standard link state sequence; immediately issue a load balancing strategy after detecting the circuit predicted to be node congestion, the load balancing strategy is to select half of all flow tables with heavy traffic demand through the congested node for the controller, and after deleting the congested node in the network topology, the source node of the corresponding flow table is recalculated by the shortest path algorithm and issued; if the circuit reason is node congestion, the node congestion recovers quickly under the influence of the load balancing strategy, otherwise the load balancing strategy will not promote the link recovery; immediately issue a new routing strategy after detecting the circuit predicted to be link failure, the new routing strategy is to update the original routing table according to the shortest path algorithm based on the network topology after removing the link by the controller; but if the actual circuit reason is node congestion, the link recovers quickly; in order to further evaluate the accuracy of the predicted fault type, record the congestion duration and link failure duration in the standard link state sequence, respectively, remove outliers by using DBSCAN algorithm, initialize the initial reaction time ΔT of determining whether the circuit is node congestion after issuing the load balancing strategy and the reaction time ΔT' of determining whether the circuit is node congestion after issuing the new routing strategy to the minimum value of the congestion duration after removing outliers, and initialize the reaction time ΔT'' of judging whether the circuit is link failure after issuing the new routing strategy to the maximum duration of link failure after removing outliers.

5. The method of claim 4, wherein: Step four implementation method is, Step 4.1: After data preprocessing of the collected original link state sequence, the Informer model is trained and predicted to obtain the link state matrix. After the circuit is predicted to be broken, the fault detection point {t1,...,t w} set in step three is sequentially sent. If no circuit break is detected, the path duration from the end of the original link state sequence to the t w moment is recorded and added to the original link state sequence for retraining and prediction. If the circuit is detected, according to the predicted circuit duration, the fault type of the circuit is judged, if the circuit is judged as node congestion, step 4.2 is performed, if the circuit is judged as link fault, step 4.3 is performed; in order to adapt to the change of link characteristics, the difference between the predicted circuit starting point and the detected circuit starting point is added to the sequence of ΔT1 in step 3.2 to recalculate the fault detection interval, and the detected point after the next prediction of the circuit {t1,...,t w} is updated. Step 4.2: If the circuit is judged to be node congestion, issue a load balancing strategy; if the circuit reason is node congestion, under the influence of the load balancing strategy, the node receives a large amount of traffic, the cache queue is continuously reduced, and the node congestion is quickly alleviated; The node monitors the queue occupancy rate at all times, and after the node memory is reduced to a suitable threshold value, it is indicated that the node enters a state not prone to congestion at this time, and the time is recorded and sent to the controller; in order to obtain the expected effect of the above mechanism, the setting of the threshold value is crucial, denoted as T1, in order to calculate this value, a value representing the node cache entering a busy state is added, denoted as T2; T1 and T2 should be able to cope with the following extreme situation: when the satellite checks the cache queue, the queue occupancy is close to but still below T1, after the check, the queue occupancy just exceeds T1, but the satellite needs to wait for the check interval Δt Q to detect the change in queue occupancy; to ensure that the queue occupancy does not exceed T2 during Δt Q , T1 and T2 should satisfy: (T2 - T1) • L • S > (I - O) max • At Q (14) where L represents the size of the buffer queue, S represents the average packet size, (I-O) max represents the maximum difference between the input and output traffic of the node during this period; If the queue occupancy happens to exceed T2 after one check, the satellite also needs to wait Δt Q to realize the change in queue occupancy; to avoid discarding packets due to buffer overflow, T2 should satisfy: (1 - T2) • L • S ≥ (I' - O') max • Δt Q (15) According to formula (15), we get: T1 and T2 are set as: If the congestion node monitors that the queue occupancy rate drops below T1 within ΔT time, it indicates that the prediction of the circuit duration is correct, and the node congestion has recovered; the original congested node needs to feed back the time point of the queue occupancy rate dropping to T1 to the controller to update the fault detection parameter ΔT, the controller obtains the congestion duration by subtracting the detected circuit starting point from the time point of the queue occupancy rate dropping to T1, and ΔT takes the maximum value of the duration of the current node congestion and the duration of the node congestion after issuing the load balancing strategy last time; the congestion duration is added to the original link state sequence as incremental data, and is trained again in the Informer model after preprocessing and prediction. If the controller does not receive the node passage notification within the ΔT time, indicating a prediction error for the duration of the outage, the outage is caused by a link failure, and a new routing strategy is issued; The adjacent nodes continuously monitor the link state, record the failure duration, and return to the Informer model for retraining and prediction; Step 4.3: If the outage is determined to be a link failure, issue a new routing strategy; If the controller receives a node passage notification within ΔT', indicating a prediction error for the duration of the outage, the outage is caused by node congestion, the new routing table removes the outage link, the traffic entering the node through the link is reduced to 0, and the node congestion is quickly alleviated; ΔT' takes the maximum of the duration of the current node congestion and the node congestion after the last new routing strategy is issued, and the congestion duration is added to the original link state sequence after preprocessing and input into the Informer model for retraining and prediction; If the controller receives a node passage notification within the ΔT" time range, it indicates that the prediction result is correct, and the node has a link failure; Update ΔT" based on the duration of the link failure, ΔT" takes the maximum of the current link failure and the duration of the last link failure; Add the link failure duration to the original link state sequence and input it into the Informer model for retraining and prediction; If the controller does not receive a node passage notification within the ΔT" time range, it is determined that the node has a permanent and unrecoverable hardware failure, the control center collects the alarm information of the failed node, and delivers it to manual processing.

Images