CN115130617B - Detection method for continuous increase of self-adaptive satellite data mode - Google Patents
Detection method for continuous increase of self-adaptive satellite data mode Download PDFInfo
- Publication number
- CN115130617B CN115130617B CN202210892285.9A CN202210892285A CN115130617B CN 115130617 B CN115130617 B CN 115130617B CN 202210892285 A CN202210892285 A CN 202210892285A CN 115130617 B CN115130617 B CN 115130617B
- Authority
- CN
- China
- Prior art keywords
- network
- new
- gcnn
- basic
- similarity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title description 22
- 238000013528 artificial neural network Methods 0.000 claims abstract description 36
- 238000000034 method Methods 0.000 claims abstract description 34
- 230000004044 response Effects 0.000 claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 12
- 230000008569 process Effects 0.000 claims abstract description 5
- 239000013256 coordination polymer Substances 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 17
- 238000010586 diagram Methods 0.000 claims description 13
- 238000011176 pooling Methods 0.000 claims description 10
- 230000003044 adaptive effect Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 210000002569 neuron Anatomy 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000003780 insertion Methods 0.000 claims description 2
- 230000037431 insertion Effects 0.000 claims description 2
- 238000013527 convolutional neural network Methods 0.000 description 7
- 238000003745 diagnosis Methods 0.000 description 6
- 230000036541 health Effects 0.000 description 6
- 238000003909 pattern recognition Methods 0.000 description 5
- 101100133212 Drosophila melanogaster NetB gene Proteins 0.000 description 4
- 241001181114 Neta Species 0.000 description 4
- 230000003068 static effect Effects 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000009825 accumulation Methods 0.000 description 2
- 238000013145 classification model Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000010485 coping Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 238000012847 principal component analysis method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Radio Relay Systems (AREA)
Abstract
A self-adaptive satellite data mode is increased continuously, a basic two-classification network A is constructed as a current network, and the current network can finish identification and judgment of a satellite data mode; constructing a basic two-class network B, and training by using a satellite time sequence data set containing a new data mode; taking the front N layer with the similarity response of the current network and the basic two-class network B larger than the similarity threshold value as a sharing layer, taking the rest layer of the basic two-class network B as a new branch, connecting the rest layer of the basic two-class network B to the sharing layer, completing the dynamic growth of the network, and taking the rest layer as a new current network; repeating the process to obtain a dynamic growth neural network GCNN; in satellite telemetry work, satellite time sequence data are sent into GCNN, each time sequence is identified, the data mode is identified and judged, the corresponding working state is compared with the working state expected by a ground instruction, and when the two working states are not matched, the satellite is indicated to have faults.
Description
Technical Field
The invention belongs to the technical field of satellite data pattern recognition and fault detection, and particularly relates to a detection method for continuously increasing self-adaptive satellite data patterns.
Background
The satellite working state is jointly influenced by ground instructions and space environment. Different operating conditions are reflected in the satellite telemetry time series data as different data patterns. If the data pattern does not match the expected operating state of the ground command, a severe fault in the satellite is indicated. Telemetry data is a fundamental basis for reflecting the operating state of each subsystem of the satellite during its in-orbit operation. The mode identification and detection of satellite telemetry data is very critical for monitoring the working state and the running state of satellites, and is one of key technologies in the applications of in-orbit satellite real-time monitoring, in-orbit satellite anomaly detection, in-orbit satellite health management, in-orbit satellite fault diagnosis and the like.
Because of the many kinds of data transmitted by satellite telemetry, the data modes corresponding to different working states are very complex. In practice, satellite telemetry time sequence data patterns cannot be completely mastered in advance by people due to the continuous development and updating of satellite models, the continuous increase of the number and variety of satellites, and various influences of in-orbit satellites in complex space environments and complex mission operations. That is, satellite telemetry time series data patterns are not all known patterns, and new pattern types beyond the existing awareness often occur.
In addition to the known pattern classes, there are data patterns or new anomalies in the satellite time series data that are not yet clear. These data cannot naturally be categorized into existing categories, so the growing problem of new data patterns in satellite time series data cannot be effectively addressed using conventional pattern recognition methods. The method has the advantages that the method is used for continuously learning and detecting fault data, unknown mode data or new data mode data in the applications of in-orbit satellite real-time monitoring, in-orbit satellite anomaly detection, in-orbit satellite health management, in-orbit satellite fault diagnosis and the like, has the capability of continuously dynamically and adaptively adapting to the new data mode and the new data, and has very important significance for the applications due to the automatic growth and expansion of the model detection capability.
The existing method for identifying the time sequence data pattern generally obtains a feature vector after extracting the features of the data, only the existing data pattern can be learned, only the learned pattern data can be identified after model training is completed, and the pattern identification work of new data pattern data generated by satellites cannot be accommodated. For example, using statistical features, spectral features, wavelet features or principal component analysis methods to obtain feature vectors; and then classifying and identifying the time sequence data mode by utilizing various classification models. For example, the time-series feature vectors are classified and identified by using classifiers such as k nearest neighbor, random forest, adaBoost, support vector machine, BP network (namely multi-layer perceptron MLP) and the like, and the mode of the input vector is judged. Or calculating the similarity of two time sequence modes by using methods such as pearson correlation coefficient (Pearson Correlation Coefficient, PCC), dynamic time warping (Dynamic Time Warping, DTW) and the like, and then judging whether the two time sequence modes belong to a certain known mode or not according to the similarity. In recent years, a cyclic neural network (Recurrent Neural Network, RNN) such as a Long Short-Term Memory (LSTM) is used in large numbers for processing time series data, and it is possible to perform time series pattern recognition well, but it also does not have a dynamic growth capability.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a detection method for continuously increasing the data mode of a self-adaptive satellite, solves the problem that new data modes are continuously generated by accumulation detection and identification in an intelligent satellite system through an incremental learning neural network technology capable of dynamically increasing a neural network structure, and realizes the functions of data mode analysis and detection of 'learning while accumulating' in an in-orbit satellite health management system, in-orbit satellite anomaly detection and in-orbit satellite fault diagnosis system.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a detection method for the continuous increase of adaptive satellite data modes comprises the following steps:
step 1, constructing a basic two-classification network A as a current network, wherein the current network can finish identification and judgment of a satellite data mode;
step 2, constructing a basic two-classification network B, training the basic two-classification network B by using a satellite time sequence data set containing a new data mode, wherein the new data mode is a data mode which cannot be identified and judged by the current network;
step 3, calculating similarity response of the trained basic two-class network B and the current network;
step 4, setting a similarity threshold, taking the front N layers of which the similarity response of the current network and the basic two-class network B is larger than the similarity threshold as a sharing layer, taking the rest layers of the basic two-class network B as a new branch, connecting the rest layers of the basic two-class network B to the sharing layer, completing the dynamic growth of the network, and taking the rest layers as new current networks;
step 5, repeating the steps 2-4 to obtain a dynamic growth neural network GCNN, wherein the dynamic growth neural network GCNN continuously expands new identifiable data mode types while the network topology structure is continuously expanded, so as to accommodate new data modes generated by satellites to reflect new working states of the satellites;
and 6, in the satellite telemetry work, satellite time sequence data are sent into a dynamic growth neural network GCNN, each time sequence is identified, the data mode is identified and judged, the corresponding working state is compared with the working state expected by a ground instruction, and if the two working states are not matched, the satellite is indicated to have faults.
In one embodiment, the basic two-classification network is VGG16 or ResNet, and a feature extractor comprising five groups of convolution modules is formed by using 13 convolution layers and 5 pooling layers, the convolution kernel of the convolution layers is 3*3, the step size is 1, the activation function is ReLU, the network loss function adopts a binary cross entropy function, the pooling layers all adopt a maximum pooling method, and the number of classifier neurons formed by 3 fully connected layers is 4096, 2048 and 2 in sequence.
In one embodiment, in step 3, a network similarity comparison method is adopted to calculate the similarity between the basic two-class network B and each branch path of the current network, and the shared portion of the basic two-class network B and the current network, that is, the branch start point of the new branch on the current network, is determined according to the similarity response value.
In one embodiment, the similarity between the basic two-category network B and each branch of the current network is obtained by comparing cosine similarities of network feature graphs layer by layer, and then finding the best branch point in the list, where the calculation steps of the network layer similarity response list S are as follows:
step 1: obtaining an output characteristic diagram of each network layer of the basic two-class network B and the current network and storing the output characteristic diagram into a list L A 、L B In (a) and (b);
step 2: output characteristic diagram L of ith layer of basic two-class network B A [i]Output feature map L with current network ith layer B [i]Converting into N (C W H) format, i.e. two-dimensional, and normalizing;
step 3: calculate L A [i]And L B [i]Cosine similarity s between i I.e. calculating the inner product of the matrix of the two features after L2 normalization;
step 4: cosine similarity s of each layer i Adding to the list S;
step 5: returning to the step 2 until all layers of the network are executed;
step 6: and outputting a similarity response list S.
In one embodiment, in the step 4, the similarity threshold is set to 0.9-0.99, that is, if the similarity is greater than the similarity threshold, the corresponding layers are considered to be sufficiently similar to be shared, if the similarity is less than or equal to the similarity threshold, the corresponding layers are considered to be dissimilar and not to be shared, and the corresponding layers are placed in a new branch and compared with the similarity threshold by traversing the network layer similarity response list S to obtain the index value index of the shared layer number.
The invention provides a detection method for continuously increasing a self-adaptive satellite data mode, which realizes a lifetime learning technology of a satellite working state and a corresponding data mode. According to the method, a new unknown data mode in satellite telemetry data is used as a new data mode corresponding to a certain satellite working state, the new unknown data mode is sent into a neural network GCNN capable of online growth learning, and the problem of 'catastrophic forgetting' is overcome through lifetime learning, so that an in-orbit satellite anomaly detection system, an in-orbit satellite fault diagnosis system and an in-orbit satellite health management system can continuously identify the new unknown data mode (new data mode) in an intelligent satellite, and simultaneously can accurately identify all the existing data modes. Compared with the prior art, the advantages of the invention can be summarized as follows:
1. the invention realizes the dynamic growth of the convolutional neural network GCNN, changes the static structure form of the traditional convolutional neural network, and ensures that the neural network structure can be dynamically grown. Conventional pattern recognition applications generally employ static neural networks, i.e., the working form is static classification, pattern recognition. Conventional convolutional neural networks can only identify patterns that have been learned, and identifying unknown patterns has a significant error rate. The invention uses the convolution neural network GCNN which can dynamically increase to identify satellite data modes, and along with the increase of the GCNN network structure, the network identification capability is continuously increased, and the newly generated satellite data modes can be continuously received without forgetting the learned existing data modes.
2. The GCNN realizes incremental learning through branch combination of the basic two-class network, and different learned modes are saved through branch paths. The new data mode forms new branches, and each branch respectively stores different modes, so that the GCNN can learn the new satellite data mode and does not forget the existing data mode, and the problem of misjudgment of the existing mode identification method on the new and old data modes is solved.
3. The prior art requires putting together the new data pattern with the big data of the existing data pattern and then retraining and tuning the entire neural network. The neural network structure used in the prior art does not change, but the number of network weights required to be trained and learned is very large. The existing neural network model with the static structure is completely trained again, so that the time cost is high, and the operation efficiency of the whole satellite system is seriously affected. The method only needs to train a basic two-class network for the new data mode, and a great deal of calculation time and calculation resources are saved.
4. The invention can flexibly customize the functions required by the in-orbit satellite abnormality detection system, the in-orbit satellite fault diagnosis system and the in-orbit satellite health management system in application, and adds and deletes the mode types according to the requirements.
Drawings
FIG. 1 is a schematic diagram of the basic structure of a dynamically growing neural network.
Fig. 2 is a diagram of a basic two-class network architecture.
Fig. 3 is a schematic diagram of a network dynamic growth update flow.
Fig. 4 is an example of a dynamically growing neural network.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples.
The rapid updating development based on satellite model, quantity, type and the like, and the working environment and working task thereof are complex, so that the current situation that a new data mode frequently appears in a satellite telemetry time sequence data mode is caused. However, the existing method for identifying the time sequence data pattern cannot identify the new data pattern data, and especially the model does not have dynamic growth capability. The invention provides a detection method capable of dynamically and adaptively increasing satellite data modes from the perspective of new data mode knowledge accumulation. The invention realizes the new data mode of continuously accommodating satellite telemetry data based on the convolutional neural network, and has the functions of continuously increasing network structures, increasing the number of identifiable data modes and continuously increasing network detection capability. The invention continuously expands the detection capability of the new data mode by adjusting the network structure, and can continuously learn the new data mode like a person. The invention not only can accurately detect and identify the newly added satellite data mode, but also can maintain the detection capability of the existing mode, avoid the defect that the mode identification network needs to be relearned every time the new data mode is found, and greatly reduce the training time and the calculated amount. The invention realizes a lifetime learning method for the working state of the satellite, and can solve the difficult problems of difficult coping with new data, new data modes, new conditions in the application of in-orbit satellite anomaly detection, in-orbit satellite fault diagnosis, in-orbit satellite health management and the like.
The method for detecting the continuous growth of the dynamic self-adaptive satellite data mode can be realized based on a dynamic growth neural network (GCNN), the dynamic growth neural network (GCNN) realizes dynamic growth through network layer sharing and tree structure, and the method can continuously receive a new satellite data mode by increasing the neural network identification capability in a mode of dynamically changing the network structure under the condition that the known satellite data mode is not forgotten, and the basic structure is shown in the attached figure 1. GCNN needs to obtain a series of basic two-class networks based on convolutional neural network, and then automatically assemble the basic two-class networks into a new network. The initial network can be a two-class network based on convolutional neural network, and can only complete identification and judgment of a satellite data mode. New satellite data pattern recognition capability is continuously added on the initial network, namely, a basic two-class network of another satellite data pattern is fused to the current network, so that network growth is realized. This is then done, the existing networks continue to grow and the recognition capabilities correspondingly expand.
Specifically, the method for detecting the continuous increase of the adaptive satellite data mode is introduced as follows:
(1) Basic function
The detection method of the invention continuously expands new identifiable data pattern categories while the network topology structure is continuously expanded through the dynamic growth neural network (Growing Convolutional Neural Network, GCNN), thereby accommodating new data patterns generated by satellites to reflect new working states of the satellites, and the dynamic growth neural network GCNN has the following formula:
wherein:dynamic growth neural network GCNN, N obtained for training d For dynamic growth of the neural network GCNN in the training process, T and W are network topology and weight, T and W are optimal network topology and optimal network topology weight, and +.>With T and W, existing data pattern classes and upcoming new data pattern classes can be identified.
For a new phase network, ++>For an old phase network, deltaT t+1 For increased network topology ΔW t+1 For increasing network weight, T t Is->W of the network topology of (a) t Is->Network weights, T t+1 Is->W of the network topology of (a) t+1 Is->Network weights of (2); is->Increasing DeltaT t+1 Meaning that a new branch can be added to +.>Therefore, the GCNN can recognize new data pattern types one by one and expand the functions thereof.
In order to not only avoid losing and forgetting the learned satellite data pattern, but also continuously expand the new data pattern, the network structure of the dynamic growth neural network GCNN is similar to a tree, and the center of the dynamic growth neural network GCNN is supposed to be shared and branched. That is, the insertion of new branches means that the GCNN obtains new knowledge, while old knowledge is retained by the old branches. A tree is a collection of all branches, where a branch refers to a connection path from a root node to a leaf node, as shown below:
T GCNN =B 1 ∨B 2 ∨…∨B k (3)
CP GCNN =CP(B 1 )+CP(B 2 )+…+CP(B k ) (4)
wherein: t (T) GCNN Representing the topology of the GCNN; b (B) 1 ~B k Represents the first to kth branches of GCNN; CP (B) 1 )~CP(B k ) Respectively represent B 1 ~B k Knowledge of representatives, knowledge of representatives of individual branches together forming the functional CP of the GCNN GCNN . That is, the topology T of GCNN GCNN Is branch B 1 ,B 2 ,…,B k Knowledge represented by each branch together constitutes the functional CP of the GCNN GCNN . Topology structure of GCNN new stageBy adding a new branch B t+1 To extend new knowledge CP (B) t+1 ) Update knowledge of the new phase of GCNN +.>
(2) Basic two-class network
The dynamic growth neural network GCNN of the invention requires training a binary classification network, called a basic binary classification network, on the premise of each new class in the growth process. The invention adopts VGG16 or ResNet as the basic two-class network of each class. The network structure of the basic two-class network is shown in fig. 2, wherein a feature extractor comprising five groups of convolution modules is formed by using 13 convolution layers and 5 pooling layers, the convolution kernel of the convolution layers is 3*3, the step size is 1, the activation function is ReLU, and the pooling layers all adopt a maximum pooling method. The number of classifier neurons formed by 3 full connection layers is 4096, 2048 and 2 in sequence. By measuring the similarity layer between the new data pattern network and the original network and using the most similar layer as the sharable layer. The remaining layers of the new data pattern network will be connected as new branches. The current network is continuously growing to expand by integrating new data pattern networks one by one, which will be combined with the current network in the same way when new data patterns are again present.
In the basic two-class network, a hidden layer activation function adopts a ReLU function, and a network loss function adopts a binary cross entropy function.
In the invention, a basic two-class network A needs to be constructed as an initial network, and the initial network is the most original current network and can finish the identification and judgment of a satellite data mode.
Then, a basic two-classification network B, namely a "new data pattern basic two-classification network" is required to be constructed, wherein the new data pattern is a data pattern which cannot be identified and judged by the current network.
Specifically, the basic two-classification network B is trained with a satellite time series data set t= { T0, T1, … …, tn } containing new data patterns, and the network structure is shown in fig. 2. When a time sequence t i After all convolution layers and pooling layers, the three layers of full-connection layers and softmax layers are sent, the number of neurons of the full-connection layers is 4096, 2048 and 2 respectively, and finally two values are output to represent the two classification probability values.
In the present invention, the new data pattern is derived from newly added satellite telemetry data that is continuously accumulated. The newly received satellite telemetry data includes both known patterns and possibly new data patterns. The method of the invention needs to firstly classify and judge the newly received satellite telemetry data to see the known mode or the unknown mode. If the mode is unknown, an expert assigns a new mode category to the unknown mode, and then the method dynamically grows the new mode category and the new received data corresponding to the new mode category into the network. The network of the invention can dynamically increase without retraining all modes, and expands the recognition capability of the new mode.
(3) Network similarity comparison
The GCNN adopts a network layer similarity comparison mode to measure the similarity layer number of the new data mode basic two-class network and the current network when in growth, so as to determine the shareable first several network layers. And replacing a corresponding front layer in the new task by using the shared network layer, adopting a branch structure after the shared layer, and taking the rest layer of the new data mode classification network as a new branch to connect the shared layer to finish the dynamic growth of the network. New data pattern two-classification networks need to automatically merge with the current network one by one. The network after growth is the current network, and when new data mode arrives, the network is combined with the current network in the same way for growth.
After training a new data mode to obtain a basic two-class network B, calculating the similarity between the basic two-class network B and each branch path of the current network by adopting a network similarity comparison method, and determining the shared part of the basic two-class network B and the current network or the branch starting point of a new branch on the current network according to a similarity response value. The similarity between the basic two-class network B and each branch of the current network is obtained by comparing cosine similarity of the network characteristic diagrams layer by layer, a network layer similarity response list S is obtained, and then the optimal branch point is found in the list. The calculation steps of the network layer similarity response list S are as follows:
step 1: inputting net A and net B; net a is the current network and net B is the new data pattern base two-class network.
Step 2: for net A and net B, respectively obtaining output characteristic diagrams L of each network layer A [i]、L B [i]Respectively store in list L A 、L B In (a) and (b);
step 3: the i-th layer of the network performs:
step 4: will L A [i]And L B [i]Converting into N (C W H) format, i.e. two-dimensional, and normalizing;
step 5: calculate L A [i]And L B [i]Cosine similarity s between i I.e. calculating the inner product of the matrix of the two features after L2 normalization;
step 6: cosine similarity s of each layer i Adding to the list S;
step 7: returning to the step 2 until all layers of the network are executed;
step 8: and outputting a similarity response list S.
(4) Dynamic network growth
When the new data mode basic two-class network is combined into the current network, the invention needs to select a proper network similarity threshold value to determine the sharing layer number. In general, 0.9 to 0.99 may be selected as a default threshold, that is, if the similarity is greater than the similarity threshold, the corresponding layers are considered to be sufficiently similar to be shared, and if the similarity is less than or equal to the similarity threshold, the corresponding layers are considered to be dissimilar and not to be shared, and should be placed in a new branch. And comparing the traversed similarity response list S with a similarity threshold value to obtain an index value index of the shared layer number. The size of the similarity threshold can be adjusted according to the data of different new satellite modes. That is, a similarity threshold is set, the first N layers of the current network and the base two-class network B with similarity response greater than the similarity threshold are used as shared layers, the rest layers of the base two-class network B are used as a new branch to connect to the shared layers, and then the dynamic growth of the network is completed and used as a new current network.
As shown in FIG. 3, let k denote the class number, N, of the new data pattern k Basic two-classification model representing new data pattern k, G t Representing the current GCNN network. The current network includes m branches, where G tf Is one of the branches, f= { a, …, m } represents the class number of that branch. The network growth is implemented as follows.
Step 1: basic two-classification network N for inputting new class number k and mode k k Current GCNN network G t 。
Step 2: if the allpath dictionary is empty, then N will be k Put into allpath [ k ]]And k is stored in the dictionary T.
Step 3: otherwise, clearing the temporary dictionary I;
step 4: is performed for each key f in allpath.
Step 5: calculate N k And G tf Similarity between them.
Step 6: acquisition of shared layer index I from similarity list kf And will I kf Put into dictionary I, the key is f.
Step 7: and returning to the step 4 until all keys are executed.
Step 8: find the largest index I in I kθ Where θ is the key of the path to be merged.
Step 9: acquiring path G of key theta in allpath dictionary tθ And G is taken up tθ I of (2) kθ All layers before a layer are shared to N k I.e. N k Updated to G tθ N k-θ 。
Step 10: will G tθ N k-θ Put into allpath, [ k, θ, I kθ ]Stored in dictionary T.
Step 11: traversing allpath dictionary and merging networks in turn by using dictionary T to obtain latest G t+1 。
Step 12: output of the increased network G t+1 。
The network growth process described above is actually: first, N needs to be calculated k And G t And obtain the corresponding shared layer index value I kf F= { a, …, m }. Then find I kf The largest shared layer index value I kθ =max{I kf I f= { a, …, m } as the number of the layer to be finally inserted, and I kθ The corresponding branch number θ is the branch to be inserted. Next, I kθ N before layer k Network layer quilt I kθ G before layer tθ The branching layer is substituted. In this way, N k Is updated to G tθ N k-θ Then G is taken up tθ N k-θ Put into an allpath dictionary. The keys of the allpath dictionary are class number k, with a value of G tθ N k-θ . The entries in the allpath dictionary are actually paths from the root to each leaf in the current network. Next, the entire tree is constructed according to the order of path generation in the allpath dictionary. The first path is the initial network. The shared layer index of the second path is I kθ 2 . I of the second path kθ 2 The following layer is from I kθ 2 A new branch where the first path branches off. Then get the shared layer index I of the third path kθ 3 And at I kθ 3 A new branch is branchedBranch, which is I of the third path kθ 3 Later layers t Updated to G t+1 。
Taking a basic two-class network for training a set of satellite time series data as an example, the data set includes two patterns a, B. The initial task is to identify { A }, the new task is to identify { B }, i.e., by growing to enable the network to identify { A, B }. The NetA and NetB sub-tables represent the basic two-class network of the initial task a and the basic two-class network of the new data pattern B. And obtaining characteristic diagrams of the i layers corresponding to the NetA and the NetB, calculating the cosine similarity of the NetA and the NetB, and outputting the similarity response output results of the NetA and the NetB as shown in table 1.
Table 1 initial mode a binary network similarity response to new data mode B binary network
Layers | Similarity | Layers | Similarity | Layers | Similarity |
Conv1 | 1.0000 | Conv6 | 0.9804 | Conv11 | 0.8539 |
Conv2 | 0.9995 | Conv7 | 0.9659 | Conv12 | 0.7689 |
Maxpool1 | 0.9996 | Maxpool3 | 0.9707 | Conv13 | 0.5318 |
Conv3 | 0.9980 | Conv8 | 0.9578 | Maxpool5 | 0.5764 |
Conv4 | 0.9929 | Conv9 | 0.9403 | Fc4096 | 0.6340 |
Maxpool2 | 0.9936 | Conv10 | 0.8826 | Fc2048 | 0.3227 |
Conv5 | 0.9913 | Maxpool4 | 0.8971 | Fc2 | 0.9864 |
As can be seen from table 1, the similarity response values of the two basic two-class networks of the mode a and the mode B are smaller and smaller with the increase of the layer number, which indicates that the two networks are more and more dissimilar with the increase of the layer number of the network. The present invention will incorporate network layers that are sufficiently similar to each other. And selecting 0.97 as a similarity threshold value for the new data mode A, wherein the Conv1-Maxpool3 layers are similar enough to be shared, and the Conv8-Fc2 is used as a new branch to be accessed into the current initial task A network.
FIG. 4 is a diagram of a dynamic growth model of a network node, in which a class A basic two-classification network is stored in an allpath dictionary as an initial mode, a shared network layer is obtained by calculating network layer similarity with a new data mode class B basic two-classification network, shared network layer parameters, namely conv1-Maxpool3 layer parameters, of the class A basic two-classification network are copied to the class B basic two-classification network, then the class B basic two-classification network reloads the parameters, finally obtained class B paths are stored in the allpath dictionary, and the current network combines all paths in the allpath dictionary together for self-merging updating; when the new data mode Class C is available, calculating network layer similarity between the Class C base two-classification network and all paths in the allpath dictionary, copying the shared network layer parameters of the paths with the maximum similarity to the Class C base two-classification network, wherein the model with the maximum similarity to the Class C base two-classification network shown in fig. 4 is a Class B path, namely copying the shared network layer parameters of the Class B path and the Class C base two-classification network in the allpath dictionary to the Class C base two-classification network, reloading the parameters by the Class C base two-classification network, storing the finally obtained Class C path in the allpath dictionary, and merging the networks in the allpath dictionary to obtain the updated current network. Finally, the current network expands the recognition capability of the new data pattern of class C.
Repeating the steps to obtain the dynamic growth neural network GCNN.
The test data containing the new data mode and the existing mode is sent into a growing dynamic growing neural network GCNN, and each time sequence is identified to obtain an identification result, so that the new identification capability expanded after the network growth is verified
In satellite telemetry work, satellite time sequence data are sent into a dynamic growth neural network GCNN, each time sequence is identified, the data mode is identified and judged, the corresponding working state is compared with the working state expected by a ground instruction, and when the two working states are not matched, the satellite is indicated to have faults.
Taking GCNN to identify a satellite data pattern a as an initial task, the pattern B-pattern J is a new data pattern that is gradually added, and the following table shows the identification accuracy result of GCNN.
Claims (7)
1. The method for detecting the continuous increase of the self-adaptive satellite data mode is characterized by comprising the following steps of:
step 1, constructing a basic two-classification network A as a current network, wherein the current network can finish identification and judgment of a satellite data mode;
step 2, constructing a basic two-classification network B, training the basic two-classification network B by using a satellite time sequence data set containing a new data mode, wherein the new data mode is a data mode which cannot be identified and judged by the current network;
step 3, calculating similarity response of the trained basic two-class network B and the current network;
step 4, setting a similarity threshold, taking the front N layers of which the similarity response of the current network and the basic two-class network B is larger than the similarity threshold as a sharing layer, taking the rest layers of the basic two-class network B as a new branch, connecting the rest layers of the basic two-class network B to the sharing layer, completing the dynamic growth of the network, and taking the rest layers as new current networks;
step 5, repeating the steps 2-4 to obtain a dynamic growth neural network GCNN, wherein the dynamic growth neural network GCNN continuously expands new identifiable data mode types while the network topology structure is continuously expanded, so as to accommodate new data modes generated by satellites to reflect new working states of the satellites;
and 6, in the satellite telemetry work, satellite time sequence data are sent into a dynamic growth neural network GCNN, each time sequence is identified, the data mode is identified and judged, the corresponding working state is compared with the working state expected by a ground instruction, and if the two working states are not matched, the satellite is indicated to have faults.
2. The method for detecting the continuous growth of the self-adaptive satellite data mode according to claim 1, wherein the basic two-classification network is VGG16 or ResNet, 13 convolution layers and 5 pooling layers are used for forming a feature extractor comprising five groups of convolution modules, the convolution kernel of the convolution layers is 3*3, the step size is 1, the activation function is ReLU, the network loss function adopts a binary cross entropy function, the pooling layers all adopt a maximum pooling method, and the number of classifier neurons formed by 3 fully connected layers is 4096, 2048 and 2 in sequence.
3. The method according to claim 1, wherein in step 3, a network similarity comparison method is used to calculate the similarity between the basic two-category network B and each branch path of the current network, and the shared portion of the basic two-category network B and the current network, i.e. the branch start point of the new branch on the current network, is determined according to the similarity response value.
4. The method for detecting the continuous growth of the adaptive satellite data mode according to claim 1, wherein the similarity between the basic two-category network B and each branch of the current network is obtained by comparing cosine similarities of network feature graphs layer by layer, and then finding the best branch point in the list, and the calculation steps of the network layer similarity response list S are as follows:
step 1: obtaining an output characteristic diagram of each network layer of the basic two-class network B and the current network and storing the output characteristic diagram into a list L A 、L B In (a) and (b);
step 2: output characteristic diagram L of ith layer of basic two-class network B A [i]Output feature map L with current network ith layer B [i]Converting into N (C W H) format, i.e. two-dimensional, and normalizing;
step 3: calculate L A [i]And L B [i]Cosine similarity s between i I.e. calculating the inner product of the matrix of the two features after L2 normalization;
step 4: cosine similarity s of each layer i Adding to the list S;
step 5: returning to the step 2 until all layers of the network are executed;
step 6: and outputting a similarity response list S.
5. The method according to claim 4, wherein in step 4, the similarity threshold is set to 0.9-0.99, that is, if the similarity is greater than the similarity threshold, the corresponding layers are considered to be sufficiently similar to be shared, and if the similarity is less than or equal to the similarity threshold, the corresponding layers are considered to be dissimilar and cannot be shared, and the method should be placed in a new branch, and the index value index of the shared layer is obtained by traversing the network layer similarity response list S and comparing with the similarity threshold.
6. The method for detecting the continuous growth of the adaptive satellite data patterns according to claim 5, wherein the formula of the dynamically growing neural network GCNN is as follows:
wherein:dynamic growth neural network GCNN, N obtained for training d For dynamic growth of the neural network GCNN in the training process, T and W are network topology and weight, T and W are optimal network topology and optimal network topology weight, and +.>Using T and W to identify existing data pattern categories and upcoming new data pattern categories;
for a new phase network, ++>For an old phase network, deltaT t+1 For increased network topology ΔW t+1 For increasing network weight, T t Is->W of the network topology of (a) t Is->Network weights, T t+1 Is->W of the network topology of (a) t+1 Is->Network weights of (2); is->Increasing DeltaT t+1 Meaning that a new branch can be added to +.>Therefore, the GCNN can recognize new data pattern types one by one and expand the functions thereof.
7. The method for detecting the continuous growth of the adaptive satellite data patterns according to claim 6, wherein the network structure of the dynamically growing neural network GCNN is a tree, the insertion of new branches means that the GCNN obtains new knowledge, and old knowledge is reserved by old branches, and the tree is a set of all branches, and the branches refer to connection paths from root nodes to leaf nodes, and the formula is as follows:
T GCNN =B 1 ∨B 2 ∨…∨B k
CP GCNN =CP(B 1 )+CP(B 2 )+…+CP(B k )
wherein: t (T) GCNN Representing the topology of the GCNN; b (B) 1 ~B k Represents the first to kth branches of GCNN; CP (B) 1 )~CP(B k ) Respectively represent B 1 ~B k Knowledge of representatives, knowledge of representatives of individual branches together forming the functional CP of the GCNN GCNN The method comprises the steps of carrying out a first treatment on the surface of the Topology of GCNN new stageFlapping structureBy adding a new branch B t+1 To extend new knowledge CP (B) t+1 ) Update knowledge of the new phase of GCNN +.>
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210892285.9A CN115130617B (en) | 2022-07-27 | 2022-07-27 | Detection method for continuous increase of self-adaptive satellite data mode |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210892285.9A CN115130617B (en) | 2022-07-27 | 2022-07-27 | Detection method for continuous increase of self-adaptive satellite data mode |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115130617A CN115130617A (en) | 2022-09-30 |
CN115130617B true CN115130617B (en) | 2024-02-27 |
Family
ID=83385763
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210892285.9A Active CN115130617B (en) | 2022-07-27 | 2022-07-27 | Detection method for continuous increase of self-adaptive satellite data mode |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115130617B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115801483B (en) * | 2023-02-10 | 2023-05-19 | 北京京能高安屯燃气热电有限责任公司 | Information sharing processing method and system |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5381513A (en) * | 1991-06-19 | 1995-01-10 | Matsushita Electric Industrial Co., Ltd. | Time series signal analyzer including neural network having path groups corresponding to states of Markov chains |
CN111860446A (en) * | 2020-08-01 | 2020-10-30 | 西安交通大学 | Detection system and method for unknown mode of satellite remote measurement time sequence data |
CN112926723A (en) * | 2019-12-05 | 2021-06-08 | 复旦大学 | Automatic network growth method based on Split LBI algorithm |
CN114743251A (en) * | 2022-05-23 | 2022-07-12 | 西北大学 | Game character facial expression recognition method based on shared integrated convolutional neural network |
-
2022
- 2022-07-27 CN CN202210892285.9A patent/CN115130617B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5381513A (en) * | 1991-06-19 | 1995-01-10 | Matsushita Electric Industrial Co., Ltd. | Time series signal analyzer including neural network having path groups corresponding to states of Markov chains |
CN112926723A (en) * | 2019-12-05 | 2021-06-08 | 复旦大学 | Automatic network growth method based on Split LBI algorithm |
CN111860446A (en) * | 2020-08-01 | 2020-10-30 | 西安交通大学 | Detection system and method for unknown mode of satellite remote measurement time sequence data |
CN114743251A (en) * | 2022-05-23 | 2022-07-12 | 西北大学 | Game character facial expression recognition method based on shared integrated convolutional neural network |
Non-Patent Citations (1)
Title |
---|
An Unknown Pattern Detection Method for Time Series Data Based on Convolutional Neural Network;Junpeng Bao等;2019 2nd International Conference on Safety Produce Informatization (IICSPI);20200519;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN115130617A (en) | 2022-09-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7362892B2 (en) | Self-optimizing classifier | |
US7769561B2 (en) | Robust sensor correlation analysis for machine condition monitoring | |
US7107254B1 (en) | Probablistic models and methods for combining multiple content classifiers | |
CN110245745B (en) | Equipment residual service life prediction method based on integrated bidirectional cyclic neural network | |
US7567878B2 (en) | Evaluating anomaly for one class classifiers in machine condition monitoring | |
CN106628097B (en) | A kind of ship equipment method for diagnosing faults based on improvement radial base neural net | |
CN110929847A (en) | Converter transformer fault diagnosis method based on deep convolutional neural network | |
CN107292097B (en) | Chinese medicine principal symptom selection method based on feature group | |
CN110263119A (en) | A kind of reasoning by cases classifier case retrieving method | |
CN112756759B (en) | Spot welding robot workstation fault judgment method | |
CN113255573B (en) | Pedestrian re-identification method based on mixed cluster center label learning and storage medium | |
CN112785015A (en) | Equipment fault diagnosis method based on case reasoning | |
CN115130617B (en) | Detection method for continuous increase of self-adaptive satellite data mode | |
CN112232370A (en) | Fault analysis and prediction method for engine | |
CN114817856A (en) | Beam-pumping unit fault diagnosis method based on structure information holding domain adaptive network | |
CN117111464A (en) | Self-adaptive fault diagnosis method under multiple working conditions | |
CN113158555A (en) | Heavy gas turbine control system analog input module BIT design method based on expert system and random forest classifier | |
CN117113078A (en) | Small sample bearing fault mode identification method and system based on multi-source data integration | |
CN109547248A (en) | Based on artificial intelligence in orbit aerocraft ad hoc network method for diagnosing faults and device | |
CN115903503A (en) | GOAP-based artificial intelligence engine control method | |
CN112015894B (en) | Text single class classification method and system based on deep learning | |
CN113326882A (en) | Model integration method and device based on classification and regression algorithm | |
CN112365014A (en) | GA-BP-CBR-based industrial equipment fault diagnosis system and method | |
Zhao et al. | Industrial fault diagnosis based on few shot learning | |
CN113988713B (en) | Diesel engine assembly quality abnormal pattern recognition method based on GNA-XALO-SVM |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |