Data synthesis method and device, computer storage medium, and electronic device

Abstract

The present disclosure relates to the technical field of artificial intelligence, and provides a data synthesis method, a data synthesis device, a computer storage medium and an electronic device, wherein the data synthesis method comprises: inputting original time series data into a trained data synthesis model, the data synthesis model comprising an encoding network and a decoding network, the decoding network comprising at least two branch decoding networks and a weight distribution network, and different branch decoding networks corresponding to different decoding strategies; performing encoding processing on the original time series data through the encoding network to obtain an encoding feature vector; performing decoding processing on the encoding feature vector through the at least two branch decoding networks to obtain at least two decoding feature vectors; and performing fusion processing on the at least two decoding feature vectors through the weight distribution network to obtain synthesis data corresponding to the original time series data. The present disclosure can reasonably explain the correlation between the synthesis data and the original time series data.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a data synthesis method, a data synthesis apparatus, a computer storage medium, and an electronic device. Background Technology

[0002] In the actual production activities of intelligent operation and maintenance by telecom operators, data is often a valuable resource. However, with the rise of AI (Artificial Intelligence) large-scale models, training deep learning models usually requires a large amount of real data. However, in the early stages of establishing an operation and maintenance platform, the lack of data often makes it difficult to achieve good results in training deep learning models, hindering the implementation of intelligent operation and maintenance.

[0003] In related technologies, synthetic data is generally generated through generative adversarial models or autoencoders. However, the above methods cannot explain the correlation between the original data and the synthetic data.

[0004] Therefore, there is an urgent need in this field to develop a new data synthesis method and apparatus.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure. Summary of the Invention

[0006] The purpose of this disclosure is to provide a data synthesis method, data synthesis apparatus, computer storage medium, and electronic device, thereby overcoming, to at least a certain extent, the technical problem of being unable to explain the correlation between raw data and synthesized data due to limitations in related technologies.

[0007] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0008] According to a first aspect of this disclosure, a data synthesis method is provided, comprising: inputting raw time-series data into a trained data synthesis model, the data synthesis model including an encoding network and a decoding network, the decoding network including at least two branch decoding networks and a weight allocation network, different branch decoding networks corresponding to different decoding strategies; encoding the raw time-series data through the encoding network to obtain an encoded feature vector; decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors; and fusing the at least two decoded feature vectors through the weight allocation network to obtain synthesized data corresponding to the raw time-series data.

[0009] In an exemplary embodiment of this disclosure, encoding the original time-series data through the encoding network to obtain an encoded feature vector includes: encoding the original time-series data through the encoding network to extract the mean vector and standard deviation vector corresponding to the original time-series data; and randomly sampling a standard normal distribution determined based on the mean vector and standard deviation vector to obtain the encoded feature vector.

[0010] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network and at least one seasonal term decomposition network; the decoding process of the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: decomposing the encoded feature vector through the trend term decomposition network to obtain a trend term component; and decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component.

[0011] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network and a residual network; the step of decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: decomposing the encoded feature vector through the trend term decomposition network to obtain trend term components; and decomposing the encoded feature vector through the residual network to obtain residual term components.

[0012] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include at least one seasonal term decomposition network and a residual network; the decoding process of the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component; and decomposing the encoded feature vector through the residual network to obtain a residual term component.

[0013] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network; the step of decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: decomposing the encoded feature vector through the trend term decomposition network to obtain a trend term component; decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component; and decomposing the encoded feature vector through the residual network to obtain a residual term component.

[0014] In an exemplary embodiment of this disclosure, the trend term decomposition network includes a fully connected layer, a dimension transformation layer, and an element permutation layer. The step of decomposing the encoded feature vector using the trend term decomposition network to obtain trend term components includes: classifying the encoded feature vector using the fully connected layer to obtain a first feature vector; performing dimension transformation on the first feature vector using the dimension transformation layer to obtain a trend factor corresponding to the encoded feature vector; the trend factor describing the changing trend of the original time-series data; constructing a target vector based on the sequence length of the original time-series data and the maximum degree of a preset polynomial matrix; determining a second feature vector based on the vector product between the trend factor and the target vector; and permuting the elements of at least two specified dimensions of the second feature vector using the element permutation layer to obtain the trend term components.

[0015] In an exemplary embodiment of this disclosure, each of the seasonal item decomposition networks includes a fully connected layer and a dimension transformation layer; the step of decomposing the encoded feature vector through the at least one seasonal item decomposition network to obtain at least one seasonal item component includes: classifying the encoded feature vector through the fully connected layer of each seasonal item decomposition network to obtain a third feature vector; performing dimension transformation on the third feature vector through the dimension transformation layer of each seasonal item decomposition network to obtain a seasonal factor corresponding to the encoded feature vector; the seasonal factor is used to describe the change period of the original time series data; performing dimension transformation on the seasonal factor again according to a pre-set seasonal index vector to obtain multiple fourth feature vectors; merging the multiple fourth feature vectors to obtain each seasonal item component, thereby obtaining the at least one seasonal item component.

[0016] In an exemplary embodiment of this disclosure, the residual network includes a dimension transformation layer, multiple convolutional layers, and a fully connected layer; the step of decomposing the encoded feature vector through the residual network to obtain residual components includes: performing dimension transformation on the encoded feature vector through the dimension transformation layer to obtain a fifth feature vector; performing feature extraction on the third feature vector through the multiple convolutional layers to obtain a sixth feature vector; and performing classification processing on the fourth feature vector through the fully connected layer to obtain the residual components.

[0017] In an exemplary embodiment of this disclosure, the step of fusing the at least two decoded feature vectors through the weight allocation network to obtain the synthetic data corresponding to the original time series data includes: outputting at least three weight coefficients through the weight allocation network; and weighting and fusing the trend component, the at least one seasonal component, and the residual component according to the at least three weight coefficients to obtain the synthetic data corresponding to the original time series data.

[0018] According to a second aspect of this disclosure, a data synthesis apparatus is provided, comprising: a data input module for inputting raw time-series data into a trained data synthesis model, the data synthesis model including an encoding network and a decoding network, the decoding network including at least two branch decoding networks and a weight allocation network, different branch decoding networks corresponding to different decoding strategies; an encoding module for encoding the raw time-series data through the encoding network to obtain an encoded feature vector; a decoding module for decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors; and a data synthesis module for fusing the at least two decoded feature vectors through the weight allocation network to obtain synthesized data corresponding to the raw time-series data.

[0019] According to a third aspect of this disclosure, a computer storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the data synthesis method described in the first aspect above.

[0020] According to a fourth aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the data synthesis method described in the first aspect by executing the executable instructions.

[0021] As can be seen from the above technical solutions, the data synthesis method, data synthesis apparatus, computer storage medium, and electronic device in the exemplary embodiments of this disclosure have at least the following advantages and positive effects:

[0022] In some embodiments of this disclosure, the technical solutions involve inputting raw time-series data into a trained data synthesis model, encoding the raw time-series data through an encoding network to obtain encoded feature vectors, decoding the encoded feature vectors through at least two branch decoding networks to obtain at least two decoded feature vectors, and fusing the at least two decoded feature vectors through a weight allocation network to obtain synthetic data corresponding to the raw time-series data. This approach enables the generation of new synthetic data based on the raw time-series data, expanding the data volume and facilitating the training of deep learning models based on the raw time-series data and synthetic data during the initial establishment of the operation and maintenance platform, thus achieving intelligent operation and maintenance. Furthermore, since the synthetic data in this disclosure is obtained by first decoding the encoded feature vectors according to at least two decoding strategies and then fusing the at least two decoded feature vectors, it can reasonably explain the correlation between the raw data and the synthetic data, providing an explanation for the synthetic data and enhancing data usability. Finally, given that the model parameters are controllable, this disclosure also allows for flexible and controllable synthetic data through parameter adjustment, enabling the synthetic data to be randomly generated within a predetermined range, thus enhancing the flexibility of data synthesis.

[0023] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0025] Figure 1 A flowchart illustrating how synthetic data is generated in related technologies;

[0026] Figure 2 A flowchart illustrating the data synthesis method in an embodiment of this disclosure is shown;

[0027] Figure 3 This diagram illustrates the structure of the coding network in an embodiment of the present disclosure.

[0028] Figure 4 This diagram illustrates how, in an embodiment of the present disclosure, raw time-series data is encoded using an encoding network to obtain encoded feature vectors.

[0029] Figure 5This illustration shows a flowchart of how to decode the encoded feature vectors through at least two branch decoding networks to obtain at least two decoded feature vectors when at least two branch decoding networks include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network in an embodiment of this disclosure.

[0030] Figure 6 This diagram illustrates the structure of the trend term decomposition network in an embodiment of this disclosure.

[0031] Figure 7 This diagram illustrates how a trend term component is obtained by decomposing the encoded feature vector using a trend term decomposition network in an embodiment of this disclosure.

[0032] Figure 8 This diagram illustrates the structure of the seasonal term decomposition network in an embodiment of this disclosure.

[0033] Figure 9 This diagram illustrates how, in an embodiment of the present disclosure, the encoded feature vector is decomposed using at least one seasonal component network to obtain at least one seasonal component.

[0034] Figure 10 This diagram illustrates the structure of the residual term decomposition network in an embodiment of this disclosure.

[0035] Figure 11 This diagram illustrates how a residual network is used to decompose the encoded feature vector and obtain residual components in an embodiment of this disclosure.

[0036] Figure 12 This diagram illustrates the structure of the weight allocation network in an embodiment of this disclosure.

[0037] Figure 13 This illustration shows a flowchart of how a weighted distribution network is used to fuse at least two decoded feature vectors to obtain synthetic data corresponding to the original time series data in an embodiment of this disclosure.

[0038] Figure 14 This diagram illustrates the overall network structure of the data synthesis method in this embodiment of the present disclosure.

[0039] Figure 15 This diagram illustrates the structure of the data synthesis apparatus in an exemplary embodiment of this disclosure.

[0040] Figure 16 A schematic diagram of the structure of an electronic device in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0041] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0042] The terms “a,” “an,” “the,” and “the” are used in this specification to indicate the presence of one or more elements / components / etc.; the terms “including” and “having” are used to indicate an open-ended inclusion and to mean that there may be other elements / components / etc. in addition to the listed elements / components / etc.; the terms “first” and “second” are used only as markings and are not a limitation on the number of objects.

[0043] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0044] In related technologies, synthetic data is typically generated through generative adversarial models, autoencoders, or variational autoencoders to expand the dataset. (Reference) Figure 1 , Figure 1 A flowchart illustrating how synthetic data is generated in related technologies is shown, such as... Figure 1 As shown:

[0045] First, the raw timing data can be input into the encoder;

[0046] The encoder can then encode the original time-series data and output the mean vector and variance vector.

[0047] Finally, synthetic data can be obtained by sampling based on the normal distribution composed of the mean vector and variance vector mentioned above.

[0048] However, the above scheme has the following drawbacks: it lacks interpretability, cannot explain the correlation between the original data and the synthetic data, and the synthetic data cannot be controlled by humans and is relatively random.

[0049] In the embodiments of this disclosure, a data synthesis method is first provided, which at least to some extent overcomes the shortcomings of related technologies that cannot explain the correlation between original data and synthesized data.

[0050] Figure 2 The diagram shows a flowchart of a data synthesis method in an embodiment of this disclosure. The execution entity of the data synthesis method can be a server that synthesizes the data.

[0051] refer to Figure 2 A data synthesis method according to an embodiment of the present disclosure includes the following steps:

[0052] Step S210: Input the original time series data into the trained data synthesis model. The data synthesis model includes an encoding network and a decoding network. The decoding network includes at least two branch decoding networks and a weight allocation network. Different branch decoding networks correspond to different decoding strategies.

[0053] Step S220: The original time-series data is encoded using an encoding network to obtain an encoded feature vector;

[0054] Step S230: Decode the encoded feature vector through at least two branch decoding networks to obtain at least two decoded feature vectors;

[0055] Step S240: At least two decoded feature vectors are fused using a weight allocation network to obtain synthetic data corresponding to the original time series data.

[0056] exist Figure 2 In the technical solution provided by the illustrated embodiment, the original time-series data is input into a trained data synthesis model. The original time-series data is encoded using an encoding network to obtain encoded feature vectors. These encoded feature vectors are then decoded using at least two branch decoding networks to obtain at least two decoded feature vectors. Finally, the at least two decoded feature vectors are fused using a weight allocation network to obtain the synthesized data corresponding to the original time-series data. On the one hand, this enables the generation of new synthesized data based on the original time-series data, expanding the data volume. This facilitates the training of a deep learning model based on the original time-series data and the synthesized data during the initial establishment of the operation and maintenance platform, achieving intelligent operation and maintenance. On the other hand, since the synthesized data in this disclosure is obtained by first decoding the encoded feature vectors according to at least two decoding strategies and then fusing the at least two decoded feature vectors, the correlation between the original data and the synthesized data can be reasonably explained, providing an explanation for the synthesized data and enhancing data usability. Furthermore, given that the model parameters are controllable, this disclosure also enables flexible and controllable synthesized data through parameter adjustment, allowing the synthesized data to be randomly generated within a predetermined range, enhancing the flexibility of data synthesis.

[0057] The following are Figure 2 The specific implementation process of each step in the process will be explained in detail:

[0058] Before step S210, it should be noted that the present disclosure can pre-train a data synthesis model. For example, training samples can be obtained, and then the data synthesis model to be trained can be iteratively trained using the training samples until the loss function of the data synthesis model to be trained tends to converge, thereby obtaining the trained data synthesis model.

[0059] In one optional implementation, service indicator data can be collected and used as the training samples. For example, the service indicator data can be network quality indicator data of the operator within a historical period. For example, network quality can be characterized by bandwidth (the "highest data rate" that can pass from one point to another in the network per unit time), latency (which can be processing latency, queuing latency, transmission latency, propagation latency, etc.), jitter (network jitter refers to the time difference between the maximum and minimum latency), and packet loss (simply put, packet loss means that one or more data packets cannot reach their destination through the network; if the receiving end detects data loss, it will send a request to the sending end based on the queue number for packet retransmission). These parameters can be set according to actual conditions, and this disclosure does not impose any special limitations on them.

[0060] In another optional implementation, after collecting the above-mentioned business indicator data, the business indicator data can be cleaned first, and the cleaned business indicator data can be sampled according to a preset sampling interval. Then, the first time series data obtained by sampling can be imputed for missing values, and the second time series data obtained after imputed for missing values ​​can be used as the above-mentioned training sample.

[0061] It should be noted that the data synthesis model to be trained mentioned above may include an encoding network and a decoding network. The encoding network may be a variational autoencoder network, used to encode the training samples to obtain encoded feature vectors. The decoding network may include at least two branch decoding networks and a weight allocation network. Different branch decoding networks are used to decode the encoded feature vectors based on different decoding strategies, thereby obtaining at least two decoded feature vectors. Then, the weight allocation network can assign weights to the at least two decoded feature vectors to fuse them and generate the synthesized sample data.

[0062] For example, the generated sample synthetic data can be compared with the pre-set label data, and the loss value of the data synthesis model to be trained can be determined based on the difference between the two. The trained data synthesis model is obtained when the loss value tends to converge.

[0063] After obtaining the trained data synthesis model, you can proceed to step S210 and input the original time series data into the trained data synthesis model.

[0064] In this step, the original time series data can be input into the trained data synthesis model. As can be seen from the relevant explanations in the above steps, the data synthesis model can include an encoding network and a decoding network. The decoding network includes at least two branch decoding networks and a weight allocation network. Different branch decoding networks correspond to different decoding strategies.

[0065] It should be noted that the above-mentioned original time series data can contain N basic data sequences (N is an integer greater than or equal to 1), and the matrix shape of each basic data sequence can be D*T, where D is the number of dimensions and is an integer greater than or equal to 1, and T is the number of time steps and is also an integer greater than or equal to 1. Therefore, the matrix shape of the above-mentioned original time series data can be N*D*T.

[0066] For example, the above-mentioned basic data sequence can be real business indicator data collected at specified intervals within a preset time period. For instance, the preset time period can be 7 days, and the specified duration can be 1 hour. Thus, the number of time steps T = m * d = 7 * 24.

[0067] In step S220, the original time-series data is encoded using an encoding network to obtain an encoded feature vector.

[0068] In this step, refer to Figure 3 , Figure 3 A schematic diagram of the structure of the coding network in an embodiment of this disclosure is shown, such as... Figure 3 As shown, this encoding network can be composed of multiple one-dimensional convolutional layers (Conv1D), flattened layers, and fully connected layers (Dense). Convolutional layers are a crucial component of deep learning neural networks; their function is to extract and map features from the input data through convolution operations, thereby enabling efficient processing of high-dimensional data such as images, speech, and text. Flattened layers transform three-dimensional matrices into two-dimensional ones. Dense layers are a commonly used type of fully connected layer in neural networks. Their function is to perform matrix multiplication of the input data with a set of weights, add a bias, and then perform a non-linear transformation through an activation function to output a new feature representation. The parameters of a Dense layer mainly include a weight matrix and a bias vector, the shapes of which depend on the dimensions of the input and output.

[0069] For details, please refer to Figure 4 , Figure 4This illustration shows a flowchart of how to encode raw time-series data through an encoding network to obtain encoded feature vectors in an embodiment of this disclosure, including steps S401-S402:

[0070] In step S401, the original time series data is encoded by an encoding network to extract the mean vector and standard deviation vector corresponding to the original time series data.

[0071] In this step, the original time series data can be encoded using the aforementioned encoding network to extract the mean vector μ and standard deviation vector σ corresponding to the original time series data.

[0072] In step S402, random sampling is performed on the standard normal distribution determined based on the mean vector and standard deviation vector to obtain the encoded feature vector.

[0073] In this step, after obtaining the mean vector μ and standard deviation vector σ, a standard normal distribution N(μ, σ) can be constructed. Then, the standard normal distribution can be randomly resampled to obtain the encoded feature vector z.

[0074] Next, refer to Figure 2 In step S230, the encoded feature vector is decoded through at least two branch decoding networks to obtain at least two decoded feature vectors.

[0075] In this step, in one optional embodiment, the at least two branch decoding networks may include a trend term decomposition network and at least one seasonal term decomposition network; in another optional embodiment, the at least two branch decoding networks may include a trend term decomposition network and a residual network; in yet another optional embodiment, the at least two branch decoding networks may include at least one seasonal term decomposition network and a residual network; in yet another optional embodiment, the two branch decoding networks may include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network.

[0076] Taking the example of the above-mentioned at least two branch decoding networks including a trend term decomposition network, at least one seasonal term decomposition network and a residual network, the above step S230 will be explained.

[0077] refer to Figure 5 , Figure 5 This embodiment of the present disclosure illustrates how, when at least two branch decoding networks include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network, the encoded feature vector is decoded using at least two branch decoding networks to obtain at least two decoded feature vectors. The flowchart includes steps S501-S503.

[0078] In step S501, the encoded feature vector is decomposed by a trend term decomposition network to obtain trend term components.

[0079] In this step, the role of the trend term decomposition network is to decompose the trend term components from the encoded feature vector. (Reference) Figure 6 , Figure 6 A schematic diagram of the trend term decomposition network in an embodiment of this disclosure is shown, as follows: Figure 6 As shown:

[0080] The trend term decomposition network can include fully connected layers (Dense), dimension transformation layers (Reshape) matrix multiplication units, and element permutation layers (Transpose). The function of the fully connected layers can be referred to the relevant explanation in step S220 above, and will not be repeated here. The function of the dimension transformation layers can be to convert two-dimensional vectors into three-dimensional vectors. The function of the element permutation layers can be to interchange the elements of the two-dimensional and three-dimensional vectors.

[0081] For details, please refer to Figure 7 , Figure 7 This illustration shows a flowchart of how to decompose the encoded feature vector using a trend term decomposition network to obtain trend term components in an embodiment of this disclosure, including steps S701-S705:

[0082] In step S701, the encoded feature vector is classified through a fully connected layer to obtain the first feature vector.

[0083] In this step, the encoded feature vector can be classified using the fully connected layer to obtain a first feature vector, the matrix shape of which can be N*D.

[0084] In step S702, the first feature vector is dimensionally transformed by the dimension transformation layer to obtain the trend factor corresponding to the encoded feature vector.

[0085] In this step, the first feature vector can be transformed in dimension by the aforementioned dimension transformation layer, so as to convert the two-dimensional first feature vector into a three-dimensional trend factor θ. t The trend factor θ t Used to describe the changing trend of raw time series data, for example, it can be the overall long-term trend of raw time series data that is continuously increasing or continuously decreasing monotonically.

[0086] It should be noted that the matrix shape of this trend factor can be N*D*P, where P is the maximum degree of the preset polynomial matrix, and the degree of the highest degree term of the polynomial is called the degree of the polynomial.

[0087] In step S703, a target vector is constructed based on the sequence length of the original time series data and the maximum degree of the preset polynomial matrix.

[0088] In this step, the target vector R can be constructed based on the sequence length (i.e., the number of time steps T) of the original time series data and the maximum degree P of the preset polynomial matrix.

[0089] For example, a time vector r can be constructed first based on the sequence length T of the original time series data. Then, based on the maximum degree P mentioned above, we can construct a power vector R of the time vector r, which is R = [r, ..., r P Furthermore, the matrix shape of the power vector R is P*T.

[0090] In step S704, the second feature vector is determined based on the vector product between the trend factor and the target vector.

[0091] In this step, the trend factor θ mentioned above can be calculated. t The second eigenvector is obtained by multiplying the vector with the target vector R.

[0092] In step S705, the elements of at least two specified dimensions of the second feature vector are permuted by the element permutation layer to obtain the trend component.

[0093] In this step, the elements of at least two specified dimensions of the second feature vector (e.g., the elements of the second and third dimensions, which can be set according to the actual situation, and this disclosure does not impose any special limitations on this) can be replaced by an element replacement layer to obtain the trend component V. t .

[0094] Next, refer to Figure 5 In step S502, the encoded feature vector is decomposed by at least one seasonal term decomposition network to obtain at least one seasonal term component.

[0095] In this step, for example, at least one seasonal term decomposition network can be set up. The function of the seasonal term decomposition network is to decompose the seasonal term components from the encoded feature vector. (See reference) Figure 8 , Figure 8 A schematic diagram of the structure of the seasonal term decomposition network in an embodiment of this disclosure is shown, as follows: Figure 8 As shown:

[0096] The trend decomposition network can contain fully connected layers (Dense) and dimension transformation layers (Reshape). The roles of the fully connected layers and dimension transformation layers can be found in the explanation of step S501 above, and will not be repeated here.

[0097] For details, please refer to Figure 9 , Figure 9 This illustration shows a flowchart of how to decompose the encoded feature vector using at least one seasonal component through at least one seasonal component decomposition network in an embodiment of this disclosure, including steps S901-S904:

[0098] In step S901, the encoded feature vector is classified through the fully connected layer of each seasonal term decomposition network to obtain the third feature vector.

[0099] In this step, taking the j-th seasonal term decomposition network as an example, the above encoded feature vector can be classified through the fully connected layer of the j-th seasonal term decomposition network to obtain the third feature vector. The matrix shape of the third feature vector can be N*D.

[0100] In step S902, the third feature vector is dimensionally transformed through the dimension transformation layer of each seasonal term decomposition network to obtain the seasonal factor corresponding to the encoded feature vector.

[0101] In this step, the third feature vector can be dimensionally transformed using the dimension transformation layer of the j-th seasonal term decomposition network, converting the two-dimensional third feature vector into a three-dimensional seasonal factor. The seasonal factor Used to describe the changing period of the original time series data, for example, it can be the seasonal or periodic fluctuation of the original time series.

[0102] It should be noted that this seasonal factor The shape of the matrix can be N*D*m.

[0103] In step S903, the seasonal factors are transformed again according to the pre-set seasonal index vector to obtain multiple fourth feature vectors.

[0104] In this step, when the time step T = m * d = 7 * 24, the pre-set seasonal index vector can be K = [0,0,0..,1,1,1…,6,6,6], with a length of T = 7 * 24. 0 appears 24 times, 1 appears 24 times, 2 appears 24 times, 6 appears 24 times, and so on, arranged in order.

[0105] Therefore, the above can be extracted. The matrix with the last dimension indexed at 0 is repeated 24 times to expand the N*D*1 matrix to N*D*24 (N*D*d); then, the above can be truncated. The matrix with the last dimension indexed as 1 is repeated 24 times to expand the N*D*1 matrix to N*D*24 (N*D*d); then, the above can be truncated. The matrix with the last dimension indexed at 2 is repeated 24 times, that is, the N*D*1 dimension matrix is ​​expanded to N*D*24 (N*D*d), and so on, until the above is truncated. The matrix with the last dimension index 6 is repeated 24 times to expand the N*D*1 matrix to N*D*24 (N*D*d), resulting in 7 fourth eigenvectors.

[0106] In step S904, multiple fourth feature vectors are merged to obtain each seasonal component, thereby obtaining at least one seasonal component.

[0107] In this step, the seven matrices obtained in step S903 can be merged in the last dimension to obtain the seasonal component output by the j-th seasonal component decomposition module (the matrix shape of the seasonal component is N*D*T), so as to obtain at least one seasonal component output by at least one seasonal component.

[0108] Next, refer to Figure 5 In step S503, the encoded feature vector is decomposed by the residual network to obtain the residual term components.

[0109] In this step, the residual network described above serves to decompose the encoded feature vector to obtain residual term components. (Reference) Figure 10 , Figure 10 This diagram illustrates the structure of the residual term decomposition network in an embodiment of this disclosure, as shown below. Figure 10 As shown:

[0110] The residual term decomposition network can include a dimension transformation layer (Reshape), a multi-layer convolutional layer (Conv1DTransp), and a fully connected layer (Dense). The function of the multi-layer convolutional layer can be referred to the relevant explanation in step S220 above, and the functions of the dimension transformation layer and the fully connected layer can be referred to the relevant explanation in step S501 above. They will not be repeated here.

[0111] For details, please refer to Figure 11 , Figure 11 This diagram illustrates how a residual network is used to decompose the encoded feature vector and obtain residual components in an embodiment of this disclosure, including steps S1101-S1103:

[0112] In step S1101, the encoded feature vector is transformed through a dimension transformation layer to obtain the fifth feature vector.

[0113] In this step, the dimensionality of the above encoded feature vector can be transformed by a dimensionality transformation layer to obtain the fifth feature vector.

[0114] In step S1102, the third feature vector is extracted through multiple convolutional layers to obtain the sixth feature vector.

[0115] In this step, the third feature vector can be extracted using the aforementioned multi-layer convolutional layers to obtain the sixth feature vector.

[0116] In step S1103, the fourth feature vector is classified through a fully connected layer to obtain the residual components.

[0117] In this step, the fourth feature vector can be classified using a fully connected layer to obtain the residual component V. r .

[0118] Next, refer to Figure 2 In step S240, at least two decoded feature vectors are fused through a weight allocation network to obtain the synthetic data corresponding to the original time series data.

[0119] In this step, the weight allocation network is used to assign weight coefficients to at least one decoded feature vector, and then to perform weighted fusion on the at least one decoded feature vector based on the assigned weight coefficients to obtain the synthetic data corresponding to the original time series data.

[0120] refer to Figure 12 , Figure 12 This diagram illustrates the structure of the weight allocation network in an embodiment of the present disclosure, as shown below. Figure 12 As shown, this weight allocation network can contain fully connected layers (Dense), dimension transformation layers (Reshape), and Softmax layers. Softmax is one of the most common activation functions in neural networks, which can transform the output of the neural network into a probability distribution.

[0121] It should be noted that when the above-mentioned at least two branch decoding networks include a trend term decomposition network and S (S is an integer greater than or equal to 1) seasonal term decomposition networks, the weight allocation network can output S+1 weight coefficients; when the above-mentioned at least two branch decoding networks include a trend term decomposition network and a residual network, the weight allocation network can output 2 weight coefficients; when the above-mentioned at least two branch decoding networks include S seasonal term decomposition networks and a residual network, the weight allocation network can output S+1 weight coefficients; when the above-mentioned two branch decoding networks include a trend term decomposition network, S seasonal term decomposition networks, and a residual network, the weight allocation network can output S+2 weight coefficients.

[0122] The above two-branch decoding network, including the trend term decomposition network, the S seasonal term decomposition network, and the residual network, will be used as an example for explanation. (Refer to...) Figure 13 , Figure 13This illustration shows a flowchart of how, in this embodiment of the present disclosure, at least two decoded feature vectors are fused using a weight allocation network to obtain synthetic data corresponding to the original time-series data, including steps S1301-S1302:

[0123] In step S1301, at least three weight coefficients of the weight allocation network output are obtained.

[0124] In this step, given that the minimum value of S is 1, the weight allocation network can output at least 1+2=3 weight coefficients.

[0125] In step S1302, the trend component, at least one seasonal component, and the residual component are weighted and fused according to at least three weighting coefficients to obtain the synthetic data corresponding to the original time series data.

[0126] In this step, the trend component, at least one seasonal component, and the residual component can be weighted and fused according to the above at least three weighting coefficients to obtain the synthetic data corresponding to the original time series data.

[0127] In the technical solution disclosed herein, since the final output result is a weighted summation of the output results corresponding to multiple different sub-modules, it is convenient to output the desired generation result in a targeted manner by increasing or decreasing the number of sub-modules, modifying the parameters of each sub-module to interfere with the output parameters of the sub-modules, etc., thereby increasing the interpretability of the model.

[0128] refer to Figure 14 , Figure 14 This diagram illustrates the overall network structure of the data synthesis method in this embodiment of the present disclosure, as shown below. Figure 14 As shown:

[0129] After encoding the original time-series data through an encoding network to obtain the encoded feature vector z, the encoded feature vector can be input into a trend term decomposition network (outputting trend term components), an S seasonal term decomposition network (outputting S seasonal term components), a residual network (outputting residual term components), and a weight allocation network. The weight allocation network outputs S+2 weight coefficients. Then, the trend term components, the S seasonal term components, and the residual term components can be weighted and fused according to these S+2 weight coefficients to obtain the synthetic data corresponding to the original time-series data.

[0130] Based on the above technical solutions, the technical solutions in this disclosure have at least the following advantages compared with the prior art:

[0131] First, the speed of generating synthetic data based on neural networks has been greatly improved;

[0132] Second, the architecture involves some relatively mature and general-purpose algorithms, resulting in low trial-and-error costs.

[0133] Therefore, this disclosure can be applied to at least the following scenarios:

[0134] First, the synthesis of multi-dimensional indicators of wireless access network equipment is used to expand data accumulation;

[0135] Second, the synthesis of multi-dimensional indicators of optical broadband network equipment is used to expand data accumulation;

[0136] Third, in other scenarios involving the synthesis of time series data, settings can be customized according to actual needs.

[0137] This disclosure also provides a data synthesis apparatus. Figure 15 This diagram illustrates the structure of the data synthesis apparatus in an exemplary embodiment of this disclosure; as shown below. Figure 15 As shown, the data synthesis device 1500 may include a data input module 1510, an encoding module 1520, a decoding module 1530, and a data synthesis module 1540. Wherein:

[0138] The data input module 1510 is used to input raw time series data into a trained data synthesis model. The data synthesis model includes an encoding network and a decoding network. The decoding network includes at least two branch decoding networks and a weight allocation network. Different branch decoding networks correspond to different decoding strategies.

[0139] Encoding module 1520 is used to encode the original time-series data through the encoding network to obtain an encoded feature vector;

[0140] Decoding module 1530 is used to decode the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors;

[0141] The data synthesis module 1540 is used to fuse the at least two decoded feature vectors through the weight allocation network to obtain the synthesized data corresponding to the original time series data.

[0142] In an exemplary embodiment of this disclosure, the encoding module 1520 uses an encoding network to encode the original time-series data to obtain an encoded feature vector, including: encoding the original time-series data through the encoding network to extract the mean vector and standard deviation vector corresponding to the original time-series data; and randomly sampling a standard normal distribution determined based on the mean vector and standard deviation vector to obtain the encoded feature vector.

[0143] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network and at least one seasonal term decomposition network; the decoding module 1530 performs decoding processing on the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors, including: decomposing the encoded feature vector through the trend term decomposition network to obtain a trend term component; and decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component.

[0144] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network and a residual network; the decoding module 1530 performs decoding processing on the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors, including: decomposing the encoded feature vector through the trend term decomposition network to obtain trend term components; and decomposing the encoded feature vector through the residual network to obtain residual term components.

[0145] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include at least one seasonal term decomposition network and a residual network; the decoding module 1530 performs decoding processing on the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors, including: decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component; and decomposing the encoded feature vector through the residual network to obtain a residual term component.

[0146] In an exemplary embodiment of this disclosure, the at least two branch decoding networks include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network; the decoding module 1530 performs decoding processing on the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors, including: decomposing the encoded feature vector through the trend term decomposition network to obtain a trend term component; decomposing the encoded feature vector through the at least one seasonal term decomposition network to obtain at least one seasonal term component; and decomposing the encoded feature vector through the residual network to obtain a residual term component.

[0147] In an exemplary embodiment of this disclosure, the trend term decomposition network includes a fully connected layer, a dimension transformation layer, and an element permutation layer. The decoding module 1530 decomposes the encoded feature vector through the trend term decomposition network to obtain trend term components, including: classifying the encoded feature vector through the fully connected layer to obtain a first feature vector; performing dimension transformation on the first feature vector through the dimension transformation layer to obtain a trend factor corresponding to the encoded feature vector; the trend factor is used to describe the changing trend of the original time series data; constructing a target vector based on the sequence length of the original time series data and the maximum degree of a preset polynomial matrix; determining a second feature vector based on the vector product between the trend factor and the target vector; and permuting the elements of at least two specified dimensions of the second feature vector through the element permutation layer to obtain the trend term components.

[0148] In an exemplary embodiment of this disclosure, each of the seasonal item decomposition networks includes a fully connected layer and a dimension transformation layer; the decoding module 1530 decomposes the encoded feature vector through the at least one seasonal item decomposition network to obtain at least one seasonal item component, including: classifying the encoded feature vector through the fully connected layer of each seasonal item decomposition network to obtain a third feature vector; performing dimension transformation on the third feature vector through the dimension transformation layer of each seasonal item decomposition network to obtain a seasonal factor corresponding to the encoded feature vector; the seasonal factor is used to describe the change period of the original time series data; performing dimension transformation on the seasonal factor again according to a pre-set seasonal index vector to obtain multiple fourth feature vectors; merging the multiple fourth feature vectors to obtain each seasonal item component, thereby obtaining the at least one seasonal item component.

[0149] In an exemplary embodiment of this disclosure, the residual network includes a dimension transformation layer, multiple convolutional layers, and a fully connected layer; the decoding module 1530 decomposes the encoded feature vector through the residual network to obtain residual term components, including: performing dimension transformation on the encoded feature vector through the dimension transformation layer to obtain a fifth feature vector; performing feature extraction on the third feature vector through the multiple convolutional layers to obtain a sixth feature vector; and performing classification processing on the fourth feature vector through the fully connected layer to obtain the residual term components.

[0150] In an exemplary embodiment of this disclosure, the data synthesis module 1540 performs fusion processing on the at least two decoded feature vectors through the weight allocation network to obtain the synthesized data corresponding to the original time series data, including: outputting at least three weight coefficients through the weight allocation network; and performing weighted fusion on the trend component, the at least one seasonal component, and the residual component according to the at least three weight coefficients to obtain the synthesized data corresponding to the original time series data.

[0151] The specific details of each module in the aforementioned data synthesis device have been described in detail in the corresponding data synthesis method, so they will not be repeated here.

[0152] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0153] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0154] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0155] This application also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device.

[0156] Computer-readable storage media can be, for example—but not limited to—electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0157] A computer-readable storage medium can be sent, propagated, or transmitted for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0158] A computer-readable storage medium carries one or more programs that, when executed by an electronic device, cause the electronic device to perform the methods described in the above embodiments.

[0159] Furthermore, this disclosure also provides an electronic device capable of implementing the above-described method.

[0160] Those skilled in the art will understand that various aspects of this disclosure can be implemented as a system, method, or program product. Therefore, various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0161] The following reference Figure 16 To describe an electronic device 1600 according to such an embodiment of the present disclosure. Figure 16 The electronic device 1600 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0162] like Figure 16As shown, the electronic device 1600 is manifested in the form of a general-purpose computing device. The components of the electronic device 1600 may include, but are not limited to: at least one processing unit 1610, at least one storage unit 1620, a bus 1630 connecting different system components (including storage unit 1620 and processing unit 1610), and a display unit 1640.

[0163] The storage unit stores program code that can be executed by the processing unit 1610, causing the processing unit 1610 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 1610 can perform actions such as... Figure 2 As shown: Step S210, input the original time-series data into the trained data synthesis model, the data synthesis model includes an encoding network and a decoding network, the decoding network includes at least two branch decoding networks and a weight allocation network, different branch decoding networks correspond to different decoding strategies; Step S220, encode the original time-series data through the encoding network to obtain an encoded feature vector; Step S230, decode the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors; Step S240, fuse the at least two decoded feature vectors through the weight allocation network to obtain the synthesized data corresponding to the original time-series data.

[0164] Storage unit 1620 may include readable media in the form of volatile storage units, such as random access memory (RAM) 16201 and / or cache memory 16202, and may further include read-only memory (ROM) 16203.

[0165] Storage unit 1620 may also include a program / utility 16204 having a set (at least one) program module 16205, such program module 16205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0166] Bus 1630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0167] Electronic device 1600 can also communicate with one or more external devices 1700 (e.g., keyboard, pointing device, Bluetooth device, etc.), one or more devices that enable a user to interact with electronic device 1600, and / or any device that enables electronic device 1600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 1650. Furthermore, electronic device 1600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 1660. As shown, network adapter 1660 communicates with other modules of electronic device 1600 via bus 1630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 1600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0168] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A data synthesis method, characterized in that, include: The raw time-series data is input into a trained data synthesis model, which includes an encoding network and a decoding network. The decoding network contains at least two branch decoding networks and a weight allocation network, with different branch decoding networks corresponding to different decoding strategies. The raw time-series data includes multiple basic data sequences composed of real business indicator data, including network quality indicator data from operators. The network quality indicator data includes bandwidth, latency, jitter, and packet loss. The original time-series data is encoded using the encoding network to obtain an encoded feature vector. The encoded feature vector is decoded by the at least two branch decoding networks to obtain at least two decoded feature vectors; when the branch decoding network includes a trend term decomposition network and at least one seasonal term decomposition network, the trend term components are decomposed by the trend term decomposition network. At least one seasonal component is decomposed using the at least one seasonal component decomposition network; When the trend term decomposition network includes a fully connected layer, a dimension transformation layer, and an element permutation layer, the fully connected layer classifies the encoded feature vector to obtain a first feature vector; the dimension transformation layer performs dimension transformation on the first feature vector to obtain a trend factor; a target vector is constructed based on the sequence length of the original time series data and the maximum degree of a preset polynomial matrix; a second feature vector is determined based on the vector product between the trend factor and the target vector; and the element permutation layer permutes the elements of at least two specified dimensions of the second feature vector to obtain the trend term components. When the seasonal item decomposition network includes a fully connected layer and a dimension transformation layer, the fully connected layer classifies the encoded feature vector to obtain a third feature vector; the dimension transformation layer performs dimension transformation on the third feature vector to obtain the seasonal factor corresponding to the encoded feature vector; the seasonal factor is then dimension-transformed again according to a pre-set seasonal index vector to obtain multiple fourth feature vectors; the multiple fourth feature vectors are merged to obtain the at least one seasonal item component. The weight allocation network is used to fuse the at least two decoded feature vectors to obtain the synthetic data corresponding to the original time series data.

2. The method according to claim 1, characterized in that, The original time-series data is encoded using the encoding network to obtain an encoded feature vector, including: The original time series data is encoded through the encoding network to extract the mean vector and standard deviation vector corresponding to the original time series data; Random sampling is performed on the standard normal distribution determined based on the mean vector and standard deviation vector to obtain the encoded feature vector.

3. The method according to claim 1, characterized in that, When the at least two branch decoding networks include a trend term decomposition network and a residual network; The step of decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: The trend term decomposition network is used to decompose the encoded feature vector to obtain trend term components; The residual network is used to decompose the encoded feature vector to obtain residual components.

4. The method according to claim 1, characterized in that, When the at least two branch decoding networks include at least one seasonal term decomposition network and a residual network; The step of decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: The encoded feature vector is decomposed using the at least one seasonal term decomposition network to obtain at least one seasonal term component; The residual network is used to decompose the encoded feature vector to obtain residual components.

5. The method according to claim 1, characterized in that, When the at least two branch decoding networks include a trend term decomposition network, at least one seasonal term decomposition network, and a residual network; The step of decoding the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors includes: The trend term decomposition network is used to decompose the encoded feature vector to obtain trend term components; The encoded feature vector is decomposed using the at least one seasonal term decomposition network to obtain at least one seasonal term component; The residual network is used to decompose the encoded feature vector to obtain residual components.

6. The method according to any one of claims 3, 4, and 5, characterized in that, The residual network includes a dimension transformation layer, multiple convolutional layers, and fully connected layers; The step of decomposing the encoded feature vector through the residual network to obtain residual components includes: The dimensionality of the encoded feature vector is transformed by the dimensionality transformation layer to obtain the fifth feature vector; The sixth feature vector is obtained by extracting features from the fifth feature vector through the multi-layer convolutional layers. The sixth feature vector is classified through the fully connected layer to obtain the residual component.

7. The method according to claim 5, characterized in that, The step of fusing the at least two decoded feature vectors through the weight allocation network to obtain the synthetic data corresponding to the original time-series data includes: The weight allocation network outputs at least three weight coefficients; The trend component, the at least one seasonal component, and the residual component are weighted and fused according to the at least three weighting coefficients to obtain the synthetic data corresponding to the original time series data.

8. A data synthesis apparatus, characterized in that, include: The data input module is used to input raw time-series data into a trained data synthesis model. The data synthesis model includes an encoding network and a decoding network. The decoding network includes at least two branch decoding networks and a weight allocation network. Different branch decoding networks correspond to different decoding strategies. The raw time-series data includes multiple basic data sequences composed of real service indicator data, including network quality indicator data from operators. The network quality indicator data includes bandwidth, latency, jitter, and packet loss. The encoding module is used to encode the original time-series data through the encoding network to obtain an encoded feature vector; A decoding module is used to decode the encoded feature vector through the at least two branch decoding networks to obtain at least two decoded feature vectors; When the branch decoding network includes a trend item decomposition network and at least one seasonal item decomposition network, the decoding module is used to decompose the trend item components through the trend item decomposition network; At least one seasonal component is decomposed using the at least one seasonal component decomposition network; When the trend term decomposition network includes a fully connected layer, a dimension transformation layer, and an element permutation layer, the decoding module is used to classify the encoded feature vector through the fully connected layer to obtain a first feature vector; to perform dimension transformation on the first feature vector through the dimension transformation layer to obtain a trend factor; to construct a target vector based on the sequence length of the original time series data and the maximum degree of the preset polynomial matrix; to determine a second feature vector based on the vector product between the trend factor and the target vector; and to permutate the elements of at least two specified dimensions of the second feature vector through the element permutation layer to obtain the trend term component. When the seasonal item decomposition network includes a fully connected layer and a dimension transformation layer, the decoding module is used to classify the encoded feature vector through the fully connected layer to obtain a third feature vector; to perform dimension transformation on the third feature vector through the dimension transformation layer to obtain the seasonal factor corresponding to the encoded feature vector; to perform dimension transformation on the seasonal factor again according to a pre-set seasonal index vector to obtain multiple fourth feature vectors; and to merge the multiple fourth feature vectors to obtain the at least one seasonal item component. The data synthesis module is used to fuse the at least two decoded feature vectors through the weight allocation network to obtain the synthesized data corresponding to the original time series data.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data synthesis method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the data synthesis method according to any one of claims 1 to 7 by executing the executable instructions.

Images