Rnn time series classification method based on deep variational information bottleneck

Abstract

The application relates to the technical field of computer data processing, and discloses an RNN time sequence classification method based on a deep variational information bottleneck, which comprises the following steps: acquiring multivariate time sequence data, processing the multivariate time sequence data through a recurrent neural network to generate a complete hidden state sequence; based on the sequence, running an optimal time segmentation algorithm to calculate optimal segmentation points, and then dividing the sequence into two parts of an early stage and a late stage; performing information bottleneck coding on the aggregated representation of the early stage sequence and the final hidden state of the complete sequence respectively to generate first and second latent representations; after fusing the two latent representations, processing through a decoder network to obtain a final classification result; calculating a composite loss function according to the classification result and a real label, and updating all trainable parameters in the model through a back propagation algorithm. The application adopts a data-driven optimal time segmentation algorithm, can adaptively identify and segment key time points at which information characteristics in the sequence change, and improves the pertinence of feature extraction.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to an RNN time series classification method based on the bottleneck of deep variational information. Background Technology

[0002] In many critical fields such as modern industrial production, equipment monitoring, and medical diagnosis, accurately interpreting and classifying time-series data is crucial. For example, the operating status of a large wind turbine may be affected by two completely different patterns simultaneously: long-term accumulated gear wear and recent sudden blade icing. How to accurately identify these intertwined gradual trends and abrupt events from a continuous data stream and make correct judgments is directly related to production safety and economic efficiency.

[0003] In existing technologies, recurrent neural networks (RNNs) and their variants, such as long short-term memory networks (LSTMs), are powerful tools for processing time-series data. By processing information step by step, they can effectively build time-dependent models between data points and capture the dynamic evolution of the sequence. To further enhance the model's capabilities, an attention mechanism has been introduced. This mechanism gives the model a global perspective, enabling it to dynamically allocate focus to any time point in the entire sequence that is most informative when making classification decisions, thereby effectively utilizing key moment information in the sequence.

[0004] However, existing technologies still have inherent limitations when processing certain types of complex sequences. First, the structure of standard recurrent neural networks makes them more sensitive to information at the end of the sequence. This recency effect weakens the model's memory of key signals from the early stages of the sequence; the information weight of a slowly developing fault symptom can easily be diluted over a long period of information transmission. Second, while attention mechanisms can review the overall picture, they weight all information within a unified feature space and do not fundamentally distinguish between long-term trends and short-term events. This mixed processing of information makes it difficult for the model to develop a specialized and pure understanding of different patterns. Finally, existing models lack an adaptive mechanism to discover potential state inflection points within the sequence. They either treat the sequence as a whole or rely on rigid, pre-defined rules for segmentation, failing to flexibly adjust the analysis strategy according to the characteristics of each data point. To address these issues, those skilled in the art have proposed an RNN time series classification method based on the bottleneck of deep variational information. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an RNN time series classification method based on the bottleneck of deep variational information. This method solves the technical problem that existing technologies, when processing time series data, cannot effectively distinguish and decouple long-term slowly changing patterns from recent abrupt changes, resulting in mixed feature representations, diluted key information, and ultimately affecting classification performance.

[0006] To achieve the above objectives, this invention provides the following technical solution: an RNN time-series classification method based on the bottleneck of deep variational information, comprising the following steps:

[0007] Acquire multivariate time series data to be classified, and process the time series data through a recurrent neural network to generate a complete hidden state sequence containing multiple time steps;

[0008] Based on the complete hidden state sequence, the optimal segmentation point is calculated using the optimal time segmentation algorithm to divide the complete hidden state sequence into an early hidden state sequence and a late hidden state sequence.

[0009] Information bottleneck encoding is performed on the aggregated representation of the early hidden state sequence and the final hidden state of the complete hidden state sequence to generate a first latent representation and a second latent representation.

[0010] The first latent representation and the second latent representation are fused to obtain a fused representation, and the fused representation is processed by a decoder network to obtain the classification result of the time series data;

[0011] Based on the classification results and the preset real labels, the value of the composite loss function is calculated, and the trainable parameters in the recurrent neural network, the information bottleneck encoding process, and the decoder network are updated through the backpropagation algorithm.

[0012] Preferably, the step of acquiring multivariate time series data to be classified and processing the time series data through a recurrent neural network to generate a complete hidden state sequence containing multiple time steps includes:

[0013] Obtain the raw multivariate time series data, and perform normalization, standardization, or missing value imputation operations on the raw multivariate time series data;

[0014] The preprocessed sequence data is fed into the recurrent neural network, which performs processing at each time step. Receive the input vector from the sequence data and according to the input vector The hidden state at the previous time step To calculate the hidden state at the current time step To generate a document containing information from... arrive Complete hidden state sequence .

[0015] Preferably, the step of calculating the optimal segmentation point based on the complete hidden state sequence using an optimal time segmentation algorithm to divide the complete hidden state sequence into an early hidden state sequence and a late hidden state sequence includes:

[0016] Receive the complete hidden state sequence ;

[0017] Traverse all possible candidate split points The The value range is from 1 to Integers;

[0018] For each candidate split point Calculate segmented quality scores The segmented quality score The calculation is based on the average similarity within the early hidden state segments, the average similarity within the late hidden state segments, and the average similarity between the early hidden state segments and the late hidden state segments;

[0019] Compare all the segment quality scores and determine the candidate segmentation point that maximizes the score as the optimal segmentation point. ;

[0020] Based on the optimal split point The complete hidden state sequence Divided into the early hidden state sequence and the late hidden state sequence .

[0021] Preferably, the step of performing information bottleneck encoding on the aggregated representation of the early hidden state sequence and the final hidden state of the complete hidden state sequence to generate a first latent representation and a second latent representation includes:

[0022] For the early hidden state sequence Perform average pooling to generate early context vectors. ;

[0023] The early context vector The input is fed into a first encoder network, which outputs a mean vector. Sum of logarithmic variance vector , to parameterize the first latent representation posterior probability distribution ;

[0024] Using reparameterization techniques, from the posterior probability distribution Generate the first latent representation ;

[0025] The last hidden state in the complete hidden state sequence is the final hidden state. The input is fed into a second encoder network, which outputs a mean vector. Sum of logarithmic variance vector , to parameterize the second latent representation posterior probability distribution ;

[0026] Using reparameterization techniques, from the posterior probability distribution The second latent representation is generated from the middle. .

[0027] Preferably, the step of fusing the first latent representation and the second latent representation to obtain a fused representation, and processing the fused representation through a decoder network to obtain the classification result of the time series data, includes:

[0028] Receive the first potential representation and the second potential representation ;

[0029] The first potential representation With the second potential representation The features are concatenated to generate a fused representation. ;

[0030] The fusion representation The input is fed into the decoder network, which processes the fused representation and feeds its original output value into the Softmax activation function to obtain a probability distribution vector. As the classification result of the time series data, This represents the total number of categories for the tasks to be classified.

[0031] Preferably, the step of calculating the value of the composite loss function based on the classification result and the preset true label, and updating the trainable parameters in the recurrent neural network, the information bottleneck encoding process, and the decoder network through the backpropagation algorithm includes:

[0032] Calculate classification loss The classification loss Used to measure the predicted probability distribution With the preset real category label Consistency between them;

[0033] Calculate with the first latent representation Related KL divergence loss The KL divergence loss Used to measure posterior probability distribution Compared with the pre-set prior distribution The differences between them;

[0034] Calculation and the second potential representation Related KL divergence loss The KL divergence loss Used to measure the posterior probability distribution Compared with the pre-set prior distribution The differences between them;

[0035] The classification loss Weighted KL divergence loss and weighted KL divergence loss Perform weighted summation to construct a composite loss function. ;

[0036] The gradient of the composite loss function with respect to the trainable parameters is calculated using the backpropagation algorithm, and the trainable parameters are updated using the gradient descent optimizer based on the gradient.

[0037] Preferably, the segmented quality score The calculation method is as follows:

[0038] ;

[0039] in, This represents the value of the entire objective function, i.e., the current candidate split point. Segmented quality score; Indicates the current candidate split point; This represents the total length of the hidden state sequence; Represents the hidden state vector in the sequence; Representing vectors and The cosine distance between them; It is a positive scaling parameter used to adjust the degree of influence of distance on similarity score; It is a function that converts distance into a similarity score.

[0040] Preferably, the probability distribution derived from the posterior probability distribution Generate the first latent representation The method is: from the standard normal distribution Random sampling noise vector and based on The calculation yielded, where This indicates element-wise multiplication.

[0041] Preferably, the probability distribution derived from the posterior probability distribution The second latent representation is generated from the middle. The method is: from the standard normal distribution Random sampling noise vector and based on The calculation yielded, where This indicates element-wise multiplication.

[0042] Preferably, the composite loss function The calculation method is as follows:

[0043] ;

[0044] in, It is the classification loss, calculated as follows: ,in It is a unique hot encoding of the real label. It is a predicted probability; Is it related to the first potential representation? The relevant KL divergence loss is calculated as follows: ,in Is the first potential representation Dimensions and These are the mean vectors. Sum of variance vectors The One element; Is with the second potential representation The relevant KL divergence loss is calculated as follows: ,in It is the second potential representation Dimensions and The mean vector output by the second encoder network is not mentioned. Sum of variance vectors The One element; and It is a hyperparameter used to control the intensity of information compression.

[0045] This invention provides a temporal classification method for RNNs based on the bottleneck of deep variational information. It has the following beneficial effects:

[0046] 1. This invention employs a data-driven optimal time segmentation algorithm, which can adaptively identify and segment key time points in the sequence where information characteristics change. Compared with the existing technology that encodes the time series as a whole without distinction, the solution of this invention solves the problem that key historical information is diluted or submerged by recent irrelevant information due to the inability to identify internal state changes, and significantly improves the targeting of feature extraction.

[0047] 2. This invention constructs a parallel dual-channel information compression structure to process the long-term historical information and recent state information of the sequence respectively, and uses the information bottleneck principle to independently purify and compress the two information channels. Unlike the existing technology that relies on a single information path for feature fusion, this invention solves the problem of mutual interference between information at different time scales in a single path, which leads to unclear feature representation, and obtains purer and more discriminative decoupled features.

[0048] 3. This invention uses a unified composite loss function to perform end-to-end collaborative optimization of sequence segmentation, dual-channel encoding, and the final classification task. It can learn how to adjust the front-end time segmentation strategy and the information compression degree of each channel according to the final classification goal. This invention changes the limitation of the separation between feature engineering and classification model in the prior art. This invention solves the defect of insufficient model generalization ability caused by the lack of a global optimization goal, and obtains more robust and accurate classification performance in complex dynamic scenarios. Attached Figure Description

[0049] Figure 1 This is a system functional module structure diagram according to an embodiment of the present invention;

[0050] Figure 2 This is a flowchart illustrating the overall execution process of the method according to an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram illustrating the principle of the optimal time segmentation steps in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram illustrating the principle of dual-channel information encoding and fusion in an embodiment of the present invention.

[0053] Explanation of icon numbers:

[0054] 10. Data receiving and preprocessing module; 20. Sequence feature extraction module; 30. Optimal time segmentation module; 40. Dual-channel information compression module; 50. Representation fusion and decoding module; 60. Model training and optimization module. Detailed Implementation

[0055] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Figure 1 This is a functional module structure diagram of a deep learning-based time series classification system according to an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a deep learning-based time series classification system for executing the RNN time series classification method based on the deep variational information bottleneck of this invention. The system can run on general-purpose computing devices including processors, memory, and input / output interfaces, and its functions are implemented through a series of cooperating functional modules.

[0057] The system includes: a data receiving and preprocessing module 10, a sequence feature extraction module 20, an optimal time segmentation module 30, a dual-channel information compression module 40, a representation fusion and decoding module 50, and a model training and optimization module 60.

[0058] The data receiving and preprocessing module 10 is used to receive externally input multivariate time series data and perform preprocessing operations such as normalization and format conversion on the data. After processing, the data is transmitted to the sequence feature extraction module 20.

[0059] The sequence feature extraction module 20 has a recurrent neural network at its core. The sequence feature extraction module 20 receives the sequence data output by the data receiving and preprocessing module 10, performs forward propagation calculation on the sequence data, generates a complete hidden state sequence containing temporal dynamic information, and transmits the hidden state sequence to the optimal time segmentation module 30 and the dual-channel information compression module 40.

[0060] The optimal time segmentation module 30 receives the hidden state sequence generated by the sequence feature extraction module 20, runs the optimal time segmentation algorithm, calculates and determines an optimal segmentation point that can divide the hidden state sequence into early segments and late segments, and outputs the location information of the optimal segmentation point.

[0061] The dual-channel information compression module 40 includes an early information encoding submodule and a final state encoding submodule. The early information encoding submodule receives an early segment of the hidden state sequence and performs information bottleneck compression on the hidden state sequence to generate a first latent representation. The final state encoding submodule receives the final hidden state of the hidden state sequence and performs information bottleneck compression on the final hidden state to generate a second latent representation. Both latent representations are transmitted to the representation fusion and decoding module 50.

[0062] The representation fusion and decoding module 50 receives the first latent representation and the second latent representation output by the dual-channel information compression module 40, concatenates the two in the feature dimension to form a fused representation, processes the fused representation through a decoder network, calculates and outputs the final classification result.

[0063] The model training and optimization module 60 calculates the value of a composite loss function based on the difference between the classification result output by the representation fusion and decoding module 50 and the true label, and updates the trainable parameters in the sequence feature extraction module 20, the dual-channel information compression module 40, and the representation fusion and decoding module 50 through backpropagation and gradient descent algorithms.

[0064] See Figure 2 , Figure 2 This is a flowchart of an RNN time series classification method based on a deep variational information bottleneck according to an embodiment of the present invention. The present invention provides an RNN time series classification method based on a deep variational information bottleneck, comprising the following steps:

[0065] S100: Obtain multivariate time series data to be classified, and process the time series data through a recurrent neural network to generate a complete hidden state sequence containing multiple time steps.

[0066] S200, based on the complete hidden state sequence, calculates an optimal split point using the optimal time segmentation algorithm to divide the complete hidden state sequence into an early hidden state sequence and a late hidden state sequence.

[0067] S300 performs information bottleneck encoding on the aggregated representation of the early hidden state sequence and the final hidden state of the complete hidden state sequence to generate a first latent representation and a second latent representation.

[0068] S400 fuses the first latent representation and the second latent representation to obtain a fused representation, and processes the fused representation through a decoder network to obtain the classification result of the time series data.

[0069] S500 calculates a composite loss function based on the classification results and preset real labels, and updates the trainable parameters in the recurrent neural network, the information bottleneck encoding process, and the decoder network through the backpropagation algorithm.

[0070] The specific implementation methods of the above steps in the embodiments of the present invention will be described in detail below.

[0071] The specific implementation method of step S100 will be described in detail below.

[0072] S101, This step involves acquiring and processing the raw multivariate time series data. The input data is a multivariate time series. ,in This represents the total time step or length of the sequence. For one A 3D real vector, representing the time step. Observed The numerical values ​​of each variable.

[0073] To adapt to the input requirements of neural network models and improve the stability and efficiency of model training, time series data undergoes one or more preprocessing operations before being fed into the network. Preprocessing operations can specifically include data normalization, such as linearly mapping the value of each variable to the interval [0,1] or [-1,1]; or data standardization, which transforms the data of each variable to make its mean 0 and standard deviation 1. For missing values ​​in the data, preprocessing operations can also include missing value imputation, which can be implemented by using valid observations from the previous time step, using the mean, or employing more complex interpolation methods.

[0074] S102, after preprocessing, the sequence data is fed into a recurrent neural network to encode temporal dependencies in the sequence and extract dynamic features. A recurrent neural network is a neural network structure capable of processing sequence data, whose internal recurrent connections allow information to be passed between time steps. In specific embodiments of this invention, the recurrent neural network can be a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU), or other variant structures well known to those skilled in the art. These structures effectively mitigate the vanishing or exploding gradient problem by introducing gating mechanisms, thereby better capturing long-term dependencies.

[0075] This recurrent neural network at each time step (from arrive ) Receive the preprocessed input vector And based on the current input and the hidden state of the previous time step To calculate the hidden state at the current time step In practical applications, the depth of a recurrent neural network (i.e., the number of stacked layers) can range from one to several layers, and the hidden state dimension... The value is typically set between 64 and 512 to balance the model's expressive power and computational complexity. This update process can be represented by the following general formula:

[0076] ;

[0077] in, In time step A generated dimensional hidden state vector, The dimension of the hidden state; In time step Input 3D data vector; It was a step in the previous time of The hidden state vector, for the initial state It is usually initialized as a zero vector; It is a dimension of The input weight matrix is ​​used to perform a linear transformation on the current input; It is a dimension of The cyclic weight matrix is ​​used to perform a linear transformation on the hidden state of the previous time step; It is A bias vector of dimension; It is a non-linear activation function, such as the Sigmoid, hyperbolic tangent (tanh), or modified linear unit (ReLU) function.

[0078] After completing the entire input sequence from arrive After the forward propagation computation, this step finally generates a complete hidden state sequence. The complete hidden state sequence Each hidden state vector in All include the sequence from its start time to the current time. The accumulated information of the entire sequence This constitutes a high-dimensional representation of the dynamic characteristics of the original time series data, and serves as the input for subsequent information analysis and compression steps.

[0079] Reference Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of the optimal time segmentation step in this embodiment of the invention; the specific implementation of step S200 will be described in detail below. The core objective of this step is to optimize the complete hidden state sequence generated by step S100. Perform an automatic, unsupervised segmentation to identify key time points in the sequence where significant changes in information characteristics occur.

[0080] S201, The calculation of this step begins with receiving the complete hidden state sequence output by the sequence feature extraction module. .

[0081] S202, to determine the optimal splitting position, this step will traverse all possible candidate splitting points. A candidate splitting point... The value range is from 1 to An integer. For each candidate split point. Hidden state sequence It is temporarily divided into two subsets: an early hidden state fragment containing from arrive All hidden states; and a late hidden state fragment containing from arrive All hidden states.

[0082] S203, for each candidate segmentation point This step will calculate a segmentation quality score. This score quantifies the merits of the current segmentation method. The core idea is that an ideal segmentation should make the hidden states within the two resulting segments as similar as possible (i.e., high cohesion), while the hidden states between the two different segments should be as dissimilar as possible (i.e., low coupling). This segmentation quality score is calculated using an optimal time segmentation (OTS) objective function.

[0083] Before calculating the objective function, it is necessary to first define any two hidden state vectors. and The similarity between them. In this embodiment, the similarity is based on cosine distance. It is calculated. The cosine distance is calculated as follows:

[0084] ;

[0085] in, Represents the dot product of two vectors. and Let L2 and L2 represent the two vectors respectively. The cosine distance ranges from [0, 2]. Then, an exponential function is used to convert the distance into a similarity score; the smaller the distance, the higher the similarity score.

[0086] Optimal Time Segmentation (OTS) Objective Function Defined as:

[0087] ;

[0088] in, This represents the value of the entire objective function, i.e., the current candidate split point. Segmented quality score; Indicates the current candidate split point; This represents the total length of the hidden state sequence; Represents the hidden state vector in the sequence; Representing vectors and The cosine distance between them; It is a positive scaling parameter used to adjust the degree of influence of distance on similarity score; It is a function that converts distance into a similarity score.

[0089] Furthermore, the meanings of the three main terms in the formula are as follows:

[0090] First item: Calculated and normalized the early fragments (from 1 to...). The average similarity within the product represents the cohesion.

[0091] Second item: Calculated and normalized the late fragments (from...) arrive The average similarity within the same region also represents cohesion.

[0092] The third item calculates and normalizes the average similarity between early and late segments, representing the coupling degree.

[0093] The first and second terms calculate and normalize the average similarity within early segments and late segments, respectively, representing cohesion.

[0094] The third term calculates and normalizes the average similarity between early and late segments, representing the coupling degree.

[0095] The entire objective function The value is the current candidate split point. The segmented quality score.

[0096] S204, for all candidate split points (from 1 to Each segment quality score was calculated. Next, this step will compare all scores and select the candidate split point that results in the highest score as the final optimal split point. .

[0097] ;

[0098] The optimal split point This marks the location in the hidden state sequence where a shift in representation content is most likely to occur. Based on this... The complete hidden state sequence Formally classified as an early hidden state sequence and late hidden state sequences Among them, the early hidden state sequence With the final hidden state This will be used as input for the subsequent step S300 to extract long-term pattern information and recent context information, respectively.

[0099] See Figure 4 , Figure 4 This is a schematic diagram illustrating the principle of dual-channel information encoding and fusion in an embodiment of the present invention; the specific implementation of step S300 will be described in detail below. This step aims to perform parallel and differentiated compression on sequence information from different time periods to extract long-term pattern information and recent context information that contribute to the classification task. This process is based on the principle of Deep Variational Information Bottleneck and is implemented through two independent encoding channels.

[0100] S301, This step involves processing the early hidden state sequence determined in step S204. Information aggregation is performed. To integrate the sequence information into a fixed-dimensional vector representation suitable for the input of the subsequent encoder network, this step employs average pooling. Specifically, all hidden state vectors in the early hidden state sequence are arithmetically averaged element-wise to generate a single early context vector. :

[0101] ;

[0102] The early context vector It integrates the sequence from the beginning to the optimal split point. All dynamic information, and used as input for early information channels.

[0103] S302, the early context vector generated in step S301 is... The input is fed into a first encoder network. The first encoder network is a parameterized neural network, such as a multilayer perceptron (MLP), whose function is to map the input vector to parameters of a predefined latent probability distribution. This MLP typically contains 1 to 3 hidden layers, with the number of neurons in each layer depending on the dimension of the hidden states. and latent representation dimension Adjustments can be made, for example, it can be set to arrive The values ​​are between [a certain range], and nonlinear activation functions such as ReLU or LeakyReLU can be used. In this embodiment, the latent distribution is defined as a Gaussian distribution. Therefore, the first encoder network receives [the data]. Then, output two vectors: a mean vector. and a log-variance vector These two vectors together parameterize a diagonal Gaussian distribution. This distribution defines the first latent representation. posterior probability .

[0104] S303 is to generate a specific first latent representation from the posterior probability distribution defined in step S302, which can be used for subsequent computation and whose gradient is propagable during training. This step employs a reparameterization technique. Specifically, it first uses a standard normal distribution... Randomly sample a noise vector with the same dimension as the latent representation. Subsequently, the first latent representation is calculated using the following formula. :

[0105] ;

[0106] in, This indicates element-wise multiplication. It is generated in this way. It can be regarded as derived from the posterior distribution This is a sampling performed under the constraint of information bottleneck, which is the result of compressing early sequence information.

[0107] S304, in parallel with the aforementioned early information channel, this step processes the complete hidden state sequence generated by step S102. The last hidden state, i.e., the final hidden state. The final hidden state. The input is fed into a separate second encoder network. The second encoder network is structurally similar to the first encoder network; the number of layers and neurons per layer in its multilayer perceptron can also follow the same principles as the first encoder network, but its parameters are trained independently. It has its own independent set of trainable parameters to ensure that the information processing of the two channels does not interfere with each other. This second encoder network also outputs a mean vector. and a log-variance vector , used to parameterize the second latent representation posterior probability distribution .

[0108] S305, also employing the reparameterization technique, generates the second latent representation from the posterior probability distribution defined in step S304. The calculation method is the same as step S303:

[0109] ;

[0110] in, It is another instance from the standard normal distribution Independent random sampling within. Generated. It is the result of refining and compressing the recent context information contained in the final hidden state under the constraint of information bottleneck.

[0111] After steps S301 to S305, the present invention ultimately obtains two complementary potential representations: It contains long-term pattern information extracted from early stages of the sequence. It contains recent state information extracted from the end stage of the sequence. These two potential representations will be fed together into the subsequent step S400 for fusion and decoding.

[0112] The specific implementation of step S400 will be described in detail below. This step receives two independent latent representations generated by the dual-channel information bottleneck encoding process in step S300, and merges them to form a unified and more comprehensive feature representation. Finally, it is mapped to the final classification result through a decoder network.

[0113] S401, This step receives the first latent representation generated by step S303. and the second latent representation generated by step S305 To integrate these two complementary representations, which respectively carry long-term historical pattern information and recent contextual information, this step performs a feature fusion operation on them.

[0114] In one specific embodiment, the feature fusion operation is concatenation. Specifically, the first latent representation vector is... (its dimensions are) ) and the second latent representation vector (its dimensions are) Perform a first-to-last join along the feature dimension to generate a dimension of fusion representation This fusion indicates Within a unified feature space, compressed representations of information from both the early and late stages of the time series are included and used as input for subsequent classification and decoding.

[0115] S402, the fused representation generated in step S401 is... The input is fed into a decoder network. The function of the decoder network is to map the high-dimensional, abstract, fused representation to a low-dimensional class probability space corresponding to the specific classification task.

[0116] In one specific implementation, the decoder network is a multi-layer perceptron (MLP). This MLP consists of at least one fully connected hidden layer and one output layer. Typically, the decoder network may contain one to two hidden layers, and the number of neurons in each layer can vary depending on the dimension of the fused representation. Gradually decrease to the total number of categories The intermediate hidden layers can use activation functions such as ReLU and LeakyReLU to ensure that the model has sufficient non-linear mapping capabilities. Each hidden layer contains multiple neurons and is followed by a non-linear activation function, such as the Modified Linear Unit (ReLU) function.

[0117] The output layer of the decoder network has a number of neurons that is equal to the total number of categories in the task to be classified. The original output values ​​(logits) produced by this output layer are equal. These logits are then fed into a Softmax activation function. The Softmax activation function transforms the original values ​​of the output layer into a... dimensional probability distribution vector Each element This indicates that the input time series belongs to the first... The probability of a class. The Softmax function is calculated as follows:

[0118] ;

[0119] in, Is the input time series belonging to the first... The predicted probability of a class; It is the corresponding number in the output layer of the decoder network. The raw output value of the neuron of the class; It is the total number of categories; .

[0120] Ultimately, the decoder network outputs a probability distribution vector. This is the classification result obtained in this step. When making the final decision, the category with the highest probability value is usually selected as the predicted label for the input time series.

[0121] The specific implementation of step S500 will be described in detail below. This step defines an objective function to guide the end-to-end training of the entire model, and iteratively updates all trainable parameters in the model using this objective function so that the model can accurately perform the temporal classification task, while learning a compressed latent representation that is beneficial to the classification task.

[0122] S501, This step calculates the classification loss of the model. This classification loss is used to measure the predicted probability distribution output by step S402. With preset real category labels The consistency between them. In one specific implementation, the classification loss uses the cross-entropy loss function. For a single time series sample, its cross-entropy loss... The calculation is as follows:

[0123] ;

[0124] in, It is the total number of categories; It is A one-hot encoded vector of dimension 1, representing the true label of the sample, if the true class of the sample is 1. ,but , and all other elements All are 0; The model predicts that the sample belongs to the category output by the decoder network in step S402. The probability of.

[0125] The smaller the value of this loss term, the closer the model's prediction is to the true label.

[0126] S502, This step calculates the first latent representation. The relevant KL divergence loss. This loss term is crucial for achieving the information bottleneck constraint in the early information channel. Its role is as a regularization term, prompting the first encoder network to learn the posterior probability distribution. It approaches a pre-defined prior distribution that contains no information. .

[0127] In this embodiment, the prior distribution It is set as a standard normal distribution Posterior probability distribution The average value output by the first encoder network in step S302 and variance The parameterized KL divergence between the two. The calculation is as follows:

[0128] ;

[0129] in, It is the first potential representation Dimensions and These are the mean vectors. Sum of variance vectors The Each element. The KL divergence loss is multiplied by an adjustable hyperparameter. It is used to control the compression intensity of information in early information channels.

[0130] S503, in a manner independent of step S502, this step calculates the second potential representation. The relevant KL divergence loss. This loss term is used to impose an information bottleneck constraint on the final state information path. It calculates the posterior probability distribution output by the second encoder network in step S304. With a prior distribution (Also set as standard normal distribution) KL divergence between ) The calculation formula is similar to step S502:

[0131] ;

[0132] in, It is the second potential representation Dimensions and The mean vector output by the second encoder network is not mentioned. Sum of variance vectors The Each element. The KL divergence loss is multiplied by another independent hyperparameter. It is used to independently control the information compression intensity of the final state information channel.

[0133] S504, this step involves weighted summation of the three aforementioned loss terms to construct a unified composite loss function for overall model optimization. The composite loss function, calculated for a single training sample after one forward propagation of the model, is as follows:

[0134] ;

[0135] This composite loss function achieves a single objective while also considering the accuracy of the classification task (through...). (item) and the degree of information compression of the two independent information channels (through two weighted items) item).

[0136] S505, this step updates the model's parameters. During training, for a batch of training samples, the forward propagation process from S100 to S400 is first executed, and then the average composite loss for that batch of samples is calculated according to S501 to S504. Subsequently, the gradient of this composite loss function with respect to all trainable parameters in the model is calculated using the backpropagation algorithm. These trainable parameters include: the weights and biases of the recurrent neural network in the sequence feature extraction module, the weights and biases of the two encoder networks in the dual-channel information compression module, and the weights and biases of the decoder network in the fusion and decoding module.

[0137] Finally, a gradient descent optimizer, such as the Adam optimizer or RMSprop optimizer, is used to update all trainable parameters once based on the calculated gradient. During training, an initial learning rate (e.g., 0.001 to 0.01) is typically set, and learning rate decay strategies (such as exponential decay, cosine annealing, or reducing the learning rate when validation set performance stagnates) can be employed to improve training stability and convergence. Furthermore, batch normalization or layer normalization techniques can be used to stabilize the training process, and dropout layers can be added between neural network layers to mitigate overfitting. Training is typically performed in mini-batch mode, with batch size set to values ​​such as 16, 32, 64, or 128. The entire process of forward propagation, loss calculation, backpropagation, and parameter updates is repeated until the model's performance on the reserved validation set reaches the convergence criterion, at which point the model training process ends.

[0138] To further illustrate the collaborative working process of the technical solution of this invention, a specific working scenario example will be used below.

[0139] This embodiment applies to predictive maintenance of large wind turbine generator sets. The specific task is to automatically diagnose the generator set's operational status within a short future time window based on time-series data collected from multiple sensors deployed on the generator set. In this scenario, preset classification labels can include three states: normal operation, gearbox wear warning, and blade icing anomaly. Gearbox wear warning is a slow-changing fault mode that relies on the accumulation of small changes in long-term operating data, while blade icing anomaly is typically a rapid change caused by recent sudden low temperatures and high humidity.

[0140] In a specific implementation, the data sample used for this task is a multivariate time series, which consists of data points collected every 10 minutes over a continuous 40-hour period by vibration sensors, temperature sensors, speed sensors, and power output sensors. Therefore, the total length of the sequence for each sample is... The specific configuration parameters of the model can be set as follows: the recurrent neural network in the sequence feature extraction module adopts an LSTM network with a hidden state dimension of 128; the two encoder networks in the dual-channel information compression module and the decoder network in the representation fusion and decoding module all adopt a multilayer perceptron containing two fully connected layers; the first latent representation With the second potential representation The dimensions are all set to 32; the hyperparameters in the composite loss function are set empirically to... and .

[0141] When a 240-bit time series sample from a wind turbine is input into the system, the system performs a classification task according to the following process:

[0142] First, step S100 is executed. The 240-byte sequence data is normalized and then fed into the LSTM network. After 240 time steps of forward propagation, a complete hidden state sequence with a dimension of 240x128 is generated. .

[0143] Next, step S200 is executed, and the optimal time segmentation module receives the hidden state sequence. The optimal time segmentation algorithm is then applied to it. Assume the algorithm calculates the optimal segmentation point as follows: This indicates that the most significant changes in the generator unit's operating state characterization occurred around this time point (i.e., the 30th hour). Based on this, the hidden state sequence is divided into an early hidden state sequence containing the first 180 hidden states. .

[0144] Subsequently, step S300 is executed, and the system processes the two information channels in parallel. In the earlier information channel, [the system processes the information through the two channels]. Average pooling yields a 128-dimensional early context vector. It is then fed into the first encoder network to generate a first potential representation of the long-term cumulative wear trend. In the final state information channel, the last hidden state of the complete hidden state sequence is... The data is fed into the second encoder network to generate a second latent representation representing recent sudden operating conditions. .

[0145] Then, step S400 is executed to convert the 32-dimensional... And 32-dimensional Pieced together to form a 64-dimensional fused representation The data is then fed into the decoder network. The decoder network ultimately outputs a three-dimensional probability vector, corresponding to the probabilities of the three states: "normal operation," "gearbox wear warning," and "blade icing anomaly."

[0146] During the model training phase, step S500 is executed, where the system calculates the composite loss based on the output probabilities and the true labels of the samples. For example, if the true label of a sample is "gearbox wear warning," but the model predicts a low probability for it, a large cross-entropy loss will occur. Simultaneously, the two KL divergence loss terms constrain the two encoder networks, ensuring they extract only the information most relevant to the classification. Through backpropagation and parameter updates, the model is guided to learn how to effectively identify gradually changing wear features from early sequences while simultaneously accurately capturing rapidly changing icing features from the final state.

[0147] Through the process described in this embodiment, the technical solution provided by the present invention can effectively distinguish between faults that depend on long-term historical patterns (such as gearbox wear) and abnormal states determined by recent key information (such as blade icing). By explicitly separating, independently compressing and finally fusing the two information streams, it provides technical support for achieving more accurate and robust industrial equipment condition monitoring and predictive maintenance.

[0148] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A temporal classification method for RNNs based on the bottleneck of deep variational information, characterized in that, Includes the following steps: Acquire multivariate time series data to be classified, and process the time series data through a recurrent neural network to generate a complete hidden state sequence containing multiple time steps; Based on the complete hidden state sequence, the optimal segmentation point is calculated using the optimal time segmentation algorithm to divide the complete hidden state sequence into an early hidden state sequence and a late hidden state sequence. Information bottleneck encoding is performed on the aggregated representation of the early hidden state sequence and the final hidden state of the complete hidden state sequence to generate a first latent representation and a second latent representation. The first latent representation and the second latent representation are fused to obtain a fused representation, and the fused representation is processed by a decoder network to obtain the classification result of the time series data; Based on the classification results and the preset real labels, the value of the composite loss function is calculated, and the trainable parameters in the recurrent neural network, the information bottleneck encoding process, and the decoder network are updated through the backpropagation algorithm.

2. The RNN time series classification method based on the bottleneck of deep variational information as described in claim 1, characterized in that, The process of acquiring multivariate time series data to be classified and processing the time series data through a recurrent neural network to generate a complete hidden state sequence containing multiple time steps includes: Obtain the raw multivariate time series data, and perform normalization, standardization, or missing value imputation operations on the raw multivariate time series data; The preprocessed sequence data is fed into the recurrent neural network, which performs processing at each time step. Receive the input vector from the sequence data and according to the input vector The hidden state at the previous time step To calculate the hidden state at the current time step To generate a document containing information from... arrive Complete hidden state sequence .

3. The RNN time series classification method based on the deep variational information bottleneck as described in claim 1, characterized in that, The step of dividing the complete hidden state sequence into an early hidden state sequence and a late hidden state sequence by calculating the optimal segmentation point using an optimal time segmentation algorithm, based on the complete hidden state sequence, includes: Receive the complete hidden state sequence ; Traverse all possible candidate split points The The value range is from 1 to Integers; For each candidate split point Calculate segmented quality scores The segmented quality score The calculation is based on the average similarity within the early hidden state segments, the average similarity within the late hidden state segments, and the average similarity between the early hidden state segments and the late hidden state segments; Compare all the segment quality scores and determine the candidate segmentation point that maximizes the score as the optimal segmentation point. ; Based on the optimal split point The complete hidden state sequence Divided into the early hidden state sequence and the late hidden state sequence .

4. The RNN time series classification method based on the deep variational information bottleneck as described in claim 1, characterized in that, The step of performing information bottleneck encoding on the aggregated representation of the early hidden state sequence and the final hidden state of the complete hidden state sequence to generate a first latent representation and a second latent representation includes: For the early hidden state sequence Perform average pooling to generate early context vectors. ; The early context vector The input is fed into a first encoder network, which outputs a mean vector. Sum of logarithmic variance vector , to parameterize the first latent representation posterior probability distribution ; Using reparameterization techniques, from the posterior probability distribution Generate the first latent representation ; The last hidden state in the complete hidden state sequence is the final hidden state. The input is fed into a second encoder network, which outputs a mean vector. Sum of logarithmic variance vector , to parameterize the second latent representation posterior probability distribution ; Using reparameterization techniques, from the posterior probability distribution The second latent representation is generated from the middle. .

5. The RNN time series classification method based on the deep variational information bottleneck as described in claim 1, characterized in that, The step of fusing the first latent representation and the second latent representation to obtain a fused representation, and then processing the fused representation through a decoder network to obtain the classification result of the time series data, includes: Receive the first potential representation and the second potential representation ; The first potential representation With the second potential representation The features are concatenated to generate a fused representation. ; The fusion representation The input is fed into the decoder network, which processes the fused representation and feeds its original output value into the Softmax activation function to obtain a probability distribution vector. As the classification result of the time series data, where This represents the total number of categories for the tasks to be classified.

6. The RNN time series classification method based on the deep variational information bottleneck as described in claim 1, characterized in that, The step of calculating the value of the composite loss function based on the classification result and the preset real labels, and updating the trainable parameters in the recurrent neural network, the information bottleneck encoding process, and the decoder network through the backpropagation algorithm includes: Calculate classification loss The classification loss Used to measure the predicted probability distribution With the preset real category label Consistency between them; Calculate with the first latent representation Related KL divergence loss The KL divergence loss Used to measure posterior probability distribution Compared with the pre-set prior distribution The differences between them; Calculation and the second potential representation Related KL divergence loss The KL divergence loss Used to measure the posterior probability distribution Compared with the pre-set prior distribution The differences between them; The classification loss Weighted KL divergence loss and weighted KL divergence loss Perform weighted summation to construct a composite loss function. ; The gradient of the composite loss function with respect to the trainable parameters is calculated using the backpropagation algorithm, and the trainable parameters are updated using the gradient descent optimizer based on the gradient.

7. The RNN time series classification method based on the deep variational information bottleneck as described in claim 3, characterized in that, The segmented quality score The calculation method is as follows: ； in, This represents the value of the entire objective function, i.e., the current candidate split point. Segmented quality score; Indicates the current candidate split point; This represents the total length of the hidden state sequence; Represents the hidden state vector in the sequence; Representing vectors and The cosine distance between them; It is a positive scaling parameter used to adjust the degree of influence of distance on similarity score; It is a function that converts distance into a similarity score.

8. The RNN time series classification method based on the deep variational information bottleneck as described in claim 4, characterized in that, The from the posterior probability distribution Generate the first latent representation The method is: from the standard normal distribution Random sampling noise vector and based on The calculation yielded, where This indicates element-wise multiplication.

9. The RNN time series classification method based on the deep variational information bottleneck as described in claim 4, characterized in that, The from the posterior probability distribution The second latent representation is generated from the middle. The method is: from the standard normal distribution Random sampling noise vector and based on The calculation yielded, where This indicates element-wise multiplication.

10. The RNN time-series classification method based on the deep variational information bottleneck according to claim 6, characterized in that, The composite loss function The calculation method is as follows: ； in, It is the classification loss, calculated as follows: ,in It is a unique hot encoding of the real label. It is a predicted probability; Is it related to the first potential representation? The relevant KL divergence loss is calculated as follows: ,in Is the first potential representation Dimensions and These are the mean vectors. Sum of variance vectors The One element; Is with the second potential representation The relevant KL divergence loss is calculated as follows: ,in It is the second potential representation Dimensions and The mean vector output by the second encoder network is not mentioned. Sum of variance vectors The One element; and It is a hyperparameter used to control the intensity of information compression.

Images