An industrial time series data enhancement method and device based on collaborative constraints

Abstract

The present application relates to data processing technology, disclose a kind of industrial time series data enhancement method based on collaborative constraint, comprising: obtaining industrial time series data, and using data preprocessing module to execute data preprocessing operation to industrial time series data, obtain standard industrial time series data;Standard industrial time series data is input into improved variational autoencoder model to generate data initial sample;Data initial sample is input into conditional generative adversarial network and is subjected to data optimization operation to obtain optimized data sample;Optimized data sample is subjected to industrial time series correlation constraint using time series feature constraint model, and obtain industrial constraint data sample;Industrial constraint data sample is subjected to sample evaluation and sample screening to obtain target generated data sample.The present application also proposes a kind of industrial time series data enhancement device based on collaborative constraint, equipment and medium.The present application can ensure the data quality of generated data and improve the sample amount of generated data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an industrial time series data enhancement method and device based on collaborative constraints. BACKGROUND

[0002] In the Industrial Internet of Things (IIoT) scenario, time series data (such as sensor readings, device state monitoring data, etc.) is a key data type. Such data usually has the characteristics of data scarcity, skewness, and the presence of signal noise and outliers, where data scarcity refers to the difficulty of obtaining key fault or extreme working condition data in some industrial devices or scenarios; skewness refers to the fact that time series data may be unbalanced, with a large amount of data concentrated in the "normal state" and difficult to cover "abnormal states" or failure scenarios. The presence of non-stable factors such as signal noise and outliers can affect accurate data analysis. In order to address the above problems, data enhancement has become an effective solution. In traditional industrial systems, data enhancement methods are mainly based on rules or simple transformations, such as stretching, shifting, etc. However, these methods are difficult to effectively capture complex time series patterns.

[0003] In recent years, deep learning-based generative models (such as conditional generative adversarial networks GAN and variational autoencoders VAE) have performed well in image generation, natural language processing, etc. However, in the field of industrial time series data enhancement, the application of these generative models is still limited, mainly facing the following three difficulties. First, the generation of time series data needs to consider the time series structure and long-term dependence characteristics. Second, the industrial field has high requirements for data quality, and the generated data needs to be highly credible and close to real data. Finally, when generating rare or abnormal data, deep generative models are easily limited by the distribution of training data. The above traditional technical methods will result in low quality of generated data and insufficient sample size of generated data. SUMMARY

[0004] The present application provides an industrial time series data enhancement method and device based on collaborative constraints, which can ensure the data quality of generated data and improve the sample size of generated data.

[0005] Obtain industrial time series data, and perform data preprocessing operations on the industrial time series data using a data preprocessing module to obtain standard industrial time series data;

[0006] Input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples;

[0007] Input the data initial samples into a conditional generative adversarial network for data optimization operations to obtain optimized data samples;

[0008] The time sequence correlation constraint model is used to constrain the optimization data samples to obtain industrial constraint data samples.

[0009] The sample evaluation is performed on the industrial constraint data samples, and the sample screening is performed to obtain target generation data samples.

[0010] Optionally, the data preprocessing module is used to perform data preprocessing operations on the industrial time sequence data to obtain standard industrial time sequence data, including:

[0011] The statistical analysis method is used to detect abnormal point data in the industrial time sequence data, and the abnormal point data is removed to obtain normal industrial time sequence data.

[0012] The normal industrial time sequence data is subjected to noise removal processing by using a smoothing filter to obtain denoised industrial time sequence data.

[0013] The dynamic time warping algorithm is used to align the time sequence of the normalized and standardized denoised industrial time sequence data to obtain normalized industrial time sequence data.

[0014] The missing values in the normalized industrial time sequence data are subjected to interpolation processing to obtain standard industrial time sequence data.

[0015] Optionally, the statistical analysis method is used to detect abnormal point data in the industrial time sequence data, and the abnormal point data is removed, including:

[0016] The mean and standard deviation of the industrial time sequence data are calculated.

[0017] The mean and standard deviation are used to construct an abnormal point range standard, and the abnormal points in the industrial time sequence data are marked and removed by using the abnormal point range standard.

[0018] Optionally, the standard industrial time sequence data is input into the improved variational autoencoder model to generate data initial samples, including:

[0019] The standard industrial time sequence data is input into the encoder of the improved variational autoencoder model to output latent distribution parameters.

[0020] The latent distribution parameters are used to construct a latent distribution, and latent variables are sampled from the latent distribution.

[0021] The loss function of the improved variational autoencoder model is constructed according to the latent variables, the weight value of the divergence expression part in the loss function is dynamically adjusted by using the hyperparameter, and the latent distribution is approximated to the standard normal prior by using the weight value of the divergence expression part in the loss function.

[0022] The latent sampling points are randomly sampled from the standard normal prior, and the latent sampling points are input into the decoder of the improved variational autoencoder model to generate data initial samples.

[0023] Optionally, the data optimization operation of inputting the data initial sample into the conditional generative adversarial network comprises:

[0024] The input layer of the generator in the conditional generative adversarial network receives the data initial sample and the condition information, and converts the condition information into a condition vector;

[0025] A random noise distribution is obtained, and the random noise distribution, the condition vector and the data initial sample are used to generate a noisy data initial sample;

[0026] The generator and the discriminator in the conditional generative adversarial network are optimized by using a perception loss function, and the optimized generator and the discriminator are used for adversarial learning to obtain an optimized data sample.

[0027] Optionally, the industrial time sequence correlation constraint on the optimized data sample by using the time sequence feature constraint model comprises:

[0028] The time sequence feature constraint model comprises: a change rate smoothing constraint of adjacent time points constructed by using a regularization algorithm, a time dependence constraint constructed by using a pre-constructed time memory network, a change range constraint and a change rate constraint constructed by using application scenario knowledge, and a multivariate correlation constraint constructed by calculating a correlation matrix in the optimized data sample and using the correlation matrix;

[0029] The change rate smoothing constraint, the time dependence constraint, the change range constraint, the change rate constraint and the multivariate correlation constraint are fused to obtain a total time sequence feature constraint;

[0030] The industrial time sequence correlation constraint on the optimized data sample is performed by using the total time sequence feature constraint to obtain an industrial constraint data sample.

[0031] Optionally, the sample evaluation on the industrial constraint data sample comprises:

[0032] A first evaluation index is calculated by using a pre-constructed statistical distribution matching formula, a second evaluation index is calculated by using a pre-constructed constraint compliance rate formula, and a third evaluation index is calculated by using a pre-constructed continuous consistency formula;

[0033] The first evaluation index, the second evaluation index and the third evaluation index are fused to obtain a comprehensive evaluation index;

[0034] A comprehensive evaluation score is calculated by using the comprehensive evaluation index, and the sample evaluation on the industrial constraint data sample is performed according to the comprehensive evaluation score to obtain a sample evaluation result.

[0035] In order to solve the above problems, the application further provides an industrial time sequence data enhancement device based on collaborative constraint, which comprises:

[0036] The data generation module is configured to obtain industrial time series data, perform data preprocessing on the industrial time series data by using a data preprocessing module to obtain standard industrial time series data, and input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples.

[0037] The data screening module is configured to input the data initial samples into a conditional generative adversarial network to perform data optimization and obtain optimized data samples, constrain the optimized data samples by using a time series feature constraint model to obtain industrial constraint data samples, perform sample evaluation on the industrial constraint data samples, and perform sample screening to obtain target generated data samples.

[0038] To solve the above problems, the present application further provides an electronic device, which comprises:

[0039] at least one processor; and

[0040] a memory in communication connection with the at least one processor; wherein

[0041] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the above-mentioned collaborative constraint-based industrial time series data enhancement method.

[0042] To solve the above problems, the present application further provides a computer readable storage medium, which stores at least one computer program, and the at least one computer program is executed by a processor in an electronic device to implement the above-mentioned collaborative constraint-based industrial time series data enhancement method.

[0043] The present application uses a data preprocessing module to perform data preprocessing on industrial time series data, which can remove abnormal data and missing data in the data, improve data quality, and provide high-quality basic data for subsequent data generation. In addition, the standard industrial time series data is input into an improved variational autoencoder model to generate data initial samples, and the use of the improved variational autoencoder model can ensure the reconstruction quality and data diversity. In addition, the data initial samples are input into a conditional generative adversarial network to perform data optimization and obtain optimized data samples, and the introduction of conditional information in the generative adversarial network can improve the authenticity of data generation. Furthermore, the optimized data samples are constrained by using a time series feature constraint model to perform industrial time series correlation constraint, which can strengthen the time series structure and long-term dependence characteristics, ensure that the generated data conforms to the industrial time series logic, and finally, the industrial constraint data samples are evaluated and screened, which can screen high-quality samples and balance data distribution, ensure the data quality of the generated data, and increase the sample quantity of the generated data. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 A flowchart of a collaborative constraint-based industrial time series data enhancement method provided by an embodiment of the present application is shown in FIG. 1.

[0045] Figure 2 A function block diagram of a collaborative constraint-based industrial time series data enhancement device provided by an embodiment of the present application is shown in FIG. 2.

[0046] Figure 3 A structural diagram of an electronic device implementing the collaborative constraint-based industrial time series data enhancement method provided by an embodiment of the present application is shown in FIG. 3.

[0047] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0048] It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0049] An embodiment of the present application provides a collaborative constraint-based industrial time series data enhancement method. The execution subject of the collaborative constraint-based industrial time series data enhancement method includes but is not limited to at least one of electronic devices such as a server and a terminal that can be configured to execute the method provided by the present application. In other words, the collaborative constraint-based industrial time series data enhancement method can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be a standalone server, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), and big data and artificial intelligence platforms, etc. basic cloud computing services.

[0050] Referring to Figure 1 A flowchart of a collaborative constraint-based industrial time series data enhancement method provided by an embodiment of the present application is shown in FIG. 1. In this embodiment, the collaborative constraint-based industrial time series data enhancement method includes:

[0051] S1, industrial time series data is obtained, and a data preprocessing module is used to perform a data preprocessing operation on the industrial time series data to obtain standard industrial time series data.

[0052] In the embodiment of the application, the industrial time series data is a group of time series data in an industrial time series data set, wherein the industrial time series data set refers to a set of continuous observation values recorded in time sequence by a temperature sensor, a humidity sensor or other industrial sensors in an industrial Internet of Things (IIoT) scene.

[0053] As an embodiment of the application, a data preprocessing module is used to perform a data preprocessing operation on the industrial time series data to obtain standard industrial time series data, including:

[0054] An abnormal point data in the industrial time series data is detected by using a statistical analysis method, and the abnormal point data is removed to obtain normal industrial time series data;

[0055] The normal industrial time series data is subjected to noise removal processing by using a smoothing filter to obtain denoised industrial time series data;

[0056] The normalized and standardized denoised industrial time series data is subjected to time series alignment by using a dynamic time warping algorithm to obtain normalized industrial time series data;

[0057] The missing values in the normalized industrial time series data are subjected to interpolation processing to obtain standard industrial time series data.

[0058] Exemplarily, the data preprocessing module is used to perform a data preprocessing operation on the industrial time series data to obtain standard industrial time series data, which can be realized by using the following implementation steps:

[0059] Input: original industrial time series data set, denoted as X = {x1, x2, …, x N}, wherein is each group of time series data (length T), and i is the serial number of the i-th group of time series data.

[0060] Output: preprocessed time series data X' = {x'1, x'2, …, x' N}, which is free of noise, eliminates abnormal points, and is normalized and completed, meeting the input requirements of a subsequent model.

[0061] Firstly, data cleaning, i.e. abnormal point detection and removal, is completed, and the core is to remove abnormal data (such as invalid data generated during equipment failure) in the industrial time series data. Statistical analysis or simple rule detection is mainly used. Firstly, the mean μ i and the standard deviation σ i of each time series data are calculated:

[0062]

[0063] Then, the abnormal point is defined as a data point beyond a certain range, and if (k = 3 in general), it is marked as an abnormal point.

[0064] Second, complete data denoising, i.e. signal smoothing and filtering. The collected industrial time series data often contains high-frequency noise, which can be smoothed by the following method. First, use the smoothing filter. The moving average defines the window size as w, and the sliding average formula is as follows:

[0065]

[0066] where j is the jth window, x i t is the sliding average value of the ith group of time series data with length t, and the sliding average effectively smooths small fluctuations. Then, signal decomposition is performed to complete wavelet denoising. Discrete wavelet transform (DWT) can be used to decompose the signal, remove high-frequency noise, and retain low-frequency and medium-frequency parts;

[0067] x i = L i + H i

[0068] where L i is the low-frequency component, and H i is the high-frequency component. L i can be selected to retain, and H i can be discarded or reduced in weight.

[0069] Third, complete data normalization and standardization. That is, scaling the data to a unified range helps improve the convergence speed of the deep learning model. There are two methods, minimum-maximum normalization and standardization. Among them, the minimum-maximum normalization is suitable for data whose input features are in a known range.

[0070]

[0071] where max(x i ) is the maximum value of each group of time series data, min(x i ) is the minimum value of each group of time series data, is the ith group of time series data with length t, and standardization is often used for time series data to remove the mean and scale by the standard deviation:

[0072]

[0073] Fourth, use dynamic time warping (DTW) to achieve time alignment. Since industrial equipment may cause time series data to be misaligned in time during operation, time alignment is a very important step. The dynamic time warping method is defined as given two time series data x m = {x 1 ​x 2 ,…,x T} and x n ={y 1 ,y 2 ,…,y T}, find the best alignment path (i.e., minimize the distance D) by DTW:

[0074] D(m,n) = |x m -x n | + min{D(m-1,n), D(m,n-1), D(m-1,n-1)}

[0075] where D(m,n) is the cumulative distance of the mth point and the nth point, and the main role of DTW is to align the time series and form a normalized sequence.

[0076] In the fifth step, the data completion work is completed, since there may be a large number of missing values in the industrial data, in order to achieve better results in the experiment, the missing points need to be completed. For some data, linear interpolation method can be used, such as inserting linear points between adjacent two known values and .

[0077]

[0078] For some complex data, model-based interpolation can be used, and the missing values are predicted by means of a deep learning model (such as a simple autoencoder or RNN network).

[0079] Further, the statistical analysis method is used to detect the abnormal point data in the industrial time series data and perform the rejection processing on the abnormal point data, including:

[0080] The mean and standard deviation of the industrial time series data are calculated;

[0081] The abnormal point range standard is constructed by using the mean and standard deviation, and the abnormal point range standard is used to mark and reject the abnormal point in the industrial time series data.

[0082] S2, input the standard industrial time series data into the improved variational autoencoder model to generate data initial samples.

[0083] In the embodiment of the application, the improved variational autoencoder model refers to a data generation model in which a hyperparameter is introduced into the variational autoencoder.

[0084] As an embodiment of the application, inputting the standard industrial time series data into the improved variational autoencoder model to generate data initial samples comprises:

[0085] inputting the standard industrial time series data into the encoder of the improved variational autoencoder model to output latent distribution parameters;

[0086] Construct a latent distribution using the latent distribution parameters, and sample latent variables from the latent distribution;

[0087] The loss function of the improved variational autoencoder model is constructed based on the latent variables, and the weight values ​​of the divergence expression part in the loss function are dynamically adjusted using hyperparameters. The weight values ​​of the divergence expression part in the loss function are then used to make the latent distribution approximate the standard normal prior.

[0088] Potential sampling points are obtained by randomly sampling from the standard normal prior, and these potential sampling points are input into the decoder of the improved variational autoencoder model to generate initial data samples.

[0089] For example, the following implementation steps can be used to input standard industrial time-series data into an improved variational autoencoder model to generate initial data samples:

[0090] (1) Encoder: The input is the preprocessed time-series data X′={x′1,x′2,…,x′ N The output is the latent distribution parameters μ and σ. The encoder structure uses a deep convolutional neural network (CNN) or an LSTM network with an attention mechanism to capture the local features and temporal dependencies of the time series.

[0091] (2) Decoder: The input is the latent variable z, and the output is the generated data. The decoder structure uses deconvolution operations or neural networks combined with attention mechanisms to reconstruct temporal features.

[0092] (3) Loss Function: ωVAE achieves more flexible distribution control in the latent space by introducing a hyperparameter ω to adjust the weights of the KL divergence. The loss function of tVAE is defined as:

[0093]

[0094] in, For the expected operation, q φ (z|x) represents the probability distribution of the encoder output, φ represents the encoder neural network parameters, and logP θ (x|z) represents the log-likelihood of the decoder, θ represents the parameters of the decoder neural network, and p(z) represents the latent distribution. It is the reconstruction error, representing the similarity between the generated data and the real data, KL(q) φ(z|x)||p(z)) is the KL divergence, representing the difference between the latent distribution and the prior distribution, and ω is a hyperparameter, used to balance the weight of reconstruction error and KL divergence. When ω>1, the weight of KL divergence increases, and the distribution of latent space is closer to the prior distribution (usually standard normal distribution), the diversity of generated data increases, but the reconstruction quality may decrease slightly; when ω<1, the weight of KL divergence decreases, and the model pays more attention to the reconstruction quality, and the generated data is closer to the real data, but the diversity may be insufficient.

[0095] The improved variational autoencoder model described in the embodiments of the present application dynamically adjusts the weight of KL divergence in the loss function through the hyperparameter ω, and by dynamically adjusting the value of ω, the reconstruction quality and data diversity are guaranteed. In the early stage of training, a smaller ω value is set to prioritize reconstruction quality; in the later stage of training, the ω value is gradually increased to enhance the distribution constraint of the latent space and improve the diversity of the generated data.

[0096] Further, the normal working mechanism of VAE can capture the overall distribution of data, while abnormal data is often less and difficult to be explicitly modeled, so before inputting the standard industrial time series data into the improved variational autoencoder model to generate initial data samples, an abnormal data priority mechanism is established, which is implemented by the following implementation steps:

[0097] First, abnormal detection guided generation is performed, and the input time series data x is divided into normal samples x normal and abnormal samples x anormaly by a preliminary abnormal detection mechanism. Abnormal samples are given higher weights in the encoding stage to highlight their distribution characteristics. The guided target is expressed as,

[0098]

[0099] where the weight w i =α(α>1) is used to amplify the importance of abnormal samples so that the optimization process pays more attention to abnormal areas. The final loss function is adjusted to,

[0100]

[0101] Then, the latent space abnormal distribution is strengthened, and the latent representation z anormaly of the abnormal sample may be in the sparse area of the data distribution in the latent space of tVAE. In order to prioritize the generation of abnormal samples, the abnormal distribution can be strengthened through abnormal prior adjustment and resampling mechanism. When adjusting the abnormal prior, the prior distribution of the latent space needs to be modified so that the abnormal area has a higher generation probability. Assuming that p(z) is usually a standard normal distribution N(0,1), the adjusted abnormal prior distribution p anomaly (z) is:

[0102]

[0103] wherein, μ anomaly and denote the distribution parameters of the abnormal samples in the latent space. A resampling mechanism is further established, and in the generation stage, the sampling weight distribution is tilted to the abnormal area (the probability of sampling from p anomaly (z) is greater than the normal distribution).

[0104] Finally, an abnormality enhancement post-processing is performed. The generated samples are further screened through abnormality characteristic matching to ensure that the generated samples can reflect the abnormal characteristics and avoid non-real situations. The abnormality characteristic matching indicators include distribution matching and time continuity. That is, the characteristics of the generated abnormal samples should be consistent with the first-order (mean, variance) and high-order characteristics (frequency components, etc.) of the original abnormal samples, and at the same time, have the continuity of real time series.

[0105] L dist =||μ gen -μ real ||+||σ gen -σ real ||

[0106]

[0107] wherein, L dist is an abnormality characteristic matching loss, μ gen is the mean of the generated abnormal sample, μ real is the mean of the original abnormal sample, σ gen is the standard deviation of the generated abnormal sample, σ real is the standard deviation of the original abnormal sample, and L smooth1 is a continuity smoothing loss.

[0108] In the embodiment of the application, the abnormality characteristic matching loss calculates the difference between the generated abnormal sample and the original abnormal sample in the mean and the standard deviation. The smaller the difference is, the closer the distribution of the generated sample is to the distribution of the real abnormal sample. The continuity smoothing loss calculates the smoothness of the generated sample in the time dimension. By minimizing this loss function, it can be ensured that the generated sample has continuity and smoothness in time, avoiding sudden changes or discontinuities. After the above screening, the truly effective abnormal samples

[0109] S3, input the initial data sample into the conditional generative adversarial network to perform data optimization operation to obtain optimized data sample.

[0110] In the embodiment of the application, the conditional generative adversarial network refers to a generative adversarial network with added conditional information, which can guide the generative adversarial network to generate data that is more in line with actual needs.

[0111] As an embodiment of the present application, the data initial sample is input into the conditional generative adversarial network to perform data optimization operation to obtain an optimized data sample, comprising:

[0112] The input layer of the generator in the conditional generative adversarial network receives the data initial sample and the condition information, and converts the condition information into a condition vector;

[0113] A random noise distribution is obtained, and the random noise distribution, the condition vector and the data initial sample are used to generate a noisy data initial sample;

[0114] The generator and the discriminator in the conditional generative adversarial network are optimized by using a perception loss function, and the optimized generator and discriminator are used for adversarial learning to obtain an optimized data sample.

[0115] Further, the condition information is converted into a condition vector, comprising:

[0116] Identifying the condition category of the condition information;

[0117] When the condition category is a discrete category, the condition information is mapped to a low-dimensional space to obtain a condition vector;

[0118] When the condition category is a continuous type, the condition information is vector transformed to obtain a condition vector.

[0119] Exemplarily, the condition information is converted into a condition vector, which can be realized by the following implementation steps:

[0120] The initial input of the generator of the conditional generative adversarial network is a sample generated based on tVAE and condition information c, and further improves the sample distribution through random noise ε,

[0121]

[0122] When the condition information is a discrete label (such as device state, category label), the discrete label is mapped to a low-dimensional embedding space, for example, the device state has 3 kinds (normal, abnormal, fault), which can be embedded in a 2-dimensional space.

[0123] c embed =E(c)

[0124] Wherein, E is an embedding matrix, which maps the discrete label c to a d-dimensional condition vector c embed .

[0125] When the condition information is a continuous value (such as timestamp, physical parameter), it can be directly input as a vector, or transformed through a simple fully connected layer, for example, the timestamp is spliced with the input data to form a new input vector.

[0126] c proj = W c c + b c

[0127] where W c is a learnable weight and b c is a bias term.

[0128] In the industrial Internet of Things scenario, the condition information may change dynamically over time. To adapt to such changes, a dynamic adjustment mechanism can be designed to combine condition information with a time window and dynamically update the condition information. For example, in a one-hour time window, update the condition information (such as the current device state, environmental temperature). This enables dynamic adjustment of condition information based on real-time characteristics of the data, and when data anomalies are detected, automatically update the condition information to generate data containing abnormal patterns. Condition information is introduced in the middle layer of the discriminator to enhance the discriminant ability of the discriminator, and condition normalization is used to adjust the mean and variance of the middle layer output through the condition information:

[0129] h cond = γ(c) · BN(h) + β(c)

[0130] where BN(h) is a batch normalization operation on feature h, γ(c) is a scaling parameter learned through condition information c, and β(c) is an offset parameter learned through condition information c.

[0131] Exemplarily, input the initial data sample into the conditional generative adversarial network for data optimization operation to obtain an optimized data sample, which can be implemented by the following implementation steps:

[0132] Use the adversarial game of generator G ψ (z) and discriminator D η (X) to further optimize the preliminary generated data and generate samples that are of high quality and closer to the real distribution. Perceptual loss is a loss function based on the high-level features of a deep learning model, which uses features extracted by a pre-trained deep neural network (usually a convolutional neural network) to measure the difference between generated data and real data. In order to better capture the semantic and structural information of the data, so that the generated data is closer to the real data at the high-level feature level, perceptual loss is introduced on this basis, therefore, the objective L G of generator optimization is:

[0133]

[0134] The objective L p of discriminator optimization is:

[0135]

[0136] wherein, Φ l is the feature map of the lth layer, C l denotes the number of channels of the lth layer feature map, H l denotes the height of the lth layer feature map, W l denotes the width of the lth layer feature map, respectively. Therefore, the cGAN-ωVAE joint optimization can be realized in view of this, by combining the objective functions of ωVAE and cGAN, the following joint optimization objective constraint L Hybrid is formed, and the process from the latent representation to the generated result forms a complete generation system.

[0137] L Hybrid =L VAE +λL G +ρL D

[0138] wherein, L VAE is used to ensure that the generated data is consistent with the latent distribution of the real data, L G is used to ensure that the generated samples are close to the real distribution, L D is used to improve the classification performance of the discriminator in the real and generated data, λ is a first weight parameter, and ρ is a second weight parameter, used to balance the contribution of the subnetworks.

[0139] S4, using a time sequence feature constraint model to constrain the industrial time sequence correlation of the optimized data samples to obtain industrial constraint data samples.

[0140] As an embodiment of the present application, the use of a time sequence feature constraint model to constrain the industrial time sequence correlation of the optimized data samples comprises:

[0141] The time sequence feature constraint model comprises: using a regularization algorithm to construct a change rate smoothing constraint of adjacent time points, using a pre-constructed time memory network to construct a time dependence constraint, using application scenario knowledge to construct a change range constraint and a change rate constraint, calculating a correlation matrix in the optimized data samples and using the correlation matrix to construct a multivariate correlation constraint;

[0142] The change rate smoothing constraint, the time dependence constraint, the change range constraint, the change rate constraint and the multivariate correlation constraint are fused to obtain a time sequence feature total constraint;

[0143] The time sequence feature total constraint is used to constrain the industrial time sequence correlation of the optimized data samples to obtain industrial constraint data samples.

[0144] Further, the regularization algorithm is used to construct the change rate smoothing constraint of adjacent time points, and the formula is as follows:

[0145] For the generated time sequence data L is a change rate smoothing constraint between adjacent time points smooth2 As follows:

[0146]

[0147] Wherein, is the value of the time series data output by the generator at time t+1, is the value of the time series data output by the generator at time t, 2 is a square norm operator.

[0148] Further, a pre-constructed time memory network is used to construct a time-dependent constraint, and the formula is as follows:

[0149] The pre-constructed time memory network is used to model the time-dependent constraint of the latent variable,

[0150] h t =LSTM(z t ,h t-1 )

[0151] Wherein, z t is the input noise or the latent variable output by the encoder, h t is the hidden state, h t-1 is the hidden state of the previous time step in the time memory network, used to maintain the time-dependent relationship. Thus, the time series output generated by the LSTM is,

[0152]

[0153] Wherein, f(·) is a decoding network.

[0154] In the embodiment of the application, the pre-constructed time memory network can adopt an LSTM (Long Short-Term Memory) network.

[0155] Further, the application scenario knowledge is used to construct a change range constraint and a change rate constraint, and the formula is as follows:

[0156] In the processing of the change range constraint L range , the generated industrial time series data should meet the normal operation range of the equipment:

[0157]

[0158] Wherein, A is the lower limit of the operation range, and B is the upper limit of the operation range. A penalty mechanism is applied to the samples exceeding the range, and the formula is as follows:

[0159]

[0160] Wherein, Lrange Constraints on the range of variation.

[0161] Under the constraint of rate of change L rate In terms of processing, the rate of change of the generated data Set physical constraints (such as the rate of change not exceeding a certain threshold C), as shown in the following formula:

[0162]

[0163] Furthermore, multivariate correlation constraints are constructed using the correlation matrix, and the implementation steps are as follows:

[0164] It is noteworthy that industrial IoT time-series data is often multivariate, and these variables have significant dependencies, such as multiple sensors influencing each other. Therefore, it is necessary to maintain the dynamic characteristics of multivariate data. Thus, in handling multivariate correlation constraints, it is assumed that the multivariate time-series data is... The correlation matrix R is calculated during training:

[0165] R[π,τ]=Corr(x π ,x τ )

[0166] Where Corr represents the correlation coefficient, x π For the complete time series of the π-th variable, x τ This represents the complete time series of the τ-th variable. To ensure the multivariate correlation matrix of the generated data... To approximate R as closely as possible to the real data, a loss function needs to be constructed:

[0167] Furthermore, by integrating the constraints on rate of change smoothing, time dependence, range of change, rate of change, and multivariate correlation, a total constraint on time-series characteristics is obtained, including:

[0168] Total constraint L of temporal features constraint The formula is as follows: L constraint =λ1L smooth2 +λ2L range +λ3L rate +λ4L corr λ1 is the first weight parameter, λ2 is the second weight parameter, λ3 is the third weight parameter, and λ4 is the fourth weight parameter.

[0169] This invention embodiment achieves joint optimization of objective constraints L by fusing and co-optimizing them. Hybrid With time series feature total constraint L constraint The final constraint L is obtained from the following formula: L = L Hybrid +α*L constraint α is the constraint weight of the total constraint of the time series features.

[0170] S5, sample evaluation is performed on the industrial constraint data sample, and sample screening is performed to obtain a target generated data sample.

[0171] As an embodiment of the present application, sample evaluation is performed on the industrial constraint data sample, including:

[0172] The first evaluation index is calculated using a pre-constructed statistical distribution matching formula, the second evaluation index is calculated using a pre-constructed constraint compliance rate formula, and the third evaluation index is calculated using a pre-constructed continuous consistency formula;

[0173] The first evaluation index, the second evaluation index, and the third evaluation index are fused to obtain a comprehensive evaluation index;

[0174] A comprehensive evaluation score is calculated using the comprehensive evaluation index, and sample evaluation is performed on the industrial constraint data sample according to the comprehensive evaluation score, to obtain a sample evaluation result.

[0175] Further, the first evaluation index is calculated using a pre-constructed statistical distribution matching formula, the second evaluation index is calculated using a pre-constructed constraint compliance rate formula, and the third evaluation index is calculated using a pre-constructed continuous consistency formula, which can adopt the following implementation steps:

[0176] (1) The first evaluation index: statistical distribution matching, that is, the generated data should be consistent with the target distribution x ~ p data . For example: The distribution of energy consumption may exhibit certain characteristics over time (such as peak and off-peak periods), and the generated sample should be limited within a similar statistical range. Therefore, the matching distribution can be defined as follows:

[0177] D KL = D KL (P data || P gen )

[0178] Where P data is the distribution of real-time series data, and P gen is the distribution of generated data. If the generated sample distribution deviates greatly from the real distribution (high KL divergence), the evaluation score of the sample is reduced.

[0179] (2) The second evaluation index: constraint compliance rate, that is, considering the compliance of industrial equipment operation data subject to physical constraints (such as parameter range constraints, maximum change rate limits, etc.). The corresponding constraint rule can be modeled as,

[0180]

[0181] At the same time, the constraint compliance rate of the generated data is defined as:

[0182]

[0183] Where m is the total number of generated samples, and C(·) indicates whether the sample satisfies the constraint rules.

[0184] (3) The third evaluation metric: contextual continuity, that is, the generated time series data should maintain smoothness and contextual consistency in the time dimension.

[0185]

[0186] For multivariate time series data, correlation consistency is calculated by comparing the Pearson correlation coefficient matrix.

[0187] L corr =||R data -R gen || 2

[0188] Among them, R data For the correlation matrix of real data, R gen To generate a correlation matrix for the data.

[0189] Based on the above indicators, a quality evaluation score for the generated samples can be obtained.

[0190]

[0191] Among them, parameters The first equilibrium parameter is... The second equilibrium parameter, parameter This is the third balancing parameter, used to balance the importance of various indicators.

[0192] Furthermore, using quality evaluation scores Set a sample screening threshold, and use the sample screening threshold to screen the industrial constraint data samples to obtain the target generated data samples.

[0193] For example, using quality assessment scores Set a sample selection threshold, and use the sample selection threshold to select industrial constraint data samples to obtain target generated data samples. The following implementation steps are adopted:

[0194] By setting quality evaluation scores threshold Q min Automatically filter out samples marked as abnormal.

[0195]

[0196] As another embodiment of the present invention, the implementation of anomaly detection can also be combined with a more complex discriminative model, such as anomaly detection (AE) based on an autoencoder:

[0197]

[0198] The anomaly score reflects whether a sample deviates from the true target distribution of the data, and represents the effective set of samples obtained after filtering. for:

[0199]

[0200] Furthermore, to avoid pattern collapse (i.e., generating samples with too small a coverage area), sample diversity assessment and resampling are necessary. Measuring sample diversity ensures that the generated data broadly covers the target distribution. Based on the latent representation of the generated data, the distribution P of the generated data in the latent space is... latent The underlying distribution Q of the real data latent Comparison:

[0201] L div =D KL (Q latent ||P latent )

[0202] The latent distribution can be achieved using an ωVAE encoder or a deep feature extraction network. If insufficient sample coverage is found in the data, the generator is used to resample (incorporating latent spatial diversity) to supplement the samples in sparse regions.

[0203] This invention utilizes a data preprocessing module to perform data preprocessing operations on industrial time-series data. This removes outliers and missing data, improving data quality and providing high-quality foundational data for subsequent data generation. Furthermore, standard industrial time-series data is input into an improved variational autoencoder model to generate initial data samples. The use of this improved model ensures reconstruction quality and data diversity. Additionally, the initial data samples are input into a conditional generative adversarial network (GAN) for data optimization, resulting in optimized data samples. Introducing conditional information into the GAN enhances the realism of the generated data. Moreover, a time-series feature constraint model is used to impose industrial time-series correlation constraints on the optimized data samples, strengthening the time-series structure and long-term dependency characteristics, ensuring that the generated data conforms to industrial time-series logic. Finally, the industrial constraint data samples are evaluated and screened to select high-quality samples and balance data distribution, ensuring data quality and increasing the sample size of the generated data.

[0204] like Figure 2As shown is a function module diagram of an industrial time series data enhancement device based on collaborative constraints provided by an embodiment of the application.

[0205] The industrial time series data enhancement device based on collaborative constraints 100 can be installed in an electronic device. According to the functions implemented, the industrial time series data enhancement device based on collaborative constraints 100 can include a data generation module 101 and a data screening module 102.

[0206] The module described in the application can also be referred to as a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and can complete a fixed function, which are stored in the memory of the electronic device.

[0207] In this embodiment, the functions of each module / unit are as follows:

[0208] The data generation module 101 is configured to obtain industrial time series data, perform data preprocessing operations on the industrial time series data using a data preprocessing module to obtain standard industrial time series data, and input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples.

[0209] In the embodiment of the application, the industrial time series data is a group of time series data in an industrial time series data set, wherein the industrial time series data set refers to a set of continuous observation values recorded by temperature sensors, humidity sensors and other industrial sensors in an industrial Internet of Things (IIoT) scenario in time sequence.

[0210] As an embodiment of the application, performing data preprocessing operations on the industrial time series data using a data preprocessing module to obtain standard industrial time series data includes:

[0211] Detecting abnormal point data in the industrial time series data using a statistical analysis method and performing rejection processing on the abnormal point data to obtain normal industrial time series data;

[0212] Removing noise from the normal industrial time series data using a smoothing filter to obtain denoised industrial time series data;

[0213] Aligning the normalized and standardized denoised industrial time series data using a dynamic time warping algorithm to obtain standardized industrial time series data;

[0214] Performing interpolation processing on missing values in the standardized industrial time series data to obtain standard industrial time series data.

[0215] Exemplarily, performing data preprocessing operations on the industrial time series data using a data preprocessing module to obtain standard industrial time series data can be implemented by the following implementation steps:

[0216] Input: original industrial time series dataset, denoted as X = {x1, x2, …, xT}, where xi is the i-th time series data, T is the length of the time series data. N}, where xi is the i-th time series data, T is the length of the time series data. is the i-th time series data, T is the length of the time series data.

[0217] Output: preprocessed time series data X' = {x'1, x'2, …, x'T}, which removes noise, eliminates outliers, and completes normalization and completion, meeting the input requirements of subsequent models. N

[0218] The first step is to complete data cleaning, i.e. outlier detection and elimination, the core of which is to eliminate abnormal data (such as invalid data generated during equipment failure) in industrial time series data. Mainly through statistical analysis or simple rule detection. First, calculate the mean μ i and standard deviation σ i of each time series data:

[0219]

[0220] Then, define outliers as data points outside a certain range, if (usually k = 3), it is marked as an outlier.

[0221] The second step is to complete data denoising, i.e. signal smoothing and filtering. Industrial time series data often contains high-frequency noise, which can be smoothed by the following method. First, use a smoothing filter. Define the window size as w, and use the sliding average formula:

[0222]

[0223] where j is the j-th window, x i ′ t is the sliding average value of the i-th time series data with length t, and the sliding average effectively smooths small fluctuations. Then, perform signal decomposition and complete wavelet denoising. Consider using Discrete Wavelet Transform (DWT) to decompose the signal, remove high-frequency noise, and retain low-frequency and medium-frequency parts;

[0224] x i = L i + H i

[0225] where L i is the low-frequency component, H i is the high-frequency component. You can choose to retain L i , discard or reduce the weight of H i .

[0226] ​Third, data normalization and standardization. That is, scaling the data to a uniform range helps to improve the convergence speed of deep learning model. There are two methods, min-max normalization and standardization. Among them, min-max normalization is suitable for data whose input features are in known limit range.

[0227]

[0228] where max(x i ) is the maximum value of each group of time series data, min(x i ) is the minimum value of each group of time series data, is the ith group of time series data with length t, and standardization is commonly used for time series data, removing the mean and scaling by standard deviation:

[0229]

[0230] Fourth, time alignment using dynamic time warping (DTW). Since industrial equipment may cause time series data to be misaligned in time during operation, time alignment is a very important step. Among them, dynamic time warping method is defined as given two time series data x m ={x 1 ,x 2 ,…,x T} and x n ={y 1 ,y 2 ,…,y T}, find the best alignment path (i.e. minimize distance D) by DTW:

[0231] D(m,n) = |x m -x n | + min{D(m-1,n),D(m,n-1),D(m-1,n-1)}

[0232] where D(m,n) is the cumulative distance of the mth point and the nth point, and the main function of DTW is to align the time series and form a standardized sequence.

[0233] Fifth, complete the data completion work. Since there may be a large number of missing values in industrial data, in order to achieve better results in the experiment, these missing points need to be completed. For some data, linear interpolation method can be used, such as inserting linear points between the adjacent two known values and :

[0234]

[0235] For some complex data, model-based interpolation can be used to predict missing values with deep learning models such as simple autoencoders or RNN networks.

[0236] Further, the statistical analysis method is used to detect the abnormal point data in the industrial time series data and perform the rejection processing on the abnormal point data, including:

[0237] Calculate the mean and standard deviation of the industrial time series data.

[0238] An abnormal point range standard is constructed using the mean and standard deviation, and the abnormal points in the industrial time series data are marked and rejected using the abnormal point range standard.

[0239] In the embodiment of the application, the improved variational autoencoder model refers to a data generation model in which a hyperparameter is introduced into the variational autoencoder.

[0240] As an embodiment of the application, the standard industrial time series data is input into the improved variational autoencoder model to generate data initial samples, including:

[0241] The standard industrial time series data is input into the encoder of the improved variational autoencoder model to output latent distribution parameters.

[0242] A latent distribution is constructed using the latent distribution parameters, and a latent variable is sampled from the latent distribution.

[0243] A loss function of the improved variational autoencoder model is constructed according to the latent variable, and the weight value of the divergence expression part in the loss function is dynamically adjusted using the hyperparameter, and the latent distribution is approximated to the standard normal prior using the weight value of the divergence expression part in the loss function.

[0244] A latent sampling point is randomly sampled from the standard normal prior, and the latent sampling point is input into the decoder of the improved variational autoencoder model to generate data initial samples.

[0245] Exemplarily, the standard industrial time series data is input into the improved variational autoencoder model to generate data initial samples, which can be implemented by the following implementation steps:

[0246] (4) Encoder: the input is the preprocessed time series data x' = {x'1, x'2, …, x'N}, and the output is the latent distribution parameters μ and σ. N The encoder structure uses a deep convolutional neural network (CNN) or an LSTM network with an attention mechanism to capture local features and time dependencies of the time series.

[0247] (5) Decoder: the input is the latent variable z, and the output is the generated data The decoder structure uses deconvolution operations or neural networks combined with attention mechanisms to restore the time series features.

[0248] (6) Loss function: ωVAE adjusts the weight of KL divergence by introducing a hyperparameter ω, achieving more flexible distribution control in latent space. The loss function of tVAE is defined as:

[0249]

[0250] where, is the expectation operation, q φ (z|x) is the probability distribution output by the encoder, φ is the encoder neural network parameter, logp θ (x|z) is the log-likelihood of the decoder, θ is the decoder neural network parameter, p(z) is the latent distribution, is the reconstruction error, representing the similarity between generated data and real data, KL(q φ (z|x)||p(z)) is the KL divergence, representing the difference between the latent distribution and the prior distribution, ω is the hyperparameter, used to balance the weight of reconstruction error and KL divergence. When ω>1, the weight of KL divergence increases, the distribution of latent space is closer to the prior distribution (usually standard normal distribution), the diversity of generated data increases, but the reconstruction quality may decrease slightly; when ω<1, the weight of KL divergence decreases, the model pays more attention to the reconstruction quality, the generated data is closer to the real data, but the diversity may be insufficient.

[0251] The improved variational autoencoder model described in the embodiments dynamically adjusts the weight of KL divergence in the loss function through the hyperparameter ω, and by dynamically adjusting the value of ω, the reconstruction quality and data diversity are guaranteed. In the early stage of training, a small ω value is set to prioritize reconstruction quality; in the later stage of training, the ω value is gradually increased to enhance the distribution constraint of the latent space and improve the diversity of generated data.

[0252] Further, the normal working mechanism of VAE captures the overall distribution of data, while abnormal data is often less and difficult to be explicitly modeled, so before inputting the standard industrial time series data into the improved variational autoencoder model to generate initial data samples, an abnormal data priority mechanism is established, which is implemented by the following implementation steps:

[0253] First, abnormal detection guided generation is performed, and the input time series data x is divided into normal samples x normal and abnormal samples x anormaly by a preliminary abnormal detection mechanism. Abnormal samples are given higher weights in the encoding stage to highlight their distribution characteristics. The guiding target is expressed as,

[0254]

[0255] where, the weight w i= α (α > 1) is used to amplify the importance of abnormal samples in order to pay more attention to abnormal areas in optimization. The final loss function is adjusted as,

[0256]

[0257] Then, the latent space anomaly distribution reinforcement is carried out, and the latent representation z anormaly In the latent space of tVAE, it can be in the sparse area of data distribution, in order to preferentially generate abnormal samples, the abnormal distribution can be reinforced by abnormal prior adjustment and resampling mechanism. When abnormal prior adjustment is carried out, the prior distribution of latent space needs to be modified, so that the abnormal area has a higher generation probability. Assuming that p(z) is usually a standard normal distribution N(0, 1), the abnormal prior distribution p anomaly (z) is adjusted as:

[0258]

[0259] Wherein, μ anomaly and represent the distribution parameters of abnormal samples in the latent space. Then, the resampling mechanism is established, and in the generation stage, the sampling weight is allocated to the abnormal area (the probability of sampling from p anomaly (z) is greater than that of normal distribution).

[0260] Finally, the abnormal enhancement post-processing is carried out. The generated samples are further screened by matching abnormal characteristics, to ensure that the generated samples can reflect abnormal characteristics and avoid non-real situations. The abnormal characteristic matching index includes distribution matching and time continuity. That is, the characteristics of the generated abnormal samples should be consistent with the first order (mean, variance) and high order characteristics (frequency component, etc.) of the original abnormal samples, and at the same time, have the continuity of real time sequence.

[0261] L dist = || μ gen - μ real || + || σ gen - σ real ||

[0262]

[0263] Wherein, L dist is the abnormal characteristic matching loss, μ gen is the mean of the generated abnormal sample, μ real is the mean of the original abnormal sample, σ gen is the standard deviation of the generated abnormal sample, σ real is the standard deviation of the original abnormal sample, and L smooth1 is the continuity smoothness loss.

[0264] In the embodiment of the present application, the abnormal characteristic matching loss calculates the difference between the generated abnormal sample and the original abnormal sample in the mean and standard deviation. The smaller the difference is, the closer the distribution of the generated sample is to the distribution of the real abnormal sample. The continuity smoothing loss calculates the smoothness of the generated sample in the time dimension. By minimizing this loss function, it can ensure that the generated sample has continuity and smoothness in time, avoiding sudden changes or discontinuities. After the above screening, the truly effective abnormal samples are retained

[0265] The data screening module 102 is configured to input the data initial sample into the conditional generative adversarial network to perform data optimization operation to obtain an optimized data sample, constrain the optimized data sample by using a time sequence feature constraint model to obtain an industrial constraint data sample, perform sample evaluation on the industrial constraint data sample, and perform sample screening to obtain a target generated data sample.

[0266] In the embodiment of the present application, the conditional generative adversarial network refers to a generative adversarial network to which condition information is added, and the condition information can guide the generative adversarial network to generate data that is more in line with actual needs.

[0267] As an embodiment of the present application, the data initial sample is input into the conditional generative adversarial network to perform data optimization operation to obtain an optimized data sample, including:

[0268] The input layer of the generator in the conditional generative adversarial network receives the data initial sample and the condition information, and converts the condition information into a condition vector;

[0269] A random noise distribution is obtained, and the random noise distribution, the condition vector and the data initial sample are used to generate a noisy data initial sample;

[0270] The generator and the discriminator in the conditional generative adversarial network are optimized by using a perception loss function, and the optimized generator and the discriminator are used for adversarial learning to obtain the optimized data sample.

[0271] Further, the condition information is converted into a condition vector, including:

[0272] The condition category of the condition information is identified;

[0273] When the condition category is a discrete category, the condition information is mapped to a low-dimensional space to obtain the condition vector;

[0274] When the condition category is a continuous type, the condition information is vector transformed to obtain the condition vector.

[0275] Exemplarily, the condition information is converted into a condition vector, which can be realized by the following implementation steps:

[0276] The initial input of the generator of the conditional generative adversarial network is a sample generated based on the tVAE and the condition information c, and further improves the sample distribution through random noise ε,

[0277]

[0278] When the condition information is a discrete label (such as a device state, a category label), the discrete label is mapped to a low-dimensional embedding space, for example, a device state has 3 kinds (normal, abnormal, fault), which can be embedded in a 2-dimensional space.

[0279] c embed =E(c)

[0280] wherein E is an embedding matrix, which maps the discrete label c to a d-dimensional condition vector c embed .

[0281] When the condition information is a continuous value (such as a timestamp, a physical parameter), it can be directly input as a vector or transformed through a simple fully connected layer, for example, the timestamp is spliced with the input data to form a new input vector.

[0282] c proj =W c c+b c

[0283] wherein W c is a learnable weight and b c is a bias term.

[0284] In the embodiments of the present application, in the industrial Internet of Things scene, the condition information may change dynamically over time. In order to adapt to such changes, a dynamic adjustment mechanism can be designed to combine the condition information with the time window and dynamically update the condition information. For example, in the time window of every hour, the condition information (such as the current device state, the environmental temperature) is updated. This enables dynamic adjustment of the condition information according to the real-time characteristics of the data, and when data anomalies are detected, the condition information is automatically updated to generate data containing abnormal patterns. The condition information is introduced in the middle layer of the discriminator, which is used to enhance the discrimination ability of the discriminator, and the condition normalization is used to adjust the mean and variance of the middle layer output through the condition information:

[0285] h cond =γ(c)·BN(h)+β(c)

[0286] wherein BN(h) is a batch normalization operation on the feature h, γ(c) is a scaling parameter learned through the condition information c, and β(c) is an offset parameter learned through the condition information c.

[0287] Exemplarily, the data initial sample is input into the conditional generative adversarial network for data optimization operation to obtain an optimized data sample, and the following implementation steps can be used to realize the same:

[0288] using the generator G ψ (z) and the discriminator D η (X) to further optimize the preliminary generated data and generate samples closer to the real distribution. The perceptual loss is a loss function based on the high-level features of the deep learning model, which uses the features extracted by the pre-trained deep neural network (usually a convolutional neural network) to measure the difference between the generated data and the real data. In order to better capture the semantic and structural information of the data, so that the generated data is closer to the real data at the high-level feature level, the perceptual loss is introduced on this basis, and therefore the optimization goal L G of the generator is:

[0289]

[0290] the optimization goal L p of the discriminator is:

[0291]

[0292] wherein, Φ l is the feature map of the lth layer, C l represents the number of channels of the lth layer feature map, H l represents the height of the lth layer feature map, and W l represents the width of the lth layer feature map. Therefore, the cGAN-ωVAE joint optimization can be realized in view of this, and by combining the objective functions of ωVAE and cGAN, the following joint optimization objective constraint L Hybrid is obtained, and the process from the latent representation to the generated result forms a complete generation system.

[0293] L Hybrid = L VAE + λL G + ρL D

[0294] wherein, L VAE is used to ensure that the generated data is consistent with the latent distribution of the real data, L G is used to ensure that the generated samples approximate the real distribution, L D is used to improve the classification performance of the discriminator in judging the real and generated data, λ is a first weight parameter, and ρ is a second weight parameter used to balance the contribution of the subnetworks.

[0295] As an embodiment of the present application, the industrial time sequence correlation constraint on the optimized data sample by using the time sequence feature constraint model comprises the following steps:

[0296] The time sequence feature constraint model comprises: constructing a change rate smoothing constraint of adjacent time points by using a regularization algorithm, constructing a time dependence constraint by using a pre-constructed time memory network, constructing a change range constraint and a change rate constraint by using application scenario knowledge, calculating a correlation matrix in the optimization data sample and constructing a multivariate correlation constraint by using the correlation matrix;

[0297] The time sequence feature total constraint is obtained by fusing the change rate smoothing constraint, the time dependence constraint, the change range constraint, the change rate constraint and the multivariate correlation constraint;

[0298] The industrial constraint data sample is obtained by using the time sequence feature total constraint to perform industrial time sequence correlation constraint on the optimization data sample.

[0299] Further, the change rate smoothing constraint of adjacent time points is constructed by using a regularization algorithm, and the formula is as follows:

[0300] For the generated time sequence data The change rate smoothing constraint L between adjacent time points is required smooth2 As follows:

[0301]

[0302] Wherein, is the value of the time sequence data output by the generator at time t+1, is the value of the time sequence data output by the generator at time t, ||·|| 2 is a square norm operator.

[0303] Further, the time dependence constraint is constructed by using a pre-constructed time memory network, and the formula is as follows:

[0304] The latent variable is time-dependently constrained and modeled by using a pre-constructed time memory network,

[0305] h t =LSTM(z t ,h t-1 )

[0306] Wherein, z t is an input noise or a latent variable output by an encoder, h t is a hidden state, h t-1 is a hidden state of a previous time step in the time memory network, used to maintain the time dependence. Thus, the time sequence output generated by the LSTM is,

[0307]

[0308] Wherein, f(·) is a decoding network.

[0309] In the embodiment of the application, the pre-constructed time memory network can adopt an LSTM (Long Short-Term Memory) network.

[0310] Further, the change range constraint and the change rate constraint are constructed by using application scenario knowledge, and the formulas are as follows:

[0311] In the processing of the change range constraint L range , the generated industrial time series data should meet the normal operation range of the equipment:

[0312]

[0313] wherein A is the lower limit of the operation range, and B is the upper limit of the operation range. A penalty mechanism is applied to the samples exceeding the range, and the formula is as follows:

[0314]

[0315] wherein L range is the change range constraint.

[0316] In the processing of the change rate constraint L rate , the change rate of the generated data is set to a physical constraint (for example, the change rate cannot exceed a threshold C), and the formula is as follows:

[0317]

[0318] Further, the multivariate correlation constraint is constructed by using the correlation matrix, and the implementation steps are as follows:

[0319] It is noted that the industrial Internet of Things time series data is often multivariate, and there is a significant dependent relationship between these variables, such as multiple sensors affecting each other, so it is necessary to maintain the dynamic characteristics of the multivariate. Thus, in the processing of the multivariate correlation constraint, it is assumed that the multivariate time series data is In the training process, the correlation matrix R is calculated:

[0320] R[pi, tau] = Corr (x π , x τ )

[0321] wherein Corr represents the correlation coefficient, x π is the complete time series of the pi-th variable, and x τ is the complete time series of the tau-th variable. In order to ensure that the multivariate correlation matrix of the generated data is as close as possible to the R of the real data, a loss needs to be constructed:

[0322] Further, the change rate smoothing constraint, the time dependence constraint, the change range constraint, the change rate constraint and the multivariate correlation constraint are fused to obtain a total time sequence feature constraint, including:

[0323] The total time sequence feature constraint L constraint is shown in the following formula: L constraint = λ1L smooth2 + λ2L range + λ3L rate + λ4L corr , wherein λ1 is a first weight parameter, λ2 is a second weight parameter, λ3 is a third weight parameter, and λ4 is a fourth weight parameter.

[0324] The embodiment of the application fuses the joint optimization target constraint L Hybrid and the total time sequence feature constraint L constraint to obtain a final constraint L, and the formula is as follows: L = L Hybrid + α * L constraint , wherein α is a constraint weight of the total time sequence feature constraint.

[0325] As an embodiment of the application, sample evaluation is performed on the industrial constraint data sample, including:

[0326] The first evaluation index is calculated by using a pre-constructed statistical distribution matching formula, the second evaluation index is calculated by using a pre-constructed constraint compliance rate formula, and the third evaluation index is calculated by using a pre-constructed continuous consistency formula.

[0327] The first evaluation index, the second evaluation index and the third evaluation index are fused to obtain a comprehensive evaluation index.

[0328] The comprehensive evaluation index is used to calculate a comprehensive evaluation score, and the sample evaluation is performed on the industrial constraint data sample according to the comprehensive evaluation score to obtain a sample evaluation result.

[0329] Further, the first evaluation index is calculated by using a pre-constructed statistical distribution matching formula, the second evaluation index is calculated by using a pre-constructed constraint compliance rate formula, and the third evaluation index is calculated by using a pre-constructed continuous consistency formula, which can adopt the following implementation steps:

[0330] (1) The first evaluation index: statistical distribution matching, that is, the generated data should be consistent with the target distribution x ~ p data . For example: the distribution of energy consumption may present certain characteristics (such as peak and valley periods) over time, and the generated sample should be limited within a similar statistical range. Therefore, the matching distribution can be defined as follows:

[0331] D KL = D KL (Pdata ||P gen )

[0332] where P data is the distribution of real time series data, P gen is the distribution of generated data. If the generated sample distribution deviates greatly from the real distribution (high KL divergence), the evaluation score of this sample is reduced.

[0333] (4) The second evaluation index: constraint compliance rate, which considers the compliance of industrial equipment operation data subject to physical constraints (such as parameter range constraints, maximum change rate limits, etc.). The corresponding constraint rule can be modeled as,

[0334]

[0335] Meanwhile, the constraint compliance rate of the generated data is defined as:

[0336]

[0337] where m is the total number of generated samples, and C(·) represents whether the sample satisfies the constraint rule.

[0338] (5) The third evaluation index: context continuity, which means that the generated time series data should maintain smoothness and context consistency in the time dimension,

[0339]

[0340] For multivariate time series data, the correlation consistency is calculated by comparing the Pearson correlation coefficient matrix,

[0341] L corr =||R data -R gen || 2

[0342] where R data is the correlation matrix of real data, and R gen is the correlation matrix of generated data.

[0343] By integrating the above indexes, the quality evaluation score of the generated sample can be defined as

[0344]

[0345] where the parameter is the first balance parameter, the parameter is the second balance parameter, and the parameter is the third balance parameter. The balance parameters are used to balance the importance of each index.

[0346] Furthermore, using quality evaluation scores Set a sample screening threshold, and use the sample screening threshold to screen the industrial constraint data samples to obtain the target generated data samples.

[0347] For example, using quality assessment scores Set a sample selection threshold, and use the sample selection threshold to select industrial constraint data samples to obtain target generated data samples. The following implementation steps are adopted:

[0348] By setting quality evaluation scores threshold Q min Automatically filter out samples marked as abnormal.

[0349]

[0350] As another embodiment of the present invention, the implementation of anomaly detection can also be combined with a more complex discriminative model, such as anomaly detection (AE) based on an autoencoder:

[0351]

[0352] The anomaly score reflects whether a sample deviates from the true target distribution of the data, and represents the effective set of samples obtained after filtering. for:

[0353]

[0354] Furthermore, to avoid pattern collapse (i.e., generating samples with too small a coverage area), sample diversity assessment and resampling are necessary. Measuring sample diversity ensures that the generated data broadly covers the target distribution. Based on the latent representation of the generated data, the distribution P of the generated data in the latent space is... latent The underlying distribution Q of the real data latent Comparison:

[0355] L div =D KL (Q latent ||P latent )

[0356] The latent distribution can be achieved using an ωVAE encoder or a deep feature extraction network. If insufficient sample coverage is found in the data, the generator is used to resample (incorporating latent spatial diversity) to supplement the samples in sparse regions.

[0357] Reference Figure 3As shown, a structural schematic diagram of an electronic device for implementing the collaborative constraint-based industrial time series data enhancement method is provided in an embodiment of the present application.

[0358] The electronic device can include a processor 10, a memory 11, a communication bus 12, and a communication interface 13, and can further include a computer program stored in the memory 11 and executable on the processor 10, such as a collaborative constraint-based industrial time series data enhancement method program.

[0359] In some embodiments, the processor 10 can be composed of an integrated circuit, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more combinations of a central processing unit (CPU), a microprocessor, a digital processing chip, a graphics processor, and various control chips. The processor 10 is the control unit of the electronic device, which connects various components of the electronic device through various interfaces and lines, executes programs or modules stored in the memory 11 (such as a collaborative constraint-based industrial time series data enhancement method program), and calls data stored in the memory 11 to perform various functions and process data of the electronic device.

[0360] The memory 11 includes at least one type of readable storage medium, including a flash memory, a mobile hard disk, a multimedia card, a card-type memory (such as an SD or DX memory), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device, such as a mobile hard disk of the electronic device. In other embodiments, the memory 11 can also be an external storage device of the electronic device, such as a plug-in mobile hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 11 can include both an internal storage unit and an external storage device of the electronic device. The memory 11 can be used not only to store application software and various data installed in the electronic device, such as a code of a collaborative constraint-based industrial time series data enhancement method program, but also to temporarily store data that has been or will be output.

[0361] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to realize the connection and communication between the memory 11 and the at least one processor 10, etc.

[0362] The communication interface 13 is configured to realize the communication between the electronic device and other devices, and includes a network interface and a user interface. Optionally, the network interface can include a wired interface and / or a wireless interface (such as a WI-FI interface, a Bluetooth interface, etc.), and is usually configured to establish a communication connection between the electronic device and other electronic devices. The user interface can be a display (Display), an input unit (such as a keyboard (Keyboard)), and optionally, the user interface can also be a standard wired interface, a wireless interface. Optionally, in some embodiments, the display can be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch screen, etc. The display can also be appropriately referred to as a display screen or a display unit, and is configured to display information processed in the electronic device and display a visualized user interface.

[0363] Figure 3 Only the electronic device with components is shown, and those skilled in the art can understand that, Figure 3 The structure shown does not constitute a limitation on the electronic device, and can include fewer or more components than shown, or combine some components, or different component arrangements.

[0364] For example, although not shown, the electronic device can further include a power supply (such as a battery) for supplying power to each component. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, so as to realize functions such as charge management, discharge management, and power consumption management through the power management device. The power supply can also include one or more direct-current or alternating-current power supplies, a recharging device, a power supply fault detection circuit, a power supply converter or inverter, a power supply status indicator, etc. The electronic device can also include various sensors, a Bluetooth module, a Wi-Fi module, etc., which are not described here.

[0365] It should be understood that the embodiments are only for illustration, and the scope of the patent application is not limited by the structure.

[0366] The memory 11 in the electronic device stores a collaborative constraint-based industrial time series data enhancement method program, which is a combination of multiple instructions and can realize the following when running in the processor 10:

[0367] Obtain industrial time series data, and perform data preprocessing operation on the industrial time series data by using a data preprocessing module to obtain standard industrial time series data;

[0368] Input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples;

[0369] Input the data initial samples into a conditional generative adversarial network to perform data optimization operation to obtain optimized data samples;

[0370] Constrain the optimized data samples by using a time series feature constraint model to obtain industrial constraint data samples;

[0371] Perform sample evaluation and sample screening on the industrial constraint data samples to obtain target generated data samples.

[0372] Specifically, the processor 10 can refer to the description of the related steps in the corresponding embodiments of the accompanying drawings for the specific implementation method of the above instructions, which will not be repeated here.

[0373] Further, the modules / units integrated in the electronic device 1 are implemented in the form of software function units and sold or used as independent products, which can be stored in a computer readable storage medium. The computer readable storage medium can be volatile or non-volatile. For example, the computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory).

[0374] The application also provides a computer readable storage medium, which stores a computer program, and the computer program can realize the following when being executed by a processor of an electronic device:

[0375] Obtain industrial time series data, and perform data preprocessing operation on the industrial time series data by using a data preprocessing module to obtain standard industrial time series data;

[0376] Input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples;

[0377] Input the data initial samples into a conditional generative adversarial network to perform data optimization operation to obtain optimized data samples;

[0378] The industrial constraint data samples are subjected to sample evaluation and sample screening to obtain target generation data samples.

[0379] The industrial constraint data samples are subjected to sample evaluation and sample screening to obtain target generation data samples.

[0380] In several embodiments provided by the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other manners. For example, the described device embodiments are merely schematic. The division of the modules is merely a logical function division. There can be another division manner in actual implementation.

[0381] The modules illustrated as separated components can or can not be physically separated, and the components illustrated as modules can or can not be physical units. They can be located in one position or distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purposes of the embodiments.

[0382] In addition, each function module in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of hardware plus software function module.

[0383] It is obvious for those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0384] Therefore, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims to which they relate.

[0385] The blockchain referred to in the present application is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm and other computer technologies. Blockchain, in essence, is a decentralized database, which is a series of data blocks associated using cryptographic methods. Each data block contains information of a batch of network transactions, used to verify the validity (anti-fake) of the information and generate the next block. The blockchain can include a blockchain underlying platform, a platform product service layer and an application service layer.

[0386] Embodiments of the present application can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (AI) is to use digital computers or computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0387] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in a system claim can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names and not to indicate any particular order.

[0388] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for industrial time series data enhancement based on collaborative constraints, characterized in that, The method comprises: acquiring industrial time series data and performing data preprocessing operation on the industrial time series data by using a data preprocessing module to obtain standard industrial time series data, wherein the industrial time series data refers to a set of continuous observation values recorded in time sequence by a temperature sensor and a humidity sensor in an industrial Internet of Things scene; inputting the standard industrial time series data into an improved variational autoencoder model to generate data initial samples, wherein the improved variational autoencoder model dynamically adjusts the weight value of the divergence expression part in the loss function by using hyperparameters; inputting the data initial samples into a conditional generative adversarial network for data optimization operation to obtain optimized data samples; constraining the optimized data samples by using a time series feature constraint model to obtain industrial constraint data samples, wherein the constraint of the optimized data samples by using the time series feature constraint model comprises: the time series feature constraint model comprises: constructing a change rate smoothing constraint of adjacent time points by using a regularization algorithm, constructing a time dependence constraint by using a pre-constructed time memory network, constructing a change range constraint and a change rate constraint by using application scenario knowledge, calculating a correlation matrix in the optimized data samples and constructing a multivariate correlation constraint by using the correlation matrix; fusing the change rate smoothing constraint, the time dependence constraint, the change range constraint, the change rate constraint and the multivariate correlation constraint to obtain a total time series feature constraint; constraining the optimized data samples by using the total time series feature constraint to obtain the industrial constraint data samples; sample evaluation and sample screening are performed on the industrial constraint data samples to obtain target generated data samples, wherein the sample evaluation on the industrial constraint data samples comprises: calculating a first evaluation index by using a pre-constructed statistical distribution matching formula, calculating a second evaluation index by using a pre-constructed constraint compliance rate formula, and calculating a third evaluation index by using a pre-constructed continuous consistency formula; fusing the first evaluation index, the second evaluation index and the third evaluation index to obtain a comprehensive evaluation index; calculating a comprehensive evaluation score by using the comprehensive evaluation index, and performing sample evaluation on the industrial constraint data samples according to the comprehensive evaluation score to obtain a sample evaluation result.

2. The method of claim 1, wherein, The method comprises: detecting abnormal point data in the industrial time series data by using a statistical analysis method and performing rejection processing on the abnormal point data to obtain normal industrial time series data; performing noise removal processing on the normal industrial time series data by using a smoothing filter to obtain denoised industrial time series data; aligning the normalized and standardized denoised industrial time series data by using a dynamic time warping algorithm to obtain standardized industrial time series data; performing interpolation processing on the missing values in the standardized industrial time series data to obtain the standard industrial time series data.

3. The method of claim 2, wherein the method is based on a co-constrained industrial time series data augmentation method, and the method comprises: The method comprises: calculating the mean and standard deviation of the industrial time series data; The mean and the standard deviation are used to construct an abnormal point range standard, and the abnormal point range standard is used to mark and remove the abnormal points in the industrial time series data.

4. The method of claim 1, wherein the method is based on a co-constrained industrial time series data augmentation method, and the method comprises: The standard industrial time series data is input into the improved variational autoencoder model to generate data initial samples, which includes: The standard industrial time series data is input into the encoder of the improved variational autoencoder model to output latent distribution parameters; The latent distribution is constructed by using the latent distribution parameters, and the latent variables are sampled from the latent distribution; The loss function of the improved variational autoencoder model is constructed according to the latent variables, and the weight value of the divergence expression part in the loss function is dynamically adjusted by using the hyperparameter, and the latent distribution is approximated to the standard normal prior by using the weight value of the divergence expression part in the loss function. The latent sampling points are randomly sampled from the standard normal prior, and the data initial samples are generated by inputting the latent sampling points into the decoder of the improved variational autoencoder model.

5. The method of claim 1, wherein, The data optimization operation is performed on the data initial samples in the conditional generative adversarial network to obtain optimized data samples, which includes: The input layer of the generator in the conditional generative adversarial network receives the data initial samples and the conditional information, and the conditional information is converted into a conditional vector; The random noise distribution is obtained, and the random noise distribution, the conditional vector and the data initial samples are used to generate a noisy data initial sample; The generator and the discriminator in the conditional generative adversarial network are optimized by using the perceptual loss function, and the optimized generator and the discriminator are used for adversarial learning to obtain the optimized data samples.

6. A device for enhancing industrial time series data based on collaborative constraints, characterized in that, The device implements the industrial time series data enhancement method based on collaborative constraint as claimed in any one of claims 1-5, and the device includes: A data generation module is configured to obtain industrial time series data, perform data preprocessing operation on the industrial time series data by using a data preprocessing module to obtain standard industrial time series data, and input the standard industrial time series data into an improved variational autoencoder model to generate data initial samples; A data screening module is configured to input the data initial samples into a conditional generative adversarial network to perform data optimization operation to obtain optimized data samples, perform industrial time series correlation constraint on the optimized data samples by using a time series feature constraint model to obtain industrial constraint data samples, perform sample evaluation on the industrial constraint data samples, and perform sample screening to obtain target generated data samples.

7. An electronic device, comprising: The electronic device includes: at least one processor; and a memory connected in communication with the at least one processor; wherein The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the industrial time series data enhancement method based on collaborative constraint as claimed in any one of claims 1-5.

8. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the industrial time series data enhancement method based on collaborative constraint as claimed in any one of claims 1-5.

Images