Interpretable process monitoring method and related equipment based on deep state space model

Abstract

The present application is applicable to the field of industrial process monitoring technology, and provides an interpretable process monitoring method and related equipment based on a deep state space model, wherein the method includes: performing modal division on industrial process data samples at multiple historical moments to obtain modal labels corresponding to each industrial process data sample; segmenting the industrial process data samples with modal labels to generate a training sample set; constructing a deep state space model; the latent variables of the deep state space model include understandable discrete parts and incomprehensible continuous parts; using the training sample set to train the deep state space model with maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modal as the loss; using the trained deep state space model to analyze and process online industrial process data to determine whether there is a fault in the industrial system. The present application can enhance the interpretability of the process monitoring model and improve the credibility of the industrial process fault monitoring results.

Description

Technical Field

[0001] The present application belongs to the technical field of industrial process monitoring, and in particular relates to an interpretable process monitoring method and related equipment based on a deep state space model. Background Art

[0002] In modern industrial production, process monitoring is crucial for maintaining production line safety and stability. Traditional model-based approaches require prior knowledge of physics and mathematics and are only applicable to systems with simple structures and clear mathematical expressions. However, with the increasing complexity of modern industrial processes, model building has become difficult. The rapid development of modern industrial hardware and big data analysis tools has led to the recording and collection of vast amounts of process data. Data-based process monitoring methods rely solely on measured process variables, requiring relatively simple design requirements, leading to rapid growth. Managing complex nonlinear relationships between process variables is crucial for accurate process monitoring. Kernel methods and deep learning are the main strategies for addressing these nonlinear relationships. Kernel methods transform nonlinear relationships from their original space into a high-dimensional space, transforming them into linear correlations. However, kernel methods are highly sensitive to the choice of kernel function and hyperparameters. Consequently, deep learning has been widely used in process monitoring due to its powerful ability to extract nonlinear features.

[0003] Deep learning-based process monitoring methods can extract features through unsupervised training, thereby constructing fault-sensitive statistics and achieving fault detection. However, the fundamental reason for the application of deep learning-based process monitoring methods is that they lack interpretability, making it difficult for on-site workers to trust the accuracy of the model. Otherwise, false positives and omissions will result in wasted manpower and economic losses. Summary of the Invention

[0004] The embodiments of the present application provide an interpretable process monitoring method and related equipment based on a deep state space model, which can solve the problem that industrial process monitoring models lack interpretability, resulting in low credibility of process fault monitoring results.

[0005] In a first aspect, an embodiment of the present application provides an interpretable process monitoring method based on a deep state space model, comprising:

[0006] Obtain industrial process data samples at multiple historical moments, and use a clustering algorithm to perform modal division on the acquired industrial process data samples to obtain the modal labels corresponding to the industrial process data samples at each historical moment;

[0007] The industrial process data samples with modality labels are divided into time windows to generate training sample sets;

[0008] Construct a deep state space model; the latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part. The discrete part is used to reflect the working conditions of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system.

[0009] The deep state space model is trained using the training sample set to obtain a trained deep state space model; wherein, during the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modalities are used as the model training loss;

[0010] The trained deep state space model is used to process online industrial process data to obtain latent variable monitoring indicators and observation value monitoring indicators;

[0011] Determine whether there is a fault in the industrial system based on the latent variable monitoring indicators and the observed value monitoring indicators.

[0012] In a second aspect, an embodiment of the present application provides an interpretable process monitoring device based on a deep state space model, comprising:

[0013] An acquisition module is used to acquire industrial process data samples at multiple historical moments, and use a clustering algorithm to perform modal division on the acquired industrial process data samples to obtain modal labels corresponding to the industrial process data samples at each historical moment;

[0014] A generation module is used to segment the industrial process data samples carrying modality labels according to time windows to generate a training sample set;

[0015] A construction module is used to construct a deep state space model. The latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part. The discrete part is used to reflect the working conditions of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system.

[0016] A training module is used to train the deep state space model using the training sample set to obtain a trained deep state space model; wherein, during the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modalities are used as the model training loss;

[0017] The processing module is used to process the online industrial process data using the trained deep state space model to obtain latent variable monitoring indicators and observation value monitoring indicators;

[0018] The fault analysis module is used to determine whether there is a fault in the industrial system based on the latent variable monitoring indicators and the observed value monitoring indicators.

[0019] In a third aspect, an embodiment of the present application provides a terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned interpretable process monitoring method when executing the computer program.

[0020] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned interpretable process monitoring method is implemented.

[0021] The above solution of the present application has the following beneficial effects:

[0022] In an embodiment of the present application, a discrete part with practical significance is introduced into the latent variables of the deep state space model to make the latent variable part understandable, and the deep state space model is trained by maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the mode as the model training loss, which significantly improves the interpretability of the model. Therefore, when the deep state space model is used to analyze and process online industrial process data, the fault monitoring results determined by the latent variable monitoring indicators and observation value monitoring indicators output by the inference network of the model are highly interpretable, thereby greatly improving the credibility of the industrial process fault monitoring results.

[0023] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0025] Figure 1 A flowchart of an interpretable process monitoring method based on a deep state space model provided in one embodiment of the present application;

[0026] Figure 2 A schematic diagram of a deep state space model provided in an embodiment of the present application;

[0027] Figure 3 In the embodiment of this application, fault 1 is based on Fault detection plots of statistics;

[0028] Figure 4 In the embodiment of this application, fault 2 is based on Fault detection plots of statistics;

[0029] Figure 5In the embodiment of this application, fault 3 is based on Fault detection plots of statistics;

[0030] Figure 6 In the embodiment of this application, fault 4 is based on Fault detection plots of statistics;

[0031] Figure 7 In the embodiment of this application, fault 5 is based on Fault detection plots of statistics;

[0032] Figure 8 In the embodiment of this application, fault 1 is based on Fault detection plots of statistics;

[0033] Figure 9 In the embodiment of this application, fault 2 is based on Fault detection plots of statistics;

[0034] Figure 10 In the embodiment of this application, fault 3 is based on Fault detection plots of statistics;

[0035] Figure 11 In the embodiment of this application, fault 4 is based on Fault detection plots of statistics;

[0036] Figure 12 In the embodiment of this application, fault 5 is based on Fault detection plots of statistics;

[0037] Figure 13 A schematic diagram of the structure of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0038] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0039] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0040] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0041] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0042] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0043] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0044] In response to the problem that current industrial process monitoring models lack interpretability, resulting in low credibility of process fault monitoring results, an embodiment of the present application provides an interpretable process monitoring method based on a deep state space model. By introducing a discrete part with practical significance into the latent variables of the deep state space model, the latent variable part is made understandable, and the deep state space model is trained by maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the mode as the model training loss, the interpretability of the model is significantly improved. Therefore, when the deep state space model is used to analyze and process online industrial process data, the fault monitoring results determined by the latent variable monitoring indicators and observation value monitoring indicators output by the inference network of the model are highly interpretable, thereby greatly improving the credibility of the industrial process fault monitoring results.

[0045] The interpretable process monitoring method based on the deep state space model provided in the embodiments of the present application is exemplarily described below with reference to specific embodiments.

[0046] like Figure 1 As shown, the interpretable process monitoring method based on the deep state space model provided in the embodiment of the present application includes the following steps:

[0047] Step 11: obtain industrial process data samples at multiple historical moments, and use a clustering algorithm to perform modal division on the obtained industrial process data samples to obtain modal labels corresponding to the industrial process data samples at each historical moment.

[0048] The specific data content of the industrial process data sample is determined by the process requirements and procedures of the industrial process. For example, if the industrial process is a potassium salt flotation process, the industrial process data sample includes real-time measurable variables such as weight, temperature, flow rate, and current at historical moments.

[0049] The industrial process data samples at the above-mentioned multiple historical moments include fault samples and normal samples. The above-mentioned modal division can be understood as working condition division, such as dividing each industrial process data sample into four working conditions: normal flotation, over-flotation, weak flotation, and under-flotation, and using the corresponding working condition as the modal label of the industrial process data sample.

[0050] In some embodiments of the present application, the industrial process data samples may be obtained by monitoring an industrial process.

[0051] In some embodiments of the present application, in order to improve training accuracy, after obtaining industrial process data samples at multiple historical moments, the obtained industrial process data samples can be preprocessed by removing abnormal data, filling in missing values, and normalizing the data. Then, the preprocessed industrial process data samples can be modally divided using the kmeans algorithm (the kmeans algorithm is a clustering algorithm) to obtain the modal labels corresponding to the industrial process data samples at each historical moment.

[0052] In some embodiments of this application, missing values and outliers are filled using the mean. Each dimension of the sample is then normalized using a standard score (Z-score) to ensure that the dimensional features follow a standard normal distribution. The preprocessed industrial process data samples are clustered, with a set of four cluster centers. Each sample is classified into one of four operating conditions.

[0053] The expression for modal division of industrial process data samples using the kmeans algorithm is:

[0054] X M =kmeans(X)

[0055] in, Represents the collected industrial process data samples, the superscript M indicates that the samples are clustered into the Mth category by the kmeans algorithm, and N represents the number of industrial process data samples.

[0056] Step 12: Segment the industrial process data samples carrying modal labels according to time windows to generate a training sample set.

[0057] In some embodiments of the present application, industrial process data samples carrying modality labels can be segmented according to time windows, where the time window length T = 20 and the window sliding step size is 1. All fault data and some normal data are selected as the test sample set, where 100 normal samples are retained before the starting point and after the ending point of each fault, and the remaining samples are used as the training sample set.

[0058] Step 13: construct a deep state space model. The latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part. The discrete part is used to reflect the working conditions of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system.

[0059] The specific form of the industrial system's operating conditions is determined by the industrial process. Incomprehensible system states are represented by high-dimensional features without specific classification or meaning. For example, if the industrial process is potash flotation, the operating conditions can be categorized as normal flotation, over-flotation, weak flotation, and under-flotation.

[0060] In some embodiments of the present application, the above-mentioned deep state space model is composed of two parts: a generation network and an inference network, where the generation network is defined as:

[0061] z t =[c t ,d t ]

[0062]

[0063]

[0064] Among them, z t represents the latent variable at time t; c t represents z t The continuous part is an incomprehensible hidden variable, which can reflect the state of the system according to the theory of deep state space model; d t represents z t The discrete part of can be assumed to be any human-understandable variable that can reflect the connotation of the industrial system, that is, the working condition (i.e., mode) corresponding to the industrial process data, such as the working condition in the industrial field (for example, the working condition of the flotation process can be divided into normal flotation, over-flotation, weak flotation, and under-flotation); Indicates is the mean; is the normal distribution of covariance; represents a transfer network, which is a deep network used to parameterize the mean μ and covariance ∑ of latent variables, which defines the nonlinear law of the evolution of latent variables over time; represents a deep network with the previous hidden variable as input and parameterized mean, represents the deep network used to parameterize the covariance matrix; μ represents the mean of the latent variable; z t1 represents the latent variable at time t1; ∑ represents the covariance of the latent variable; x t represents the industrial process data at the t-th moment, t = 1, 2, …, T, and T represents the total number of industrial process data; Indicates is a normal distribution with a mean; Represents the emission network, which is used to parameterize the deep network of the mean of the observed variable, which defines the transition from the latent variable to the observed value x t (i.e. the nonlinear mapping of the industrial process data at time t above), The covariance matrix of the output random variables does not participate in the training and testing process, so the covariance matrix is omitted; Represents a deep network that takes the current hidden variable as input and is used to parameterize and reconstruct the mean of the observation distribution.

[0065] It should be noted that c t Obeying Gaussian distribution, d t It obeys the 0 / 1 distribution, and the subscript t of the variable indicates the time corresponding to the variable.

[0066] Specifically, A single hidden layer multi-layer perceptron (MLP) with a gated structure is used, and its definition expression is as follows:

[0067]

[0068]

[0069]

[0070]

[0071] Among them, g t represents the gated feature, σ R,S Denotes the activation function σ R ,σ S is applied sequentially to the hidden layers of the MLP; z t1 represents the latent variable at time t1, h trepresents the candidate features extracted by MLP, σ R,T Denotes the activation function σ R ,σ T is applied sequentially to the hidden layers of the MLP, σ R ,σ S , T Represent "ReLU", "Sigmoid" and "Tanh" functions respectively; represents the mean of latent variables at time t, represents the corresponding mean of continuous latent variables, Indicates the mean value of the discrete latent variable before normalization; Represents the normalized discrete latent variable mean; weight matrix W μ and W ∑ Used to map hidden layer features into mean and covariance; b σ and b ∑ is the bias term; represents the deep network used to parameterize the covariance matrix. The addition of the gating mechanism makes Some linear features are retained, which is beneficial for flexible modeling of dynamic systems, especially when linear and nonlinear dynamics coexist. It is worth noting that if g t All elements are equal to 1, and the system can be degenerated into a linear system. In addition, After normalization by the softmax function It is still continuous, so after calculating the classification loss, its largest element still needs to be set to 1 and the other elements to 0.

[0072] Used to parameterize z t to x t The nonlinear mapping relationship is defined as:

[0073]

[0074]

[0075] Among them, p(x t |z t ) indicates that given z t Under the condition x t The distribution of σ S,T Denotes the activation function σ S ,σ T are applied sequentially to the hidden layers of the MLP.

[0076] The posterior latent variables obtained by the inference network of the above deep state space model It contains two parts, discrete and continuous, and uses bidirectional recurrent neural network parameterization in the inference network. The bidirectional recurrent neural network is:

[0077]

[0078]

[0079]

[0080] in, represents the features extracted by the backward recurrent neural network at time t, represents a backward recurrent neural network, represents the features extracted by the backward recurrent neural network at time t+1, represents the features extracted by the forward recurrent neural network at time t, represents a forward recurrent neural network, represents the features extracted by the forward recurrent neural network at time t-1, σ S,T Denotes the activation function σ S ,σ T is applied sequentially to the hidden layers of the MLP, Indicates is the mean, is the normal distribution of covariance, Represents the concatenation of the hidden layer features of the forward recurrent neural network, the hidden layer features of the backward recurrent neural network, and the observation value at the current moment. and Represents the feature h extracted by MLP t Perform a linear mapping, represents the mean of latent variables at time t, represents the corresponding mean of continuous latent variables, Represents the mean corresponding to the discrete latent variable; represents the parameter matrix used to map features to means, represents the bias vector, W ∑ represents the parameter matrix used to map features to covariance, b ∑ Represents the bias vector.

[0081] The main objective of model training is to maximize the likelihood function of the observation, which can be derived through the variational Bayesian method to obtain the variational lower bound:

[0082]

[0083] Among them, p θ (X) represents the likelihood function of the observation, X represents all observation samples, represents the derived variational lower bound, Expressing hope, represents the variational posterior, p θ (X|Z) represents the probability of observing X under given latent variables, KL represents the calculation of KL divergence, p θ (Z) represents the prior distribution of latent variables, represents the set of all latent variables. θ and They represent the parameters that need to be trained in the generation network and the inference network respectively.

[0084] Inference network is used to parameterize the variational posterior It is the key to optimizing the likelihood function using the variational Bayesian method. During the training process, the variational lower bound needs to be factorized into:

[0085]

[0086] in, represents z t The posterior distribution of θ (x t |z t ) indicates that given z t Observed x t The probability of represents z t The posterior distribution of θ (z1) represents the prior distribution of z1, represents z t The posterior distribution of represents z t The posterior distribution of θ (z t |z t-1 ) represents z t The prior distribution of , regularization means regularization, and KL means calculating KL divergence. Since the concept of KL divergence does not exist for the discrete part of the latent variable, the variational lower bound needs to be rewritten as:

[0087]

[0088] Therefore, during training, the joint distribution of the latent variables posterior is not of interest. Instead, we focus on how to parameterize the factorization of the joint distribution, namely:

[0089]

[0090] in, represents z t The posterior distribution of z 1:t-1 represents the latent variable at the previous t-1 moment.

[0091] As mentioned above, a bidirectional recurrent neural network is used to parameterize Through this network, an accurate approximation of the variational posterior can be obtained. Its specific definition is:

[0092]

[0093]

[0094]

[0095] in, represents the features extracted by the backward recurrent neural network at time t, represents a backward recurrent neural network, represents the features extracted by the backward recurrent neural network at time t+1, represents the features extracted by the forward recurrent neural network at time t, represents a forward recurrent neural network, represents the features extracted by the forward recurrent neural network at time t-1, σ S,T Denotes the activation function σ S ,σ T is applied sequentially to the hidden layers of the MLP, Indicates is the mean, is the normal distribution of covariance, and the arrow above indicates the direction of the recurrent neural network. Represents the concatenation of the hidden layer features of the forward recurrent neural network, the hidden layer features of the backward recurrent neural network, and the observation value at the current moment. and Represents the feature h extracted by MLP t Perform a linear mapping, where in, represents the mean of latent variables at time t, represents the corresponding mean of continuous latent variables, Represents the mean corresponding to the discrete latent variable; represents the parameter matrix used to map features to means, represents the bias vector, W ∑ represents the parameter matrix used to map features to covariance, b ∑ Represents the bias vector. So far, the construction of the deep state space model has been completed.

[0096] In step 14, the deep state space model is trained using the training sample set to obtain a trained deep state space model. During the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modality are used as the model training loss.

[0097] That is, the model's loss function consists of two parts: maximizing the log-likelihood of observations and minimizing the categorical cross-entropy loss of the modalities. Maximizing the log-likelihood loss of observations builds the overall framework for model parameter identification and state estimation. Minimizing the categorical cross-entropy loss of the modalities aims to ensure that the discrete latent variables estimated by the generative network and the inference network can correctly reflect the working conditions of the system. It is defined as:

[0098]

[0099] in, represents the cross entropy loss for working condition classification, Denotes the discrete part d t The dimension is also the number of kmeans modal clustering centers, d t Indicates the mode to which the sample at the tth moment obtained by kmeans belongs, which is represented in the form of one-hot encoding, T represents the length of the time window, Indicates d t The i-th element of express The i-th element of Indicates d t The estimated value of represents the probability that the true working condition label is i, represents the prior estimate of the working condition obtained by the transfer network in the generative network of the deep state space model, represents the posterior estimation condition obtained by the inference network, and X represents all observation samples in the time window.

[0100] The continuous latent variables (i.e., the continuous part) in the traditional deep state space model are incomprehensible. The introduction of discrete variables with practical significance (i.e., the discrete part) can make the latent variable part understandable. If the generative network can accurately predict the system working condition at the current moment based on the latent variables at the previous moment, then the generative network can accurately model the evolution law of the system. Similarly, if the inference network can accurately divide the current working condition, then the accuracy of the posterior probability distribution is higher. At the same time, the discrete latent variables corresponding to different modes will cause the inference network and the generative network to be in different excitation states, which is conducive to its modeling of process data with different distributions under multiple working conditions. In summary, in the process of training the deep state space model using the training sample set, the model training loss can be calculated by the following formula

[0101]

[0102] Among them, the discrete part of the latent variable d t There is no concept of KL divergence, and the variational lower bound only considers the continuous part of the latent variable c t , for the discrete part of the latent variable d t The constraint is Denotes the optimization of θ and Minimize the training loss, θ represents the learnable parameters of the generative network, represents the learnable parameters of the inference network, represents the variational lower bound.

[0103] In some embodiments of the present application, the trained deep state space model can process multimodal industrial process data. It can be understood as a state estimator that can determine whether the system (i.e., the industrial system to which the industrial process data belongs) evolves according to the original state or has a serious deviation by quantifying the difference between the estimated value and the true value.

[0104] Specifically, after the industrial process data obtained offline is input into the inference network of the trained deep state space model for processing, two monitoring indicators can be obtained to characterize whether the industrial system evolves according to the original state or has a serious deviation: the latent variable monitoring indicator and observation monitoring indicators

[0105]

[0106]

[0107] Among them, u t|t represents the mean of the variational posterior distribution of the inference network output, u t|t1 represents the mean of the prior distribution obtained by passing the network, It represents the mean of the parameterized reconstructed observation distribution using the latent variable posterior at the current moment as input. Quantifies the difference between the prior and posterior distributions of the latent variable, corresponding to the KL divergence term in the training loss. When the system is able to accurately model the dynamics, Quantify the difference between the reconstructed observations and the actual observations, which corresponds to the reconstruction term in the training loss. When the system is able to accurately model the dynamics, therefore, and It is a suitable detection indicator to reflect the adaptability of the model.

[0108] Step 15: Use the trained deep state space model to process the online industrial process data to obtain latent variable monitoring indicators and observation value monitoring indicators.

[0109] In some embodiments of the present application, industrial process data at the current monitoring moment and T1 moments before the current monitoring moment can be obtained. Specifically, fault analysis of the industrial system can be performed based on industrial process data at a preset number of moments (such as T, where T can be understood as the time window length of the industrial process data sample when training the deep state space model) including the current monitoring moment. It can be understood that the fault analysis results obtained based on these industrial process data can be understood as the fault analysis results at the current monitoring moment.

[0110] The specific data content of the industrial process data is determined by the process requirements and procedures of the industrial process. For example, if the industrial process is a potassium salt flotation process, the industrial process data includes real-time measurable variables such as weight, temperature, flow rate, and current.

[0111] In some embodiments of the present application, the above-mentioned online industrial process data can be obtained by monitoring the industrial process.

[0112] It should be noted that in order to improve the accuracy of the model output results, after obtaining the industrial process data, the industrial process data can be preprocessed by removing abnormal data, filling in missing values, and normalizing the data. Then, the preprocessed industrial process data can be input into the inference network of the trained deep state space model for processing to obtain latent variable monitoring indicators and observation value monitoring indicators.

[0113] Step 16: Determine whether there is a fault in the industrial system based on the latent variable monitoring indicators and the observed value monitoring indicators.

[0114] In some embodiments of the present application, an evaluation statistic of an industrial system can be determined based on latent variable monitoring indicators and observation value monitoring indicators; if the evaluation statistic meets a preset fault condition, it is determined that there is a fault in the industrial system; if the evaluation statistic does not meet the preset fault condition, it is determined that there is no fault in the industrial system.

[0115] The above evaluation statistics are mainly used to evaluate the performance of the system, including latent variation evaluation statistics and the observed evaluation statistic

[0116] In some embodiments of the present application, the formula Calculate the latent variation evaluation statistic By formula Calculate the observed evaluation statistic

[0117] Among them, ∑ z express The covariance matrix of x express The covariance matrix of .

[0118] In fact, the above formula and The Mahalanobis distance between the estimated value and the true value was calculated, and this evaluation statistic is optimal under the Gaussian distribution assumption.

[0119] In some embodiments of the present application, after calculating the evaluation statistics of the industrial system, it can be determined whether the evaluation statistics meet a preset fault condition, thereby facilitating the determination of whether the industrial system has a fault.

[0120] Specifically, the latent variation evaluation statistic can be determined Is it less than the hidden change monitoring threshold η z , and the observed evaluation statistic Is it less than the observation monitoring threshold η x ; If the latent variation evaluation statistic Less than the hidden change detection threshold η z , and the observed evaluation statistic Less than the observation monitoring threshold η x , it is determined that the evaluation statistic does not meet the preset fault condition; otherwise, it is determined that the evaluation statistic meets the preset fault condition. That is, as long as Greater than or equal to η z or Greater than or equal to η x , it is determined that the evaluation statistic meets the preset fault condition.

[0121] It should be noted that the above-mentioned hidden change monitoring threshold η z and observation monitoring threshold η x It can be determined using the kernel density estimation method. Kernel density estimation is a non-parametric statistical method used to estimate the probability density function from sample data. In the threshold determination process, kernel density estimation is used to model the probability density function of the test statistic of the training set and obtain the corresponding distribution function. Usually, the test statistic corresponding to the false alarm rate of 0.5% is taken as the threshold based on the distribution function. The test statistic is based on or The Mahalanobis distance from the origin is calculated, and under normal circumstances and Therefore, the evaluation statistics of normal industrial systems should be smaller than the monitoring thresholds of the above statistics.

[0122] It is worth mentioning that the interpretable process monitoring method provided in the embodiment of the present application significantly improves the interpretability of the model by introducing a discrete part with practical significance into the latent variables of the deep state space model to make the latent variable part understandable. As a result, when the deep state space model is used to analyze and process the industrial process data at the monitoring moment, the evaluation statistics of the industrial system determined by the latent variable monitoring indicators and the observation value monitoring indicators output by the inference network of the model are highly interpretable, thereby greatly improving the credibility of the industrial process fault monitoring results analyzed based on the evaluation statistics.

[0123] The above-mentioned explainable process monitoring method is exemplified below with reference to specific examples.

[0124] In this example, the industrial process is a potassium salt flotation process. In this industrial process, the above-mentioned interpretable process monitoring method includes the following steps:

[0125] Step S1, obtaining historical data of the potassium salt flotation process.

[0126] In this example, 24,348 temporal industrial process samples were acquired, each containing 89 features. Five types of faults were included, with 1,079, 445, 312, 611, and 267 fault samples, respectively. As shown in Table 1, industrial process data includes real-time measurable variables such as weight, temperature, flow rate, and current.

[0127]

[0128]

[0129] Table 1

[0130] In step S2, the historical data is preprocessed by removing abnormal data, filling missing values, and normalizing the data; the preprocessed historical data is modally divided using the kmeans algorithm to obtain the modal label corresponding to the data at each moment.

[0131] In this example, missing values and outliers are filled with the mean. Z-score normalization is then performed on each dimension of the sample to ensure that the dimensional features follow a standard normal distribution. The preprocessed historical industrial process data is clustered, with a set number of cluster centers of four. Each sample is classified into one of four operating conditions.

[0132] In step S3, the industrial process data with modality labels are segmented according to time windows to generate a training sample set.

[0133] In this example, the preprocessed historical industrial process data is segmented into time windows, where the time window length is T = 20 and the window sliding step is 1. The above fault data and some normal data are selected as the test sample set, where 100 normal samples are retained before the starting point and after the end point of each fault, and the remaining 20,615 samples are used as the training sample set.

[0134] Step S4, offline construction and training Figure 2 The interpretable deep state-space model shown is trained based on the training sample set by maximizing the log-likelihood of the observation and minimizing the classification error of the modality as the loss.

[0135] in, Figure 2 x1 represents the initial observation value, c1 represents the initial continuous latent variable, d1 represents the initial discrete latent variable, x t-1 represents the observed variable at time t-1, c t -1 represents the discrete latent variable at time t-1, d t-1 represents the discrete part of the latent variable at time t-1, x t represents the observed variable at time t, c t represents the discrete latent variable at time t, d t Represents the discrete part of the hidden variables at the moment, x t+1 represents the observed variable at time t+1, c t+1 represents the discrete latent variable at time t+1, d t+1 Represents the discrete part of the latent variable at time t+1, S represents the sampling operation, and the value range of t is 1~T.

[0136] Step S41, in this example, Used to parameterize z t-1 to z t The nonlinear evolution law of is defined as:

[0137]

[0138] Among them, p(z t |z t-1 ) represents the latent variable z t Specifically, the transfer network adopts a single hidden layer MLP with a gating structure, which is defined as follows:

[0139]

[0140]

[0141]

[0142]

[0143] Among them, σ R,S Denotes the activation function σ R ,σ S is applied sequentially to the hidden layers of the MLP; the weight matrix and Initialized to the unit matrix; b μ and b ∑ is the bias term. The addition of the gating mechanism allows the propagation matrix to retain certain linear characteristics, which is conducive to the flexible modeling of dynamic systems, especially when linear and nonlinear dynamics coexist. It is worth noting that if g t If all elements are equal to 1, the system can be degenerated into a linear system. The dimension of the latent variable is usually smaller than the dimension of the observed variable. In this embodiment, the dimension of the latent variable is defined as 34, of which the dimension of the discrete latent variable is 4, and the rest are continuous latent variables. The specific structure is The specific structure is in Represents the connection of the neural network, and the number represents the number of neurons. It should be noted that d′ t After softmax normalization, it is still continuous. Therefore, after calculating the classification loss, it is still necessary to set its maximum element to 1 and other elements to 0 to convert it into a one-hot encoding form.

[0144] Transmitter Network Used to parameterize z t to x t The nonlinear mapping relationship is defined as:

[0145]

[0146]

[0147] Among them, MLPσ s,t The specific structure is

[0148] In step S42, in this example, a bidirectional recurrent neural network (RNN) is used to parameterize the variational posterior distribution ( Figure 2 The middle dashed line represents the RNN used to parameterize the variational posterior distribution, which is specifically defined as:

[0149]

[0150]

[0151]

[0152] The RNN hidden layer contains 100 neurons, and the arrow above indicates the direction of the recurrent neural network. Represents the concatenation of the hidden layer features of the forward recurrent neural network, the hidden layer features of the backward recurrent neural network, and the observation value at the current moment. The specific structure is and Represents h t A linear map of

[0153] Step S43, in this example, the loss function guiding network training is defined as:

[0154]

[0155] After factorization, the analytical form of the variational lower bound exists, so the gradient descent method can be used directly to optimize the parameters of the transfer network and the inference network. In this example, the Adam optimizer is selected with a learning rate of 1e -4 , the batch size is 64.

[0156] Step S5: construct statistics based on the latent variables of the trained model and the reconstructed observations and

[0157] In step S51, in this example, the deep state space model is a state estimator that can determine whether the system is evolving according to the original state or has seriously deviated by quantifying the difference between the estimated value and the true value. Therefore, two monitoring indicators can be defined as:

[0158]

[0159]

[0160] where u t|t is the mean of the variational posterior distribution output by the inference network, u t|t-1 is the mean of the prior distribution obtained by passing it through the network. Quantifies the difference between the prior and posterior distributions of the latent variable, corresponding to the KL divergence term in the training loss. When the system is able to accurately model the dynamics, Quantify the difference between the reconstructed observations and the actual observations, which corresponds to the reconstruction term in the training loss. When the system is able to accurately model the dynamics, therefore, and It is a suitable detection indicator to reflect the adaptability of the model.

[0161] Step S52: construct test statistics based on these detection indicators to further evaluate the performance of the system:

[0162]

[0163]

[0164] where ∑ z and ∑ x Respectively represent and The covariance matrix of .

[0165] Step S6, using kernel density estimation to determine the statistics and The monitoring threshold η z ,η x ;

[0166] In this example, kernel density estimation is used to determine the control limits η for two statistics. z ,η x The false alarm rate on the training set is set to 0.5% when using the two statistics for fault diagnosis. This embodiment uses 20615 samples as the training set, but the determination of the threshold is only related to the target domain samples.

[0167] Step S61: Input the samples in the training set into the trained deep state space model to obtain prior latent variables, posterior latent variables and estimated observation values.

[0168] Step S62 , calculating two test statistics according to the difference between the prior latent variable and the posterior latent variable and the difference between the estimated observation value and the true observation value.

[0169] Step S63 , sorting the two test statistics respectively and using kernel density estimation to calculate the probability density function of the test statistic corresponding to each sample.

[0170] Step S64: Calculate the probability density distribution for each probability density sequence, and determine the control limit according to the probability density distribution and the set false alarm rate.

[0171] Step S65 , input the five fault samples in the test set into the common feature extraction network and the special feature extraction network to calculate two test statistics respectively.

[0172] Step S7: Preprocess the actual industrial process online data according to the above method, input it into the trained model, and calculate the statistics online. like and It is judged that there is no fault in the current production process, otherwise there is a fault.

[0173] In this example, Figure 3 , 4, 5, 6, 7 respectively show the five types of faults based on Statistics monitoring result chart, Figure 8 , 9, 10, 11, 12 respectively show the five types of faults based on The monitoring result diagram of the statistics. The vertical line represents the dividing line between normal samples and fault samples, the middle part of the two vertical lines is the fault sample, and the two sides are normal samples. The horizontal dotted line represents the control limit (i.e., monitoring threshold) set according to the kernel density estimation. Table 2 shows the comparison of the fault detection rate of the comparison method and the present application. Among them, VAE (Variational Autoencoder) in Table 2 is a variational autoencoder, AAE (Adversarial Autoencoder) is an adversarial autoencoder, LSTM (Long Short-Term Memory) is a long short-term memory network, DMM (Deep Markov Model) is a deep Markov model, T 2 A statistic representing the comparison of means.

[0174]

[0175] Table 2

[0176] Prior working condition classification accuracy Posterior working condition classification accuracy This application 0.996 0.996

[0177] Table 3

[0178] Table 2 specifically records the fault detection comparison results of this application and four other advanced methods. The fault false alarm rate is determined by the normal samples in the test sample set. In this embodiment, the average false alarm rate of dynamic VAE is higher than the expected 0.5%, and the fault omission rate is high, and the overall fault diagnosis effect is the worst. The fault diagnosis effect of this application is the best, followed by LSTM, DMM and dynamic AAE. The experimental results show that with the assistance of working condition information, this application can better model the evolution law of the system.

[0179] In this example, latent variables are defined to represent the working conditions at the current moment. Table 3 shows the accuracy of working condition division through the generation network and the inference network of this application. Both the generation network and the inference network simultaneously complete the tasks of maximizing the observation likelihood and working condition division, and have the advantages of multi-task learning. The prior working condition classification accuracy and the posterior working condition classification accuracy in Table 3 are both 0.996, representing the working condition division accuracy of the generation network and the inference network, respectively. It can be seen that both the generation network and the inference network can well estimate the working condition of the system at the current moment, which explains why the fault detection effect of this application is better.

[0180] In summary, the interpretable process monitoring method provided by the embodiments of the present application has the following advantages:

[0181] (1) The model makes the latent variables understandable by introducing discrete variables with practical significance. The applicability of the model can then be judged by observing these discrete latent variables in real time, which significantly improves the credibility of the model.

[0182] (2) The meaning of the discrete latent variable in this application is defined as the current working condition, which is represented in the form of one-hot encoding; the discrete latent variables corresponding to different working conditions put the generative model and the inference model into different initial excitation states in order to model different distribution data;

[0183] (3) This application follows the framework of the state-space model, using a deep network to replace the linear matrix in the traditional model, and taking noise into account in the model construction process, which can effectively deal with the dynamics and high noise of industrial process data;

[0184] (4) Based on the characteristics of the proposed deep state space model, this application establishes two different statistical data to monitor the state space and observation space, which can obtain more fault information.

[0185] The interpretable process monitoring device based on the deep state space model provided by the present application is exemplarily described below in conjunction with specific embodiments.

[0186] The interpretable process monitoring device provided in the embodiment of the present application includes:

[0187] An acquisition module is used to acquire industrial process data samples at multiple historical moments, and use a clustering algorithm to perform modal division on the acquired industrial process data samples to obtain modal labels corresponding to the industrial process data samples at each historical moment;

[0188] A generation module is used to segment the industrial process data samples carrying modality labels according to time windows to generate a training sample set;

[0189] A construction module is used to construct a deep state space model. The latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part. The discrete part is used to reflect the working conditions of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system.

[0190] A training module is used to train the deep state space model using the training sample set to obtain a trained deep state space model; wherein, during the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modalities are used as the model training loss;

[0191] The processing module is used to process the online industrial process data using the trained deep state space model to obtain latent variable monitoring indicators and observation value monitoring indicators;

[0192] The fault analysis module is used to determine whether there is a fault in the industrial system based on the latent variable monitoring indicators and the observed value monitoring indicators.

[0193] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0194] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0195] like Figure 13 As shown, an embodiment of the present application provides a terminal device, such as Figure 13 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 13 Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above method embodiments when executing the computer program D102.

[0196] Specifically, when the processor D100 executes the computer program D102, it introduces a discrete part with practical significance into the latent variable of the deep state space model to make the latent variable part understandable, and trains the deep state space model by maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the mode as the model training loss, thereby significantly improving the interpretability of the model. Therefore, when the deep state space model is used to analyze and process online industrial process data, the fault monitoring results determined by the latent variable monitoring indicators and observation value monitoring indicators output by the inference network of the model are highly interpretable, thereby greatly improving the credibility of the industrial process fault monitoring results.

[0197] The processor D100 may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0198] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card, etc. equipped on the terminal device D10. Furthermore, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program. The memory D101 may also be used to temporarily store data that has been output or is to be output.

[0199] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of this application. Their specific functions and technical effects can be found in the method embodiment section and will not be repeated here.

[0200] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0201] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned various method embodiments can be implemented.

[0202] An embodiment of the present application provides a computer program product. When the computer program product is run on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0203] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to an interpretable process monitoring device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0204] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0205] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0206] In the embodiments provided in this application, it should be understood that the disclosed devices / network equipment and methods can be implemented in other ways. For example, the device / network equipment embodiments described above are merely illustrative. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0207] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0208] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An interpretable process monitoring method based on a deep state space model, characterized in that: include: Acquire industrial process data samples at multiple historical moments, and perform modal division on the acquired industrial process data samples using a clustering algorithm to obtain a modal label corresponding to each industrial process data sample at the historical moment; Segmenting the industrial process data samples carrying the modality labels according to time windows to generate a training sample set; Constructing a deep state space model; the latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part, the discrete part is used to reflect the working condition of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system; The deep state space model is trained using the training sample set to obtain a trained deep state space model; wherein, during the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modalities are used as the model training loss; The trained deep state space model is used to process online industrial process data to obtain latent variable monitoring indicators and observation value monitoring indicators; determining whether the industrial system has a fault according to the latent variable monitoring indicator and the observed value monitoring indicator; The generative network of the deep state space model is: z t =[c t ,d t ] Among them, z t represents the latent variable at time t, c t represents z t The continuous part, d t represents z t The discrete part of Indicates is the mean, is the normal distribution of covariance, represents the delivery network, represents a deep network with the previous hidden variable as input and parameterized mean, represents the deep network used to parameterize the covariance matrix, μ represents the mean of the latent variable, z t-1 represents the latent variable at time t-1, ∑ represents the latent variable covariance, x t represents the industrial process data at the t-th moment, t=1,2,…,T, T represents the total number of industrial process data, Indicates is a normal distribution with a mean of represents the transmitting network, represents a deep network that takes the current hidden variable as input and is used to parameterize and reconstruct the mean of the observation distribution; A single hidden layer MLP with a gating structure is adopted.

2. The interpretable process monitoring method based on deep state space model according to claim 1 is characterized in that: The definition expression is: Among them, g t represents the gated feature, σ R,S Denotes the activation function σ R ,σ S is applied sequentially to the hidden layers of the MLP; z t-1 represents the latent variable at time t-1, h t represents the candidate features extracted by MLP, σ R,T Denotes the activation function σ R ,σ T is applied sequentially to the hidden layers of the MLP, σ R ,σ S ,σ T Represent ReLU, Sigmoid and Tanh functions respectively; represents the mean of latent variables at time t, represents the corresponding mean of continuous latent variables, Indicates the mean value of the discrete latent variable before normalization; Represents the normalized discrete latent variable mean; weight matrix W μ and W ∑ Used to map hidden layer features into mean and covariance; b μ and b ∑ is the bias term; represents a deep network for parameterizing the covariance matrix; The posterior latent variables obtained by the inference network of the deep state space model It contains two parts, discrete and continuous, and uses bidirectional recurrent neural network parameterization in the inference network. The bidirectional recurrent neural network is: in, represents the features extracted by the backward recurrent neural network at time t, represents a backward recurrent neural network, represents the features extracted by the backward recurrent neural network at time t+1, represents the features extracted by the forward recurrent neural network at time t, represents a forward recurrent neural network, represents the features extracted by the forward recurrent neural network at time t-1, σ S,T Denotes the activation function σ S ,σ T is applied sequentially to the hidden layers of the MLP, Indicates is the mean, is the normal distribution of covariance, Represents the concatenation of the hidden layer features of the forward recurrent neural network, the hidden layer features of the backward recurrent neural network and the current moment observation value, and Represents the feature h extracted by MLP t Perform a linear mapping, represents the mean of latent variables at time t, represents the corresponding mean of continuous latent variables, Represents the mean corresponding to the discrete latent variable; represents the parameter matrix used to map features to means, represents the bias vector, W ∑ represents the parameter matrix used to map features to covariance, b ∑ Represents the bias vector.

3. The interpretable process monitoring method based on deep state space model according to claim 2, characterized in that: The determining whether the industrial system has a fault according to the latent variable monitoring indicator and the observed value monitoring indicator includes: Determining an evaluation statistic of the industrial system according to the latent variable monitoring indicator and the observed value monitoring indicator; If the evaluation statistic satisfies a preset fault condition, determining that a fault exists in the industrial system; If the evaluation statistic does not meet the preset fault condition, it is determined that there is no fault in the industrial system.

4. The interpretable process monitoring method based on deep state space model according to claim 3 is characterized in that: The evaluation statistics include latent variation evaluation statistics and observation evaluation statistics; Determining the evaluation statistic of the industrial system according to the latent variable monitoring indicator and the observed value monitoring indicator includes: By formula Calculate the latent variation evaluation statistic By formula Calculate the observed evaluation statistic in, represents the latent variable monitoring indicator, u t|t represents the mean of the variational posterior distribution of the inference network output, u t|t-1 represents the mean of the prior distribution obtained by the transfer network, ∑z represents The covariance matrix of represents the observation value monitoring indicator, represents a deep network that takes the current hidden variable posterior as input and is used to parameterize and reconstruct the mean of the observation distribution, ∑ x express The covariance matrix of .

5. The interpretable process monitoring method based on deep state space model according to claim 4 is characterized in that: After the step of determining the evaluation statistic of the industrial system based on the latent variable monitoring indicator and the observed value monitoring indicator, the interpretable process monitoring method further includes: Determine the latent variation evaluation statistic Is it less than the hidden change monitoring threshold η z , and the observation evaluation statistic Is it less than the observation monitoring threshold η x ; If the latent variation evaluation statistic Less than the hidden change detection threshold η z , and the observation evaluation statistic Less than the observation monitoring threshold η x , it is determined that the evaluation statistic does not meet the preset fault condition; Otherwise, it is determined that the evaluation statistic meets a preset fault condition.

6. The interpretable process monitoring method based on deep state space model according to claim 5, characterized in that: In the process of training the deep state space model using the training sample set, the formula Calculate model training loss Among them, the discrete part of the latent variable d t There is no concept of KL divergence, and the variational lower bound only considers the continuous part of the latent variable c t , for the discrete part of the latent variable d t The constraint is Denotes the optimization of θ and Minimize the training loss, θ represents the learnable parameters of the generative network, represents the learnable parameters of the inference network, represents the variational lower bound, M represents the discrete part d t The dimension of , T represents the time window length, Indicates d t The i-th element of express The i-th element of Indicates d t The estimated value of represents the probability that the true working condition label is i, represents the prior estimated working condition obtained by the transfer network in the generative network of the deep state space model, represents the posterior estimation condition obtained by the inference network, and X represents all observation samples in the time window.

7. The interpretable process monitoring method based on deep state space model according to claim 1, characterized in that: The method of performing modal division on the acquired industrial process data samples using a clustering algorithm to obtain modal labels corresponding to the industrial process data samples at each historical moment includes: Preprocessing the acquired industrial process data samples; The kmeans algorithm is used to perform modal division on the preprocessed industrial process data samples to obtain the modal labels corresponding to the industrial process data samples at each historical moment.

8. An interpretable process monitoring device based on a deep state space model, characterized in that: include: An acquisition module is used to acquire industrial process data samples at multiple historical moments, and perform modal division on the acquired industrial process data samples using a clustering algorithm to obtain a modal label corresponding to each industrial process data sample at the historical moment; A generation module, configured to segment the industrial process data samples carrying the modality labels according to time windows to generate a training sample set; A construction module for constructing a deep state space model; the latent variables of the generative network of the deep state space model include an understandable discrete part and an incomprehensible continuous part, the discrete part is used to reflect the working condition of the industrial system to which the industrial process data belongs, and the continuous part is used to reflect the state of the industrial system; A training module, configured to train the deep state space model using the training sample set to obtain a trained deep state space model; wherein, during the training process, maximizing the log-likelihood of the observation and minimizing the classification cross entropy of the modalities are used as the model training loss; The processing module is used to process the online industrial process data using the trained deep state space model to obtain latent variable monitoring indicators and observation value monitoring indicators; a fault analysis module, configured to determine whether the industrial system has a fault based on the latent variable monitoring indicator and the observed value monitoring indicator; The generative network of the deep state space model is: z t =[c t ,d t ] Among them, z t represents the latent variable at time t, c t represents z t The continuous part, d t represents z t The discrete part of Indicates is the mean, is the normal distribution of covariance, represents the delivery network, represents a deep network with the previous hidden variable as input and parameterized mean, represents the deep network used to parameterize the covariance matrix, μ represents the mean of the latent variable, z t-1 represents the latent variable at time t-1, ∑ represents the latent variable covariance, x t represents the industrial process data at the t-th moment, t=1,2,…,T, T represents the total number of industrial process data, Indicates is a normal distribution with a mean of represents the transmitting network, represents a deep network that takes the current hidden variable as input and is used to parameterize and reconstruct the mean of the observation distribution; A single hidden layer MLP with a gating structure is adopted.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the interpretable process monitoring method based on the deep state space model according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the interpretable process monitoring method based on the deep state space model according to any one of claims 1 to 7 is implemented.

Images