Quality-related fault detection method and device based on dual-task decoupling network

Abstract

The application relates to the technical field of industrial process detection, and discloses a quality correlation fault detection method and device based on a double-task decoupling network, which comprises the following steps: collecting observation variable time series data of a continuous industrial process; performing time series feature extraction on the observation variable time series data to obtain global context features; inputting the global context features into a double-task decoupling network which has been pre-trained offline; projecting the global context features through a quality nonlinear mapping layer and a process nonlinear mapping layer which are arranged in parallel in the double-task decoupling network to output decoupled quality-related features and process-related features; determining online monitoring statistics based on the quality-related features; and determining whether a quality correlation fault occurs based on the online monitoring statistics and pre-stored normal working condition control limits. The application can weaken the influence of normal process disturbance on fault detection results, is helpful to reduce false positives, and improves fault detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial process inspection technology, specifically to a method and apparatus for detecting quality-related faults based on a dual-task decoupling network. Background Technology

[0002] In continuous industrial production processes, key quality indicators that directly determine product quality are often constrained by factors such as testing equipment and testing time, making online real-time monitoring difficult. In actual production, it is often necessary to rely on process observation variables such as temperature, pressure, and flow rate to indirectly characterize the real-time quality status of the product, such as purity and concentration. This data-driven fault detection and diagnosis approach has become an important technical route for online quality monitoring of complex industrial processes.

[0003] However, modern continuous industrial systems generally exhibit multiple complex characteristics, such as strong variable coupling, process nonlinearity, dynamic and time-varying operating conditions, and long process time delays, posing significant challenges to quality status identification, quality-related anomaly identification, and fault tracing and diagnosis. Especially under closed-loop control conditions, normal process disturbance characteristic signals and quality-related fault characteristic signals in sensor-acquired data are easily mixed. That is, under normal disturbances such as raw material fluctuations and production load switching, the closed-loop control system adjusts the manipulated variables to suppress process fluctuations, resulting in the sensor-acquired observation data simultaneously containing normal adjustment responses and abnormal operating condition information.

[0004] In such scenarios, if we still follow the traditional method and monitor based solely on the amplitude changes of observed variables or low-dimensional statistical characteristics, it is easy to lead to a high false alarm rate in fault detection. Even if anomalies are identified, it is impossible to accurately trace the source of the fault that caused the quality defects, making root cause diagnosis difficult. Summary of the Invention

[0005] This invention provides a quality-related fault detection method and device based on a dual-task decoupling network to solve the problem in the prior art that it is difficult to separate the characteristics of normal process disturbances from the characteristics of quality-related faults, resulting in a high false alarm rate and difficulty in accurately locating the root cause of the fault.

[0006] In a first aspect, the present invention provides a quality correlation fault detection method based on a dual-task decoupling network, the method comprising: Collect time-series data of observed variables in continuous industrial processes; Temporal features are extracted from the time-series data of the observed variables to obtain global context features; The global context features are input into a pre-trained offline dual-task decoupling network. The global context features are then projected onto the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, and the decoupled quality-related features and process-related features are output. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss and geometric orthogonal constraint loss to promote the separation and representation of quality-related features and process-related features in the latent space. Based on quality-related characteristics, online monitoring statistics are determined; Based on online monitoring statistics and pre-stored normal operating condition control limits, it is determined whether a quality-related fault has occurred.

[0007] In one alternative implementation, the construction of the pre-trained offline dual-task decoupling network includes: Construct a training sample set, which includes: a sequence of process observation variables under normal operating conditions of a continuous industrial process and the corresponding real quality state vector; The network infrastructure is constructed, which includes an initial quality nonlinear mapping layer and an initial process nonlinear mapping layer set in parallel, as well as a quality supervision branch network and a process reconstruction branch network set in parallel. The output of the initial quality nonlinear mapping layer is connected to the quality supervision branch network and the process reconstruction branch network, respectively, and the output of the initial process nonlinear mapping layer is connected to the process reconstruction branch network. Construct a joint loss function composed of a weighted average of quality supervision loss, reconstruction loss, and geometric orthogonality constraint loss; Iterative training of the network infrastructure is carried out based on the training sample set until the joint loss function converges, resulting in a dual-task decoupled network.

[0008] In one optional implementation, the network infrastructure is iteratively trained based on the training sample set until the joint loss function converges, resulting in a dual-task decoupled network, including: Output latent quality features through an initial quality nonlinear mapping layer; output latent process features through an initial process nonlinear mapping layer; The quality supervision regression constraint task is executed. The potential quality features are input into the quality supervision branch network to obtain the quality state estimate. The quality supervision loss is calculated by combining the real quality state vector. The reconstruction constraint task is executed. The latent features of quality and the latent features of process are concatenated and then input into the process reconstruction branch network to obtain the reconstruction result of the original input matrix. The reconstruction loss is then calculated by combining the original input matrix. Geometric orthogonal constraint loss is constructed based on the cosine similarity between latent quality features and latent process features; Based on the joint loss function consisting of quality supervision loss, reconstruction loss, and geometric orthogonal constraint loss term, all parameters of the network infrastructure are iteratively updated until the joint loss function converges, resulting in a dual-task decoupled network.

[0009] In one optional implementation, online monitoring statistics are determined based on quality-related characteristics, including: After the dual-task decoupled network completes offline training, normal quality-related reference features are extracted for all normal working condition samples. Based on the reference features related to normal quality, determine the feature mean vector and covariance matrix; Based on quality-related features, feature mean vectors, and covariance matrices, online monitoring statistics are calculated.

[0010] In an optional implementation, when a quality-related failure is determined to have occurred, the method further includes: The quality-related features and process-related features output by the dual-task decoupling network are concatenated and then input into the process reconstruction branch network to obtain the reconstruction result of the dual-task decoupling network input matrix. Based on the reconstruction results, the absolute contribution index of each process observation variable during the fault period is determined, and the relative contribution ratio of each preset process observation variable is calculated based on the absolute contribution index. All process observation variables are sorted from largest to smallest according to their relative contribution percentage, and the top K process observation variables are selected to form a candidate variable set, where K > 0 and K is a positive integer; Based on the candidate variable set, determine the process variables of the root cause of the fault.

[0011] In one alternative implementation, the root cause process variables are determined based on the candidate variable set, including: For any pair of candidate variables in the candidate variable set, construct a restricted prediction model and a full prediction model based on a long short-term memory network, respectively. Based on the comparison of the prediction residuals between the restricted prediction model and the full prediction model, the causal strength between the candidate variables is calculated. Determine the net causal flow based on the causal strength among candidate variables; Based on the net causal flow, determine the net influence score of each candidate variable; The candidate variable with the largest net influence score is identified as the root cause process variable of the failure.

[0012] In one alternative implementation, the method further includes: Based on the net causal flow of each candidate variable, the fault propagation path is determined.

[0013] In one optional implementation, time-series features are extracted from the observed variable time-series data to obtain global context features, including: The time series data of the observed variables are divided into multivariate time windows of fixed duration, and a multivariate time window input matrix is ​​constructed. The multivariate time window input matrix is ​​input into the shared temporal encoder to obtain the temporal hidden features corresponding to multiple time steps; Temporal attention weights are assigned to the temporal hidden features corresponding to each time step based on the temporal attention mechanism, and the temporal hidden features corresponding to each time step are weighted and aggregated based on the temporal attention weights to obtain global context features.

[0014] Secondly, the present invention provides a quality correlation fault detection device based on a dual-task decoupling network, the device comprising: The data acquisition module is used to collect time-series data of observed variables in continuous industrial processes; The extraction module is used to extract time-series features from the observed variable time-series data to obtain global context features; The separation module is used to input global context features into a pre-trained offline dual-task decoupling network. The global context features are projected onto the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, and the decoupled quality-related features and process-related features are output. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss and geometric orthogonal constraint loss to promote the separation of quality-related features and process-related features in the latent space. The statistics module is used to determine online monitoring statistics based on quality-related characteristics; The diagnostic module is used to determine whether a quality-related fault has occurred based on online monitoring statistics and pre-stored normal operating condition control limits.

[0015] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the quality association fault detection method based on a dual-task decoupling network described in the first aspect or any corresponding embodiment thereof.

[0016] The beneficial effects of this invention are as follows: The quality-related fault detection method provided by this invention is based on a dual-task decoupling network. During online monitoring, real-time process data is constructed into a real-time input matrix and input into the trained dual-task decoupling network to obtain the quality-related features at the current moment. Based on these quality-related features, an online monitoring statistic is determined and compared with the normal operating condition control limit. When the online monitoring statistic exceeds the normal operating condition control limit, a quality-related fault is determined to have occurred, and an alarm is triggered. This invention promotes the separation and representation of normal process disturbance-related information and quality fault-related information in the latent space through the dual-task decoupling network. Online monitoring uses features with stronger quality correlation to construct statistics, thereby reducing the impact of process background disturbances such as raw material fluctuations and controller adjustments on the fault detection results, helping to reduce false alarms and improve fault detection accuracy.

[0017] The quality-related fault detection method provided by this invention, after triggering an alarm, calculates the reconstruction residual contribution of each process variable based on the reconstruction results output by the process reconstruction branch, and screens a set of candidate variables; performs causal inference based on long short-term memory networks on the variable pairs in the candidate variable set, calculates the causal strength and net causal flow between the candidate variables, and further calculates the net influence score of each candidate variable; identifies the candidate variable with the largest net influence score as the root cause variable of the fault, and outputs the fault propagation path analysis results based on the causal relationship between the candidate variables. Compared with relying solely on a single quality prediction model or traditional statistical monitoring methods, this invention can not only identify whether an anomaly has occurred, but also continue to complete candidate variable screening, propagation relationship analysis, and source variable location after an alarm, thereby improving the method's comprehensive diagnostic capability in complex industrial scenarios. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the first process of a quality correlation fault detection method based on a dual-task decoupling network according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the data dimensionality reduction and mapping process of the shared feature encoder according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a dual-task decoupling network according to an embodiment of the present invention; Figure 4This is a schematic diagram of the fault propagation map analysis and root cause localization procedure based on LSTM causal inference according to an embodiment of the present invention; Figure 5 This is the PCA model under the IDV(10) condition according to an embodiment of the present invention. Schematic diagram of monitoring statistics; Figure 6 This is the QD-DTDN model under the IDV(10) condition according to an embodiment of the present invention. Schematic diagram of monitoring statistics; Figure 7 The mass subspace under IDV(11) perturbation according to an embodiment of the present invention Schematic diagram of monitoring statistics; Figure 8 This is the process subspace under IDV(11) perturbation according to an embodiment of the present invention. Schematic diagram of monitoring statistics; Figure 9 This is a schematic diagram of the IDV(1) fault propagation path topology according to an embodiment of the present invention; Figure 10 This is a structural block diagram of a quality correlation fault detection device based on a dual-task decoupling network according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0022] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] Currently, industrial production lines are all controlled by automated closed-loop control systems. Under normal circumstances, slight changes in raw material quality and adjustments in production capacity are considered normal production fluctuations. The automated closed-loop control system automatically adjusts operating parameters such as valves and speeds to bring the process back to a stable range. The resulting changes in sensor data are considered normal process disturbances. However, hardware / process faults on the production line, such as pipe blockages and reaction parameter drift, which directly lead to substandard products, are considered quality-related faults, and sensor data will also change. For example, if the raw material concentration decreases, the controller automatically opens the feed valve, causing the flow sensor reading to rise; simultaneously, if a partial blockage occurs in the feed pipe (a real fault that will cause substandard product purity), the flow sensor reading will fall. Both increases and decreases will trigger an alarm. In this case, both effects are present simultaneously, and the data detected by the flow sensor is mixed. Simply looking at the numerical changes makes it impossible to distinguish whether the cause is normal controller adjustment or equipment failure. This not only leads to false alarms but also makes it impossible to trace the source of the quality defect.

[0024] In view of this, according to an embodiment of the present invention, a quality correlation fault detection method based on a dual-task decoupling network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] This embodiment provides a quality correlation fault detection method based on a dual-task decoupling network, which can be used in servers, terminals, mobile terminals, etc. Figure 1 This is a flowchart of a quality correlation fault detection method based on a dual-task decoupling network according to an embodiment of the present invention, as shown below. Figure 1 As shown, the process includes the following steps: Step S101: Collect time series data of observed variables for continuous industrial processes.

[0026] Continuous industrial processes can be closed-loop continuous industrial production processes such as food processing, chemical processing, and metallurgy. For continuous industrial processes, distributed sensor arrays can be deployed to collect real-time time-series measurements of process variables throughout the entire process. These observed variables can be physical process parameters that can be collected in real-time at the production site, including temperature, pressure, flow rate, liquid level, stirring speed, valve opening, etc., but excluding finished product quality indicators. By continuously collecting data at a fixed sampling period t, the original time-series sequence of the observed variables can be obtained. .

[0027] Step S102: Extract time series features from the observed variable time series data to obtain global context features.

[0028] In some optional implementations, step S102 above includes: Step S1021: Divide the time series data of the observed variables into multivariate time windows of fixed duration and construct the multivariate time window input matrix.

[0029] A fixed sliding time window length T is set, and the continuous observation time series data is segmented using a sliding truncation method. Each window contains all observed variable values ​​from historical time t-T+1 to the current time t. The observation vectors of the T times within the window are concatenated row by row to form the multivariate time window input matrix corresponding to time t. , Defined as , where m is the total number of observed variables.

[0030] Step S1022: Input the multivariate time window input matrix into the shared temporal encoder to obtain the temporal hidden features corresponding to multiple time steps.

[0031] To achieve a nonlinear mapping from the original observation space to the potential context space, a shared temporal encoder is constructed. ,in, For shared encoder parameters. The shared temporal encoder uses a Long Short-Term Memory (LSTM) network to recursively encode the multivariate time window input matrix along the time axis to extract the temporal dependencies and process dynamic correlation information between time steps within the time window.

[0032] Reference Figure 2 As shown, the multivariate time window is input into the matrix. (i.e., the original high-dimensional input matrix) Input a shared timing encoder, and input a multivariable time window matrix along the time axis. Recursive encoding is performed to obtain the temporal hidden features within the time window: .

[0033] in, Representing a Long Short-Term Memory (LSTM) network, with a multivariable time window input matrix. For input, use encoder parameters Perform forward computation; Represents the set of temporal hidden features at all times within the window; This represents the temporal hidden features of each time step within the window.

[0034] Step S1023: Based on the time attention mechanism, assign time attention weights to the temporal hidden features corresponding to each time step, and perform weighted aggregation on the temporal hidden features corresponding to each time step based on the time attention weights to obtain global context features.

[0035] In the backend of the shared temporal encoder, a set of temporal attention weights, learned offline, are assigned to the temporal hidden features corresponding to each time step in the temporal hidden state sequence. Based on the weighted summation and fusion of the temporal hidden features corresponding to each time step, a global context feature that incorporates full-window and multi-level temporal dynamic information is obtained. , .

[0036] in, This represents the shared temporal encoder and temporal attention mechanism for the multivariable time window input matrix. The mapping process for extracting temporal features; This represents the temporal attention weight corresponding to the i-th time step; This represents the temporal hidden feature corresponding to the i-th time step; This represents the global context feature corresponding to time t. By weighted aggregation of the temporal hidden features corresponding to each time step within the time window, the global context feature is... It carries dynamic correlation information of the process within a multivariable time window.

[0037] Step S103: The global context features are input into a pre-trained offline dual-task decoupling network. The global context features are projected through the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, outputting decoupled quality-related features and process-related features. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss, and geometric orthogonal constraint loss to reduce information overlap between the quality-related features and the process-related features, promoting their separate representation in the latent space. The quality-related features carry fault information that induces product quality exceeding standards, while the process-related features carry normal process disturbance information caused by raw material fluctuations, closed-loop controller self-adjustment, and other equipment variations.

[0038] global context features After inputting the dual-task decoupling network, it is projected onto the quality-related feature subspace and the process-related feature subspace using two sets of parallel nonlinear mappings to perform global context feature processing. After performing two-stream decoupled projection, quality-related features and process-related features are directly obtained.

[0039] In some alternative implementations, the construction of a pre-trained, offline-trained dual-task decoupling network includes: Step a1: Construct a training sample set, which includes: process observation variable sequence samples under normal operating conditions of a continuous industrial process and the corresponding real quality state vector.

[0040] Historical operating data of a continuous industrial process is obtained as the training sample set. This historical data includes process observation vectors formed by multiple process observation variables at consecutive time points, and their corresponding true quality state vectors. Assume the continuous industrial process includes... process observation variables and A quality variable, defined at time... The process observation vector is represented as The corresponding true quality state vector is represented as Furthermore, time points are constructed based on sliding time windows. input matrix and corresponding mass state vector The input matrix is ​​located here. Equivalent to the above multivariable time window input matrix .

[0041] Step a2: Build the network infrastructure. The network infrastructure includes an initial quality nonlinear mapping layer and an initial process nonlinear mapping layer set in parallel, as well as a quality supervision branch network and a process reconstruction branch network set in parallel. The output of the initial quality nonlinear mapping layer is connected to the quality supervision branch network and the process reconstruction branch network, respectively, and the output of the initial process nonlinear mapping layer is connected to the process reconstruction branch network.

[0042] The initial quality nonlinear mapping layer is used to generate quality latent features based on the global context features of the input, through subspace projection of nonlinear weights and activation functions. One path of the quality latent features is input into the quality supervision branch network for quality regression constraints, and the other path is concatenated with the process latent features and input into the process reconstruction branch network to participate in the reconstruction of the multivariate time window input matrix.

[0043] The initial process nonlinear mapping layer is used to generate process latent features based on the global context features of the input through nonlinear projection. The process latent features are used to characterize the dynamic background of normal operating conditions such as small fluctuations in raw materials, closed-loop regulation of the controller, and slow changes in equipment. The information overlap between process latent features and quality latent features is reduced by the joint loss function. The output is only connected to the process reconstruction branch and is concatenated with the quality latent features to jointly restore the original observation input matrix.

[0044] The quality supervision branch network is only effective during the offline training phase, providing supervisory constraints for the initial quality nonlinear mapping layer. Specifically, it receives the latent quality features output by the initial quality nonlinear mapping layer, outputs predicted quality state estimates through a regression predictor, calculates the quality supervision loss by subtracting the predicted quality state estimates from the true quality state vector, and uses backpropagation to ensure that the latent quality features retain more information related to the true quality state. This loss, along with the reconstruction loss and geometric orthogonality constraint loss, reduces the residual quality-related information in the latent features.

[0045] The process reconstruction branch network receives the spliced ​​quality latent features and process latent features, and generates a reconstruction result of the original multivariate time window input matrix through a decoder. The reconstruction loss is calculated based on the difference between the reconstruction result and the original input matrix to improve the joint representation ability of the two types of latent features on the original process time series information, while ensuring that the process latent features retain dynamic information from the normal operating condition background. In the fault tracing stage, this branch is reused to obtain the reconstruction results of the observed variables, and candidate variables are selected based on the reconstruction residual contribution of each variable.

[0046] Step a3: Construct a joint loss function consisting of a weighted sum of quality supervision loss, reconstruction loss, and geometric orthogonal constraint loss.

[0047] To achieve joint optimization of the shared temporal encoder, dual-task decoupling network, and constraint mechanism, a joint loss function consisting of quality supervision loss, reconstruction loss, and geometric orthogonal constraint loss is constructed.

[0048] Among them, quality supervision loss is defined as: ; Reconstruction loss is defined as: ; The geometric orthogonal constraint loss is defined as: ; in, By constraining potential quality characteristics With process potential characteristics By leveraging directional similarity, the information overlap between the two classes of potential features is reduced. This leads to the global joint loss function: .

[0049] in, , and The weights for the corresponding loss terms are denoted by N, which represents the total number of samples in a training batch. During the offline training phase, normal operating condition data is used to jointly train the entire model, completing parameter learning for the shared temporal encoder, dual-task branches, and geometric orthogonal constraints, and establishing the normal statistical distribution of the quality-related feature space.

[0050] Step a4: Iteratively train the network infrastructure based on the training sample set until the joint loss function converges, thus obtaining the dual-task decoupled network.

[0051] Specifically, step a4 above includes: Step b1: Output latent quality features through the initial quality nonlinear mapping layer; output latent process features through the initial process nonlinear mapping layer.

[0052] Referring to steps S1021 to S1023 above, let the length of the sliding time window be... Then at time input matrix Represented as The input matrix is ​​used to represent the dynamic evolution information of the process over a period of time prior to the current moment. Based on the input matrix... Determine global context features.

[0053] Reference Figure 3 As shown, global context features are input into the network infrastructure. Through the initial quality nonlinear mapping layer and the initial process nonlinear mapping layer, the global context features are respectively... Projecting onto the quality-related feature subspace and the process-related feature subspace yields the output quality latent features. With process potential characteristics : ; ; in, Used for quality status monitoring and prediction. and The concatenated result is used for input matrix reconstruction.

[0054] Step b2: Perform the quality supervision regression constraint task, input the potential quality features into the quality supervision branch network to obtain the quality state estimate, and calculate the quality supervision loss by combining it with the real quality state vector.

[0055] In Task 1 (Quality Supervision Regression Constraint Task), potential quality characteristics will be... Input quality supervision branch The quality supervision branch employs a regression predictor, utilizing latent quality features. For the true quality state vector Perform supervised regression to generate quality status estimates. , Then, combined with the true quality state vector Calculate quality supervision loss , .

[0056] Step b3 involves executing a reconstruction constraint task. The latent quality features and process features are concatenated and input into the process reconstruction branch network to obtain the original input matrix reconstruction result. The reconstruction loss is then calculated using the original input matrix. The geometric orthogonal constraint loss is constructed based on the cosine similarity between the latent quality features and the latent process features.

[0057] In Task 2 (Reconstructing Constraints), the potential features of quality will be... With process potential characteristics Reconstructing the branch network after concatenation of input processes The process reconstruction branch network uses a reconstruction decoder to reconstruct the original input matrix, obtaining the input matrix reconstruction result. , .

[0058] in, For quality supervision branch parameters, These are the parameters for the process refactoring branch. Through the complementary constraints and synergistic effects of the quality supervision branch and the process refactoring branch, the high-dimensional latent features are gradually separated into quality-related features and process-related features, that is, the information closely related to the quality status and the background process dynamic information are separated and represented.

[0059] Geometric orthogonal constraint loss is an implicit regularization term used for regularization constraints. It participates in the summation of joint losses during the backpropagation phase, constraining two eigenvectors to be as orthogonal as possible.

[0060] This invention uses complementary constraints between the quality-related supervision branch and the process reconfiguration branch to concentrate information closely related to the quality status in the quality-related features, while retaining information such as background fluctuations, control adjustments, and process disturbances in the process-related features, thereby reducing the interference of closed-loop control behavior on fault identification results.

[0061] Step b4: Based on the joint loss function consisting of quality supervision loss, reconstruction loss and geometric orthogonal constraint loss term, iteratively update all parameters of the network infrastructure until the joint loss function converges, thus obtaining the dual-task decoupled network.

[0062] The joint loss function is expressed as: .

[0063] During the iterative update process, the Adam optimization algorithm is used to update the network parameters, resulting in the core parameter set of the model: ,in, This represents all network parameters of the shared timing encoder, the initial quality nonlinear mapping layer, and the initial process nonlinear mapping layer. This represents all parameters of the quality supervision branch network; This represents all parameters of the process reconstructing the branch network.

[0064] Based on the dual-task constraint, a geometric orthogonal constraint loss is further introduced to enhance the separability of the two types of potential features by reducing the directional overlap. This design can effectively alleviate the aliasing problem of quality-related fault features and conventional process disturbance features in the latent space, improve the identification capability of the quality-related feature subspace, and thus help reduce the false alarm rate in fault detection and improve monitoring robustness.

[0065] Step S104: Determine online monitoring statistics based on quality-related characteristics.

[0066] Specifically, step S104 above includes: Step S2041: After the dual-task decoupled network completes offline training, extract the normal quality-related reference features corresponding to all normal working condition samples.

[0067] Step S2042: Determine the feature mean vector based on the reference features related to normal quality. With covariance matrix .

[0068] Step S2043, based on quality-related features , eigenvalue vector and covariance matrix The online monitoring statistics Hotelling were calculated. .

[0069] Online monitoring statistics .

[0070] This invention constructs online monitoring statistics based on quality-related characteristics. The control limits were determined using kernel density estimation. This avoids the strict assumptions about statistical distributions made by traditional parameterization methods. Compared to constructing monitoring indicators directly in the original variable space, this invention focuses monitoring on quality-related feature channels, which can more accurately reflect the shift in system quality status, has higher response sensitivity to quality-related faults, and can reduce the impact of ordinary process disturbances unrelated to the final quality on alarm results.

[0071] Step S105: Based on online monitoring statistics and pre-stored normal operating condition control limits, determine whether a quality-related fault has occurred.

[0072] After the dual-task decoupled network completes offline training, normal quality-related reference features are extracted from all normal operating condition samples, and the feature mean vector is estimated. With covariance matrix Subsequently, considering that nonlinear latent characteristics usually do not satisfy the Gaussian distribution assumption, the kernel density estimation method is used to perform nonparametric fitting of the probability density of the monitoring statistics under normal operating conditions, in order to obtain the control limits under the pre-set confidence level under normal operating conditions. When the following conditions are met: When this occurs, the system determines that a quality-related fault has occurred and triggers an alarm.

[0073] The quality-related fault detection based on a dual-task decoupling network provided in this embodiment involves constructing a real-time input matrix from real-time process data and inputting it into the trained dual-task decoupling network during online monitoring to obtain the quality-related features at the current moment. Based on these quality-related features, an online monitoring statistic is determined and compared with the normal operating condition control limit. When the online monitoring statistic exceeds the normal operating condition control limit, a quality-related fault is determined to have occurred, and an alarm is triggered. This invention promotes the separation and representation of information related to normal process disturbances and information related to quality faults in the latent space through a dual-task decoupling network. Online monitoring uses features with stronger quality correlation to construct statistics, thereby reducing the impact of process background disturbances such as raw material fluctuations and controller adjustments on the fault detection results, helping to reduce false alarms and improve fault detection accuracy.

[0074] Furthermore, while traditional analysis methods based on reconstructed residual contributions can identify outliers to some extent, they essentially reflect the intensity of a variable's response to an abnormal state, rather than the causal driving relationship between variables. In strongly coupled continuous industrial processes, local faults often propagate along the process flow and control loop to multiple variables, causing multiple downstream variables to simultaneously exhibit large residual responses. Therefore, relying solely on reconstructed residuals is insufficient to accurately distinguish between root cause variables and affected variables. Thus, current technologies still lack a unified method that can simultaneously achieve dynamic time-series modeling, separate characterization of quality-related features and process-related features, online quality-related fault detection, and fault root cause localization.

[0075] Therefore, this invention also provides a quality correlation fault detection method based on a dual-task decoupling network, which can be used in servers, terminals, mobile terminals, etc. Figure 1 This is a flowchart of a quality correlation fault detection method based on a dual-task decoupling network according to an embodiment of the present invention, as shown below. Figure 1 As shown, the process includes the following steps: Step S201: Collect time-series data of observed variables for the continuous industrial process. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0076] Step S202 involves extracting time-series features from the observed variable's time-series data to obtain global context features. For details, please refer to [link to relevant documentation]. Figure 1Step S102 of the illustrated embodiment will not be described again here.

[0077] Step S203: The global context features are input into a pre-trained offline dual-task decoupling network. The global context features are projected onto the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, outputting decoupled quality-related features and process-related features. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss, and geometric orthogonality constraint loss to promote the separation and representation of quality-related features and process-related features in the latent space. For details, please refer to [link to details]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0078] Step S204: Based on quality-related characteristics, determine the online monitoring statistics. For details, please refer to [link to relevant documentation]. Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0079] Step S205: Based on online monitoring statistics and pre-stored normal operating condition control limits, determine whether a quality-related fault has occurred.

[0080] If a quality-related failure is determined to have occurred, perform the following steps: Step S206: The quality-related features and process-related features output by the dual-task decoupling network are concatenated and then input into the process reconstruction branch network to obtain the reconstruction result of the dual-task decoupling network input matrix. For details on the process reconstruction branch network, please refer to the construction process of the dual-task decoupling network in the previous embodiment; it will not be repeated here.

[0081] Step S207: Based on the reconstruction results, determine the absolute contribution index of each process observation variable during the fault period, and calculate the relative contribution ratio of each preset process observation variable based on the absolute contribution index.

[0082] Step S208: Sort all process observation variables from largest to smallest according to their relative contribution ratio, and select the top K process observation variables to form a candidate variable set, where K > 0 and K is a positive integer.

[0083] After triggering a quality-related fault alarm, based on the reconstruction results output by the process reconfiguration branch network, the contribution analysis of the reconstruction error of each process variable during the fault period is performed. For the first... Each variable, its absolute contribution index Defined as: ; in, The actual measured value of the variable. For the corresponding reconstructed structure, Let [the standard deviation of this variable under normal operating conditions] be [the standard deviation of the variable]. [This refers to the period during which the fault occurred.]

[0084] To eliminate the influence of different variable units and fluctuation ranges, the relative contribution percentage was further calculated. : ; according to The variables are sorted by their magnitude, and the top K variables are selected to form a candidate feature set. This narrows the scope of subsequent root cause analysis and reduces the complexity of causal inference.

[0085] After a fault alarm, a contribution analysis method based on the reconstructed residual is introduced, using the absolute contribution index. and relative contribution percentage The high-dimensional process variables are sorted and filtered to obtain a set of candidate variables. This step can narrow the scope of root cause diagnosis from the entire variable space to a small number of key variables closely related to the current abnormal state, effectively reducing the computational complexity of subsequent causal inference and improving the pertinence of root cause diagnosis.

[0086] Step S209: For any pair of candidate variables in the candidate variable set, construct a restricted prediction model and a full prediction model based on a long short-term memory network, respectively.

[0087] Reference Figure 4 As shown, the process of performing pairwise Granger causal analysis on the candidate variable set through the restricted model and the full model, and then calculating the causal strength, net causal flow and node net influence score, and completing the root cause localization of the failure is illustrated.

[0088] Root cause localization is achieved through causal inference based on long short-term memory networks, i.e., targeting a set of candidate variables. Any pair of candidate variables Construct constrained prediction models respectively ( Figure 4 Constrained models and full prediction models Figure 4 (The full model in the text).

[0089] Step S2010: Calculate the causal strength between candidate variables based on the comparison results of the prediction residuals of the restricted prediction model and the full prediction model.

[0090] If variables are introduced After obtaining historical information, the full prediction model adjusts the variables... If the prediction error residuals are significantly reduced compared to the restricted model, then the variable is considered to be... For variables Granger causality exists, and its causal strength is defined as: ; in, This indicates the prediction results of the constrained model. This represents the prediction result of the full model.

[0091] Step S2011: Determine the net causal flow based on the causal strength between candidate variables.

[0092] To reduce the bidirectional coupling effect caused by closed-loop feedback, the net causal flow direction is further defined as follows: .

[0093] Step S2012: Determine the net influence score of each candidate variable based on the net causal flow.

[0094] In the candidate variable set After traversing all variable pairs, the net influence score of each candidate variable is further calculated. : .

[0095] Step S2013: The candidate variable with the largest net influence score is identified as the root cause process variable. The root cause process variable is a data-driven localization result obtained based on the historical prediction relationship between candidate variables, Granger causality strength, and net influence score, used to characterize process variables with strong upstream driving effects within the candidate variable set.

[0096] Step S2014: Determine the fault propagation path based on the net causal flow of each candidate variable.

[0097] The variable with the largest net influence score is taken as the root cause variable of the failure, and the failure propagation path analysis results are output by combining the net causal flow relationship between the variables.

[0098] By combining a nonlinear causal inference method based on long short-term memory networks, the Granger causal strength among candidate variables is calculated. Pure Cause and Effect Flow and net influence score This enables the identification of fault sources and analysis of fault propagation paths in strongly coupled, feedback-driven industrial scenarios. Compared to methods that rely solely on the magnitude of reconstruction residuals or the contribution of a single variable to determine the root cause, this invention can further distinguish between root cause variables and affected variables, improving the accuracy and interpretability of fault source localization.

[0099] In this embodiment, after an alarm is triggered, the reconstruction residual contribution of each process variable is calculated based on the reconstruction results output by the process reconstruction branch, and a candidate variable set is selected. Causal inference based on a long short-term memory network is performed on the variable pairs in the candidate variable set to calculate the causal strength and net causal flow between candidate variables, and further calculate the net influence score of each candidate variable. The candidate variable with the largest net influence score is determined as the root cause variable of the fault, and the fault propagation path analysis results are output based on the causal relationship between the candidate variables. Compared with relying solely on a single quality prediction model or traditional statistical monitoring methods, this invention can not only identify whether an anomaly has occurred, but also continue to complete candidate variable selection, propagation relationship analysis, and source variable location after an alarm is triggered, thereby improving the method's comprehensive diagnostic capability in complex industrial scenarios.

[0100] The present invention also provides a complete verification embodiment, as detailed below.

[0101] To verify the effectiveness of the method described in this invention in quality-related fault detection and root cause diagnosis in continuous industrial processes, this embodiment uses the Tennessee-Eastman (TE) process as the object of experimental analysis. The TE process, characterized by nonlinearity, strong coupling, and dynamic time-varying features, is a commonly used standard test platform in fault detection and root cause diagnosis research. In the experiment, 33 continuously monitored variables were selected as the process input space, and the reactor outlet product composition G (XMEAS 35) was selected as the quality variable. The training set consisted of 480 sample points from 24 hours of continuous operation under normal conditions. The test set included 960 sample points for each fault mode, comprising samples from the first 8 hours of normal operation and subsequent fault samples. To match the time-series input requirements, a sliding window technique was used to construct the sample tensor, with a window length of T=15. The Adam optimizer was used during the training phase, with an initial learning rate of 0.001, a batch size of 64, and a maximum number of training epochs of 200.

[0102] The experimental analysis is as follows: To evaluate the fault detection performance of the method of this invention, Principal Component Analysis (PCA), Dynamic Partial Least Squares (DPLS), Long Short-Term Memory Autoencoder (LSTM-AE), and Supervised LSTM are introduced as comparative methods. Fault Detection Rate (FDR) and False Alarm Rate (FAR) are used as evaluation metrics, where FDR measures the proportion of samples triggering alarms during fault periods, and FAR measures the proportion of false alarm samples during normal periods. Experimental analysis mainly focuses on three aspects: overall fault detection performance, typical fault response characteristics, and root cause diagnosis effectiveness.

[0103] First, regarding the overall fault detection performance, Table 1 presents the comparison results of FDR and FAR of different models under typical quality-related faults and quality-independent disturbances. For quality-related faults IDV(6), IDV(8), and IDV(10), the FDRs corresponding to the method of this invention are 99.85%, 99.58%, and 92.04%, respectively, which are significantly higher than DPLS's 98.12%, 99.49%, and 75.09% under the corresponding conditions, while maintaining a low FAR of 11.62%, 3.75%, and 11.25%, respectively. Among them, under the IDV(10) condition, although the FDRs of LSTM-AE and Supervised-LSTM reach 94.56% and 95.83%, respectively, slightly higher than the 92.04% of the method of this invention, their FARs are as high as 51.25% and 46.25%, respectively, while the method of this invention is only 11.25%. This indicates that the method of the present invention can not only maintain a high fault detection capability, but also effectively suppress false alarms caused by closed-loop regulation and background disturbances.

[0104] Table 1. Performance comparison of FDR and FAR under typical operating conditions of the TE process (%)

[0105] For quality-independent disturbance conditions, Table 1 further demonstrates the advantages of the method of the present invention in disturbance suppression. For example, under IDV(4) conditions, the FDR and FAR of the method of the present invention are 45.51% and 10.62%, respectively, while the FAR of LSTM-AE and Supervised-LSTM reaches 63.75% and 56.25%, respectively; under IDV(11) conditions, the FDR of the method of the present invention is 93.81%, and the FAR is 16.12%, which is lower than the 57.50% of LSTM-AE and the 63.12% of Supervised-LSTM; under IDV(15) conditions, the FAR of the method of the present invention is only 5.26%, which is significantly lower than the 16.88% of DPLS, the 51.25% of LSTM-AE, and the 50.62% of Supervised-LSTM. Therefore, it can be seen that in complex closed-loop operating conditions where quality-related faults and quality-independent disturbances coexist, the method of the present invention can better balance detection rate and false alarm suppression capability, demonstrating superior comprehensive fault detection performance.

[0106] Furthermore, the IDV(10) condition is analyzed as a typical case of quality-related failure. This failure corresponds to random and drastic fluctuations in the feed temperature (C), which directly affects the reactor's thermodynamic equilibrium and further impacts the final quality. See [link to relevant documentation]. Figure 5 and Figure 6 As shown, Figure 5The PCA method is given. The statistical curve is used to characterize the online monitoring response of the traditional PCA method under quality-related failure conditions; Figure 6 The present invention provides the method for constructing a mass-related feature subspace. The statistical curves characterize the online detection response of the present invention to IDV(10) quality-related faults in the quality-related feature subspace. The results show that PCA's... Although the statistics change somewhat after fault injection, the overall response is relatively lagging, and it repeatedly falls below the control limits during fault evolution; in contrast, the method of this invention... The statistics can exceed the limit more quickly after a fault occurs and maintain a significant shift over a longer monitoring interval. Combined with the 92.04% FDR and 11.25% FAR of the method of this invention under the IDV(10) condition in Table 1, it can be further illustrated that after constructing the monitoring statistics in the quality-related feature subspace, the method of this invention can more effectively reflect the quality state shift and improve the response sensitivity to quality-related faults.

[0107] To verify the ability of this invention to suppress mass-independent disturbances, IDV(11) was selected as a typical disturbance condition for analysis. This condition corresponds to random fluctuations in the reactor cooling water inlet temperature, which will cause significant process fluctuations under feedback regulation, but its impact is not directly transmitted to the final quality state. See [link to relevant documentation]. Figure 7 and Figure 8 , Figure 7 Mass subspace statistics are given. The variation curve is used to characterize the response of the quality-related feature channel of the present invention to non-quality fault disturbances under process disturbance conditions; Figure 8 The process subspace statistics are given. The variation curves are used to characterize the response of the process-related feature channels of this invention to background disturbances and control regulation behavior under process disturbance conditions. The results show that in the process subspace... Significant changes occurred after the introduction of the disturbance, exceeding the control limits, while the mass subspace... The overall performance remained within the control limits, with no significant exceedances. This demonstrates that the dual-task decoupling and geometrically orthogonal constraint mechanism constructed in this invention can effectively isolate quality-related changes and process background disturbances at the feature level. (Refer to Table 1 for details.) The result of FAR=16.12% under operating conditions shows that the method of the present invention can effectively suppress false alarms caused by closed-loop disturbances.

[0108] In terms of verifying the root cause diagnosis capability, the IDV(1) fault was used as an example for analysis. This fault corresponds to a step change in the A / C feed ratio, and the mechanism is relatively clear. First, the relative contribution of each variable was calculated using the reconstruction residual contribution analysis, and the top 10 variables in terms of contribution were selected to form a candidate feature subset. As shown in Table 2, the contribution of XMEAS(1) reached 43.9%, which was significantly higher than that of other variables. The contribution of XMEAS(18), XMEAS(4), XMEAS(9) and XMEAS(16) were 8.9%, 7.8%, 7.4% and 5.2%, respectively. In addition, the contribution of XMEAS(7), XMEAS(13), XMEAS(20), XMV(9) and XMV(7) were 4.0%, 3.7%, 3.0%, 2.6% and 1.8%, respectively. The results show that the present invention can effectively screen out key variables closely related to the current abnormal state through the RBC method, thereby achieving preliminary compression of the high-dimensional process variable space.

[0109] Table 2 shows the statistical results of the RBC contribution of the top 10 key variables under the IDV(1) operating condition.

[0110] Based on the candidate variable screening, nonlinear Granger causal inference based on long short-term memory networks was further performed on the variable pairs in the candidate variable set. The analysis results are shown in Table 3. Table 3 shows that there are obvious two-way interaction relationships between the candidate variables. For example, the causal strength between XMEAS(18) and XMEAS(19) is respectively... and This indicates a strong dynamic coupling between the two. Meanwhile, the effect of XMEAS(20) on XMEAS(19) reaches 0.6912, and XMEAS(1) also shows a significant driving effect on XMEAS(18) and XMEAS(19). This suggests that the source of the fault cannot be directly determined solely based on the magnitude of the residual contribution; further analysis combining the causal propagation structure is necessary.

[0111] Table 3. Granger causality matrix among candidate variables under IDV(1) conditions.

[0112] Furthermore, analysis combining net influence scores and fault propagation topology reveals that the highest statistical contribution does not necessarily mean that the variable is the source of the fault. In strongly coupled closed-loop processes, some downstream response variables, due to the combined effects of fault propagation and feedback regulation, may exhibit larger residual shifts, thus ranking higher in the RBC ranking; while the true source variable may not have the largest instantaneous response amplitude, but it typically exhibits a stronger outward driving effect in causal propagation networks. (Refer to...) Figure 9 As shown, Figure 9 A fault propagation path topology diagram of IDV(1) is presented, showing the main propagation paths, secondary propagation paths, and feedback coupling relationships among candidate variables, illustrating the topological structure of fault propagation from upstream driving nodes to intermediate propagation nodes and downstream response nodes. The results show that this causal network exhibits a clear hierarchical propagation characteristic, with XMEAS(7) and XMEAS(13) showing strong upstream driving effects, XMEAS(18) and XMEAS(19) located in the intermediate propagation layer, and XMEAS(4) and XMEAS(35) closer to the downstream response layer. These results demonstrate that the method of this invention can not only screen out key abnormal variables after a fault alarm but also further reveal the fault propagation relationship, achieving effective location of the fault source.

[0113] This invention constructs a unified analysis framework consisting of shared temporal coding, dual-task feature decoupling, online monitoring, and root cause diagnosis. It enables online identification of quality-related anomalies in continuous industrial processes, candidate variable compression, root cause variable localization, and propagation path analysis. It can simultaneously achieve quality-related fault detection and root cause localization in closed-loop control scenarios of continuous industrial processes. It is suitable for complex operating conditions in continuous industrial processes where key quality indicators are difficult to obtain directly online, process variables are strongly coupled, and closed-loop regulation exists, demonstrating good engineering adaptability and application promotion value.

[0114] This invention addresses the problems of overlapping characteristics between conventional process disturbances and quality-related faults in continuous industrial processes under closed-loop control conditions, the high false alarm rate of traditional methods, and the difficulty in root cause localization. It proposes a unified diagnostic method for quality-related anomaly identification and propagation analysis. This method includes: constructing a multivariate time window input matrix; extracting process dynamic context features using a shared time encoder; obtaining quality-related and process-related features through a dual-task decoupling network; constructing online monitoring statistics and control limits based on the quality-related features to achieve quality-related fault detection; after an alarm, using reconstruction residual contribution analysis to screen candidate variables, and combining causal inference based on a long short-term memory network and net influence score to complete fault root cause localization and propagation path analysis. This invention can effectively reduce the interference of closed-loop regulation and background disturbances on fault detection results, improve the ability to identify quality-related faults, and enhance the accuracy and interpretability of root cause diagnosis.

[0115] This embodiment also provides a quality correlation fault detection device based on a dual-task decoupling network. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0116] This embodiment provides a quality correlation fault detection device based on a dual-task decoupling network, such as... Figure 10 As shown, it includes: The acquisition module 301 is used to acquire time-series data of observed variables in a continuous industrial process; Extraction module 302 is used to extract time series features from the time series data of observed variables to obtain global context features; The separation module 303 is used to input global context features into a pre-trained offline dual-task decoupling network. The global context features are projected through the quality nonlinear mapping layer and the process nonlinear mapping layer set in parallel in the dual-task decoupling network, respectively, and the decoupled quality-related features and process-related features are output. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss and geometric orthogonal constraint loss to promote the separation and representation of quality-related features and process-related features in the latent space. Statistical module 304 is used to determine online monitoring statistics based on quality-related characteristics; The diagnostic module 305 is used to determine whether a quality-related fault has occurred based on online monitoring statistics and pre-stored normal operating condition control limits.

[0117] In some alternative implementations, the apparatus further includes: a construction module; the construction module is specifically used for: Construct a training sample set, which includes: a sequence of process observation variables under normal operating conditions of a continuous industrial process and the corresponding real quality state vector; The network infrastructure is constructed, which includes an initial quality nonlinear mapping layer and an initial process nonlinear mapping layer set in parallel, as well as a quality supervision branch network and a process reconstruction branch network set in parallel. The output of the initial quality nonlinear mapping layer is connected to the quality supervision branch network and the process reconstruction branch network, respectively, and the output of the initial process nonlinear mapping layer is connected to the process reconstruction branch network. Construct a joint loss function composed of a weighted average of quality supervision loss, reconstruction loss, and geometric orthogonality constraint loss; Iterative training of the network infrastructure is performed based on the training sample set until the joint loss function converges, resulting in a dual-task decoupled network. This includes: outputting latent quality features through an initial quality nonlinear mapping layer; outputting latent process features through an initial process nonlinear mapping layer; performing a quality-supervised regression constraint task, inputting the latent quality features into the quality-supervised branch network to obtain quality state estimates, and calculating the quality-supervised loss based on the true quality state vector; performing a reconstruction constraint task, concatenating the latent quality features and process latent features and inputting the concatenated results into the process reconstruction branch network to obtain the original input matrix reconstruction result, and calculating the reconstruction loss based on the original input matrix; a geometric orthogonal constraint loss is constructed based on the cosine similarity between the latent quality features and the process latent features; and iteratively updating all parameters of the network infrastructure based on the joint loss function composed of the quality-supervised loss, reconstruction loss, and geometric orthogonal constraint loss term until the joint loss function converges, resulting in the dual-task decoupled network.

[0118] In one alternative implementation, the statistics module 304 is specifically used for: After the dual-task decoupled network completes offline training, normal quality-related reference features are extracted for all normal working condition samples. Based on the reference features related to normal quality, determine the feature mean vector and covariance matrix; Based on quality-related features, feature mean vectors, and covariance matrices, online monitoring statistics are calculated.

[0119] In an optional implementation, the diagnostic module 305 is further specifically used for: The quality-related features and process-related features output by the dual-task decoupling network are concatenated and then input into the process reconstruction branch network to obtain the reconstruction result of the dual-task decoupling network input matrix. Based on the reconstruction results, the absolute contribution index of each process observation variable during the fault period is determined, and the relative contribution ratio of each preset process observation variable is calculated based on the absolute contribution index. All process observation variables are sorted from largest to smallest according to their relative contribution percentage, and the top K process observation variables are selected to form a candidate variable set, where K > 0 and K is a positive integer; Based on the candidate variable set, the process variables for the root cause of the fault are determined. These include: For any pair of candidate variables in the candidate variable set, construct a restricted prediction model and a full prediction model based on a long short-term memory network, respectively. Based on the comparison of the prediction residuals between the restricted prediction model and the full prediction model, the causal strength between the candidate variables is calculated. Determine the net causal flow based on the causal strength among candidate variables; Based on the net causal flow, determine the net influence score of each candidate variable; The candidate variable with the largest net influence score is identified as the root cause process variable of the failure.

[0120] Based on the net causal flow of each candidate variable, the fault propagation path is determined.

[0121] In one optional implementation, the extraction module 302 includes: The time series data of the observed variables are divided into multivariate time windows of fixed duration, and a multivariate time window input matrix is ​​constructed. The multivariate time window input matrix is ​​input into the shared temporal encoder to obtain the temporal hidden features corresponding to multiple time steps; Temporal attention weights are assigned to the temporal hidden features corresponding to each time step based on the temporal attention mechanism, and the temporal hidden features corresponding to each time step are weighted and aggregated based on the temporal attention weights to obtain global context features.

[0122] The quality correlation fault detection device based on a dual-task decoupling network provided in this invention can execute the quality correlation fault detection method based on a dual-task decoupling network provided in any embodiment of this invention, and has the corresponding functional modules and beneficial effects of the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.

[0123] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0124] The following is a detailed reference. Figure 11 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0125] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 11 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0126] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the quality association fault detection method based on a dual-task decoupling network according to embodiments of the present invention.

[0127] Figure 11 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0128] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the quality-related fault detection method based on a dual-task decoupling network shown in the above embodiments is implemented.

[0129] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0130] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A quality correlation fault detection method based on a dual-task decoupling network, characterized in that, The method includes: Collect time-series data of observed variables in continuous industrial processes; Temporal features are extracted from the time-series data of the observed variables to obtain global context features; The global context features are input into a pre-trained offline dual-task decoupling network. The global context features are then projected onto the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, to output decoupled quality-related features and process-related features. The dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss, and geometric orthogonality constraint loss to promote the separation and representation of the quality-related features and process-related features in the latent space. Based on the aforementioned quality-related characteristics, online monitoring statistics are determined; Based on the online monitoring statistics and the pre-stored normal operating condition control limits, determine whether a quality-related fault has occurred; In the event of a quality-related fault, the quality-related features and process-related features output by the dual-task decoupling network are input into the process reconstruction branch network to obtain the reconstruction result of the dual-task decoupling network input matrix. Based on the reconstruction result, the absolute contribution index of each process observation variable during the fault period is determined. Based on the absolute contribution index, a candidate variable set is obtained. Causal inference based on a long short-term memory network is performed on the variable pairs in the candidate variable set to calculate the causal strength and net causal flow between the candidate variables and determine the net influence score of each candidate variable. The candidate variable with the largest net influence score is determined as the root cause variable of the fault. The dual-task decoupling network includes the process reconstruction branch.

2. The method according to claim 1, characterized in that, The construction of the pre-trained offline dual-task decoupling network includes: Construct a training sample set, which includes: a sequence of process observation variables under normal operating conditions of a continuous industrial process and the corresponding real quality state vector; A network infrastructure is established; the network infrastructure includes an initial quality nonlinear mapping layer and an initial process nonlinear mapping layer set in parallel, as well as a quality supervision branch network and a process reconstruction branch network set in parallel, wherein the output of the initial quality nonlinear mapping layer is connected to the quality supervision branch network and the process reconstruction branch network respectively, and the output of the initial process nonlinear mapping layer is connected to the process reconstruction branch network. Construct a joint loss function composed of a weighted average of quality supervision loss, reconstruction loss, and geometric orthogonality constraint loss; Iterative training of the network infrastructure is performed based on the training sample set until the joint loss function converges, thus obtaining the dual-task decoupled network.

3. The method according to claim 2, characterized in that, The iterative training of the network infrastructure based on the training sample set until the joint loss function converges, resulting in the dual-task decoupled network, includes: The initial quality nonlinear mapping layer outputs potential quality features; the initial process nonlinear mapping layer outputs potential process features. The quality supervision regression constraint task is executed, and the potential quality features are input into the quality supervision branch network to obtain the quality state estimate. The quality supervision loss is then calculated by combining the actual quality state vector. The reconstruction constraint task is executed. The quality latent features and the process latent features are concatenated and then input into the process reconstruction branch network to obtain the original input matrix reconstruction result. The reconstruction loss is then calculated by combining the original input matrix. The geometric orthogonal constraint loss is constructed based on the cosine similarity between the latent quality features and the latent process features; Based on the joint loss function consisting of the quality supervision loss, the reconstruction loss, and the geometric orthogonal constraint loss term, all parameters of the network infrastructure are iteratively updated until the joint loss function converges, thus obtaining the dual-task decoupled network.

4. The method according to claim 1, characterized in that, The determination of online monitoring statistics based on the aforementioned quality-related characteristics includes: After the dual-task decoupling network completes offline training, normal quality-related reference features corresponding to all normal working condition samples are extracted. Based on the aforementioned normal quality-related reference features, determine the feature mean vector and covariance matrix; The online monitoring statistics are calculated based on the quality-related features, the feature mean vector, and the covariance matrix.

5. The method according to claim 1, characterized in that, The candidate variable set obtained based on the absolute contribution index includes: The relative contribution percentage of each preset process observation variable is calculated based on the absolute contribution index. All process observation variables are sorted from largest to smallest according to their relative contribution percentage, and the top K process observation variables are selected to form a candidate variable set, where K > 0 and K is a positive integer.

6. The method according to claim 5, characterized in that, The process of performing causal inference based on long short-term memory networks on the variable pairs in the candidate variable set, calculating the causal strength and net causal flow between candidate variables, and determining the net influence score of each candidate variable includes: For any pair of candidate variables in the candidate variable set, a restricted prediction model and a full prediction model based on a long short-term memory network are constructed respectively. The causal strength between candidate variables is calculated based on the comparison results of the prediction residuals of the restricted prediction model and the full prediction model. The net causal flow is determined based on the causal strength between candidate variables; Based on the net causal flow, the net influence score of each candidate variable is determined.

7. The method according to claim 6, characterized in that, The method further includes: Based on the net causal flow of each candidate variable, the fault propagation path is determined.

8. The method according to claim 1, characterized in that, The step of extracting time-series features from the observed variable's time-series data to obtain global context features includes: The time series data of the observed variables are divided into multivariate time windows of fixed duration, and a multivariate time window input matrix is ​​constructed. The multivariable time window input matrix is ​​input into a shared temporal encoder to obtain temporal hidden features corresponding to multiple time steps; The temporal attention mechanism assigns temporal attention weights to the temporal hidden features corresponding to each time step, and the temporal hidden features corresponding to each time step are weighted and aggregated based on the temporal attention weights to obtain the global context features.

9. A quality correlation fault detection device based on a dual-task decoupling network, characterized in that, The device includes: The data acquisition module is used to collect time-series data of observed variables in continuous industrial processes; The extraction module is used to extract time-series features from the observed variable time-series data to obtain global context features; The separation module is used to input the global context features into a pre-trained offline dual-task decoupling network, and project the global context features through the parallel quality nonlinear mapping layer and process nonlinear mapping layer in the dual-task decoupling network, respectively, to output decoupled quality-related features and process-related features; wherein, the dual-task decoupling network is jointly trained with quality supervision loss, reconstruction loss and geometric orthogonal constraint loss to promote the separation and representation of quality-related information and process-related information in the latent space; The statistics module is used to determine online monitoring statistics based on the aforementioned quality-related characteristics; The diagnostic module is used to determine whether a quality-related fault has occurred based on the online monitoring statistics and pre-stored normal operating condition control limits. If a quality-related fault is determined to have occurred, the quality-related features and process-related features output by the dual-task decoupling network are input into the process reconstruction branch network to obtain the reconstruction result of the dual-task decoupling network input matrix. Based on the reconstruction result, the absolute contribution index of each process observation variable during the fault period is determined. Based on the absolute contribution index, a candidate variable set is obtained. Causal inference based on a long short-term memory network is performed on the variable pairs in the candidate variable set to calculate the causal strength and net causal flow between the candidate variables, and to determine the net influence score of each candidate variable. The candidate variable with the largest net influence score is determined as the root cause variable of the fault. The dual-task decoupling network includes the process reconstruction branch.

10. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the quality correlation fault detection method based on a dual-task decoupling network as described in any one of claims 1 to 8.

Images