A multi-source data fusion-based whole-process fault diagnosis system and early warning method

Abstract

This invention belongs to the field of fault prediction technology and discloses a full-process fault diagnosis system and early warning method based on multi-source data fusion. This invention collects multi-dimensional heterogeneous operational data from all sections of the production line. After unified preprocessing, it adopts a two-level processing mode combining single-section feature-level fusion with cross-section correlation feature fusion throughout the entire process. Coupled with a unique section fault transmission coupling quantitative calculation method, it dynamically optimizes the cross-process feature correlation weights. It performs layered anomaly identification of single sections, cross-section fault matching, and root cause localization. Combining time-series features, it infers the evolution trend of operating conditions, classifies multiple early warning levels, and generates process adjustment and equipment protection commands in a linked manner. It also utilizes on-site feedback data to achieve iterative optimization of the analysis logic. This invention solves the industry problems of limited monitoring scope, isolated data, difficulty in fault tracing, and delayed early warning in existing technologies, improving the operational stability and anomaly control capabilities of film extrusion production.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, specifically to a full-process fault diagnosis system and early warning method based on multi-source data fusion. Background Technology

[0002] Plastic film extrusion production is a continuous assembly line operation with closely linked processes. Materials and operating conditions gradually transmit their effects along the production flow. Currently, most production control methods in the industry focus on data collection and status monitoring in a single, independent section. Data from different processes is fragmented and fails to reflect the correlation and impact between upstream and downstream sections.

[0003] Traditional monitoring methods focus only on single equipment operating parameters or local quality indicators, and the data sources are limited, failing to comprehensively cover various information such as equipment operating status, process execution conditions, on-site working environment, and finished product quality. Faced with slowly evolving hidden anomalies in the production process, as well as chain reactions of problems caused by fluctuations in upstream operating conditions, conventional monitoring methods struggle to identify trends in advance, and problems can only be discovered after a significant malfunction has occurred.

[0004] During the troubleshooting phase, most existing methods can only pinpoint the current work section where the anomaly occurred, failing to trace the transmission path of the anomaly between different processes, making it difficult to distinguish between primary and secondary issues, resulting in low efficiency in tracing the root cause. Furthermore, most diagnostic models only identify the anomaly, lacking trend prediction and proactive control. After a problem is discovered, relying on manual experience to adjust parameters leads to delayed actions, potentially causing material waste, equipment wear and tear, and fluctuations in the quality of batch products.

[0005] Conventional data fusion methods often use fixed calculation weights, which cannot fit the process characteristics of the rheological properties of film extrusion materials coupled with temperature. They cannot dynamically adjust the correlation analysis logic according to changes in production conditions, and the cross-process anomaly propagation patterns cannot be quantitatively analyzed, further limiting the practical effectiveness of fault identification and risk prediction.

[0006] A search revealed that Chinese patent CN120722874A discloses a self-diagnostic system for industrial production line faults based on digital twins. This system continuously collects equipment operation data through a status monitoring module, uses raw material characteristic data to correct historical operation data, and trains a fault self-diagnosis model to reduce the impact of raw material characteristic differences on the diagnostic results. However, this technical solution is primarily geared towards general industrial production lines and fails to specifically address the continuous material transfer and strong coupling of multiple stages in plastic film extrusion production lines. Its data acquisition mainly relies on equipment operation data, lacking synchronous fusion and collection of heterogeneous data from multiple dimensions such as process parameters, product quality, and the on-site environment. The diagnostic model depends on offline training using historical operation data and lacks the dynamic quantitative calculation capability for cross-stage fault transmission correlation based on real-time operation data, making it difficult to cope with the chain-like fault evolution pattern of upstream operating condition fluctuations propagating downstream in plastic film extrusion production. Furthermore, this solution does not establish a complete control chain from fault identification to closed-loop regulation, exhibiting deficiencies in early trend warning and adaptive control.

[0007] To address this, a full-process fault diagnosis system and early warning method based on multi-source data fusion is proposed. Summary of the Invention

[0008] The present invention aims to solve the problems mentioned in the background art by providing a full-process fault diagnosis system and early warning method based on multi-source data fusion.

[0009] The specific technical solution is as follows:

[0010] A full-process fault diagnosis system based on multi-source data fusion, belonging to the field of fault prediction technology, includes a full-process multi-source data acquisition unit, a multi-source data collaborative preprocessing unit, a two-level fusion feature engineering unit, a hierarchical fault diagnosis root cause localization unit, a fault evolution early warning unit, and a closed-loop execution unit, wherein:

[0011] The full-process multi-source data acquisition unit is used to collect multi-dimensional heterogeneous operation data of the entire production line process and output it to the multi-source data collaborative preprocessing unit.

[0012] The multi-source data collaborative preprocessing unit is used to perform time synchronization alignment, outlier removal, missing value completion and normalization processing on the received heterogeneous running data to obtain a standardized full-process dataset, which is then output to the two-level fusion feature engineering unit.

[0013] The two-level fusion feature engineering unit is used to perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section associations in the entire process. The coupling coefficient is calculated using the section fault transmission coupling association quantification equation, and the cross-section feature association weights are dynamically optimized to obtain the single-section fusion feature set and the full-process association fusion feature set. The single-section fusion feature set is output to the hierarchical fault diagnosis root cause localization unit, and the full-process association fusion feature set is output to the hierarchical fault diagnosis root cause localization unit and the fault evolution early warning unit.

[0014] The hierarchical fault diagnosis root cause localization unit is used to complete single-section anomaly identification based on single-section fusion feature set, complete cross-section fault transmission association matching and fault root cause localization based on full-process association fusion feature set, and output fault diagnosis results to fault evolution early warning unit and closed-loop execution unit.

[0015] The fault evolution early warning unit is used to complete early fault trend warning under fault-free operating conditions based on the full-process correlation and fusion feature set, complete the evolution trend warning of the faults that have occurred in combination with the fault diagnosis results, and output the corresponding level of fault warning signal to the closed-loop execution unit.

[0016] The closed-loop execution unit is used to generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding section based on the received fault diagnosis results and fault warning signals, and send them to the local control system of the production line. At the same time, it collects the operation feedback data after the instruction is executed and outputs it to the multi-source data collaborative preprocessing unit to complete the online iterative optimization of feature distribution adaptation and diagnostic model.

[0017] The above-mentioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a multi-source data acquisition unit for the entire process, which includes section acquisition sub-units deployed sequentially in the raw material batching section, extrusion plasticizing section, melt filtering section, film forming and cooling section, traction stretching section, and winding section of the production line, as well as a supporting environmental parameter acquisition sub-unit.

[0018] Each section's data acquisition subunit includes a process parameter acquisition module, an equipment status acquisition module, and a product quality online detection module;

[0019] The process parameter acquisition module is used to collect temperature, pressure, melt flow rate, motor speed and material ratio parameters of the corresponding process section, and the acquisition frequency is set to 1Hz-100Hz.

[0020] The equipment status acquisition module is used to collect the current, voltage, vibration, and noise parameters of the drive motor of the corresponding section, as well as the running displacement and fit clearance parameters of the transmission components. The acquisition frequency is set to 100Hz-2000Hz.

[0021] The product quality online detection module is used to collect material melt index parameters in the raw material batching section and the extrusion plasticizing section, and to collect film thickness, number of crystal points, tensile strength, haze and heat shrinkage parameters in the film forming and cooling section to the winding section. The collection frequency is set to 0.5Hz-50Hz.

[0022] The environmental parameter acquisition subunit is used to collect environmental temperature, humidity, and dust concentration parameters of the production line operating environment, with the acquisition frequency set to 0.1Hz-5Hz.

[0023] The above-mentioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a multi-source data collaborative preprocessing unit comprising a data time synchronization module, a data cleaning module, and a data normalization module.

[0024] The data time synchronization module uses the system clock of the production line main controller as a reference. It adopts linear interpolation for high-frequency equipment status data of 100Hz-2000Hz and zero-order hold method for process, quality and environmental data of 0.1Hz-100Hz. It performs timestamp synchronization and alignment on heterogeneous operating data with different acquisition frequencies. The time step after alignment is uniformly set to 0.0005s. This time step is equal to the sampling period corresponding to the highest sampling frequency (2000Hz) among all acquisition modules.

[0025] The data cleaning module uses a combination of Grubbs' criterion and the sliding window mean method to remove outliers and fill in missing values ​​in the synchronized data. The length of the sliding window is set to 5-100 time steps, and the significance level of Grubbs' criterion is set to 0.05.

[0026] The data normalization module uses the min-max normalization method to map the cleaned data to the 0-1 range, resulting in a standardized full-process dataset.

[0027] The aforementioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a two-level fusion feature engineering unit comprising a single-section feature-level fusion module and a cross-section related feature fusion module for the entire process.

[0028] The single-section feature-level fusion module extracts high-dimensional hidden layer features from the multi-source data corresponding to each section using a stacked autoencoder with 3 to 8 layers for the standardized dataset of each section. The number of neurons in the encoder and decoder layers of the stacked autoencoder is symmetrically set. The hidden layer activation function is the ReLU function, and the output layer uses the Sigmoid function. Training is completed by unsupervised pre-training combined with supervised fine-tuning with few samples. The pre-training stage uses steady-state operation data of the production line, and the fine-tuning stage uses the class imbalance loss function combined with the hard sample mining strategy. Then, the attention mechanism is used to dynamically assign weights to features of different types of parameters. The weight assignment values ​​range from 0 to 1 to obtain the single-section fusion features of each section. These are summarized to form a single-section fusion feature set, which is output to the hierarchical fault diagnosis root cause localization unit and the full-process cross-section correlation feature fusion module.

[0029] The full-process cross-segment correlation feature fusion module constructs a directed acyclic graph (DAG) segment correlation topology based on the sequence of segments and material transfer logic of the production line. Each segment is treated as a graph node, the material and energy transfer relationships between segments are directed edges, and the fusion features of individual segments are node features. The prior weights of the edges are dynamically adjusted using the segment fault transmission coupling coefficient. A 2-4 layer graph attention neural network is used to extract full-process correlation features from the single-segment fusion features of all segments. The final edge weights are calculated using a weighted fusion of the prior coupling coefficient and the network-learned attention weights. The fusion formula is as follows: ,in Let be the edge weight, and λ be a preset prior weight coefficient, with a value range of [0.4, 0.8]. The coupling coefficient for fault propagation in the work section. The attention weights learned by the graph attention neural network are used to obtain a full-process correlation fusion feature set containing cross-section fault transmission correlation information, which is then output to the hierarchical fault diagnosis root cause localization unit and fault evolution early warning unit.

[0030] The above-mentioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a hierarchical fault diagnosis root cause localization unit comprising a single-section anomaly identification module, a cross-section fault transmission association matching module, and a root cause localization module.

[0031] The single-section anomaly identification module, based on the feature components of the corresponding section in the single-section fusion feature set, uses a pre-trained single-class support vector machine anomaly detection model to identify whether there is an operational anomaly in the corresponding section. The single-class support vector machine uses a Gaussian kernel function, and the kernel function bandwidth is determined by combining grid search method with 5-fold cross-validation. The anomaly judgment threshold is calibrated by the 3σ criterion. The single-section anomaly identification result is output to the cross-section fault transmission association matching module.

[0032] The cross-section fault transmission association matching module, based on the section association topology of the production line and combined with the single section anomaly identification results, uses the maximum mutual information coefficient to calculate the nonlinear correlation degree between the abnormal features of different sections, filters out the abnormal section combinations with a correlation degree greater than a preset correlation threshold, and outputs them to the root cause localization module. The preset correlation threshold ranges from 0.5 to 0.7 and is calibrated by combining the receiver operating characteristic curve (ROC curve) with the production line process requirements.

[0033] The root cause localization module, based on the fusion feature set of abnormal work section combinations and the whole process association, adopts a Bayesian network fault causal reasoning model constructed based on the work section topology to locate the source work section, fault type, and fault impact range of the fault, and generate fault diagnosis results. The Bayesian network uses the abnormal state of each work section as the root node, the fault type as the leaf node, and the work section coupling relationship as the directed edge. The conditional probability table is determined by the maximum likelihood estimation method based on expert experience and historical fault data. The reasoning algorithm adopts the connection tree algorithm. The fault types include melt temperature runaway, filter screen blockage, melt fracture, uneven cooling, traction roller slippage, winding tension runaway, die mouth accumulation, roller surface adhesion, and melt degradation.

[0034] The aforementioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a fault evolution early warning unit comprising a feature trend extraction module, a fault evolution prediction module, and an early warning level classification module.

[0035] The feature trend extraction module, based on the full-process correlation and fusion feature set of continuous time period, uses the sliding window least squares method to fit the evolution trend of each feature component to obtain the feature trend sequence. The length of the sliding window is set to 10-200 time steps.

[0036] The fault evolution prediction module, for fault-free operating conditions, uses a 2-6 layer long short-term memory network time series prediction model based on feature trend sequences and a historical fault sample library. The model input is a feature trend sequence of 20-100 consecutive time steps, and the output is the feature change trend and fault occurrence probability within a preset future time period. The mean squared error loss function combined with the Adam optimizer is used to complete the training, predicting the feature change trend and fault occurrence probability within the preset future time period. For operating conditions where faults have occurred, the module combines fault diagnosis results and feature trend sequences to predict the spread range and deterioration probability of the fault. The preset time period ranges from 5 min to 30 min.

[0037] The warning level classification module classifies warning levels according to the predicted failure probability. When the failure probability is <30%, normal production is maintained and no warning is triggered; a failure probability of 30% to <60% is a low-risk warning, 60% to <90% is a medium-risk warning, and 90% and above is a high-risk warning. The warning probability threshold is calibrated by combining the ROC curve with the safety control requirements of the production line, and the corresponding level of failure warning signal is output.

[0038] The aforementioned fault diagnosis system for the entire process of plastic film extrusion production line based on multi-source data fusion includes a closed-loop execution unit comprising an instruction generation module, an instruction issuance module, and an execution feedback module.

[0039] The instruction generation module generates corresponding control instructions based on the fault diagnosis results and the warning level of the fault warning signal. Specifically, for low-risk warnings, it generates process parameter fine-tuning instructions for the corresponding section; for medium-risk warnings, it generates process parameter adjustment instructions and equipment inspection reminder instructions for the corresponding section; for high-risk warnings, it generates load reduction operation instructions or equipment shutdown protection instructions for the corresponding section; and for located faults, it generates corresponding process correction instructions.

[0040] The instruction issuing module sends the generated control instructions to the local programmable logic controller of the production line via industrial Ethernet or fieldbus. Before the instructions are issued, they must be verified by process safety boundaries. Instructions that exceed the process safety limits will be intercepted and manual review will be triggered.

[0041] The execution feedback module is used to collect production line operation data after the control command is executed and output it to the multi-source data collaborative preprocessing unit. The preprocessed feedback data is used for feature distribution adaptation of the two-level fusion feature engineering unit and online incremental learning and iterative optimization of the diagnostic model in the hierarchical fault diagnosis root cause localization unit. The update cycle of incremental learning is set to 12h-24h, and the rapid update cycle under emergency conditions is set to 1h-12h. Incremental learning adopts a sliding time window dataset update method to retain nearly 3 months of valid operation data.

[0042] This invention also provides a method for fault diagnosis and early warning of the entire process of a plastic film extrusion production line based on multi-source data fusion, applied to the aforementioned fault diagnosis system for the entire process of a plastic film extrusion production line based on multi-source data fusion, comprising the following steps:

[0043] S1: Collect multi-dimensional heterogeneous operational data of the entire process of the plastic film extrusion production line;

[0044] S2: Perform time synchronization alignment, outlier removal, missing value completion, and normalization on the collected heterogeneous operational data to obtain a standardized full-process dataset;

[0045] S3: Perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section association in the entire process. Calculate the coupling coefficient based on the quantification equation of fault transmission coupling association in the section, and dynamically optimize the cross-section feature association weights to obtain the single-section fusion feature set and the full-process association fusion feature set.

[0046] S4: Based on the single-section fusion feature set, complete the single-section anomaly identification; based on the full-process association fusion feature set, complete the cross-section fault transmission association matching and fault root cause location, and generate fault diagnosis results.

[0047] S5: Based on the full-process correlation and fusion feature set, complete the early fault trend warning under fault-free working conditions, combine the fault diagnosis results to complete the evolution trend warning of the faults that have occurred, divide the warning level according to the non-overlapping probability boundary, do not trigger the warning if the probability is <30%, and output the corresponding level of fault warning signal.

[0048] S6: Based on the fault diagnosis results and fault warning signals, generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding work section, and send them to the local control system of the production line for execution. At the same time, collect the operation feedback data after the instruction is executed to complete the online iterative optimization of feature distribution adaptation and diagnostic model.

[0049] The above-mentioned method for fault diagnosis and early warning of the entire process of plastic film extrusion production line based on multi-source data fusion, wherein step S3 specifically includes:

[0050] S31: For each work section's standardized dataset, a stacked autoencoder with 4 to 6 layers is preferred to extract high-dimensional hidden layer features from the multi-source data corresponding to each work section. Then, an attention mechanism is used to dynamically assign weights to features of different types of parameters to obtain the single-work section fusion features of each work section, which are then summarized to form a single-work section fusion feature set.

[0051] S32: Based on the sequential timing and material transfer logic of the plastic film extrusion production line, a directed acyclic graph (DAG) topology structure for the connection between work sections is constructed. Each work section is a graph node, the material and energy transfer relationship between work sections is a directed edge, and the fusion features of a single work section are node features. The coupling coefficient is calculated by substituting the quantification equation of the coupling relationship of the fault transmission between work sections and the prior weights of the edges are updated. It is preferred to use a 2-3 layer graph attention neural network to extract the full-process association features of the fusion features of the single work sections of all work sections. Finally, the edge weights are calculated by weighted fusion of the prior coupling coefficient and the network learning attention weights to obtain a full-process association fusion feature set containing cross-work section fault transmission association information.

[0052] The aforementioned method for fault diagnosis and early warning of the entire process of a plastic film extrusion production line based on multi-source data fusion, wherein step S4 specifically includes:

[0053] S41: Based on the feature components of the corresponding work section in the single-section fusion feature set, a pre-trained single-class support vector machine anomaly detection model is used to identify whether there is an operational anomaly in the corresponding work section and generate the single-section anomaly identification result.

[0054] S42: Based on the production line section association topology, combined with the single section anomaly identification results, the nonlinear correlation degree between different section anomaly features is calculated using the maximum mutual information coefficient. 0.5-0.7 is selected as the preferred correlation threshold, and abnormal section combinations with a correlation degree greater than the preset correlation threshold are screened out.

[0055] S43: Based on the feature set of abnormal section combination and full process association, a Bayesian network fault causal reasoning model based on the section topology is adopted to locate the source section, fault type and fault impact range of the fault. It covers all fault types such as melt temperature runaway, filter screen blockage, melt rupture, uneven cooling, traction roller slippage, winding tension runaway, die mouth material accumulation, roller surface material sticking and melt degradation, and generates fault diagnosis results.

[0056] The aforementioned method for fault diagnosis and early warning of the entire process of a plastic film extrusion production line based on multi-source data fusion, wherein step S5 specifically includes:

[0057] S51: Based on the full-process correlation fusion feature set of continuous time period, the evolution trend of each feature component is fitted by the sliding window least squares method to obtain the feature trend sequence.

[0058] S52: For fault-free operating conditions, based on feature trend sequences and historical fault sample databases, a 2-6 layer long short-term memory network time series prediction model is used to predict the feature change trend and fault occurrence probability within the next 5-30 minutes; for operating conditions with faults, the fault diagnosis results and feature trend sequences are combined to predict the spread range and deterioration probability of the fault.

[0059] S53: Based on the predicted probability of failure, classify the corresponding early warning levels and output the corresponding level of early warning signal. The early warning levels include low risk warning, medium risk warning and high risk warning, with the corresponding failure probability ranges being 30% to <60%, 60% to <90% and 90% and above, respectively.

[0060] The present invention has the following beneficial effects:

[0061] 1. Establish information exchange channels across all processes in the production line, break the limitations of isolated data in a single work section, comprehensively integrate multiple types of operational information, fully cover the operational status of the entire production chain, and reduce misjudgments and omissions of problems caused by incomplete information coverage.

[0062] 2. Standardize and organize raw operating data of different frequencies and types, and adopt differentiated interpolation methods for data of different frequencies to eliminate time series deviations and spurious feature interference, reduce the interference of messy data on analysis and judgment, improve the quality of data utilization, and make subsequent feature analysis and status judgment more in line with the actual working conditions on site.

[0063] 3. Layered data feature mining and cross-process correlation analysis are completed. Through the original section fault transmission coupling correlation quantification equation, the working condition linkage law between upstream and downstream sections is fully explored. Combined with the coupling characteristics of material transfer and process temperature, the abnormal propagation relationship is quantified, the analysis interference caused by irrelevant data is weakened, and the ability to express key correlation features is strengthened.

[0064] 4. Complete anomaly identification, correlation matching, and source tracing step by step. By separating upstream and downstream disturbances and Bayesian causal reasoning, clearly distinguish between direct anomalies and chain-derived anomalies, accurately locate the location and scope of the problem, simplify the on-site troubleshooting process, and reduce the time and manpower required for troubleshooting.

[0065] 5. By relying on continuous operation data to capture the slow changes in operating conditions, early detection of subtle anomalies can be achieved, and the deterioration of operating conditions and the spread of anomalies can be predicted in advance. This changes the passive handling mode of faults and reduces the impact of sudden problems on continuous production.

[0066] 6. By combining graded early warning and adaptive control command output, and adding a process safety boundary verification mechanism, differentiated control under different risk levels can be achieved, taking into account both the continuity of normal production and equipment safety protection under special working conditions, thereby improving the rationality and timeliness of production control.

[0067] 7. An operational data feedback optimization mechanism is added, adopting a dual-mode incremental learning strategy of fixed cycle + emergency conditions. This allows for continuous adjustment of the analysis logic in response to changes in production conditions, adapting to various production and operational fluctuation scenarios, maintaining stable analysis and judgment capabilities over the long term, and improving the overall environmental adaptability of the solution.

[0068] 8. Reduce the number of unplanned downtimes during production, stabilize the overall quality uniformity of film products, reduce losses caused by long-term abnormal operation of equipment, reasonably control material losses and maintenance costs during production, and improve the operational stability and comprehensive management level of the entire extrusion production line. Attached Figure Description

[0069] Figure 1 This is a schematic diagram illustrating the connection relationships between various units in a full-process fault diagnosis system based on multi-source data fusion provided in an embodiment of the present invention.

[0070] Figure 2 This is a graph showing how fault identification accuracy changes with the severity of the fault.

[0071] Figure 3 A curve showing the lead time for early warning;

[0072] Figure 4 A dual-axis comparison graph of false alarm rate and fault response time;

[0073] Figure 5 A graph showing the accuracy of root cause localization. Detailed Implementation

[0074] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0075] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0076] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0077] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0078] Reference Figures 1-5 This specific implementation provides the following Examples 1-2 and comparative experimental examples, wherein... Figure 1 The interconnections between units in a full-process fault diagnosis system based on multi-source data fusion are demonstrated. Figure 2 The trends in accuracy of fault identification are shown for mild (93.2%), moderate (97.8%), and severe (99.1%) faults. Figure 3This demonstrates that our system provides an average early warning time of 12.6 minutes, compared to the other system's mere 2.3 minutes. Figure 4 The left Y-axis displays the false alarm rate and the right Y-axis displays the fault response time, showing the comparison of the false alarm rate and fault response time between this system and the comparison system. Figure 5 The results show that our system has a root cause localization accuracy of 94.5%, compared to only 52.3% for the comparison system.

[0079] Example 1

[0080] The full-process fault diagnosis system based on multi-source data fusion provided in this embodiment belongs to the field of fault prediction technology. It includes a full-process multi-source data acquisition unit, a multi-source data collaborative preprocessing unit, a two-level fusion feature engineering unit, a hierarchical fault diagnosis root cause localization unit, a fault evolution early warning unit, and a closed-loop execution unit.

[0081] The full-process multi-source data acquisition unit is used to collect multi-dimensional heterogeneous operation data of the entire process of the plastic film extrusion production line and output it to the multi-source data collaborative preprocessing unit.

[0082] The multi-source data collaborative preprocessing unit is used to perform time synchronization alignment, outlier removal, missing value completion and normalization on the received heterogeneous running data to obtain a standardized full-process dataset, which is then output to the two-level fusion feature engineering unit.

[0083] The two-level fusion feature engineering unit is used to perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section association in the entire process. This results in a single-section fusion feature set and a full-process association fusion feature set. The single-section fusion feature set is output to the hierarchical fault diagnosis root cause localization unit, and the full-process association fusion feature set is output to the hierarchical fault diagnosis root cause localization unit and the fault evolution early warning unit.

[0084] The hierarchical fault diagnosis root cause localization unit is used to complete single-section anomaly identification based on single-section fusion feature set, complete cross-section fault transmission association matching and fault root cause localization based on full-process association fusion feature set, and output fault diagnosis results to fault evolution early warning unit and closed-loop execution unit.

[0085] The fault evolution early warning unit is used to complete early fault trend warning under fault-free conditions based on the full-process correlation and fusion feature set, and to complete the evolution trend warning of the faults that have occurred by combining the fault diagnosis results, and output the corresponding level of fault warning signal to the closed-loop execution unit.

[0086] The closed-loop execution unit is used to generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding section based on the received fault diagnosis results and fault warning signals, and send them to the local control system of the plastic film extrusion production line. At the same time, it collects the operation feedback data after the instruction is executed and outputs it to the multi-source data collaborative preprocessing unit to complete the online iterative optimization of feature distribution adaptation and diagnostic model.

[0087] The above-mentioned solution has constructed a complete control system covering the entire production process, which has enabled data flow and information linkage in all aspects of production, realized full-process tracking and control of production operation status, and can simultaneously complete the identification of operational anomalies, the prediction of development trends and the coordinated execution of handling actions, forming a complete closed loop of operation control. This effectively reduces the problems of missed detection, misjudgment and delayed handling of anomalies in the production process, ensures the continuous and stable operation of the production process, and reduces the risks to product quality and equipment operation caused by production fluctuations.

[0088] Specifically, in this embodiment, the full-process multi-source data acquisition unit includes section acquisition sub-units deployed sequentially in the raw material batching section, extrusion plasticizing section, melt filtration section, film forming and cooling section, traction stretching section, and winding section of the plastic film extrusion production line, as well as a supporting environmental parameter acquisition sub-unit.

[0089] Each section's data acquisition subunit includes a process parameter acquisition module, an equipment status acquisition module, and a product quality online detection module;

[0090] The process parameter acquisition module is used to collect temperature, pressure, melt flow rate, motor speed, and material ratio parameters for the corresponding process section. The acquisition frequency is set to 1Hz-100Hz.

[0091] The equipment status acquisition module is used to collect the current, voltage, vibration, and noise parameters of the drive motor of the corresponding section, as well as the running displacement and fit clearance parameters of the transmission components. The acquisition frequency is set to 100Hz-2000Hz.

[0092] The online product quality monitoring module is used to collect melt index parameters of materials in the raw material batching and extrusion plasticizing sections, and to collect parameters of film thickness, number of crystal points, tensile strength, haze, and heat shrinkage rate in the film forming and cooling section to the winding section. The sampling frequency is set to 0.5Hz-50Hz.

[0093] The environmental parameter acquisition subunit is used to collect parameters such as ambient temperature, ambient humidity, and dust concentration in the production line operating environment. The acquisition frequency is set to 0.1Hz-5Hz.

[0094] By adopting the above solution, comprehensive coverage of the operation information of the entire production process is achieved. It can simultaneously acquire multi-dimensional information on equipment operation, process execution, product status and working environment during the production process, avoiding the status judgment bias caused by single-dimensional information. It provides a comprehensive and complete information foundation for subsequent operation status control and reduces the abnormal omissions caused by missing information.

[0095] Specifically, in this embodiment, the multi-source data collaborative preprocessing unit includes a data time synchronization module, a data cleaning module, and a data normalization module;

[0096] The data time synchronization module uses the system clock of the production line main controller as a reference. It adopts linear interpolation for high-frequency equipment status data of 100Hz-2000Hz and zero-order hold method for process, quality and environmental data of 0.1Hz-100Hz. It performs timestamp synchronization and alignment on heterogeneous operating data with different acquisition frequencies. The time step after alignment is uniformly set to 0.0005s. This time step is equal to the sampling period corresponding to the highest sampling frequency (2000Hz) among all acquisition modules.

[0097] The data cleaning module uses a combination of Grubbs' criterion and the sliding window mean method to remove outliers and fill in missing values ​​in the synchronized data. The length of the sliding window is set to 5-100 time steps, and the significance level of Grubbs' criterion is set to 0.05.

[0098] The data normalization module uses the min-max normalization method to map the cleaned data to the 0-1 range, resulting in a standardized full-process dataset.

[0099] By adopting the above scheme, the operational information collected from different sources was processed in a unified and standardized manner, eliminating the time sequence deviation and invalid interference information between different information, avoiding the false feature problem caused by linear interpolation of low-frequency data, providing a unified processing basis for information of different dimensions, improving the accuracy and consistency of subsequent information processing, and avoiding the interference of deviations in the original information on subsequent state judgments.

[0100] Specifically, in this embodiment, the two-level fusion feature engineering unit includes a single-section feature-level fusion module and a full-process cross-section associated feature fusion module;

[0101] The single-section feature-level fusion module extracts high-dimensional hidden layer features from the multi-source data corresponding to each section using a stacked autoencoder with 3 to 8 layers for the standardized dataset of each section. The number of neurons in the encoder and decoder layers of the stacked autoencoder is symmetrically set. The hidden layer activation function is the ReLU function, and the output layer uses the Sigmoid function. Training is completed by unsupervised pre-training combined with supervised fine-tuning of small samples. The pre-training stage uses steady-state operation data of the production line, and the fine-tuning stage uses the class imbalance loss function combined with the hard sample mining strategy. Then, the attention mechanism is used to dynamically assign weights to features of different types of parameters. The weight assignment values ​​range from 0 to 1 to obtain the single-section fusion features of each section. These are summarized to form a single-section fusion feature set, which is output to the hierarchical fault diagnosis root cause localization unit and the full-process cross-section correlation feature fusion module.

[0102] The full-process cross-section association feature fusion module, based on the sequential timing and material transfer logic of the plastic film extrusion production line, constructs a directed acyclic graph (DAG) topology for section association. Each section is treated as a graph node, the material and energy transfer relationships between sections are represented as directed edges, and the fusion features of individual sections are used as node features. The prior weights of the edges are dynamically adjusted using the section fault propagation coupling coefficient. A 2-4 layer graph attention neural network is employed to extract full-process association features from the fusion features of individual sections across all sections. The final edge weights are calculated using a weighted fusion of the prior coupling coefficient and the network-learned attention weights. The fusion formula is as follows: ,in Let be the edge weight, and λ be a preset prior weight coefficient, with a value range of [0.4, 0.8]. The coupling coefficient for fault propagation in the work section. The attention weights learned by the graph attention neural network are used to obtain a full-process correlation fusion feature set containing cross-section fault transmission correlation information, which is then output to the hierarchical fault diagnosis root cause localization unit and fault evolution early warning unit.

[0103] By adopting the above scheme, the operational information of each section was processed in a layered and progressive manner. First, the internal information of each section was integrated and refined. Then, by combining the sequence of production sections and the material flow relationship, the information association and in-depth mining between each section of the entire production process were completed. Through the integration of process prior rules and data-driven learning, the inherent correlation information between each link of production was fully explored, avoiding the limitations of state judgment caused by the isolation of information in each section, and effectively improving the overall control capability of production operation status.

[0104] Specifically, in this embodiment, the hierarchical fault diagnosis root cause localization unit includes a single-section anomaly identification module, a cross-section fault transmission association matching module, and a root cause localization module;

[0105] The single-section anomaly identification module, based on the feature components of the corresponding section in the single-section fusion feature set, uses a pre-trained single-class support vector machine anomaly detection model to identify whether there are operational anomalies in the corresponding section. The single-class support vector machine uses a Gaussian kernel function, and the kernel function bandwidth is determined by grid search combined with 5-fold cross-validation. The anomaly judgment threshold is calibrated by the 3σ criterion. The single-section anomaly identification result is output to the cross-section fault transmission association matching module.

[0106] The cross-section fault transmission association matching module, based on the section association topology of the production line and combined with the single section anomaly identification results, uses the maximum mutual information coefficient to calculate the nonlinear correlation degree between the abnormal features of different sections, filters out the abnormal section combinations with a correlation degree greater than a preset correlation threshold, and outputs them to the root cause localization module. The preset correlation threshold ranges from 0.5 to 0.7 and is calibrated by combining the receiver operating characteristic curve (ROC curve) with the production line process requirements.

[0107] The root cause localization module, based on the fusion feature set of abnormal work section combinations and the whole process association, adopts a Bayesian network fault causal reasoning model constructed based on the work section topology to locate the source work section, fault type and fault impact range of the fault, and generate fault diagnosis results. The Bayesian network uses the abnormal state of each work section as the root node, the fault type as the leaf node, and the work section coupling relationship as the directed edge. The conditional probability table is determined by the maximum likelihood estimation method based on expert experience and historical fault data. The reasoning algorithm adopts the connection tree algorithm. Fault types include melt temperature runaway, filter screen blockage, melt fracture, uneven cooling, traction roller slippage, and winding tension runaway.

[0108] By adopting the above scheme, a hierarchical and progressive identification and location of production operation anomalies was achieved. First, the independent operating status of each section was judged. Then, combined with the inherent correlation of the production process, the relationship between the anomalies in each section was sorted out. Finally, the source and scope of the anomaly were accurately located, avoiding confusion between the source of the anomaly and secondary impacts. This effectively improved the accuracy of anomaly location and provided an accurate basis for subsequent handling actions.

[0109] Specifically, in this embodiment, the fault evolution early warning unit includes a feature trend extraction module, a fault evolution prediction module, and an early warning level classification module;

[0110] The feature trend extraction module, based on the full-process correlation and fusion feature set of continuous time period, uses the sliding window least squares method to fit the evolution trend of each feature component to obtain the feature trend sequence. The length of the sliding window is set to 10-200 time steps.

[0111] The fault evolution prediction module, for fault-free operating conditions, uses a 2-6 layer long short-term memory network time series prediction model based on feature trend sequences and a historical fault sample library. The model input is a feature trend sequence of 20-100 consecutive time steps, and the output is the feature change trend and fault occurrence probability within a preset future time period. The mean squared error loss function combined with the Adam optimizer is used to complete the training, predicting the feature change trend and fault occurrence probability within the preset future time period. For operating conditions with existing faults, the module combines fault diagnosis results and feature trend sequences to predict the spread range and deterioration probability of the fault. The preset time period ranges from 5 min to 30 min.

[0112] The early warning level classification module classifies the corresponding early warning levels based on the predicted probability of failure and outputs the corresponding early warning signals. The early warning levels include low-risk, medium-risk, and high-risk warnings, with corresponding failure probability ranges of 30% to <60%, 60% to <90%, and 90% and above, respectively.

[0113] By adopting the above scheme, it is possible to fit the changing trend of the operating status based on continuous operating information, combine historical operating conditions, predict the subsequent changes in the operating status and the possibility of anomalies, and make advance predictions for different operating states when no anomalies have occurred and predictions of the development trend when anomalies have occurred. Different alert levels are set according to the probability of anomalies, thereby realizing advance control of production operation risks and reducing the impact of sudden anomalies on production.

[0114] Specifically, in this embodiment, the closed-loop execution unit includes an instruction generation module, an instruction issuance module, and an execution feedback module;

[0115] The instruction generation module generates corresponding control instructions based on the fault diagnosis results and the warning level of the fault warning signal. Specifically, for low-risk warnings, it generates process parameter fine-tuning instructions for the corresponding section; for medium-risk warnings, it generates process parameter adjustment instructions and equipment inspection reminder instructions for the corresponding section; for high-risk warnings, it generates load reduction operation instructions or equipment shutdown protection instructions for the corresponding section; and for the located faults, it generates corresponding process correction instructions.

[0116] The command issuing module sends the generated control commands to the local programmable logic controller of the plastic film extrusion production line via industrial Ethernet or fieldbus. Before the commands are issued, they must be verified by the process safety boundary. Commands that exceed the process safety limit will be intercepted and manual review will be triggered.

[0117] The execution feedback module is used to collect production line operation data after the control command is executed and output it to the multi-source data collaborative preprocessing unit. The preprocessed feedback data is used for feature distribution adaptation of the two-level fusion feature engineering unit and online incremental learning and iterative optimization of the diagnostic model in the hierarchical fault diagnosis root cause localization unit. The update cycle of incremental learning is set to 12h-24h, and the rapid update cycle under emergency conditions is set to 1h-12h. Incremental learning adopts a sliding time window dataset update method to retain nearly 3 months of valid operation data.

[0118] By adopting the above solution, matching production adjustment and control instructions can be generated based on the judgment results of the operating status and the risk prediction results. These instructions are then sent to the control equipment on the production site for execution. Safety verification avoids production accidents caused by out-of-control instructions. At the same time, production operation feedback information after instruction execution is collected for subsequent operation status control and judgment logic optimization. This achieves closed-loop control of the entire chain from status identification and risk prediction to disposal execution and effect feedback, allowing production control actions to adapt to real-time changes in production operation status and improving the effectiveness and timeliness of production control.

[0119] Example 2

[0120] This embodiment provides a method for fault diagnosis and early warning of the entire process of a plastic film extrusion production line based on multi-source data fusion, which is applied to the full-process fault diagnosis system based on multi-source data fusion in Embodiment 1, and includes the following steps:

[0121] S1: Collect multi-dimensional heterogeneous operational data of the entire process of the plastic film extrusion production line;

[0122] S2: Perform time synchronization alignment, outlier removal, missing value completion, and normalization on the collected heterogeneous operational data to obtain a standardized full-process dataset;

[0123] S3: Perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section association in the entire process. Based on the section fault transmission coupling quantization equation with limited boundaries, dynamically optimize the cross-section feature association weights to obtain the single-section fusion feature set and the full-process association fusion feature set.

[0124] S4: Based on the single-section fusion feature set, complete the single-section anomaly identification; based on the full-process association fusion feature set, complete the cross-section fault transmission association matching and fault root cause location, and generate fault diagnosis results.

[0125] S5: Based on the full-process correlation and fusion feature set, complete the early fault trend warning under fault-free working conditions, combine the fault diagnosis results to complete the evolution trend warning of the faults that have occurred, divide the warning level according to the non-overlapping probability boundary, and output the corresponding level of fault warning signal.

[0126] S6: Based on the fault diagnosis results and fault warning signals, generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding work section. After abnormal instruction interception and verification, the instructions are sent to the local control system of the production line for execution. At the same time, the operation feedback data after instruction execution is collected, and the feature distribution adaptation and online iterative optimization of the diagnostic model are completed in a dual mode of fixed cycle + emergency working condition.

[0127] By adopting the above solution, a complete control process covering the entire production process has been formed. The operation logic of all steps from production information acquisition to disposal execution has been standardized, and the information flow and action linkage of each link in the production process have been opened up. It has realized the full-process tracking and control of the production operation status, and can simultaneously complete the identification of operational anomalies, the prediction of development trends and the linkage execution of disposal actions, forming a complete operation control closed loop. This effectively reduces the problems of missed judgment, misjudgment and delayed disposal of anomalies in the production process, and ensures the continuous and stable operation of the production process.

[0128] Specifically, in this embodiment, step S3 includes:

[0129] S31: For each work section's standardized dataset, a stacked autoencoder with 4 to 6 layers is preferred to extract high-dimensional hidden layer features from the multi-source data corresponding to each work section. Then, an attention mechanism is used to dynamically assign weights to features of different types of parameters to obtain the single-work section fusion features of each work section, which are then summarized to form a single-work section fusion feature set.

[0130] S32: Based on the sequential timing and material transfer logic of the plastic film extrusion production line, a directed acyclic graph is constructed to establish a section association topology. Each section is used as a graph node, the material and energy transfer relationship between sections is used as a directed edge, and the single section fusion feature is used as the node feature. A 2-3 layer graph attention neural network is preferably used to extract the full-process association features of the single section fusion features of all sections, resulting in a full-process association fusion feature set containing cross-section fault transmission association information.

[0131] The above scheme standardizes the hierarchical and progressive processing flow of production operation information. First, it integrates and refines the information within each work section. Then, it combines the sequence of production sections and the material flow relationship to complete the information association and in-depth mining between each work section of the entire production process. This fully explores the inherent correlation information between each link of production, avoids the limitations of status judgment caused by the isolation of information in each work section, effectively improves the overall control capability of production operation status, and provides a more production-realistic information foundation for subsequent operation status judgment.

[0132] Specifically, step S32 is further defined as follows: Based on the sequential timing and material transfer logic of the plastic film extrusion production line, a directed acyclic graph is constructed to establish a section association topology. Each section is used as a graph node, the material and energy transfer relationship between sections is used as a directed edge, and the single section fusion feature is used as a node feature. The coupling coefficient is calculated by substituting the section fault transmission coupling association quantification equation and updating the prior weight of the edge. Preferably, a 2-3 layer graph attention neural network is used to extract the full-process association features of the single section fusion features of all sections. Finally, the edge weight is calculated by weighted fusion of the prior coupling coefficient and the network learning attention weight to obtain a full-process association fusion feature set containing cross-section fault transmission association information.

[0133] The quantitative equation for the fault transmission coupling correlation of the work section is expressed as follows:

[0134]

[0135] In the formula:

[0136] : Fault propagation coupling coefficient between the i-th and j-th work sections, dimensionless, with a value range of (0,1). The larger the value, the stronger the correlation between the abnormal propagation of the two work sections.

[0137] : The real-time operating condition fluctuation offset of upstream section i is calculated by fusion of multiple operating parameters such as temperature, pressure and speed through minimum-maximum normalization to achieve same-scale equilibrium. It represents the degree of deviation of the operation of a single section from the steady state. It is dimensionless and has a value range of [0,1].

[0138] The interpretable fluctuation component generated by the downstream j section affected by the operating condition fluctuation of the upstream i section is obtained by separating the downstream self-disturbance and the upstream transmitted disturbance through partial least squares method. It is dimensionless and takes values ​​in the range of [0,1] after minimum-maximum normalization.

[0139] : Material conveying disturbance attenuation coefficient from section i to section j, dimensionless, value range (0,1]. The larger the value, the more stable the material conveying, the smaller the attenuation of upstream operating condition fluctuations transmitted to downstream, and the stronger the fault transmission capability. It is obtained by statistical analysis of melt conveying rate and material residence time under rated operating conditions of the production line, and after minimum-maximum normalization processing.

[0140] : The absolute difference between the actual temperature of the melt at the outlet of section i and the actual temperature of the melt at the inlet of section j, in °C;

[0141] The maximum allowable melt temperature deviation between process section i and process section j, in °C, is used to normalize the temperature deviation term and ensure overall dimensional uniformity.

[0142] The derivation of the equation is as follows:

[0143] Step 1. Plastic film extrusion production is a continuous material transfer process. Upstream operating condition fluctuations will gradually propagate downstream along the material flow direction. The intensity of abnormal propagation depends first on the amplitude of the upstream fluctuations and the degree of response of the downstream affected by the upstream. Therefore, the basic correlation term adopts the product relationship between upstream fluctuations and downstream explainable responses to reflect the positive transmission law of operating condition disturbances. At the same time, the causal inversion is avoided by separating the downstream disturbances themselves.

[0144] Step 2. The stability of material conveying directly restricts the efficiency of fault propagation. The more stable the conveying state, the weaker the cross-section abnormal attenuation and the stronger the fault transmission capability. Therefore, the material conveying disturbance attenuation coefficient is introduced as the numerator to realize the positive constraint of transmission intensity, which conforms to the physical law of continuous extrusion material conveying.

[0145] Step 3. In the film extrusion process, the matching degree of melt temperature between adjacent sections is a key condition affecting the melt rheological state and fault generation. The greater the actual temperature deviation of the melt between sections, the more obvious the difference in material properties, and the weaker the fault transmission ability across sections.

[0146] Step 4. Based on the temperature matching law, construct a temperature deviation normalization correction term, and use the actual melt temperature difference between adjacent sections to constrain the maximum allowable deviation limit of the process, so as to linearly quantify the coupling weakening effect caused by temperature deviation and avoid calculation distortion caused by extreme temperature difference.

[0147] Step 5. By using a limiting function, the coupling coefficients are strictly constrained within the (0,1) interval to avoid numerical overflow and ensure the stability of the weight calculation of the graph attention neural network.

[0148] Step 6. Integrate the basic transmission terms, transport constraint terms, process temperature correction terms, and limiting functions to finally obtain the quantitative equation for the coupling correlation of fault transmission in the work section.

[0149] Example:

[0150] 1. Select the extrusion plasticizing section as section i and the melt filtration section as the adjacent downstream section j, and collect the operating parameters and quality inspection data of the two sections in real time.

[0151] 2. The upstream operating condition fluctuation offset, the downstream explained fluctuation component affected by the upstream, and the cross-section material transport disturbance attenuation coefficient were calculated respectively.

[0152] 3. Collect the actual temperature values ​​of the melt at the inlet and outlet of the two sections, and calculate the temperature correction factor in combination with the maximum allowable temperature deviation limit of the process.

[0153] 4. Substitute the coupling correlation quantization equation to calculate the real-time fault transmission coupling coefficient between the two work sections, and use the amplitude limiting function to ensure that the value falls within the (0,1) interval.

[0154] 5. Write this coefficient as a priori weight into the directed edge of the graph attention neural network, and perform weighted fusion with the attention weights learned by the network. For strongly coupled work sections, increase the cross-work section feature fusion weight, and for weakly coupled work sections, reduce the feature interference weight.

[0155] 6. Based on the revised weights, the entire process of related features is fused, making the fused features more consistent with the actual fault propagation path and improving the accuracy of root cause localization.

[0156] The working principle and process of the equation:

[0157] 1. After the system completes the preprocessing of multi-source data and the fusion of single-section features, it divides the upstream and downstream related nodes based on the topology of the production line section.

[0158] 2. Calculate the fault transmission coupling coefficient between adjacent work sections group by group according to the time period. For non-adjacent long-distance coupled work sections, the coupling coefficient is calculated by multiplying the path transmission coefficients of the directed acyclic graph.

[0159] 3. Using the coupling coefficient as the basis for dynamic prior weights of topological edges, the association weights of cross-segment feature fusion throughout the entire process are updated in real time.

[0160] 4. The weighted and optimized associated feature set is synchronously input into the fault diagnosis unit and the early warning unit.

[0161] 5. During the diagnosis process, priority should be given to strengthening the analysis of the correlation characteristics of strongly coupled sections to identify material transmission-related cascading faults.

[0162] 6. In the early warning stage, the scope and speed of the anomaly's downstream spread are predicted based on the coupling and correlation laws, which assists in the determination of graded early warning.

[0163] 7. Finally, combine the results of the coupling and correlation to output accurate diagnostic conclusions and treatment instructions.

[0164] Equation Technique Effects:

[0165] 1. Combining the practical characteristics of continuous material conveying and multi-stage process coupling in plastic film extrusion, this method enables quantitative evaluation of the degree of abnormal propagation in adjacent stages and long-distance stages, thus overcoming the limitations of scenario adaptation caused by fixed weight settings.

[0166] 2. Dynamically adjust the proportion of cross-section feature fusion based on the actual coupling strength between sections, reduce the interference of invalid information from non-related sections, and strengthen the expression of related features on the material transfer path.

[0167] 3. Taking into account the impact of the actual temperature matching characteristics of the melt in adjacent sections on the melt state and fault propagation, and conforming to the material rheological laws of film extrusion production, the multi-source data fusion results are more in line with the actual on-site operating conditions.

[0168] 4. Effectively suppresses feature redundancy caused by large-scale data fusion across multiple work sections, narrows the analysis scope of fault tracing, and reduces the difficulty of identifying chain-like production anomalies.

[0169] 5. Provide quantitative evidence for the evolution and spread of faults across work sections, improve the accuracy of medium- and long-term risk prediction, and make graded early warning and production response adjustments more targeted.

[0170] 6. Strict amplitude constraints ensure stable coefficient values, adapt to the weight calculation logic of graph attention neural networks, and avoid numerical explosion and non-convergence problems during model training.

[0171] Specifically, in this embodiment, step S4 includes:

[0172] S41: Based on the feature components of the corresponding work section in the single-section fusion feature set, a pre-trained single-class support vector machine anomaly detection model is used to identify whether there is an operational anomaly in the corresponding work section and generate the single-section anomaly identification result.

[0173] S42: Based on the production line section association topology, combined with the single section anomaly identification results, the nonlinear correlation degree between different section anomaly features is calculated using the maximum mutual information coefficient. 0.5-0.7 is selected as the preferred correlation threshold, and abnormal section combinations with a correlation degree greater than the preset correlation threshold are screened out.

[0174] S43: Based on the feature set of abnormal section combination and full process association, a Bayesian network fault causal reasoning model based on the section topology is adopted to locate the source section, fault type and fault impact range of the fault. It covers all fault types such as melt temperature runaway, filter screen blockage, melt rupture, uneven cooling, traction roller slippage, winding tension runaway, die mouth material accumulation, roller surface material sticking and melt degradation, and generates fault diagnosis results.

[0175] The above scheme standardizes the hierarchical and progressive identification and location process for production operation anomalies. First, the independent operating status of each section is judged. Then, combined with the inherent correlation of the production process, the relationship between the anomalies in each section is sorted out. Finally, the source and scope of the anomaly are accurately located, avoiding confusion between the source of the anomaly and secondary impacts. This effectively improves the accuracy of anomaly location and provides an accurate basis for subsequent handling actions.

[0176] Specifically, in this embodiment, step S5 includes:

[0177] S51: Based on the full-process correlation fusion feature set of continuous time period, the evolution trend of each feature component is fitted by the sliding window least squares method to obtain the feature trend sequence.

[0178] S52: For fault-free operating conditions, based on feature trend sequences and historical fault sample databases, a 2-6 layer long short-term memory network time series prediction model is used to predict the feature change trend and fault occurrence probability within the next 5-30 minutes; for operating conditions with faults, the fault diagnosis results and feature trend sequences are combined to predict the spread range and deterioration probability of the fault.

[0179] S53: Based on the predicted probability of failure, classify the corresponding early warning levels and output the corresponding level of early warning signal. The early warning levels include low risk warning, medium risk warning and high risk warning, with the corresponding failure probability ranges being 30% to <60%, 60% to <90% and 90% and above, respectively.

[0180] The above-mentioned scheme standardizes the process of predicting and alerting production operation risks. It can fit the changing trend of the operating status based on continuous operating information, and predict the subsequent changes in the operating status and the possibility of anomalies by combining historical operating data. At the same time, it can make advance predictions when no anomalies have occurred and predictions of the development trend when anomalies have occurred for different operating states. Different alert levels are set according to the probability of anomalies, so as to realize the early control of production operation risks and reduce the impact of sudden anomalies on production.

[0181] Comparative experimental cases

[0182] The performance of this system and the control system was compared on a plastic film extrusion production line.

[0183] To verify the technical superiority of this invention, a three-month on-site comparative test was conducted on the same plastic film extrusion production line using a comparative system (a digital twin-based industrial production line fault self-diagnosis system, publication number CN120722874A). The production line consisted of six main stages: raw material batching, extrusion plasticizing, melt filtration, molding and cooling, traction stretching, and winding. During the test, both systems were deployed, and identical full-process operation data were collected to compare key performance indicators.

[0184] 1. Test Condition Settings

[0185] Test subject: A company's MDO-1200 plastic film extrusion production line, producing BOPP film with a daily output of approximately 15 tons.

[0186] Fault types: Six types of faults that occur frequently in the industry were selected, including melt temperature runaway, filter screen blockage, melt rupture, uneven cooling, traction roller slippage, and winding tension runaway.

[0187] Fault injection method: Without affecting normal production, three levels of faults—mild, moderate, and severe—are artificially created through process parameter offset or equipment status simulation, with a total of 72 effective fault injections (12 per level).

[0188] Evaluation indicators:

[0189] Fault identification accuracy (%); Root cause location accuracy (%); Average warning lead time (minutes).

[0190] False alarm rate (times / 24 hours); average fault response and handling time (minutes).

[0191] 2. Comparison Results Data Table:

[0192] Fault identification accuracy (overall) 96.7% 71.2% +25.5% minor fault identification accuracy 93.2% 48.6% +44.6% Accuracy of medium fault identification 97.8% 76.3% +21.5% Accuracy of severe fault identification 99.1% 88.7% +10.4% Root cause localization accuracy (source section localization) 94.5% 52.3% +42.2% Average warning lead time 12.6 minutes 2.3 minutes +10.3 minutes False alarm rate (times / 24 hours) 0.8 times 3.5 times -77.1% Mean time to failure 3.2 minutes 14.5 minutes -77.9%

[0193] 3. Comparative Analysis of Typical Failure Scenarios

[0194] Scenario 1: A chain reaction of failures caused by slow temperature drift of the upstream melt (extrusion plasticizing section → melt filtration section → molding cooling section)

[0195] Comparison system: When the temperature in the extrusion plasticizing section exceeded the threshold by approximately 8°C, an abnormal temperature alarm was triggered in the extruder. However, the system could not identify the impact of this abnormality on the risk of subsequent filter clogging and film thickness deviation. The fault was not detected until the pressure differential in the melt filtration section increased by 30%, at which point approximately 200 kg of defective film had already been produced.

[0196] This system, when the temperature deviates from the steady-state value by only 3°C, predicts the risk of filter clogging and the trend of film thickness deviation 18 minutes in advance by extracting cross-process correlation features and coupling quantitative calculation of process fault transmission. It issues a medium-risk warning and automatically adjusts the extruder heating power and melt pump speed, successfully preventing the fault from spreading downstream and producing no defective products.

[0197] Scenario 2: Slight slippage of the traction roller causing tension fluctuations during winding.

[0198] The comparison system only identifies anomalies when the current fluctuation of the traction roller drive motor exceeds 15%, but it cannot distinguish whether the anomaly is caused by slippage of the traction roller itself or by changes in the coefficient of friction of the film surface due to uneven cooling upstream. It misjudges it as a motor failure, leading to unnecessary shutdowns for maintenance.

[0199] This system: By using a Bayesian network fault causal reasoning model, combined with the crystallinity characteristics of the thin film in the upstream cooling section and the vibration spectrum characteristics of the traction roller, it accurately locates the cause of the fault as micro-slippage caused by wear on the surface of the traction roller, and generates process correction instructions (fine-tuning the clamping force of the traction roller and the outlet air temperature of the cooling section), allowing the production line to run continuously without stopping.

[0200] 4. Data Analysis and Conclusions

[0201] The above comparative data shows that, compared with the general digital twin diagnostic solution of the comparison system CN120722874A, this system has achieved significantly better results by improving the following key technologies to address the multi-stage, strongly coupled, and continuous material transfer characteristics of plastic film extrusion production lines:

[0202] 1. Multi-dimensional heterogeneous data fusion capability: This system collects four types of data: process, equipment, quality, and environment. The comparison file is mainly based on equipment operation data. Therefore, the system improves the accuracy of identifying minor faults (early weak signs) by 44.6%.

[0203] 2. Cross-section fault transmission correlation quantification: The system’s unique cross-section fault transmission coupling correlation quantification equation and graph attention neural network can dynamically capture the propagation patterns of upstream and downstream anomalies, improving the root cause location accuracy by 42.2% and extending the early warning lead time from 2.3 minutes to 12.6 minutes.

[0204] 3. Closed-loop execution and iterative optimization: This system continuously adapts to changes in operating conditions by incrementally learning online through instruction execution and feedback data, reducing the false alarm rate by 77.1% and shortening the fault response and handling time by 77.9%.

[0205] In summary, the present invention is significantly superior to the prior art in terms of fault identification sensitivity, root cause location accuracy, early warning timeliness, and field adaptability, and is particularly suitable for the intelligent operation and maintenance of the entire process of plastic film extrusion production line.

[0206] In summary, the overall working principle of this solution is as follows:

[0207] This solution leverages the complete process layout of a plastic film extrusion production line to establish a full-process data acquisition system. It simultaneously collects various operational information, including process status, equipment status, product quality, and on-site environmental data, from each stage. All heterogeneous data is uniformly aligned in time sequence, impurity data is removed, and standardized to eliminate data discrepancies arising from different acquisition methods, providing a well-organized data foundation for subsequent unified analysis.

[0208] A layered and progressive fusion processing model is adopted. First, feature integration and refinement of multiple parameters are completed within a single work section. Then, combined with the material conveying sequence and process layout logic of the production line, a work section association topology is built to achieve feature association mining across processes throughout the entire process. Relying on a unique quantitative calculation method for work section fault propagation coupling association, the strength of anomaly propagation between adjacent processes is quantified, and the matching ratio of cross-work section feature fusion is dynamically adjusted to conform to the objective laws of working condition coupling and material transfer in actual production.

[0209] Based on the fused information after hierarchical processing, the independent operating status of each work section is first determined. Then, cross-work section anomaly matching analysis is carried out by combining the correlation between work processes. The source of the problem and the scope of impact are determined by combining causal reasoning logic. The system continuously captures the operating characteristic data of consecutive time periods, fits the trend of state changes, and predicts the evolution trend of subsequent working conditions by combining historical operating patterns. Different early warning levels are divided according to the level of risk.

[0210] The system combines diagnostic results with trend warnings to generate corresponding control commands for process fine-tuning, equipment inspection load adjustment, or safety protection. These commands are then sent to the local control devices on the production line for automatic execution. Simultaneously, the system collects on-site operational data after command execution, feeding it back to the data processing and analysis module. This continuously optimizes the feature fusion logic and discriminant analysis rules, forming a complete operational logic for data acquisition, processing, fusion, diagnosis, positioning, trend warning, control, execution, and iterative optimization.

[0211] Overall usage instructions:

[0212] During the equipment deployment phase, following the complete process of raw material batching, extrusion, plasticizing, melt filtering, molding, cooling, traction stretching, and winding, corresponding acquisition modules are deployed in each section. These modules, along with environmental acquisition components, complete the on-site full coverage deployment, establish communication connections between each module and the local control device on the production line, and build a complete data transmission link.

[0213] In the early stage of system operation, basic parameter settings are completed, the data statistics window range of the allowable deviation range of process temperature is defined, and the associated judgment benchmark conditions are determined. Based on the steady-state operation data of production, the stable benchmark parameters of material transportation between each section are generated, and the initial configuration and parameter initialization of the analysis model are completed.

[0214] During daily production operations, various acquisition modules continuously collect on-site operational data in real time, automatically entering the preprocessing stage to complete unified correction and standardization. The system continuously performs two-level feature fusion calculations, synchronously calculates the coupling correlation coefficients of adjacent work sections in real time, dynamically updates the cross-process feature correlation weights, and continuously conducts full-process status analysis.

[0215] During production, real-time status monitoring of work sections and cross-process correlation analysis are performed. Once an operational anomaly is identified, correlation matching and source location are completed step by step, and status trend projections are conducted simultaneously. Corresponding alerts are output based on the risk level. Operators can combine the diagnostic content and early warning prompts provided by the system with the automatic control commands issued by the system to complete process parameter adjustments and equipment management operations.

[0216] During long-term use, based on feedback data after on-site control, the analysis logic and fusion rules are continuously updated to adapt to the changes in operating conditions brought about by different production batches and different raw material specifications. This ensures that the overall analysis and control logic continuously matches the actual operating conditions of the production line, guaranteeing long-term stable operation and adaptability.

[0217] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A full-process fault diagnosis system based on multi-source data fusion, characterized in that, It includes a full-process multi-source data acquisition unit, a multi-source data collaborative preprocessing unit, a two-level fusion feature engineering unit, a hierarchical fault diagnosis and root cause localization unit, a fault evolution early warning unit, and a closed-loop execution unit, among which: The full-process multi-source data acquisition unit is used to collect multi-dimensional heterogeneous operation data of the entire production line process and output it to the multi-source data collaborative preprocessing unit. The multi-source data collaborative preprocessing unit is used to perform time synchronization alignment, outlier removal, missing value completion and normalization processing on the received heterogeneous running data to obtain a standardized full-process dataset, which is then output to the two-level fusion feature engineering unit. The two-level fusion feature engineering unit is used to perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section associations in the entire process. The coupling coefficient is calculated using the section fault transmission coupling association quantification equation, and the cross-section feature association weights are dynamically optimized to obtain the single-section fusion feature set and the full-process association fusion feature set. The single-section fusion feature set is output to the hierarchical fault diagnosis root cause localization unit, and the full-process association fusion feature set is output to the hierarchical fault diagnosis root cause localization unit and the fault evolution early warning unit. The hierarchical fault diagnosis root cause localization unit is used to complete single-section anomaly identification based on single-section fusion feature set, complete cross-section fault transmission association matching and fault root cause localization based on full-process association fusion feature set, and output fault diagnosis results to fault evolution early warning unit and closed-loop execution unit. The fault evolution early warning unit is used to complete early fault trend warning under fault-free operating conditions based on the full-process correlation and fusion feature set, complete the evolution trend warning of the faults that have occurred in combination with the fault diagnosis results, and output the corresponding level of fault warning signal to the closed-loop execution unit. The closed-loop execution unit is used to generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding section based on the received fault diagnosis results and fault warning signals, and send them to the local control system of the production line. At the same time, it collects the operation feedback data after the instruction is executed and outputs it to the multi-source data collaborative preprocessing unit to complete the online iterative optimization of feature distribution adaptation and diagnostic model.

2. The end-to-end fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The full-process multi-source data acquisition unit includes section acquisition sub-units deployed sequentially in the production line for the raw material batching section, extrusion plasticizing section, melt filtration section, film forming and cooling section, traction stretching section, and winding section, as well as a supporting environmental parameter acquisition sub-unit. Each section's data acquisition subunit includes a process parameter acquisition module, an equipment status acquisition module, and a product quality online detection module; The process parameter acquisition module is used to collect temperature, pressure, melt flow rate, motor speed and material ratio parameters of the corresponding process section, and the acquisition frequency is set to 1Hz-100Hz. The equipment status acquisition module is used to collect the current, voltage, vibration, and noise parameters of the drive motor of the corresponding section, as well as the running displacement and fit clearance parameters of the transmission components. The acquisition frequency is set to 100Hz-2000Hz. The product quality online detection module is used to collect material melt index parameters in the raw material batching section and the extrusion plasticizing section, and to collect film thickness, number of crystal points, tensile strength, haze and heat shrinkage parameters in the film forming and cooling section to the winding section. The collection frequency is set to 0.5Hz-50Hz. The environmental parameter acquisition subunit is used to collect environmental temperature, humidity, and dust concentration parameters of the production line operating environment, with the acquisition frequency set to 0.1Hz-5Hz.

3. The full-process fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The multi-source data collaborative preprocessing unit includes a data time synchronization module, a data cleaning module, and a data normalization module. The data time synchronization module uses the system clock of the production line main controller as a reference. It adopts linear interpolation for high-frequency equipment status data of 100Hz-2000Hz and zero-order hold method for process, quality and environmental data of 0.1Hz-100Hz. It performs timestamp synchronization and alignment on heterogeneous operating data with different acquisition frequencies. The time step after alignment is uniformly set to 0.0005s, which is equal to the sampling period corresponding to the highest sampling frequency among all acquisition modules. The data cleaning module uses a combination of Grubbs' criterion and the sliding window mean method to remove outliers and fill in missing values ​​in the synchronized data. The length of the sliding window is set to 5-100 time steps, and the significance level of Grubbs' criterion is set to 0.

05. The data normalization module uses the min-max normalization method to map the cleaned data to the 0-1 range, resulting in a standardized full-process dataset.

4. The full-process fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The two-level fusion feature engineering unit includes a single-section feature-level fusion module and a full-process cross-section associated feature fusion module; The single-section feature-level fusion module extracts high-dimensional hidden layer features from the multi-source data corresponding to each section using a stacked autoencoder with 3 to 8 layers for the standardized dataset of each section. The number of neurons in the encoder and decoder layers of the stacked autoencoder is symmetrically set. The hidden layer activation function is the ReLU function, and the output layer uses the Sigmoid function. Training is completed by unsupervised pre-training combined with supervised fine-tuning with few samples. The pre-training stage uses steady-state operation data of the production line, and the fine-tuning stage uses the class imbalance loss function combined with the hard sample mining strategy. Then, the attention mechanism is used to dynamically assign weights to features of different types of parameters. The weight assignment value ranges from 0 to 1 to obtain the single-section fusion features of each section. These are summarized to form a single-section fusion feature set, which is output to the hierarchical fault diagnosis root cause localization unit and the full-process cross-section correlation feature fusion module. The full-process cross-segment correlation feature fusion module constructs a directed acyclic graph (DAG) segment correlation topology based on the sequence of segments and material transfer logic of the production line. Each segment is treated as a graph node, the material and energy transfer relationships between segments are directed edges, and the fusion features of individual segments are node features. The prior weights of the edges are dynamically adjusted using the segment fault transmission coupling coefficient. A 2-4 layer graph attention neural network is used to extract full-process correlation features from the single-segment fusion features of all segments. The final edge weights are calculated using a weighted fusion of the prior coupling coefficient and the network-learned attention weights. The fusion formula is as follows: ,in Let be the edge weight, and λ be a preset prior weight coefficient, with a value range of [0.4, 0.8]. The coupling coefficient for fault propagation in the work section. The attention weights learned by the graph attention neural network are used to obtain a full-process correlation fusion feature set containing cross-section fault transmission correlation information, which is then output to the hierarchical fault diagnosis root cause localization unit and fault evolution early warning unit.

5. The full-process fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The hierarchical fault diagnosis root cause localization unit includes a single-section anomaly identification module, a cross-section fault transmission association matching module, and a root cause localization module. The single-section anomaly identification module, based on the feature components of the corresponding section in the single-section fusion feature set, uses a pre-trained single-class support vector machine anomaly detection model to identify whether there is an operational anomaly in the corresponding section. The single-class support vector machine uses a Gaussian kernel function, and the kernel function bandwidth is determined by combining grid search method with 5-fold cross-validation. The anomaly judgment threshold is calibrated by the 3σ criterion. The single-section anomaly identification result is output to the cross-section fault transmission association matching module. The cross-section fault transmission association matching module, based on the section association topology of the production line and combined with the single section anomaly identification results, uses the maximum mutual information coefficient to calculate the nonlinear correlation degree between the abnormal features of different sections, filters out the abnormal section combinations with a correlation degree greater than a preset correlation threshold, and outputs them to the root cause localization module. The preset correlation threshold ranges from 0.5 to 0.7 and is calibrated by combining the subject working characteristic curve with the production line process requirements. The root cause localization module, based on the fusion feature set of abnormal work section combinations and the whole process association, adopts a Bayesian network fault causal reasoning model constructed based on the work section topology to locate the source work section, fault type, and fault impact range of the fault, and generate fault diagnosis results. The Bayesian network uses the abnormal state of each work section as the root node, the fault type as the leaf node, and the work section coupling relationship as the directed edge. The conditional probability table is determined by the maximum likelihood estimation method based on expert experience and historical fault data. The reasoning algorithm adopts the connection tree algorithm. The fault types include melt temperature runaway, filter screen blockage, melt fracture, uneven cooling, traction roller slippage, winding tension runaway, die mouth accumulation, roller surface adhesion, and melt degradation.

6. The full-process fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The fault evolution early warning unit includes a feature trend extraction module, a fault evolution prediction module, and an early warning level classification module; The feature trend extraction module, based on the full-process correlation and fusion feature set of continuous time period, uses the sliding window least squares method to fit the evolution trend of each feature component to obtain the feature trend sequence. The length of the sliding window is set to 10-200 time steps. The fault evolution prediction module, for fault-free operating conditions, uses a 2-6 layer long short-term memory network time series prediction model based on feature trend sequences and historical fault sample database. The model input is a feature trend sequence of 20-100 consecutive time steps, and the output is the feature change trend and fault occurrence probability within a preset time period. The mean squared error loss function combined with the Adam optimizer is used to complete the training and predict the feature change trend and fault occurrence probability within a preset time period. For the fault conditions that have occurred, the spread range and the probability of deterioration of the fault are predicted by combining the fault diagnosis results and the characteristic trend sequence. The preset time range is 5min-30min. The warning level classification module classifies the corresponding warning levels based on the predicted failure probability. No warning is triggered when the failure probability is <30%, 30% to <60% is a low-risk warning, 60% to <90% is a medium-risk warning, and 90% and above is a high-risk warning. The warning probability threshold is calibrated by combining the ROC curve with the production line safety control requirements, and outputs the corresponding level of fault warning signal.

7. The end-to-end fault diagnosis system based on multi-source data fusion according to claim 1, characterized in that, The closed-loop execution unit includes an instruction generation module, an instruction issuance module, and an execution feedback module; The instruction generation module generates corresponding control instructions based on the fault diagnosis results and the warning level of the fault warning signal. For low-risk warnings, fine-tuning instructions for the corresponding process parameters of the work section are generated. For medium-risk warnings, generate corresponding process parameter adjustment instructions and equipment inspection prompts for the relevant work sections; For high-risk warnings, generate load reduction operation instructions or equipment shutdown protection instructions for the corresponding work section; for identified faults, generate corresponding process correction instructions. The instruction issuing module sends the generated control instructions to the local programmable logic controller of the production line via industrial Ethernet or fieldbus. Before the instructions are issued, they must be verified by process safety boundaries. Instructions that exceed the process safety limits will be intercepted and manual review will be triggered. The execution feedback module is used to collect production line operation data after the control command is executed and output it to the multi-source data collaborative preprocessing unit. The preprocessed feedback data is used for feature distribution adaptation of the two-level fusion feature engineering unit and online incremental learning and iterative optimization of the diagnostic model in the hierarchical fault diagnosis root cause localization unit. The update cycle of incremental learning is set to 12h-24h, and the rapid update cycle under emergency conditions is set to 1h-12h. Incremental learning adopts a sliding time window dataset update method to retain nearly 3 months of valid operation data.

8. A full-process fault diagnosis and early warning method based on multi-source data fusion, characterized in that, The system applied to the full-process fault diagnosis system based on multi-source data fusion as described in any one of claims 1-7 includes the following steps: S1: Collect multi-dimensional heterogeneous operational data of the entire process of the plastic film extrusion production line; S2: Perform time synchronization alignment, outlier removal, missing value completion, and normalization on the collected heterogeneous operational data to obtain a standardized full-process dataset; S3: Perform two-level feature extraction and fusion processing on the standardized full-process dataset: first, feature-level fusion of multi-source data in a single section, and then feature-level fusion of cross-section association in the entire process. Calculate the coupling coefficient based on the quantification equation of fault transmission coupling association in the section, and dynamically optimize the cross-section feature association weights to obtain the single-section fusion feature set and the full-process association fusion feature set. S4: Based on the single-section fusion feature set, complete the single-section anomaly identification; based on the full-process association fusion feature set, complete the cross-section fault transmission association matching and fault root cause location, and generate fault diagnosis results. S5: Based on the full-process correlation and fusion feature set, complete the early fault trend warning under fault-free working conditions, and combine the fault diagnosis results to complete the evolution trend warning of the faults that have occurred. If the probability of fault occurrence is <30%, the warning will not be triggered, and the corresponding level of fault warning signal will be output. S6: Based on the fault diagnosis results and fault warning signals, generate process parameter adjustment instructions or equipment shutdown protection instructions for the corresponding work section, and send them to the local control system of the production line for execution. At the same time, collect the operation feedback data after the instruction is executed to complete the online iterative optimization of feature distribution adaptation and diagnostic model.

9. The full-process fault diagnosis and early warning method based on multi-source data fusion according to claim 8, characterized in that, Step S3 specifically includes: S31: For each work section's standardized dataset, a stacked autoencoder with 3 to 8 layers is used to extract high-dimensional hidden layer features of the corresponding multi-source data for each work section. Then, an attention mechanism is used to dynamically assign weights to features of different types of parameters to obtain the single-work section fusion features of each work section. These features are then aggregated to form a single-work section fusion feature set. S32: Based on the sequential timing and material transfer logic of the production line sections, a directed acyclic graph section association topology is constructed. Each section is used as a graph node, the material and energy transfer relationship between sections is used as a directed edge, and the single section fusion feature is used as the node feature. The coupling coefficient is calculated by substituting into the section fault transmission coupling association quantification equation and the prior weight of the edge is updated. A 2-4 layer graph attention neural network is used to extract the full-process association features of the single section fusion features of all sections. Finally, the edge weight is calculated by weighted fusion of the prior coupling coefficient and the network learning attention weight to obtain the full-process association fusion feature set containing cross-section fault transmission association information.

10. The full-process fault diagnosis and early warning method based on multi-source data fusion according to claim 8, characterized in that, Step S4 specifically includes: S41: Based on the feature components of the corresponding work section in the single-section fusion feature set, a pre-trained single-class support vector machine anomaly detection model is used to identify whether there is an operational anomaly in the corresponding work section and generate the single-section anomaly identification result. S42: Based on the section association topology of the production line, combined with the single section anomaly identification results, the nonlinear correlation degree between different section anomaly features is calculated using the maximum mutual information coefficient, and abnormal section combinations with a correlation degree greater than the preset correlation threshold are selected. S43: Based on the feature set of abnormal section combination and whole process association, a Bayesian network fault causal reasoning model based on the section topology is adopted to locate the source section, fault type and fault impact range of the fault. Fault types include melt temperature runaway, filter screen blockage, melt rupture, uneven cooling, traction roller slippage, winding tension runaway, die mouth material accumulation, roller surface material sticking, melt degradation, and generate fault diagnosis results.

Images