A VR-based monitoring method for flame-retardant nonwoven fabric production management

Abstract

This invention relates to the fields of industrial monitoring and VR technology, specifically a method for managing the production of flame-retardant nonwoven fabrics based on VR technology. The method includes: collecting temperature, rotation speed, and tension data streams from the production line and performing multi-scale decomposition; identifying data offset events for each component using adaptive thresholds; and then fusing these data to generate a composite abnormal event descriptor. Based on this descriptor, fault cause chains and equipment relationship networks are retrieved from a production process knowledge graph; a virtual sensing data stream is constructed and compared with processed real data for verification; and a structured report containing fault location and maintenance suggestions is generated accordingly. The report and event descriptor are mapped to a 3D equipment model in the VR monitoring scene. This method achieves accurate tracing and early warning of complex coupled faults in the production process, and significantly improves the intuitiveness of monitoring and decision-making efficiency through immersive 3D visualization.

Description

Technical Field

[0001] This invention relates to the fields of industrial monitoring and VR technology, and in particular to a method for production management of flame-retardant nonwoven fabrics based on VR technology monitoring. Background Technology

[0002] In the production of flame-retardant nonwoven fabrics, the coordinated stability of multiple physical quantities such as temperature, rotation speed, and tension is crucial to product quality. Existing production monitoring technologies primarily rely on sensor networks deployed at key nodes of the production line to collect data in real time and issue alarms for anomalies using preset fixed thresholds or simple statistical process control methods. While these methods can respond to obvious parameter exceedances, their sensitivity and diagnostic depth are insufficient for early, latent faults caused by multivariate coupling, gradual drift, or complex equipment interactions. Alarm information is often isolated and lacks causal correlation, making it difficult to accurately trace the root cause of the fault, let alone predict its propagation path.

[0003] Current production management systems rely heavily on operator experience to interpret alarm information during anomaly handling. Fault location and equipment status assessment are typically performed through two-dimensional configuration interfaces or data reports. This approach has limitations in presenting complex equipment spatial relationships, the scope of fault impact, and maintenance operation guidelines; the information is not intuitive enough, and decision support efficiency needs improvement. Technicians need to frequently switch between multiple system interfaces and interpret large amounts of data to form a holistic understanding of the production status, resulting in slow response times and a high risk of misjudgment. Therefore, achieving deep correlation and accurate diagnosis of multi-source anomaly information, and presenting the diagnostic results to operators in a highly intuitive and immersive manner, is crucial to improving the level of intelligent management in flame-retardant nonwoven fabric production. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a VR-based monitoring method for the production management of flame-retardant nonwoven fabrics.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for production management of flame-retardant nonwoven fabric based on VR technology monitoring, comprising:

[0006] Collect multi-dimensional real-time data streams containing temperature, rotation speed, and tension from the flame-retardant nonwoven fabric production line;

[0007] The multidimensional real-time data stream is decomposed into trend components, periodic components and residual components.

[0008] For the data of the trend component, the periodic component and the residual component, independent state evaluation windows are established respectively, and an adaptive threshold algorithm is used to identify data offset events in each state evaluation window;

[0009] The data offset events are aligned and merged along the time dimension to form a composite anomaly event descriptor;

[0010] Based on the composite abnormal event descriptor, the associated fault cause chain and equipment interaction relationship network are retrieved from the pre-constructed production process knowledge graph;

[0011] By utilizing the fault cause chain and the device interaction network, a virtual sensing data stream is constructed for the current production scenario;

[0012] The virtual sensing data stream is compared and verified with the processed original multidimensional real-time data stream;

[0013] Based on the comparative verification results, a structured report is generated that includes fault location, cause inference, and maintenance recommendations;

[0014] The structured report and the composite anomaly event descriptor are mapped together onto the 3D device model in the VR monitoring scene.

[0015] Preferably, the step of performing multi-scale decomposition on the multi-dimensional real-time data stream to separate the trend component, periodic component, and residual component specifically includes:

[0016] The multidimensional real-time data stream is processed using the empirical mode decomposition algorithm to obtain a series of intrinsic mode function components.

[0017] By calculating the average period and energy distribution of the intrinsic mode function components, data representing the trend component that characterizes the long-term trend, the period component that characterizes the inherent period of the equipment, and the residual component that characterizes noise and transient anomalies are selected.

[0018] For the trend component, a local weighted regression algorithm is applied for smoothing to extract its change envelope. For the periodic component, a spectrum analysis and principal period identification algorithm is used to extract its dominant frequency and amplitude. For the residual component data, an independent component analysis algorithm is applied to separate the singular components that are unrelated to known interference sources.

[0019] Preferably, for the data of the trend component, the periodic component, and the residual component, independent state evaluation windows are established respectively. Within each state evaluation window, an adaptive threshold algorithm is used to identify data offset events, specifically including:

[0020] A long-term state evaluation window is established for the changing envelope of the trend component. Based on the statistical distribution of historical data within the long-term state evaluation window, the healthy operating range boundary of the trend component is dynamically calculated. When the changing envelope continuously exceeds the healthy operating range boundary, it is marked as a trend deviation event.

[0021] A mid-term state assessment window is established for the dominant frequency and amplitude of the periodic component. Within the mid-term state assessment window, periodic distortion events are identified by comparing the difference between the current period and the historical average period, as well as the deviation between the current amplitude and the historical amplitude spectrum.

[0022] A short-term state evaluation window is established for the singular components of the residual components. Within the short-term state evaluation window, a dynamic alarm threshold is set using extreme value theory to capture transient spike events.

[0023] Preferably, aligning and fusing the data offset events along the time dimension to form a composite anomaly event descriptor specifically includes:

[0024] The trend offset event, the periodic distortion event, and the transient spike event are each assigned a unified timestamp; an aligned time window is established centered on the event occurrence time.

[0025] Within the alignment time window, the temporal proximity and trend correlation between different trend offset events, periodic distortion events, and transient spike events are calculated;

[0026] Based on the correlation between the time proximity and the change trend, events that meet the fusion conditions are aggregated into a composite event unit.

[0027] A feature vector is extracted for each composite event unit. The feature vector contains a combination of event types, a time distribution pattern, and a weighted value of the severity of each component, thereby forming the composite abnormal event descriptor.

[0028] Preferably, the step of retrieving associated fault cause chains and equipment interaction networks from a pre-constructed production process knowledge graph based on the composite abnormal event descriptor specifically includes:

[0029] The feature vector of the composite abnormal event descriptor is converted into a knowledge graph query statement. The knowledge graph query statement is then used to traverse the entity nodes and relation edges in the production process knowledge graph. The entity nodes include equipment components, process parameters, and material batches.

[0030] The relation edges include causal relationships, spatial connectivity relationships, and process logic dependencies; all fault mode nodes that match the event types in the feature vector are retrieved;

[0031] Tracing back multiple upstream causal nodes connected to the fault mode node forms a possible fault cause chain;

[0032] Simultaneously, other equipment or process segment nodes associated with the fault mode node are located, and a network of equipment interaction relationships is constructed.

[0033] Preferably, the step of constructing a virtual sensing data stream for the current production scenario using the fault cause chain and the device interaction network specifically includes:

[0034] Select the most likely causative node from the fault cause chain as a virtual sensing point;

[0035] Based on the device interaction network, the physical location and measurement parameter type of the virtual sensing point are inferred;

[0036] By combining the temporal distribution pattern of the composite anomaly event descriptor and using a physical model or a data-driven model, the data sequence that should be observed at the most likely causal node can be simulated in reverse.

[0037] The data sequence constitutes the virtual sensing data stream, which should theoretically be able to explain the abnormal phenomena represented by the composite abnormal event descriptor.

[0038] Preferably, the step of comparing and verifying the virtual sensing data stream with the processed original multidimensional real-time data stream specifically includes:

[0039] In the device interaction network, find the actual sensor measurement point that is closest to the physical location of the virtual sensing point; extract the processed raw multidimensional real-time data stream of the actual sensor measurement point in the corresponding time period; calculate the similarity index between the virtual sensing data stream and the processed raw multidimensional real-time data stream of the actual sensor measurement point;

[0040] The similarity metrics include dynamic time-normalized distance and cross-correlation coefficient;

[0041] If the similarity index is higher than the preset verification threshold, it is determined that the virtual sensing data stream is supported by the data of the actual sensor measurement points, and the probability of the fault cause chain increases; otherwise, the probability of the fault cause chain decreases.

[0042] Preferably, based on the comparison and verification results, a structured report containing fault location, cause inference, and maintenance recommendations is generated, specifically including:

[0043] Integrate the comparative verification results with the credibility score of the fault cause chain; based on the equipment interaction network, accurately label the physical equipment or process links where the fault may occur, and complete the fault location;

[0044] By combining the production process manual and the historical maintenance case library, natural language descriptions of cause inferences are generated for the most likely cause nodes in the fault cause chain.

[0045] Based on the aforementioned cause inference, the maintenance procedure knowledge base is retrieved to generate maintenance recommendations that include specific operating steps, required spare parts, and safety precautions.

[0046] The fault location, the cause inference, and the maintenance recommendations are integrated into the structured report.

[0047] Preferably, mapping the structured report and the composite anomaly event descriptor together onto the 3D device model in the VR monitoring scene specifically includes:

[0048] Load 3D equipment models corresponding one-to-one with the physical production line into the VR monitoring platform; based on the fault location information, highlight the corresponding virtual equipment or components in the 3D scene;

[0049] The event type combination and time distribution pattern information in the composite abnormal event descriptor are transformed into dynamic, visually animated icons attached to the highlighted component for display.

[0050] The key content of the structured report, including the summary of the cause inference and the key points of the maintenance recommendations, will be displayed in an interactive floating information panel next to the highlighted component.

[0051] Preferably, after loading the three-dimensional equipment models corresponding one-to-one with the physical production line into the VR monitoring platform, the process further includes:

[0052] Receive the latest multi-dimensional real-time data stream from the industrial control system;

[0053] The motion state of each moving part in the 3D device model is synchronized with the physical device in real time.

[0054] When a new composite abnormal event descriptor is detected, the VR view is automatically switched to the location of the virtual device corresponding to the fault location, and an alarm sound is triggered. At the same time, a floating information panel containing the latest structured report content is automatically popped up.

[0055] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0056] By pre-constructing a production process knowledge graph, the identified composite anomaly event descriptors are matched and retrieved with the fault cause chains and equipment interaction relationships stored in the graph, dynamically constructing a virtual sensing data stream that should be generated under the current situation. This virtual data stream serves as a reference benchmark and is compared and verified online with the original real data stream after multi-scale decomposition processing. This method changes the traditional judgment mode based on a single data threshold, improving the accuracy of anomaly diagnosis through bidirectional verification of "real-virtual" data. It can effectively distinguish between random noise and real fault symptoms, reveal the inherent causal logic between multi-parameter anomalies, thereby achieving accurate source tracing and early warning of complex coupled faults, transforming the monitoring process from passive alarm to proactive, knowledge-based deep analysis.

[0057] Validated structured diagnostic reports and composite anomaly event descriptors are directly mapped onto a 3D production line equipment model built using virtual reality technology. Fault locations are highlighted on the 3D model, and related equipment interactions are displayed as a visual network. Specific cause analysis text and maintenance guidance information are spatially anchored and overlaid next to the corresponding equipment. This solution breaks down the information isolation of traditional 2D monitoring screens, providing operators with a unified and immersive monitoring environment. It shortens the time required to understand the full picture of the fault from receiving an alarm, reduces the cognitive load of information interpretation, and makes remote inspections, collaborative consultations, and maintenance training more intuitive and efficient, improving the intuitiveness of production management and emergency response capabilities. Attached Figure Description

[0058] Figure 1 This is a flowchart of the flame-retardant nonwoven fabric production management method based on VR technology monitoring according to the present invention;

[0059] Figure 2 A flowchart for multi-scale decomposition of multi-dimensional real-time data streams;

[0060] Figure 3 A flowchart for generating a composite exception event descriptor;

[0061] Figure 4 A bar chart comparing the stability scores of flame-retardant nonwoven fabric production equipment before and after maintenance.

[0062] Figure 5 A bar chart comparing the response times of different operation types in a VR monitoring system for flame-retardant nonwoven fabric production. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the 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, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0065] Please see Figure 1 A VR-based monitoring method for flame-retardant nonwoven fabric production management is proposed, with the following overall implementation scheme: Multi-dimensional real-time data streams, including temperature, rotational speed, and tension, are collected from a sensor network deployed on the flame-retardant nonwoven fabric production line. The collected multi-dimensional real-time data streams are decomposed into trend components representing the direction and rate of long-term changes, periodic components representing the inherent repetitive patterns of the equipment, and residual components containing noise and transient fluctuations. For the decomposed trend, periodic, and residual components, independent state evaluation windows are established, and an adaptive threshold algorithm is used to identify data deviation events within each state evaluation window. The identified data deviation events are aligned and fused along the time dimension to form a composite anomaly event descriptor that comprehensively reflects multi-scale anomalies. Based on this composite anomaly event descriptor, associated fault cause chains and equipment interaction networks are retrieved from a pre-constructed production process knowledge graph containing relationships between equipment, processes, and materials. Using the retrieved fault cause chains and equipment interaction networks, a virtual sensor data stream is constructed for the current production scenario. The virtual sensor data stream is compared and verified with the processed original multidimensional real-time data stream. Based on the comparison and verification results, a structured report containing fault location, cause inference, and maintenance suggestions is generated. The generated structured report, along with the composite anomaly event descriptor, is mapped onto the 3D device model in the VR monitoring scene to achieve visualized monitoring and interactive diagnosis of anomalies.

[0066] In one embodiment of the present invention, see [reference] Figure 2In practice, a multi-dimensional real-time data stream containing temperature, rotation speed, and tension, collected from the flame-retardant nonwoven fabric production line, is used as input. This multi-dimensional real-time data stream is processed using an empirical mode decomposition (EMD) algorithm. The EMD algorithm decomposes the original signal into a series of intrinsic mode function (IMF) components arranged from high to low frequency through an iterative screening process. After the IMF component decomposition is completed, the average period and energy distribution parameters of each IMF component are calculated, and component screening is performed. IMF components with the longest average period and energy distribution concentrated in the lowest 5% frequency band are selected as trend components representing long-term trends. IMF components with average periods matching known inherent physical periods of the equipment, such as the spinning machine spindle rotation period or the drying oven fan vibration period, and with stable energy distributions, are selected as periodic components representing the inherent period of the equipment. The remaining IMF components with extremely short average periods and random energy distributions are classified as residual components representing noise and transient anomalies.

[0067] In some embodiments, for the selected trend components, a locally weighted regression algorithm is applied for smoothing to extract their change envelope. The locally weighted regression algorithm assigns a weight to each data point based on its distance from neighboring points, and a smooth local polynomial is obtained through weighted least squares fitting, thereby forming an envelope depicting the overall trend of the trend component. In specific implementations, for the selected periodic components, spectrum analysis and principal period identification algorithms are used for processing. Spectrum analysis transforms the periodic components to the frequency domain through fast Fourier transform, while the principal period identification algorithm identifies the frequency component with the highest energy in the spectrum as the dominant frequency and extracts its corresponding amplitude information. In some embodiments, for the data of the classified residual components, an independent component analysis algorithm is applied to separate singular components that are unrelated to known interference sources. The independent component analysis algorithm assumes that the residual component data is a linear mixture of several statistically independent source signals, and estimates these source signals through the unmixing matrix. Among them, components that are unrelated to known interference modes such as electromagnetic noise in the workshop environment and power fluctuations are identified as singular components. It is understandable that the above screening process can be implemented using a quantitative criterion. For example, a screening function combining the average period and energy concentration can be set to automatically classify the intrinsic modulus components into trend components, periodic components, or residual components. This screening function can be expressed as:

[0068]

[0069] in: Indicates the first The classification discriminant value of each intrinsic modulus function component. This indicates the average period of the component. This indicates a preset reference cycle (such as the standard production cycle of a production line). This indicates the proportion of that component's energy in the total energy. and These are weighting coefficients set based on prior knowledge. By setting different... Automated classification of components can be achieved using threshold intervals, for example, when... When the value is below the first threshold, the component is classified as residual component data; when When the value is between the first threshold and the second threshold, the component is classified as a periodic component; when When the value exceeds the second threshold, the component is classified as a trend component. Optionally, in specific implementations, for the boundary effects generated by the empirical mode decomposition algorithm, preprocessing methods such as mirror continuation or signal splicing can be used to reduce decomposition errors at the endpoints. It is understood that the selection of the kernel function and the setting of the bandwidth parameter in the local weighted regression algorithm need to be adjusted according to the density and drastic change of the trend component data points. Too small a bandwidth may lead to insufficient smoothing, while too large a bandwidth may mask the true trend changes. Optionally, before separating singular components, the independent component analysis algorithm can use the signal patterns of known interference sources as reference inputs to enhance the algorithm's ability to identify and separate irrelevant components.

[0070] In one embodiment of the present invention, see [reference] Figure 3 In specific implementation, a long-term state assessment window is established for the change envelope extracted from the trend component. The time length of the long-term state assessment window is set to be sufficient to cover multiple complete batches of flame-retardant nonwoven fabric production or several hours of continuous operation. Based on the statistical distribution of historical change envelope data within the long-term state assessment window, the healthy operating interval boundary of the trend component is dynamically calculated. The healthy operating interval boundary is usually defined as a range of plus or minus three standard deviations of the historical data mean or an interval based on quantiles. When the change envelope exceeds the healthy operating interval boundary for more than five consecutive sampling periods, a trend deviation event is marked and recorded. In specific implementation, a medium-term state assessment window is established for the dominant frequency and amplitude extracted from the periodic component. The time length of the medium-term state assessment window is set to cover dozens to hundreds of inherent operating cycles of the equipment. Within the medium-term state assessment window, by calculating the relative difference between the current cycle and the historical average cycle, and the Mahalanobis distance deviation between the current amplitude and the historical amplitude spectrum distribution, when the relative difference or Mahalanobis distance deviation exceeds their respective set static thresholds, a periodic distortion event is identified and recorded.

[0071] In some embodiments, a short-term state assessment window is established for the singular components separated from the residual component data. The duration of the short-term state assessment window is set to be much shorter than the equipment operating cycle, possibly only a few seconds or tens of sampling points. Within the short-term state assessment window, an alarm threshold is dynamically set based on the tail characteristics of the recent singular component data using an extreme value theory model based on the generalized Pareto distribution. When the instantaneous amplitude of the singular component exceeds the dynamic alarm threshold, a transient spike event is captured and recorded. In specific implementations, all marked trend offset events, periodic distortion events, and transient spike events are tagged with a precise timestamp in a uniform format. The timestamp contains date and millisecond-level time information synchronized from the data acquisition source. An aligned time window is established centered on the occurrence time of a single event. The width of the aligned time window is predefined according to the dynamic characteristics of the production process, for example, a few seconds for a high-speed spinning process and tens of seconds for a slow heat setting process. In some embodiments, within an aligned time window, the temporal proximity and trend correlation between different events are calculated. Temporal proximity is quantified by the reciprocal of the absolute difference between the timestamps of the events, and trend correlation is evaluated by calculating the Pearson correlation coefficient of the original data segments associated with the events within a short time window. It is understood that the logic of event aggregation requires a quantitative fusion metric as a basis for judgment. It can be calculated using the following formula:

[0072]

[0073] in: Indicates the first The event and the first A fusion metric between events and These represent the times when the two events occurred. Represents the absolute value of the time difference. This represents a preset system time resolution or minimum effective time unit. The correlation coefficient represents the changing trends of data segments related to two events. and Weighting coefficients to adjust the importance of temporal proximity and trend correlation. Based on the calculated fusion metric. This will satisfy the fusion metric. Events exceeding a preset threshold are aggregated into a composite event unit. Optionally, the width of the alignment time window can be adaptive, dynamically adjusted based on the average event occurrence rate of the current production line; the higher the occurrence rate, the narrower the alignment time window. It can be understood that for events occurring in the same equipment unit but of different types, when calculating the correlation of changing trends, a process parameter data segment directly related to that equipment unit can be selected for calculation. For example, for events in the drying roller area, the temperature and tension data of that area can be used for correlation analysis.

[0074] In practical implementation, a multi-dimensional feature vector is extracted for each aggregated composite event unit. The feature vector's dimensions include the event type combination code, the distribution pattern descriptor of the event timestamps within the alignment time window, and the severity weighting values ​​of each trend-shifting event, periodic distortion event, and transient spike event constituting the composite event unit. The event type combination code uses binary bits to represent the existence of different event types. The distribution pattern descriptor includes the time span of the first and last events and the density of the event cluster. The severity weighting values ​​are calculated based on the extent to which each event exceeds its own threshold, thus forming a structured composite abnormal event descriptor. Optionally, the feature vector may also include the sensor or device location identifier of the event source to facilitate subsequent fault localization. It can be understood that the aggregation process of the composite event unit is iterative. When a new event occurs, it attempts to merge with existing events within the still-open alignment time window. If it cannot merge with any existing event, it may open a new alignment time window starting from itself.

[0075] In one embodiment of the present invention, the feature vector of the composite abnormal event descriptor is converted into a knowledge graph query statement. The conversion process involves mapping the event type combination encoding in the feature vector to a predefined fault mode category identifier in the production process knowledge graph, mapping the time distribution pattern descriptor to time constraints, and mapping the severity weighting value to a relevance weighting coefficient. In the specific implementation, the converted knowledge graph query statement is used to traverse the entity nodes and relation edges in the production process knowledge graph. The entity nodes in the production process knowledge graph include equipment components such as "drying roller bearing" and "tension sensor", process parameters such as "hot air temperature" and "linear velocity", and material batches such as "flame retardant coating batch number". The relation edges include causal relationships such as "bearing wear causes speed fluctuation", spatial connection relationships such as "motor A drives drying roller B", and process logic dependency relationships such as "temperature setpoint affects tension threshold". In practice, all fault mode nodes that match the event types in the feature vector are retrieved. For example, when the feature vector indicates a combination of "periodic distortion" and "transient spikes", it may match fault mode nodes such as "periodic tension fluctuations caused by roller eccentricity" or "sensor reading spikes caused by electrical interference" in the knowledge graph.

[0076] In some embodiments, multiple upstream causal nodes connected to the retrieved fault mode node are traced to form a possible fault cause chain. The tracing process performs a depth-first or breadth-first search along the causal relationship edges. For example, starting from the fault mode node "abnormal temperature rise in the drying zone," it is possible to trace back to upstream causal nodes such as "heating tube aging," "thermocouple calibration drift," or "insufficient heat transfer oil flow," thereby forming multiple candidate fault cause chains. In specific implementations, other equipment or process segment nodes associated with the fault mode node are simultaneously located to construct an equipment interaction relationship network. The location process is achieved by traversing spatial connection relationship edges and process logic dependency relationship edges. For example, the fault mode node "winding tension loss of control" may be associated with upstream nodes such as "tension detection roller," "servo motor controller," and downstream nodes such as "winding machine." These nodes and their connections together constitute a local equipment interaction relationship network. It can be understood that the construction and traversal of knowledge graph query statements require a quantitative relevance scoring mechanism to rank the search results. It can be calculated using the following formula:

[0077]

[0078] in: This represents the relevance score between the query result nodes and the composite exception event descriptor. The feature vector representing the composite exception event descriptor, The feature vector representation of a node in a knowledge graph. Function to compute vector with vector Cosine similarity between them This indicates the relation edge connected to this node. The function represents the edge of the relation. Confidence weights and This is a weighted coefficient used to balance feature similarity and relationship confidence. Based on relevance scoring. The retrieved fault mode nodes are sorted, and the nodes with the highest scores are prioritized for subsequent causal chain tracing. Optionally, when constructing the device interaction relationship network, a hop limit can be set, such as tracing only relevant nodes within a two- or three-hop range, to ensure the focus and manageability of the relationship network.

[0079] In practical implementation, the most probable causal chain is selected from multiple possible fault causal chains. The selection criteria can be based on the frequency of occurrence of the causal node in historical cases, the node's reliability score, or its consistency with other parameters in the current production context. The root node or key intermediate node in this most probable causal chain is identified as the virtual sensing point. In some embodiments, the physical location and measurement parameter type of the virtual sensing point are inferred based on the constructed equipment interaction network. For example, if the virtual sensing point is determined to be "the flame retardant concentration in the upstream impregnation tank," then based on the spatial connection relationship in the equipment interaction network, its physical location is located near the "liquid inlet pipe" of "impregnation tank A," and the measurement parameter type is "concentration." It can be understood that the construction of the virtual sensing data stream requires combining the temporal distribution pattern of the composite anomaly event descriptor. The temporal distribution pattern of the composite anomaly event descriptor describes the temporal characteristics of the occurrence, duration, and dissipation of the anomaly event. Using a physical model or a data-driven model, with the final anomaly phenomenon represented by the composite anomaly event descriptor as the output boundary condition, the data sequence that should be observed at the virtual sensing point and could lead to the anomaly phenomenon is simulated in reverse.

[0080] Optionally, the physical model can be a differential equation or transfer function describing the process, and the data-driven model can be a trained neural network or regression model. The reverse simulation process may involve iterative adjustment of model parameters or optimization of the input sequence. The data sequence generated by the reverse simulation constitutes the virtual sensing data stream. Theoretically, this virtual sensing data stream should be able to explain the abnormal phenomena characterized by the composite anomaly event descriptor. For example, the simulated sequence of "slow decrease in heat transfer oil flow" can explain the anomaly of "trend shift in drying zone temperature".

[0081] In one embodiment of the present invention, in a specific implementation, the actual sensor measurement point closest to the physical location of the virtual sensing point is found in the constructed device interaction relationship network. This process is achieved by comparing the spatial coordinates or topological connection distances of nodes in the device interaction relationship network. For example, the virtual sensing point is inferred to be "concentration of the inlet pipe of immersion tank A", while the nearest actual sensor measurement point identified in the device interaction relationship network is "concentration sensor C-102 at the outlet of immersion tank A". In a specific implementation, the processed original multidimensional real-time data stream of the actual sensor measurement point within the corresponding time period is extracted. The corresponding time period is the period of anomaly occurrence covered by the composite anomaly event descriptor. The extracted data is data that has been processed by the empirical mode decomposition algorithm and separated into corresponding components. The similarity index between the virtual sensing data stream and the processed original multidimensional real-time data stream of the actual sensor measurement point is calculated. The similarity index includes dynamic time warping distance and cross-correlation coefficient. Dynamic time warping distance is used to measure the similarity in shape between two time series that may have different lengths, and cross-correlation coefficient is used to measure the degree of linear correlation between the two time series in terms of their changing trends.

[0082] Understandably, a comprehensive metric is needed to determine whether virtual sensing data streams are supported by actual data, such as an overall similarity metric. It can be calculated using the following formula:

[0083]

[0084] in: This represents the overall similarity index. This represents the dynamic time-warped distance between the virtual sensing data stream and the actual sensor data stream. Indicates the length of the time series reference used for normalization. This represents the cross-correlation coefficient between the two. and Weighting coefficients to adjust the importance of dynamic time-warped distance and cross-correlation coefficient. Overall similarity index. The value range is typically between 0 and 1, with values ​​closer to 1 indicating higher similarity. If the overall similarity index... If the data stream exceeds the preset verification threshold, it is determined that the virtual sensing data stream is supported by data from the actual sensor measurement points, increasing the likelihood of the corresponding fault cause chain; otherwise, the likelihood of the corresponding fault cause chain decreases. The preset verification threshold can be calibrated based on the comparison results of historical normal data, see Table 1.

[0085] Table 1: Similarity Index Calculation Table

[0086]

[0087] In some embodiments, the verification results are integrated with the credibility score of the fault cause chain. The credibility score may be derived from the relevance score during the knowledge graph retrieval stage. Including expert prior weights, the integration method can be weighted averaging or rule-based decision fusion. Based on the equipment interaction network, the physical equipment or process links where the fault may occur are accurately labeled to complete the fault location. For example, when the virtual sensing point "feedforward tension controller output" is verified and its corresponding cause chain has high credibility, the fault is located in the "feedforward tension control module" or its connected "servo drive". Combining the production process manual and the historical maintenance case library, the most likely cause node in the fault cause chain is generated with natural language descriptions of cause inferences. The production process manual provides theoretical knowledge of equipment and processes, and the historical maintenance case library provides the mapping relationship between past similar anomalies and the final root cause. The generation process can use template filling or sequence-to-sequence natural language generation models. Optionally, when generating natural language descriptions, readings from actual sensor measurement points can be introduced as supporting information.

[0088] In practical implementation, based on the generated cause inference, the maintenance procedure knowledge base is retrieved to generate maintenance recommendations that specify operational steps, required spare parts, and safety precautions. The maintenance procedure knowledge base stores standardized maintenance operation guidelines for different equipment and different failure modes in a structured format. It is understood that the final generated structured report needs to integrate fault location, cause inference, and maintenance recommendations. The structured report uses a predefined XML or JSON format and contains fixed fields to record the faulty equipment identifier, location time, possible cause description, recommended maintenance action list, required tool and spare part codes, and operational risk warnings. Optionally, the structured report may also include details of the similarity index calculation and verification conclusions as additional explanations of diagnostic reliability. In some embodiments, if the virtual sensing data stream fails verification, i.e., the overall similarity index... If the result is below the verification threshold, the cause inference section of the structured report will include statements that the cause chain of the failure may be inaccurate, and may recommend checking other potential cause paths or initiating a more extensive diagnostic procedure.

[0089] See Figure 4This is a bar chart comparing the stability scores of flame-retardant nonwoven fabric production equipment before and after maintenance. It clearly shows the changes in stability scores for four core equipment parameters: impregnation tank concentration stability, tension control accuracy, drying zone temperature stability, and setting machine speed stability, across four stages: before maintenance, one week after maintenance, two weeks after maintenance, and one month after maintenance. The stability scores for all four parameters steadily improved with the maintenance cycle, validating the effectiveness of maintenance recommendations generated based on VR monitoring and demonstrating high accuracy in fault location and cause inference. This chart visually proves that the maintenance scheme based on VR monitoring and virtual sensing verification can significantly improve equipment stability, providing data support for optimizing the system algorithm. The improvement in equipment stability directly corresponds to the improvement in the production quality of flame-retardant nonwoven fabric; this chart can serve as a core basis for production quality traceability and continuous improvement.

[0090] In one embodiment of the present invention, in a specific implementation, a three-dimensional equipment model corresponding one-to-one with the physical production line is loaded into the VR monitoring platform. The three-dimensional equipment model is constructed based on the actual equipment CAD drawings and layout survey data of the flame-retardant nonwoven fabric production line. The model includes the three-dimensional geometric structure, material texture, and kinematic relationship of key equipment such as drying ovens, impregnation tanks, tension rollers, and winding machines. According to the fault location information in the structured report, the corresponding virtual equipment or component is highlighted in the three-dimensional scene. The highlighting method can be to change the component rendering color to a warning color (such as red), add a pulsating light effect, or overlay a semi-transparent outline frame around the component. For example, when the fault location information is "drying roller bearing B-203", the corresponding drying roller bearing model in the three-dimensional scene will be surrounded by a red pulsating light halo.

[0091] The event type combinations and time distribution patterns in the composite anomaly event descriptors are transformed into dynamic, visually animated icons attached to the highlighted components. The design of these animated icons is directly related to the event type; for example, a "periodic distortion" event is represented by a rotating stripe animation around the component, and a "transient spike" event is represented by a burst of starlight flashes on the component's surface. The playback speed or intensity of the animation can be correlated with the event density or severity in the time distribution pattern. Key content from the structured report, including causal inference summaries and maintenance recommendations, is displayed as an interactive floating information panel near the highlighted component. The floating information panel is anchored to the highlighted component by a leader line, and the panel presents information in a structured list of text and symbols. Users can use a VR controller to click buttons on the panel to expand to more detailed report content or link to related documents.

[0092] Understandably, in order to effectively integrate and differentiate multiple abnormal events from different times or of different severity in a VR scene, a visual coding rule is needed to manage the overlay and display priority of icons, and a visual priority coefficient. It can be calculated using the following formula:

[0093]

[0094] in: This indicates the display priority of the visual animated icons in the VR scene. This represents the sum of the severity weights of all event types in the composite exception event descriptor. This represents the time decay factor from the time the event occurred to the current moment (e.g., the more recent the event, the lower the decay factor). (The larger the value) and A coefficient used to adjust the weighting of severity and time urgency. Based on visual priority coefficients. The size, transparency, and whether the icon is placed on top when multiple events overlap are determined. In some embodiments, the motion state of each moving part in the 3D equipment model is synchronized with the physical equipment in real time. The VR monitoring platform receives the latest multi-dimensional real-time data stream from the industrial control system in real time through industrial communication protocols such as OPCUA or MQTT, and parses the status parameters such as equipment speed and valve opening from the data stream, thereby driving the corresponding parts in the 3D model (such as rotating rollers and swinging guide rods) to make matching movements.

[0095] In practical implementation, when the VR monitoring system detects the generation of a new composite anomaly event descriptor, it automatically switches the operator's VR perspective to the virtual device location corresponding to the fault location. The perspective switch uses a smooth flight animation transition to ensure that the operator does not experience dizziness, and simultaneously triggers a spatial alarm sound. The sound source of the alarm sound is set at the virtual fault device, and the tone and volume can be modulated according to the severity of the composite anomaly event descriptor. At the same time, a floating information panel containing the latest structured report content automatically pops up. Optionally, a global alarm list floating panel can be provided in the VR scene, listing all active composite anomaly event descriptors and their corresponding fault locations in chronological order. The operator can quickly navigate to the corresponding device location by clicking on the items in the list. It can be understood that for monitoring scenarios with multiple parallel production lines, the VR monitoring platform can load multiple independent 3D production line models and lay them out in a unified virtual monitoring hall. The operator can walk freely or teleport in the hall to view the status of different production lines. Each production line's 3D model independently executes the above-mentioned data-driven, event mapping, and alarm processes.

[0096] See Figure 5This is a bar chart comparing the response times of different operation types in a VR monitoring system for flame-retardant nonwoven fabric production. The red dashed line marks the VR system's latency threshold (300ms), reflecting the real-time performance of each core operation. All operation response times are below the 300ms system latency threshold, indicating that the VR monitoring system's real-time performance meets the basic requirements of industrial production scenarios and will not affect fault handling efficiency due to excessive latency. This chart visually verifies the real-time performance indicators of the VR monitoring system, providing a quantitative basis for system stability and reliability. By identifying the operation with the longest response time, the 3D rendering algorithm, data transmission protocol, or server computing power allocation can be optimized to further improve system real-time performance. Stable low-latency response is core to the user experience of a VR monitoring system, and this chart can serve as a key reference for subsequent system iterations and user satisfaction evaluation.

[0097] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring, characterized in that, The method includes: Collect multi-dimensional real-time data streams containing temperature, rotation speed, and tension from the flame-retardant nonwoven fabric production line; The multidimensional real-time data stream is decomposed into trend components, periodic components and residual components. For the data of the trend component, the periodic component and the residual component, independent state evaluation windows are established respectively, and an adaptive threshold algorithm is used to identify data offset events in each state evaluation window; The data offset events are aligned and merged along the time dimension to form a composite anomaly event descriptor; Based on the composite abnormal event descriptor, the associated fault cause chain and equipment interaction relationship network are retrieved from the pre-constructed production process knowledge graph; By utilizing the fault cause chain and the device interaction network, a virtual sensing data stream is constructed for the current production scenario; The virtual sensing data stream is compared and verified with the processed original multidimensional real-time data stream; Based on the comparative verification results, a structured report is generated that includes fault location, cause inference, and maintenance recommendations; The structured report and the composite anomaly event descriptor are mapped together onto the 3D device model in the VR monitoring scene.

2. The method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 1, characterized in that, The step of performing multi-scale decomposition on the multidimensional real-time data stream to separate the trend component, periodic component, and residual component specifically includes: The multidimensional real-time data stream is processed using the empirical mode decomposition algorithm to obtain a series of intrinsic mode function components. By calculating the average period and energy distribution of the intrinsic mode function components, data representing the trend component that characterizes the long-term trend, the period component that characterizes the inherent period of the equipment, and the residual component that characterizes noise and transient anomalies are selected. For the trend component, a local weighted regression algorithm is applied for smoothing to extract its change envelope. For the periodic component, a spectrum analysis and principal period identification algorithm is used to extract its dominant frequency and amplitude. For the residual component data, an independent component analysis algorithm is applied to separate the singular components that are unrelated to known interference sources.

3. The method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 2, characterized in that, For the data of the trend component, the periodic component, and the residual component, independent state evaluation windows are established respectively. Within each state evaluation window, an adaptive threshold algorithm is used to identify data offset events, specifically including: A long-term state evaluation window is established for the changing envelope of the trend component. Based on the statistical distribution of historical data within the long-term state evaluation window, the healthy operating range boundary of the trend component is dynamically calculated. When the changing envelope continuously exceeds the healthy operating range boundary, it is marked as a trend deviation event. A mid-term state assessment window is established for the dominant frequency and amplitude of the periodic component. Within the mid-term state assessment window, periodic distortion events are identified by comparing the difference between the current period and the historical average period, as well as the deviation between the current amplitude and the historical amplitude spectrum. A short-term state evaluation window is established for the singular components of the residual components. Within the short-term state evaluation window, a dynamic alarm threshold is set using extreme value theory to capture transient spike events.

4. The method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 3, characterized in that, The step of aligning and fusing the data offset events along the time dimension to form a composite anomaly event descriptor specifically includes: The trend offset event, the periodic distortion event, and the transient spike event are each assigned a unified timestamp; an aligned time window is established centered on the event occurrence time. Within the alignment time window, the temporal proximity and trend correlation between different trend offset events, periodic distortion events, and transient spike events are calculated; Based on the correlation between the time proximity and the change trend, events that meet the fusion conditions are aggregated into a composite event unit. A feature vector is extracted for each composite event unit. The feature vector contains a combination of event types, a time distribution pattern, and a weighted value of the severity of each component, thereby forming the composite abnormal event descriptor.

5. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 4, characterized in that, The step of retrieving associated fault cause chains and equipment interaction networks from a pre-constructed production process knowledge graph based on the composite abnormal event descriptor specifically includes: The feature vector of the composite abnormal event descriptor is converted into a knowledge graph query statement. The knowledge graph query statement is then used to traverse the entity nodes and relation edges in the production process knowledge graph. The entity nodes include equipment components, process parameters, and material batches. The relation edges include causal relationships, spatial connectivity relationships, and process logic dependencies; all fault mode nodes that match the event types in the feature vector are retrieved; Tracing back multiple upstream causal nodes connected to the fault mode node forms a possible fault cause chain; Simultaneously, other equipment or process segment nodes associated with the fault mode node are located, and a network of equipment interaction relationships is constructed.

6. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 5, characterized in that, The step of constructing a virtual sensing data stream for the current production scenario by utilizing the fault cause chain and the device interaction network specifically includes: Select the most likely causative node from the fault cause chain as a virtual sensing point; Based on the device interaction network, the physical location and measurement parameter type of the virtual sensing point are inferred; By combining the temporal distribution pattern of the composite anomaly event descriptor and using a physical model or a data-driven model, the data sequence that should be observed at the most likely causal node can be simulated in reverse. The data sequence constitutes the virtual sensing data stream, which should theoretically be able to explain the abnormal phenomena represented by the composite abnormal event descriptor.

7. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 6, characterized in that, The step of comparing and verifying the virtual sensing data stream with the processed original multidimensional real-time data stream specifically includes: In the device interaction network, find the actual sensor measurement point that is closest to the physical location of the virtual sensing point; extract the processed raw multidimensional real-time data stream of the actual sensor measurement point in the corresponding time period; calculate the similarity index between the virtual sensing data stream and the processed raw multidimensional real-time data stream of the actual sensor measurement point; The similarity metrics include dynamic time-normalized distance and cross-correlation coefficient; If the similarity index is higher than the preset verification threshold, it is determined that the virtual sensing data stream is supported by the data of the actual sensor measurement points, and the probability of the fault cause chain increases; otherwise, the probability of the fault cause chain decreases.

8. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 7, characterized in that, Based on the comparison and verification results, a structured report is generated that includes fault location, cause inference, and maintenance recommendations, specifically including: Integrate the comparative verification results with the credibility score of the fault cause chain; based on the equipment interaction network, accurately label the physical equipment or process links where the fault may occur, and complete the fault location; By combining the production process manual and the historical maintenance case library, a cause inference in natural language is generated for the most likely cause node in the fault cause chain; Based on the aforementioned cause inference, the maintenance procedure knowledge base is retrieved to generate maintenance recommendations that include specific operating steps, required spare parts, and safety precautions. The fault location, the cause inference, and the maintenance recommendations are integrated into the structured report.

9. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 8, characterized in that, The step of mapping the structured report and the composite anomaly event descriptor onto the 3D device model in the VR monitoring scene specifically includes: Load 3D equipment models corresponding one-to-one with the physical production line into the VR monitoring platform; based on the fault location information, highlight the corresponding virtual equipment or components in the 3D scene; The event type combination and time distribution pattern information in the composite abnormal event descriptor are transformed into dynamic, visually animated icons attached to the highlighted component for display. The key content of the structured report, including the summary of the cause inference and the key points of the maintenance recommendations, will be displayed in an interactive floating information panel next to the highlighted component.

10. A method for production management of flame-retardant nonwoven fabric based on VR technology monitoring according to claim 9, characterized in that, After loading the 3D equipment models corresponding one-to-one with the physical production line into the VR monitoring platform, the process further includes: Receive the latest multi-dimensional real-time data stream from the industrial control system; The motion state of each moving part in the 3D device model is synchronized with the physical device in real time. When a new composite abnormal event descriptor is detected, the VR view is automatically switched to the location of the virtual device corresponding to the fault location, and an alarm sound is triggered. At the same time, a floating information panel containing the latest structured report content is automatically popped up.

Images