A deep learning-based cotton combing machine multi-source data fusion abnormal early warning system

Abstract

The application discloses a kind of based on deep learning's cotton carding machine multi-source data fusion abnormal early warning system, it is related to textile machinery intelligent monitoring technical field, the present application is packaged single-head monitoring data by combing head identification and process cycle identification, form combing head cycle object, and by material transfer delay and delay correlation record, combing head cycle object is associated with the corresponding combined sliver quality data of quality arrival cycle identification;In the case where the cotton web of single combing head is abnormal after guide cotton, combination and drafting are covered by combined sliver result, the system can still output abnormal early warning result with machine identification, combing head identification and process cycle identification, so that abnormal source is limited to corresponding combing head and corresponding process cycle.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for textile machinery, and in particular to an anomaly early warning system based on deep learning for multi-source data fusion in cotton combing machines. Background Technology

[0002] The cotton combing machine is a key piece of equipment in the spinning preparation process. It is used to comb, separate, join, guide, combine, and draft cotton laps, and output combed sliver. In actual production, a cotton combing machine typically has multiple combing heads arranged horizontally along the machine frame. Each combing head periodically combs its corresponding cotton lap, forming a single-headed cotton web that enters the guide channel. The multiple outputs converge into a combined combed sliver during subsequent conveying, combining, and drafting processes. The combined combed sliver then enters the quality inspection, sliver winding, and sliver can output stages, serving as the feed for subsequent drawing or spinning processes.

[0003] Current methods for monitoring anomalies in cotton combing machines include sliver weight detection, photoelectric detection of web breakage or cotton spillage, transmission current detection, vibration detection, temperature detection, PLC alarm data acquisition, and data aggregation on a workshop platform. Sliver weight detection is typically installed on the output side after merging and drafting to reflect the weight deviation and evenness fluctuation of the merged slivers. Photoelectric detection is usually installed near the guide plate, small table, or trumpet opening to detect missing, tangled, or locally accumulated webs. Equipment-side monitoring typically collects the operating status of the motor, transmission mechanism, suction mechanism, and key moving parts. PLC alarms and workshop platform records summarize data on machine shutdowns, cotton blockages, sliver breaks, roll changes, can changes, and changes in process parameters.

[0004] When multiple combing heads operate in parallel, if a single combing head experiences issues such as top combing snagging, lap unwinding fluctuations, insufficient localized air suction, abnormal separation and bonding, or obstructed cotton guiding, the abnormal condition may initially manifest as edge breakage, abrupt changes in thickness, intermittent overlaps, localized accumulation, or disordered texture in the corresponding cotton web. These localized abnormalities mix with the normal output of other combing heads after entering the merging and drafting process. The combined combed sliver quality inspection point yields the result of the superposition of multiple outputs. The abnormal components from a single head are averaged in the combined sliver quantitative value and evenness value, making it difficult to directly determine the source of the abnormality from the output-side inspection results.

[0005] In centralized monitoring scenarios in workshops, maintenance personnel typically only receive alarms for the entire machine, information on combined combed sliver quality fluctuations, or trigger information for a single detection point. This type of information lacks correlation records with specific combing heads, cotton laps, specific nipping intervals, and abnormal cotton web images. Maintenance personnel need to inspect multiple combing heads one by one, and the anomaly investigation process is affected by the time of machine downtime, the state of residual cotton web, and manual experience. It is difficult to establish a continuous correspondence between anomaly handling records and subsequent process adjustments.

[0006] Furthermore, the overall machine current, vibration, and temperature signals primarily reflect shared transmission loads, overall machine operating status, or environmental disturbances. When used individually, they are difficult to distinguish between abnormalities in a single combed head's web and normal fluctuations in the overall machine. While visual or photoelectric detection can obtain localized conditions close to the web formation location, its results are easily affected by fly waste, random web texture, lighting changes, roll changes, start-up / stop processes, and can changes. The acquisition locations and response times of different data sources are also inconsistent, with differences in material transfer distance and clamping cycles between the web formation location, equipment status sampling points, PLC events, and sliver quality detection points. If multi-source data is aggregated solely based on acquisition time, corresponding offsets will occur between abnormal webs, equipment status, and subsequent sliver quality fluctuations.

[0007] Therefore, the following technical problems exist in the parallel production process of multiple combing heads in a cotton combing machine: the abnormality of the cotton web of a single combing head is masked by the merging result in the subsequent merging and drafting; it is difficult to locate the source of the abnormality in the equipment-side signals and the quality signals on the output side; visual or photoelectric signals are triggered by vibration due to on-site interference; and multi-source data lacks associated early warning objects for the combing head under the conditions of clamping cycle and material transfer delay. Summary of the Invention

[0008] This application provides a deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines, which solves the problem that the source of anomalies in multi-combed cotton webs is masked after merging and drafting, and that there is a misalignment in the multi-source data.

[0009] This invention provides a deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines. The cotton combing machine includes multiple combing heads, and the cotton webs output by each combing head are combined and stretched to form a combined combed sliver. The system includes a multi-source data access module, a process cycle alignment module, a time-series heterogeneous graph generation module, a deep learning early warning module, and an early warning handling module.

[0010] The multi-source data access module collects single-head monitoring data, shared mechanism data, merged combed sliver quality data, and process event data.

[0011] The process cycle alignment module generates a process cycle identifier based on the clamping cycle count. When the clamping cycle count is missing, it generates a process cycle identifier based on the driving phase. It then encapsulates the single-head monitoring data into a combing head cycle object by combining the combing head identifier and the process cycle identifier.

[0012] After the delay calibration trigger event occurs, the process cycle alignment module generates a material transfer delay based on the material path length between the comb head output position and the quality detection position of the combined comb strip, the current conveying speed, the draft ratio, and the clamping cycle count, and generates a delay association record containing the original process cycle identifier and the quality arrival cycle identifier.

[0013] The time-series heterogeneity diagram generation module generates a time-series heterogeneity diagram based on the combing head cycle object, shared mechanism data, merged combing strip quality data, process event data, and delay association records. The time-series heterogeneity diagram includes at least the combing head node carrying the combing head cycle object, the merged combing strip quality node carrying the quality cycle characteristics, and the material flow edge connecting the combing head node and the merged combing strip quality node according to the delay association records.

[0014] The deep learning early warning module encodes the temporal heterogeneous graph corresponding to the continuous process cycle identifier, and generates the abnormal contribution value of each combing head cycle object based on the single head monitoring data in the combing head cycle object, the quality cycle characteristics obtained by statistical analysis of the merged combing strip quality data within the time window corresponding to the quality arrival cycle identifier, and the edge relationship of the temporal heterogeneous graph.

[0015] When the abnormal contribution value is greater than or equal to the trigger threshold, the early warning and handling module outputs an abnormal early warning result that includes the machine identifier, combing head identifier, process cycle identifier, and abnormal source category.

[0016] In some embodiments, the material path length includes the front material path length between the comb head output position and the drafting inlet, and the rear material path length between the drafting outlet and the quality detection position of the combined comb strip. The material transfer delay is generated based on the front material path length, the rear material path length, the current conveying speed, the drafting ratio, and the clamping cycle.

[0017] The delay calibration trigger events include machine initialization, process formula switching, speed increase, speed decrease, and shutdown recovery; the delay association record also includes machine identifier, combing head identifier, combined combing strip quality inspection point identifier, and association status identifier.

[0018] In some embodiments, the multi-source data access module includes a single-head access unit, a shared mechanism access unit, a quality access unit, and an event access unit;

[0019] The single-head access unit collects at least one of the following for each combing head: cotton web image, cotton web edge continuity value, cotton web coverage value, cotton roll unwinding value, local suction value, and cotton web breakage or cotton gushing state, and writes it into the machine identifier and combing head identifier.

[0020] The shared mechanism access unit collects at least one of the following: main motor current, separating roller speed, drafting roller speed, suction main pipe pressure, cotton dropping mechanism status, or machine vibration value;

[0021] The quality access unit collects and merges at least one of the following: combed sliver quantitative value, sliver fluctuation value, sliver breakage status, or sliver coiling status;

[0022] The event access unit collects at least one of the following events: start-up, stop-up, change of drum, change of cotton roll, change of batch, speed reduction, speed increase, or cleaning and maintenance.

[0023] In some embodiments, the process cycle alignment module includes a cycle recognition unit and a window encapsulation unit;

[0024] The cycle recognition unit determines the process cycle identifier from the clamp action signal, cylinder drive signal, separation roller drive signal, or spindle encoding signal;

[0025] The window encapsulation unit segments the single-head monitoring data, shared mechanism data, and process event data with different sampling frequencies according to the process cycle identifier. It determines the quality arrival time window of the merged combed strip quality data according to the quality arrival cycle identifier, and performs statistics on the merged combed strip quality data that falls within the quality arrival time window to obtain the quality cycle characteristics. Then, it encapsulates the single-head monitoring data under the same machine identifier, the same combing head identifier, and the same process cycle identifier into a combing head cycle object.

[0026] When multiple sampling points exist within the same process cycle identifier, the window encapsulation unit extracts at least one of the following as a cycle statistic: mean, extreme value, duration, and number of state changes.

[0027] In some embodiments, the temporal heterogeneous graph generation module includes a node generation unit and an edge generation unit;

[0028] The node generation unit generates combing head nodes, shared mechanism nodes, merged combing strip quality nodes, and process event nodes. The combing head node carries the combing head cycle object.

[0029] The edge generation unit generates material flow edges, shared mechanism edges, event constraint edges, and time continuity edges. Material flow edges connect combing head nodes and merged combing strip quality nodes according to material transfer delays. Shared mechanism edges connect shared mechanism nodes and combing head nodes affected by the shared mechanism. Event constraint edges connect process event nodes and nodes affected by the process event. Time continuity edges connect combing head nodes of the same combing head under adjacent process cycle identifiers.

[0030] In some embodiments, the multi-source data access module further includes a single-head feature generation unit;

[0031] The single-head feature generation unit performs combing head region cropping on the cotton web image to obtain the cotton web region image corresponding to the combing head, and generates at least one of the following values ​​based on the cotton web region image: edge breakage value, overlap continuity value, local accumulation value, or texture fluctuation value; the single-head feature generation unit splices the data generated from the cotton web region image with the cotton roll unwinding value, local air suction value, web breakage state, or cotton gushing state to form a single-head monitoring feature.

[0032] In some embodiments, the deep learning early warning module includes a visual encoder, a temporal encoder, a heterogeneous graph encoding network, and an anomaly decoding network; the visual encoder receives single-head monitoring data carried by the combing head node or single-head monitoring features generated from the single-head monitoring data, and outputs a single-head cotton web representation;

[0033] The timing encoder receives the single-head web representation of the same combing head under continuous process cycle identifiers and outputs the single-head timing representation; the heterogeneous graph encoding network receives the combing head node representation, shared mechanism node representation, merged combing sliver quality node representation, process event node representation, and edge type identifier, and outputs the fused node representation; the anomaly decoding network receives the fused node representation and outputs the anomaly contribution value and anomaly source category.

[0034] In some embodiments, the edge type identifier includes a material flow edge identifier, a shared mechanism edge identifier, an event constraint edge identifier, and a time-continuous edge identifier;

[0035] The heterogeneous graph coding network converts edge type identifiers into edge type embeddings, generates edge type weights based on source node representations, target node representations, and edge type embeddings, and aggregates adjacent node representations according to edge type weights to obtain fused node representations that output the contribution value of participating anomalies and the anomaly source category.

[0036] The deep learning early warning module also includes a process event masking unit. The process event masking unit adds event tags to the corresponding process cycle identifier based on the process event data and converts the event tags into process event node representations. The anomaly decoding network outputs process transition source category, single-head cotton web source category, shared mechanism source category, merged combed sliver source category, or sensor status source category based on the process event node representation. The process transition source category is used to record the early warning trigger source caused by the process event.

[0037] In some embodiments, the anomaly warning result includes an anomaly location field, an anomaly source field, an anomaly link field, and a handling field; the anomaly location field records the machine identifier, combing head identifier, and process cycle identifier;

[0038] The anomaly source field includes anomaly source category and anomaly source details. The anomaly source category records at least one of the following: single-head cotton web source category, shared mechanism source category, process transition source category, combined combed sliver source category, or sensor status source category. The anomaly source details record at least one of the following: cotton web edge breakage, discontinuous cotton web overlap, local accumulation, single-head suction anomaly, cotton roll unwinding anomaly, separation and joining anomaly, shared mechanism anomaly, combined combed sliver anomaly, or sensor status anomaly.

[0039] The Abnormal Link field records the node and edge type that participated in generating the abnormal contribution value; the Handling field records the inspection object, the sending object, the review status, and the maintenance status; when the Abnormal Link field contains quality missing test association, conservative association, or low confidence flag, the Abnormal Source field retains the candidate abnormal source category, and the Handling field records the pending review status.

[0040] In some embodiments, the deep learning early warning module further includes a sample construction unit and a model update unit;

[0041] The sample construction unit extracts training samples from historical data according to machine identifier, combing head identifier and process cycle identifier. The training samples include normal samples, single-head abnormal samples, shared mechanism abnormal samples, process transition samples and merged combing strip abnormal samples.

[0042] The sample construction unit generates sample labels based on maintenance records, manual review records, quality traceability records, or shutdown alarm records;

[0043] The model update unit updates the model parameters of the visual encoder, temporal encoder, heterogeneous graph encoding network and anomaly decoding network with training samples and sample labels, and uses the anomaly source category as the training classification target, and the machine identifier, combing head identifier and process cycle identifier as localization supervision information.

[0044] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0045] This invention encapsulates single-head monitoring data through combing head identification and process cycle identification to form a combing head cycle object. By using material transfer delay and delay correlation records, it establishes a correlation between the combing head cycle object and the quality data of the merged combed sliver corresponding to the quality arrival cycle identification. Even when a single combing head's web anomaly is masked by the merged combed sliver result after cotton guiding, doubling, and drafting, the system can still output anomaly warning results using machine identification, combing head identification, and process cycle identification, thus limiting the source of the anomaly to the corresponding combing head and the corresponding process cycle.

[0046] Because material transfer delays are generated based on material path length, current conveying speed, draft ratio, and clamping cycle time, and are re-established after machine initialization, process formula switching, speed increase, speed decrease, or shutdown recovery, the time misalignment between the single-head cotton web formation position and the quality detection position of the combined combed sliver is converted into a quality arrival cycle identifier. Therefore, single-head monitoring data, shared mechanism data, process event data, and combined combed sliver quality data all participate in anomaly contribution value generation under the same process cycle time caliber, reducing mismatches in anomaly sources caused by direct splicing of data collection times.

[0047] The temporal heterogeneity graph generation module connects the combing head node, shared mechanism node, combined combed sliver quality node, and process event node with material flow edges, shared mechanism edges, event constraint edges, and time continuity edges. The deep learning early warning module performs representation transmission between nodes with different edge types. Thus, local anomalies in single-head cotton webs, changes in the state of shared mechanisms, process transition events, and fluctuations in the quality of combined combed slivers can enter the anomaly contribution value generation process according to their respective action links, avoiding the mixed judgment of anomaly sources by single visual detection, single photoelectric detection, or whole-machine status detection.

[0048] The process event masking unit converts start-up, shutdown, can change, lap change, batch change, speed increase, speed decrease, and cleaning / maintenance events into process event node representations. The anomaly decoding network combines edge type weights, process event node representations, and delay association records to output the anomaly source category. Thus, planned process transitions resulting in changes in web texture, evenness fluctuations, or missing detections are recorded as early warning triggers for process events, while unplanned single-head web anomalies are recorded as single-head web source categories. The anomaly link field and handling field in the early warning result can respectively point to the combing head, shared mechanism, quality inspection point, or the inspection object corresponding to the process event. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0050] Figure 1 This is a framework diagram of the multi-source data fusion anomaly early warning system for a cotton combing machine based on deep learning, as described in the embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0053] Example 1:

[0054] like Figure 1As shown, this embodiment employs a deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines. The cotton combing machine includes multiple combing heads operating in parallel. The cotton web formed by each combing head is guided, combined, and drafted to form a combined combed sliver.

[0055] The system includes a multi-source data access module, a process cycle alignment module, a time-series heterogeneous graph generation module, a deep learning early warning module, and an early warning handling module. Specifically, these modules transmit data in the order of data access, cycle alignment, graph structure generation, anomaly identification, and early warning output. Data records generated in the previous processing stage carry machine identifiers and data generation times. Subsequent processing stages read the corresponding data records based on these machine identifiers and generation times, ensuring a continuous data source relationship between the combing head cycle object, delay-related records, time-series heterogeneous graph, and anomaly early warning results.

[0056] The multi-source data access module collects single-head monitoring data, shared mechanism data, merged combed sliver quality data, and process event data.

[0057] The process cycle alignment module generates a process cycle identifier based on the clamping cycle count, and generates a process cycle identifier based on the drive phase when the clamping cycle count is missing. It encapsulates single-head monitoring data into a combing head cycle object by encapsulating the combing head identifier and the process cycle identifier. After machine initialization, process formula switching, speed increase, speed decrease, or shutdown recovery, it generates a material transfer delay based on the material path length from the combing head output position to the quality inspection position of the combined combed strip, the current conveying speed, the draft ratio, and the clamping cycle count. Based on the material transfer delay, it generates a delay association record that includes machine identifier, combing head identifier, original process cycle identifier, quality arrival cycle identifier, combined combed strip quality inspection point identifier, and associated status identifier.

[0058] The timing heterogeneity diagram generation module generates a timing heterogeneity diagram based on the combing head cycle object, shared mechanism data, merged combing strip quality data, process event data, and delay association records.

[0059] The deep learning early warning module encodes the temporal heterogeneous graph corresponding to the continuous process cycle identifier. It uses the single-head monitoring data in the combing head cycle object, the quality cycle characteristics obtained by statistical analysis of the merged combing strip quality data within the time window corresponding to the quality arrival cycle identifier, and the edge relationships of the temporal heterogeneous graph to generate the abnormal contribution value of each combing head cycle object.

[0060] The early warning and handling module outputs an anomaly warning result when the anomaly contribution value is greater than or equal to the trigger threshold. The anomaly warning result includes the machine identifier, combing head identifier, process cycle identifier, and anomaly source category. The trigger threshold is an implementation parameter used to determine whether to output an anomaly warning result. The trigger threshold is tuned by the distribution of anomaly contribution values ​​of normal samples and confirmed anomaly samples on the validation set and is bound to the model parameter version.

[0061] Multi-source data refers to a data set originating from single combing head detection locations, shared mechanism detection locations, combined combed sliver quality detection locations, and process event record locations. A combing head cycle object is a data record encapsulated using machine identifier, combing head identifier, and process cycle identifier as indexes, representing single-head monitoring data of the same combing head within a single process cycle. A process cycle identifier is a cycle number generated after segmenting the continuous operation of the cotton combing machine based on clamping cycle time or drive phase. Material transfer delay refers to the time interval or number of process cycles corresponding to the single-head cotton web reaching the combined combed sliver quality detection location from the combing head output location. A temporal heterogeneous graph refers to a graph data structure containing different node types, different edge types, and connections between adjacent process cycles. Anomaly contribution value refers to the anomaly response quantity generated by the deep learning early warning module for the combing head cycle object, used to trigger anomaly early warning results. Anomaly source category refers to the source type recorded in the anomaly early warning results, including single-head cotton web source category, shared mechanism source category, process transition source category, combined combed sliver source category, or sensor status source category.

[0062] The process transition source category is used to record the source of early warnings caused by process events. The recorded object is the impact of process events on detection data and node characterization, and is not recorded as a single-head fault source.

[0063] The "Abnormal Source Category" records the upper-level category of abnormal sources, while the "Abnormal Source Details" records the specific abnormal states under each abnormal source category. The abnormal source details corresponding to the "Single-Headed Web Source Category" include edge breakage, discontinuous overlap, local accumulation, local suction abnormalities, cotton lap unwinding abnormalities, separation and joining abnormalities, web breakage, and cotton gushing. The abnormal source details corresponding to the "Shared Mechanism Source Category" include main motor current abnormalities, suction main pipe pressure abnormalities, roller speed abnormalities, and machine vibration abnormalities. The abnormal source details corresponding to the "Process Transition Source Category" include start-up, shutdown, can change, cotton lap change, batch change, speed increase, speed decrease, and cleaning and maintenance. The abnormal source details corresponding to the "Merged Combed Sliver Source Category" include quantitative abnormalities, evenness abnormalities, sliver breakage, and sliver coiling. The abnormal source details corresponding to the "Sensor Status Source Category" include missing data acquisition, data acquisition exceeding limits, and abnormal data acquisition channel status.

[0064] The data acquisition time is the time information written when the acquisition end forms the original data record, and the data generation time is the time information written when the corresponding processing stage generates the intermediate data record. The data acquisition time and the data generation time use the same clock reference of the machine control system; when there is a deviation between the acquisition end time and the machine control system time, the receiving time recorded by the machine control system is used as the alignment time, and the time correction status is written into the corresponding data record.

[0065] The delay calibration unit is a data processing unit in the process cycle alignment module that performs the generation of material transfer delays, while the correlation correction unit is a data processing unit in the process cycle alignment module that performs the writing of delay correlation records. The delay calibration unit receives the material path length from the combing head output position to the merged combed sliver quality detection position, the current conveying speed, the draft ratio, and the clamping cycle, and outputs the material transfer delay; the correlation correction unit receives the material transfer delay, the combing head cycle object, and the merged combed sliver quality data, and outputs the delay correlation record.

[0066] Furthermore, the association status identifier is used to record the generation status of delayed association records. The association status identifier includes normal association, temporary association, conservative association, quality missing test association, and association pending review. When there is a missing test in the quality data of the merged combed strip corresponding to the quality arrival cycle identifier, the association correction unit generates a quality missing test association and writes the quality detection point identifier, quality arrival cycle identifier, and missing test status identifier into the delayed association record. The quality missing test association participates in the input of the anomaly decoding network, but its corresponding merged combed strip quality data is not used as the priority output basis for a single anomaly source category.

[0067] In one embodiment, the multi-source data access module includes a single-head access unit, a shared mechanism access unit, a quality access unit, and an event access unit. The single-head access unit collects at least one of the following for each combing head: web image, web edge continuity value, web coverage value, lap unwinding value, local suction value, web breakage state, or cotton slugging state, and writes the collected data into the machine identifier and combing head identifier. The shared mechanism access unit collects at least one of the following: main motor current, separating roller speed, drafting roller speed, suction main pipe pressure, cotton dropping mechanism state, or machine vibration value. The quality access unit collects at least one of the following: combined combing sliver weight value, sliver fluctuation value, sliver breakage state, or sliver looping state. The event access unit collects at least one of the following: start-up, stop-up, can change, lap change, batch change, speed reduction, speed increase, or cleaning and maintenance events.

[0068] Single-head monitoring data includes data acquisition time, machine identifier, combing head identifier, acquisition channel identifier, detection location identifier, and single-head detection value. Shared mechanism data includes data acquisition time, machine identifier, shared mechanism identifier, operating status value, and alarm status value. Combined combed sliver quality data includes data acquisition time, machine identifier, quality inspection point identifier, quantitative detection value, sliver detection value, and sliver breakage status value. Process event data includes event occurrence time, machine identifier, event type, event duration, and event source. After the acquired data enters the multi-source data access module, it is written to the data buffer according to the machine identifier and acquisition channel identifier, and the data buffer provides data records with source identifiers to the process cycle alignment module.

[0069] Specifically, the source identifiers in the data buffer include the acquisition channel identifier, the detection location identifier, and the data category identifier. The data category identifier is used to distinguish between single-head monitoring data, shared mechanism data, merged combed sliver quality data, and process event data. After the merged combed sliver quality data enters the data buffer, it forms quality inspection records according to the quality inspection point identifier and the data acquisition time. The quality inspection records participate in delay correlation in the process cycle alignment module based on the quality arrival cycle identifier.

[0070] In one embodiment, the process cycle alignment module includes a beat recognition unit and a window encapsulation unit. The beat recognition unit determines the process cycle identifier from the clamp action signal, cylinder drive signal, separation roller drive signal, or spindle encoding signal. The window encapsulation unit segments the single-head monitoring data, shared mechanism data, and process event data with different sampling frequencies according to the process cycle identifier, segments the merged combing strip quality data according to the quality arrival cycle identifier, and encapsulates the single-head monitoring data under the same machine identifier, the same combing head identifier, and the same process cycle identifier into a combing head cycle object. When multiple sampling points exist within the same process cycle identifier, the window encapsulation unit extracts at least one of the mean, extreme value, duration, and number of state changes as cycle statistics.

[0071] The process cycle alignment module uses the clamping beat signal as the cycle segmentation reference. When the clamping beat signal is missing, it uses one of the following as the cycle segmentation reference: the spindle encoding signal, the cylinder drive signal, or the separation roller drive signal. The drive phase is used to indicate the phase position of the spindle, cylinder, or separation roller within one mechanical operation cycle. When the clamping beat signal is missing, the process cycle alignment module considers one preset phase cycle as a process cycle and generates a process cycle identifier upon completion of each preset phase cycle. The window encapsulation unit aggregates data with sampling frequencies higher than the process cycle frequency into cycle statistics and writes data with sampling frequencies lower than the process cycle frequency into the corresponding process cycle identifier according to the holding time. For single-head monitoring data not collected within the same process cycle identifier, the window encapsulation unit writes a missing measurement status identifier, which participates in the subsequent anomaly source category determination.

[0072] In one example, the missing data status is identified as the complete collection status of single-head monitoring data, shared mechanism data, merged combed sliver quality data, or process event data within the corresponding process cycle identifier. When the cotton web area image is missing, the window encapsulation unit retains the already collected data from the cotton lap unwinding value, local suction value, web breakage status, and cotton spurting status, and writes it into the image missing data status identifier; when the cotton lap unwinding value, local suction value, web breakage status, and cotton spurting status are all missing, the corresponding combing head cycle object still retains the machine identifier, combing head identifier, process cycle identifier, and missing data status identifier, and participates in the subsequent time-series heterogeneous diagram generation as a low-confidence data record.

[0073] In one implementation, the temporal heterogeneous graph generation module includes a node generation unit and an edge generation unit. The node generation unit generates combing head nodes, shared mechanism nodes, merged combing strip quality nodes, and process event nodes. The combing head nodes carry combing head cycle objects. The edge generation unit generates material flow edges, shared mechanism edges, event constraint edges, and time-continuous edges. The material flow edges connect the combing head nodes and merged combing strip quality nodes according to the material delivery delay. The shared mechanism edges connect the shared mechanism nodes and the combing head nodes affected by the shared mechanism. The event constraint edges connect the process event nodes and the nodes affected by the process event. The time-continuous edges connect the combing head nodes of the same combing head under adjacent process cycle identifiers.

[0074] The data fields for the combing head node include machine identifier, combing head identifier, process cycle identifier, and single-head monitoring characteristics. The data fields for the shared mechanism node include machine identifier, shared mechanism identifier, process cycle identifier, and shared mechanism characteristics. The data fields for the merged combed strip quality node include machine identifier, quality inspection point identifier, quality arrival cycle identifier, and merged combed strip quality characteristics. The data fields for the process event node include machine identifier, event category, event occurrence cycle identifier, and event duration status. Material flow records carry delay-related records, shared mechanisms carry the scope of shared mechanism action, event constraints carry the event's impact cycle range, and time continuity carries the relationship between adjacent cycles.

[0075] In one embodiment, the multi-source data access module further includes a single-head feature generation unit; the single-head feature generation unit performs combing head region cropping on the cotton web image to obtain the cotton web region image corresponding to the combing head, and generates at least one of edge breakage value, overlap continuity value, local accumulation value or texture fluctuation value based on the cotton web region image; the single-head feature generation unit splices the data generated from the cotton web region image with the cotton roll unwinding value, local air suction value, web breakage state or cotton gushing state to form a single-head monitoring feature.

[0076] The cotton web area image is bound to the combing head identifier. Edge breakage values ​​are formed by edge gaps, discontinuities, and spread in the cotton web area image. Overlap continuity values ​​are formed by the overlap, breakage, and texture continuity of adjacent cotton web segments. Local accumulation values ​​are formed by sudden changes in local brightness, thickness, or coverage area in the cotton web area image. Texture fluctuation values ​​are formed by the texture changes in the cotton web area image over consecutive processing cycles. Single-head monitoring features include image features, single-head sensing features, and missing measurement status identifiers.

[0077] Single-head sensing characteristics are formed from collected data on cotton lap unwinding value, local suction value, web breakage state, and cotton slugging state. Shared mechanism characteristics are formed from collected data on main motor current, separating roller speed, drafting roller speed, suction main pipe pressure, cotton dropping mechanism state, or machine vibration value. Combined combed sliver quality characteristics are formed from collected data on combined combed sliver basis weight, evenness fluctuation value, sliver breakage state, or sliver coiling state.

[0078] In one implementation, the deep learning early warning module further includes a process event masking unit. The process event masking unit adds event tags to the corresponding process cycle identifiers based on the process event data and converts the event tags into process event node representations. The process event node representations are entered into the heterogeneous graph encoding network as the node representations corresponding to the event constraint edges and participate in the anomaly source category output in the anomaly decoding network.

[0079] Event tags include event category, event start cycle identifier, event end cycle identifier, and event source. After the process event node representation enters the anomaly decoding network, it participates in the anomaly source category determination along with the combing head node representation under the same process cycle identifier. Process transition status records corresponding to start-up, stop-up, can change, lap change, batch change, speed reduction, speed increase, or cleaning / maintenance events are classified as process transition source categories; anomaly status records not covered by process events and jointly pointed to by single-head monitoring characteristics, delay correlation records, and adjacent cycle relationships are classified as single-head web source categories.

[0080] In one implementation, the anomaly warning result includes an anomaly location field, an anomaly source field, an anomaly link field, and a handling field. The anomaly location field records the machine identifier, combing head identifier, and process cycle identifier. The anomaly source field includes an anomaly source category and an anomaly source detail. The anomaly source category records at least one of the following: single-head cotton web source category, shared mechanism source category, process transition source category, merged combing sliver source category, or sensor status source category. The anomaly source detail records at least one of the following: cotton web edge breakage, discontinuous cotton web overlap, local accumulation, single-head suction anomaly, cotton lap unwinding anomaly, separation and joining anomaly, shared mechanism anomaly, merged combing sliver anomaly, or sensor status anomaly. The anomaly link field records the nodes and edge types that participate in the generation of the anomaly contribution value. The handling field records the inspection object, the sending object, the review status, and the maintenance status.

[0081] The anomaly link field includes the combing head node identifier, shared mechanism node identifier, merged combing sliver quality node identifier, process event node identifier, material flow edge identifier, shared mechanism edge identifier, event constraint edge identifier, and time continuity edge identifier. The inspection object in the handling field includes at least one of the following: combing head, top comb, cotton lap, cotton guide channel, suction branch, separation and joining mechanism, drafting mechanism, or quality inspection point. The review status includes pending review, confirmed, and ruled out. The maintenance status includes pending processing, processing, and processed. When anomaly warning results are written to the quality traceability record, the machine identifier, combing head identifier, and process cycle identifier are used as indexes.

[0082] Furthermore, the sample labels include machine identifier, combing head identifier, process cycle identifier, anomaly source category, anomaly source details, and review status. When the review status is "confirmed," the corresponding training sample participates in anomaly source category identification and anomaly location field generation; when the review status is "excluded," the corresponding training sample participates in training as a non-anomaly sample or process transition sample; when the review status is "pending review," the corresponding training sample enters the sample library but does not participate in the training of the current model parameter version.

[0083] In a preferred embodiment of Example 1, when the delay calibration unit determines the material transfer delay, it uses the actual material channel between the comb head output position and the combined comb strip quality detection position as the calibration object.

[0084] The actual material flow path is defined as the material travel path between the combing head output position, the cotton guide path, the drafting inlet, the drafting outlet, and the quality inspection position of the combined combed sliver. The length of the initial material path corresponds to the path from the combing head output position to the drafting inlet, and the length of the subsequent material path corresponds to the path from the drafting outlet to the quality inspection position of the combined combed sliver. The lengths of the initial material paths for multiple combing heads are recorded separately, while the lengths of the subsequent material paths are recorded according to the common path after merging.

[0085] For each combing head, first measure the length of the front material path from the output position of the combing head to the drafting inlet, and the length of the rear material path from the drafting outlet to the quality inspection position of the combined combed strips, under the shutdown and calibration state; under the running state, read the conveying speed, drafting ratio and clamping cycle corresponding to the current process cycle, and calculate the material arrival time according to the relationship between the speed change of the material before and after drafting.

[0086] The material arrival time is calculated as follows:

[0087] ,

[0088] in, The label for finely combed hair is: And the process cycle is marked as The material arrival time corresponding to the combing head cycle object; This is a label for finely combed hair; For process cycle identification; superscript This serves as a material transfer identifier, used to distinguish quantities related to material transfer delays. The label for finely combed hair is: The length of the material path from the combing head output position to the drafting inlet, with a value greater than or equal to 0; The process cycle is identified as The conveying speed of the material before stretching, with a value greater than 0; The length of the material path from the traction outlet to the quality inspection position of the merged combed strip is greater than or equal to 0, and is recorded according to the common channel after merging; The process cycle is identified as The draw ratio at that time takes a value greater than 0.

[0089] The delay calibration unit converts the material arrival time into the number of process cycles, enabling the same combing head cycle object to correspond with the quality data of the merged combed strips when it actually arrives at the quality inspection position. The conversion method is as follows:

[0090] ,

[0091] in, The label for finely combed hair is: And the process cycle is marked as The number of material transfer delay cycles corresponding to the combing head cycle object. It is a non-negative integer; , and The aforementioned definition will be used; The process cycle is identified as The single-cycle time corresponding to the clamping beat at that time is greater than 0; This is a rounding-up operator. The number of material delivery delay cycles obtained by rounding up is used to determine the quality inspection cycle after the material arrival time. The quality arrival cycle identifier indicates that the quality inspection position of the combined combed strip can obtain the first inspection cycle affected by the material corresponding to the combing head cycle object. When the quality inspection data is recorded using a continuous sampling method, the correlation correction unit statistically analyzes the combined combed strip quality data using the time window corresponding to the quality arrival cycle identifier.

[0092] Before generating the quality arrival cycle identifier, the correlation correction unit converts the original process cycle identifiers within the same machine into consecutive integer cycle numbers that increment over time, and uses these consecutive integer cycle numbers as the process cycle identifiers in subsequent cycle offset calculations. When the original process cycle identifier is a timestamp, coded string, or composite number, the association correction unit establishes a one-to-one mapping relationship between the original process cycle identifier and the consecutive integer cycle number according to the acquisition time sequence, and writes the mapping relationship into the delay association record for backtracking between the quality arrival cycle identifier and the original process cycle identifier.

[0093] The correlation correction unit generates a quality arrival cycle identifier based on the number of material transfer delay cycles, and writes the quality data of the combined combed strips collected under this quality arrival cycle identifier into the delay correlation record of the corresponding combing head cycle object. The method for determining the quality arrival cycle identifier is as follows:

[0094] ,

[0095] in, The label for finely combed hair is: And the process cycle is marked as The quality arrival cycle identifier corresponding to the combing head cycle object. Represented by consecutive integer periodic numbers that increase with time within the same machine; , and The aforementioned definition will be used.

[0096] When the current process cycle is within a stable operating window, the delay calibration unit calculates the number of material transfer delay cycles based on the conveying speed, draft ratio, and clamping cycle time corresponding to the current process cycle. A stable operating window refers to a window where the conveying speed, draft ratio, and clamping cycle time are all within the stable range of process parameters defined by the equipment installation calibration record and historical quality traceability record for multiple consecutive process cycles. When the current process cycle is in a transition phase of speed increase, speed decrease, or shutdown recovery, the delay calibration unit marks the corresponding combing head cycle object as a delayed transition object and establishes a temporary delay association record using the larger value between the number of material transfer delay cycles formed in the most recent stable operating phase and the currently recalculated number of material transfer delay cycles. This temporary delay association record is used for low-confidence early warning calculations during the transition phase, using a larger number of material transfer delay cycles to prevent single-head cotton webs that have not yet reached the quality inspection position of the merged combed sliver from being prematurely associated with the quality inspection data. After the stable operating window is restored, the association correction unit regenerates formal delay association records for combing head cycle objects that have not yet completed quality traceability. Once the conveying speed, draw ratio, and clamping cycle time have returned to a stable operating window, a formal delay correlation record will be generated.

[0097] The stable calibration interval is a numerical range formed by historical stable operation data corresponding to normal samples under the same machine, process type, drafting process configuration, and data category. Historical stable operation data comes from historical process cycles that did not record abnormal warning results, shutdown alarm records, or maintenance confirmation of abnormalities. The stable calibration interval is used to determine the deviation status of merged combed sliver quality data, shared mechanism data, and single-head monitoring characteristics, and is bound to the process type or drafting process configuration version.

[0098] The delay calibration unit checks the variation range of conveyor speed, draft ratio, and clamping cycle time using the process cycle as the refresh unit. If the conveyor speed, draft ratio, or clamping cycle time of the current process cycle deviates from the calibration value of the most recent stable operating window, and the deviation reaches the preset recalculation trigger condition, the material transfer delay cycle number is recalculated starting from the current process cycle, and the correlation correction unit regenerates the delay correlation record for subsequent combing head cycle objects. The preset recalculation trigger condition is determined by the equipment installation calibration record and historical quality traceability record, and is version-bound according to the process type or drafting process configuration switch.

[0099] The preset recalculation trigger condition is a parameter used to determine whether material transfer delay needs to be recalculated. When any parameter among conveying speed, draft ratio, and clamping cycle time deviates from the calibration value of the most recent stable operating window within a continuous process cycle, the delay calibration unit reads the parameter deviation boundary from the equipment installation calibration record and historical quality traceability record; the process cycle that reaches the corresponding parameter deviation boundary is taken as the recalculation start cycle. The parameter deviation boundary is saved according to the same process type and the same drafting process configuration.

[0100] Recalculation trigger conditions include speed-up events, speed-down events, stop recovery events, batch change events, draft ratio adjustment events, and situations where the fluctuation range of the conveying speed reaches a preset speed fluctuation threshold over multiple consecutive process cycles. During recalculation, delay association records that have been written and completed for quality traceability maintain their original association relationships, while combing head cycle objects that have not yet reached the quality inspection position establish new delay association records according to the number of material transfer delay cycles after recalculation.

[0101] The boundary of the material transfer delay cycle number is jointly defined by the equipment installation calibration record and historical quality traceability record. When there are missing, out-of-bounds, or abnormal conditions in the upstream material path length, downstream material path length, conveying speed, draft ratio, or clamping cycle time, the delay calibration unit preferentially uses the number of material transfer delay cycles formed during the most recent stable operation phase under the same machine, the same combing head, the same process type, and the same drafting process configuration as a conservative value. If there are no stable operation records under the same process type and the same drafting process configuration, the upper boundary of the material transfer delay cycle number defined by the equipment installation calibration record is used as the conservative value, and the corresponding delay-related record is marked as a conservative-related record. After the abnormality is resolved, the delay calibration unit recalculates the number of material transfer delay cycles according to the new conveying speed, draft ratio, and clamping cycle time.

[0102] The upper and lower boundaries of the number of material transfer delay cycles are implementation parameters determined by the shortest material path, longest material path, minimum conveying speed, maximum conveying speed, draw ratio range, and clamping cycle range in the equipment installation calibration record. When the number of material transfer delay cycles generated by the delay calibration unit is lower than the lower boundary, the lower boundary is used; when it is higher than the upper boundary, the upper boundary is used, and the boundary processing status is written to the associated status identifier.

[0103] With other parameters remaining constant, increasing the conveyor speed reduces the material arrival time and tends to decrease the number of material transfer delay cycles; increasing the draft ratio reduces the arrival time of the material in the later stages and also tends to decrease the number of material transfer delay cycles; when the single-cycle time corresponding to the clamping cycle increases, the number of material transfer delay cycles tends to decrease according to the ratio of material arrival time to single-cycle time. In this way, the correlation between the combing head cycle object and the combined combed sliver quality data can be updated synchronously with changes in process speed.

[0104] Specifically, the units for the length of the front and rear material paths are meters, the unit for the conveying speed is meters per second, the unit for the single-cycle time corresponding to each clamping cycle is seconds, and the draft ratio is a dimensionless quantity. The lengths of the front and rear material paths are obtained from equipment installation calibration records or shutdown calibration records, while the conveying speed, draft ratio, and clamping cycle are recorded by the machine control system within the corresponding process cycle. The original acquired values ​​and calibration source identifiers are retained before the above data enters the delay calibration unit.

[0105] Example 2:

[0106] Based on Embodiment 1, the deep learning early warning module in the multi-source data fusion anomaly early warning system for cotton combing machines provided in this embodiment further includes a visual encoder, a temporal encoder, a heterogeneous graph encoding network, and an anomaly decoding network.

[0107] The visual encoder receives images of the cotton web area or single-head monitoring features and outputs a single-head cotton web representation; the temporal encoder receives single-head cotton web representations of the same combing head under continuous process cycle identifiers and outputs a single-head temporal representation; the heterogeneous graph coding network receives combing head node representations, shared mechanism node representations, merged combing sliver quality node representations, process event node representations, and edge type identifiers and outputs a fused node representation; the anomaly decoding network receives the fused node representations and outputs the anomaly contribution value and anomaly source category.

[0108] When the visual encoder receives images of the cotton web region, it performs region normalization on the images and outputs a fixed-dimensional single-head cotton web representation. When the visual encoder receives single-head monitoring features, it performs feature mapping on the single-head monitoring features and outputs a fixed-dimensional single-head cotton web representation. The time-series encoder reads the single-head cotton web representation of the same combing head according to the continuous process cycle identifier and outputs a single-head time-series representation containing the changes between adjacent cycles. After receiving the fused node representations, the anomaly decoding network generates an anomaly contribution value and an anomaly source category; the anomaly contribution value is used for comparison with the trigger threshold, and the anomaly source category is used to write into the anomaly source field.

[0109] In one implementation, the edge type identifier includes material flow edge identifier, shared mechanism edge identifier, event constraint edge identifier, and time continuous edge identifier; the heterogeneous graph coding network converts the edge type identifier into edge type embedding, and generates edge type weights based on the source node representation, target node representation, and edge type embedding, and aggregates the representations of adjacent nodes according to the edge type weights to obtain the fused node representation of the contribution value of participating anomalies and the anomaly source category output.

[0110] In one example, edge type identifiers are categorized as material flow edges, shared mechanism edges, event constraint edges, and time-continuous edges, and edge type embeddings are vectorized representations mapped from the edge type identifiers. The source node representation is the node representation of the edge's starting node in the current encoding layer, and the target node representation is the node representation of the edge's terminating node in the current encoding layer. The edge type weight is a normalized weight within the range of edges entering the same target node, used to represent the degree of participation of adjacent node representations in the target node aggregation process.

[0111] In a preferred embodiment of Example 2, the heterogeneous graph coding network sets edge type embeddings for material flow edges, shared mechanism edges, event constraint edges, and time-continuous edges. The edge type embeddings corresponding to material flow edges are used to characterize the delayed correlation between the combing head cycle object and the merged combing strip quality data; the edge type embeddings corresponding to shared mechanism edges are used to characterize the common influence of the shared mechanism state on multiple combing heads; the edge type embeddings corresponding to event constraint edges are used to characterize the constraint influence of start-up, stop, can change, lap change, batch change, speed increase, speed decrease, or cleaning and maintenance events on the node state; and the edge type embeddings corresponding to time-continuous edges are used to characterize the state continuity relationship of the same combing head between adjacent process cycles.

[0112] Before generating edge connection scores, the combing head node representation includes deviation features from the relatively stable calibration interval of single-head monitoring features; the merged combing strip quality node representation includes deviation features from the relatively stable calibration interval of quality cycle features; the shared mechanism node representation includes deviation features from the relatively stable calibration interval of shared mechanism features; and the process event node representation includes event labels within the process event influence window. These deviation features and event labels are included in the edge connection score calculation along with the corresponding node representations. In each layer of the heterogeneous graph coding network, the edge connection score is generated using source node representations, target node representations, and edge type embeddings. The edge connection score is calculated as follows:

[0113] ,

[0114] in, For the first Source node in layer coding Meridian type Point to target node The edge connection score; This serves as the identifier for the coding layer in a heterogeneous graph coding network. This serves as the source node identifier. Identify the target node; This is used to identify the edge type, which includes material flow edges, shared mechanism edges, event constraint edges, and time-continuous edges. edge type The corresponding score projection vector; It is the transpose symbol; It is the hyperbolic tangent activation function; edge type The corresponding source node representation mapping matrix; For the source node In the Node representation in layer coding; edge type The corresponding target node representation mapping matrix; For the target node In the Node representation in layer coding; edge type The corresponding edge type embedding mapping matrix; edge type Corresponding edge type embedding; edge type The corresponding score bias vector.

[0115] After edge connection scores are generated, the heterogeneous graph coding network is normalized within the range of incoming edges to the same target node to obtain edge type weights:

[0116] ,

[0117] in, For the first Source node in layer coding Meridian type Point to target node Edge type weight; It is an exponential function; , , , and The aforementioned definition will be used; Identifier for candidate edge types participating in normalization; A set of edge types; For the target node The candidate neighboring source node identifier; For edge type Point to target node The set of adjacent source nodes; For the first Candidate adjacent source nodes in layer coding Candidate edge type Point to target node The edge connection score.

[0118] When the target node When there is at least one incoming edge in the current temporal heterogeneous graph, the heterogeneous graph coding network calculates the edge type weights as described above; when the target node... When there are no incoming edges, the heterogeneous graph coding network does not perform incoming edge normalization calculation and sets the neighboring node aggregation term of the target node to a zero value vector, so that the target node only updates its node representation through the target node self-connection mapping matrix.

[0119] "No incoming edges to the target node" means that no node in the current temporal heterogeneous graph points to the target node through material flow edges, shared mechanism edges, event constraint edges, or time-continuous edges. "Missing data" for an edge type means that an edge has been generated, but the characteristics of the source node, target node, or the record carried by the edge are missing, out of bounds, or in an abnormal state. The former processes the adjacent node aggregation item as a zero-value vector, while the latter generates an edge connection score based on conservative input; these correspond to missing graph connections and missing data content, respectively.

[0120] The edge type weight ranges from 0 to 1. When a target node has at least one incoming edge in the current temporal heterogeneous graph, the sum of the edge type weights of all incoming edges encoded at the same layer for the same target node is 1.

[0121] During the pre-launch verification of model parameter versions, the sample construction unit verifies the direction of change of edge type weights based on the training samples with verification results of abnormal source categories. Among them, the verification samples corresponding to material flow edges include deviation features of the relatively stable calibration interval of the merged combed sliver quality data; the verification samples corresponding to shared mechanism edges include deviation features of the relatively stable calibration interval of the shared mechanism data; the verification samples corresponding to event constraint edges include the process event influence window; and the verification samples corresponding to time continuity edges include the continuous deviation features of the single-head cotton web characterization in adjacent process cycles.

[0122] The variation range of the single-head cotton web here represents the degree of deviation from the relatively stable calibration range of the single-head cotton web representation of the same combing head within adjacent process cycles. This degree of deviation is jointly determined by the edge breakage value, overlap continuity value, local accumulation value, texture fluctuation value, and missing measurement status indicator. The edge type weight of the time-continuous edge is used to characterize the continuity relationship of abnormal states between adjacent process cycles.

[0123] The aforementioned increasing trend serves as a pre-launch verification condition for the new model parameter version, calibrated by the sample construction unit based on training samples with verification results for anomaly source categories. If the edge type weights generated by the new model parameter version during historical verification sample playback are inconsistent with the verification direction of the corresponding edge type, the model update unit will not switch the model parameter version formed in this update and will continue to use the previous effective model parameter version and the previous effective source determination threshold version for early warning calculation. If the number of verification samples corresponding to a certain edge type does not reach the preset lower limit of the number of samples, the direction verification result of that edge type will not be used to increase its source determination priority. The preset lower limit of the number of samples is an implementation parameter used to determine whether a sample participates in statistics, verification, or threshold tuning, and it is recorded separately according to data category, anomaly source category, and model parameter version.

[0124] The heterogeneous graph coding network aggregates the representations of adjacent nodes based on edge type weights to obtain the updated node representation of the target node:

[0125] ,

[0126] in, For the target node In the Node representation in layer coding; Update the activation function for node representation; , , , , , , , and The aforementioned definition will be used; edge type The corresponding aggregated message mapping matrix; This is the self-connection mapping matrix for the target node.

[0127] Before performing adjacency node aggregation, heterogeneous graph coding networks map node representations of different node types to the same representation dimension. When the original dimensions of the source node representation, target node representation, or edge type embedding are different, the dimensions are aligned using the corresponding source node representation mapping matrix, target node representation mapping matrix, edge type embedding mapping matrix, or aggregation message mapping matrix before weighted aggregation.

[0128] To ensure the traceability of anomaly source categories, the heterogeneous graph coding network also calculates the edge type aggregation weights corresponding to the target node:

[0129] ,

[0130] in, For the target node In the corresponding edge type in layer encoding Edge type aggregation weight; when When it is an empty set, Set to 0; , , , , and The aforementioned definition will be used.

[0131] The anomaly decoding network uses the aggregated comb head node representation as input for the anomaly contribution value, and combines edge type aggregation weights to determine the anomaly source category. When the edge type aggregation weight corresponding to the material flow edge has the highest proportion in the abnormal contribution value formation process, and the quality data of the merged combed strip in the delayed association record deviates from the stable calibration range, the abnormal source category output is the merged combed strip source category; when the edge type aggregation weight corresponding to the shared mechanism edge has the highest proportion, and the shared mechanism data deviates from the stable calibration range, the abnormal source category output is the shared mechanism source category; when the edge type aggregation weight corresponding to the event constraint edge has the highest proportion, and the corresponding process cycle identifier is within the process event influence window, the abnormal source category output is the process transition source category; when the edge type aggregation weight corresponding to the time continuous edge has the highest proportion, and the single-head monitoring feature shows edge breakage, discontinuous overlap, local accumulation, local air suction abnormality, cotton roll unwinding abnormality, separation and joining abnormality, web breakage state, or cotton surge state, the abnormal source category output is the single-head cotton web source category; when the edge type aggregation weights corresponding to the material flow edge, shared mechanism edge, event constraint edge, and time continuous edge do not reach the source judgment threshold, and the single-head monitoring feature carried by the combing head node shows the above-mentioned single-head abnormal state, the abnormal source category output is the single-head cotton web source category.

[0132] The anomaly contribution value is the anomaly response quantity corresponding to the combing head cycle object, and its value is determined by the training samples and sample labels. The trigger threshold is an implementation parameter, which is tuned by the output distribution of normal samples and confirmed abnormal samples on the validation set and is bound to the model parameter version. The early warning and handling module compares the anomaly contribution value with the trigger threshold. When the anomaly contribution value is greater than or equal to the trigger threshold, an anomaly early warning result is generated. When the anomaly contribution value does not reach the trigger threshold, the inference record of the corresponding process cycle is retained.

[0133] When the convergence weights of two or more edge types simultaneously reach the source determination threshold, and the difference between the convergence weight of the highest edge type and at least one of the second-highest edge types is within a preset parallel determination range, the anomaly decoding network records multiple existing anomaly source categories participating in the determination in the anomaly source field, and simultaneously records the edge types participating in the determination in the anomaly link field. If the data used to support the convergence weight of the highest edge type is missing, the sensor status is abnormal, or there are conservative association record markers, the anomaly decoding network will not use that edge type as the priority output basis for a single anomaly source category; if a single anomaly source category cannot be formed after processing, the anomaly source field retains the candidate existing anomaly source categories that have reached the source determination threshold, and records the pending review status in the handling field.

[0134] When the number of normal samples under the same machine, the same process type, and the same drawing process configuration does not reach the preset lower limit of the number of samples, the stability calibration interval uses the previous effective threshold version and does not directly replace it with the statistical results under other process types or other drawing process configurations.

[0135] The impact window of a process event is determined by the event start period and event end period recorded by the event access unit. If the event end period is missing, it is truncated according to the preset event valid window. The preset event valid window is the valid period range used when the event end period is missing, and it is determined by historical process event records or equipment control records of the same event category. The source determination threshold is obtained by verifying the abnormal source categories in the training samples and is updated synchronously when the model update unit updates the model parameters. The preset parallel determination range is obtained by verifying the results of multiple abnormal source categories simultaneously participating in the generation of abnormal contribution values ​​in the training samples, and it is managed using the same threshold version as the source determination threshold. The synchronously updated source determination threshold is bound to the corresponding model parameter version, and the same model parameter version and the same source determination threshold version are used in the encoding and abnormal decoding processes of the same time-series heterogeneous graph.

[0136] The inference input after model deployment includes combing head cycle objects within a continuous process cycle, shared mechanism data, merged combing strip quality data, process event data, delay association records, and edge type identifiers. The inference output includes anomaly contribution values, anomaly source categories, anomaly source details, and anomaly link fields. During inference, the model parameter versions, stability calibration interval versions, trigger threshold versions, and source determination threshold versions bound during encoding are used.

[0137] Edge type embedding, score projection vector, node representation mapping matrix, edge type embedding mapping matrix, aggregate message mapping matrix, and target node self-connection mapping matrix are obtained by training the training samples formed by the sample construction unit.

[0138] In one implementation, the deep learning early warning module further includes a sample construction unit and a model update unit. The sample construction unit extracts training samples from historical data according to machine identifier, combing head identifier, and process cycle identifier. The training samples include normal samples, single-head abnormal samples, shared mechanism abnormal samples, process transition samples, and merged combing strip abnormal samples. The sample construction unit generates sample labels based on maintenance records, manual review records, quality traceability records, or shutdown alarm records. The model update unit updates the model parameters of the visual encoder, temporal encoder, heterogeneous graph encoding network, and anomaly decoding network with the training samples and sample labels, and uses the anomaly source category as the training classification target, and the machine identifier, combing head identifier, and process cycle identifier as anomaly location supervision information.

[0139] Training samples are derived from historical single-head monitoring data, historical shared facility data, historical merged combing strip quality data, historical process event data, and historical single-head monitoring features generated from historical single-head monitoring data. Sample labels are formed by anomaly source category, anomaly location field, and review status. Model training objectives include anomaly source category identification, anomaly location supervision, and anomaly contribution value ranking. Anomaly location supervision corresponds to machine identifier, combing head identifier, and process cycle identifier. During model deployment, the multi-source data access module, process cycle alignment module, and time-series heterogeneous graph generation module generate inference input based on online collected data. The deep learning early warning module outputs anomaly contribution value and anomaly source category, and the early warning handling module generates anomaly early warning results based on trigger thresholds. The model update unit generates updated samples based on newly added review records and maintenance records, and performs periodic or triggered updates to the deployed model.

[0140] In one implementation, model training employs a training target category comprised of classification loss, localization loss, and ranking loss. The classification loss corresponds to anomaly source category identification; the localization loss corresponds to the generation of machine identifiers, combing head identifiers, and process cycle identifiers; and the ranking loss corresponds to the ranking relationship of anomaly contribution values ​​among normal samples, process transition samples, and confirmed anomaly samples. After training, the model parameter version is validated through historical review sample playback and then bound to the corresponding stable calibration interval version, trigger threshold version, and source determination threshold version.

[0141] After the model update unit generates a new version of model parameters, it switches to the new version after the current time series heterogeneous graph has completed encoding and anomaly decoding. Time series heterogeneous graphs that have already entered the encoding process continue to use the model parameter version, stable calibration interval version, and source determination threshold version bound when they entered the encoding process. If the new model parameter version fails the historical review sample playback verification, it will not participate in the online early warning calculation.

[0142] During operation, each time a new temporal heterogeneous graph is generated, the heterogeneous graph encoding network recalculates the edge type weights according to the current node representation and the current edge connection relationship; the model parameters are only adjusted when the model update unit performs the update.

[0143] If data is missing, sensor status is abnormal, or process event duration exceeds the preset event validity window for a certain edge type, the edge connection score corresponding to that edge type will use the median value of the edge connection scores formed during the historical stable operation phase of the same edge type, the same source node type, and the same target node type as a conservative input, and participate in the normalization calculation according to the conservative input; after data recovery, the edge connection score corresponding to that edge type will be recalculated using the recovered node representation and edge type embedding.

[0144] The low-confidence flag is a status record in the abnormal link field, used to indicate that the input data for the corresponding edge type does not meet the normal inference conditions. The low-confidence flag includes the edge type identifier, missing data category, conservative input source, and generation process cycle identifier. When an anomaly warning result contains a low-confidence flag, the review status in the handling field is written as pending review, and the comb head node identifier and edge type identifier corresponding to the anomaly warning result are retained.

[0145] If the number of edge connection score samples corresponding to the historical stable operation phase does not reach the preset minimum number of samples, the edge connection score corresponding to the edge type is set to the preset low confidence score, so that the edge type weight after normalization is not used as the priority basis for source determination, and a low confidence mark is recorded in the abnormal link field.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0147] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A deep learning-based multi-source data fusion anomaly early warning system for a cotton combing machine, the cotton combing machine comprising multiple combing heads, wherein the cotton web output from each combing head is combined and drafted to form a combined combed sliver, characterized in that, It includes a multi-source data access module, a process cycle alignment module, a time-series heterogeneous graph generation module, a deep learning early warning module, and an early warning handling module; The multi-source data access module collects single-head monitoring data, shared mechanism data, merged combed sliver quality data, and process event data. The process cycle alignment module generates a process cycle identifier based on the clamping cycle count. When the clamping cycle count is missing, it generates a process cycle identifier based on the driving phase. It then encapsulates the single-head monitoring data into a combing head cycle object by combining the combing head identifier and the process cycle identifier. After the delay calibration trigger event occurs, the process cycle alignment module generates a material transfer delay based on the material path length between the comb head output position and the quality detection position of the combined comb strip, the current conveying speed, the draft ratio, and the clamping cycle count, and generates a delay association record containing the original process cycle identifier and the quality arrival cycle identifier. The time-series heterogeneity diagram generation module generates a time-series heterogeneity diagram based on the combing head cycle object, shared mechanism data, merged combing strip quality data, process event data, and delay association records. The time-series heterogeneity diagram includes at least the combing head node carrying the combing head cycle object, the merged combing strip quality node carrying the quality cycle characteristics, and the material flow edge connecting the combing head node and the merged combing strip quality node according to the delay association records. The deep learning early warning module encodes the temporal heterogeneous graph corresponding to the continuous process cycle identifier, and generates the abnormal contribution value of each combing head cycle object based on the single head monitoring data in the combing head cycle object, the quality cycle characteristics obtained by statistical analysis of the merged combing strip quality data within the time window corresponding to the quality arrival cycle identifier, and the edge relationship of the temporal heterogeneous graph. When the abnormal contribution value is greater than or equal to the trigger threshold, the early warning and handling module outputs an abnormal early warning result that includes the machine identifier, combing head identifier, process cycle identifier, and abnormal source category.

2. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 1, characterized in that, The material path length includes the front material path length from the comb head output position to the drafting inlet, and the rear material path length from the drafting outlet to the quality inspection position of the combined comb strip. The material transfer delay is generated based on the front material path length, the rear material path length, the current conveying speed, the drafting ratio, and the clamping cycle. The delay calibration trigger events include machine initialization, process formula switching, speed increase, speed decrease, and shutdown recovery; the delay association record also includes machine identifier, combing head identifier, combined combing strip quality inspection point identifier, and association status identifier.

3. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 1, characterized in that, The multi-source data access module includes a single-head access unit, a shared mechanism access unit, a quality access unit, and an event access unit; The single-head access unit collects at least one of the following for each combing head: cotton web image, cotton web edge continuity value, cotton web coverage value, cotton roll unwinding value, local suction value, and cotton web breakage or cotton gushing state, and writes it into the machine identifier and combing head identifier. The shared mechanism access unit collects at least one of the following: main motor current, separating roller speed, drafting roller speed, suction main pipe pressure, cotton dropping mechanism status, or machine vibration value; The quality access unit collects and merges at least one of the following: combed sliver quantitative value, sliver fluctuation value, sliver breakage status, or sliver coiling status; The event access unit collects at least one of the following events: start-up, stop-up, change of drum, change of cotton roll, change of batch, speed reduction, speed increase, or cleaning and maintenance.

4. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 1, characterized in that, The process cycle alignment module includes a beat recognition unit and a window encapsulation unit; The cycle recognition unit determines the process cycle identifier from the clamp action signal, cylinder drive signal, separation roller drive signal, or spindle encoding signal; The window encapsulation unit segments the single-head monitoring data, shared mechanism data, and process event data with different sampling frequencies according to the process cycle identifier. It determines the quality arrival time window of the merged combed strip quality data according to the quality arrival cycle identifier, and performs statistics on the merged combed strip quality data that falls within the quality arrival time window to obtain the quality cycle characteristics. Then, it encapsulates the single-head monitoring data under the same machine identifier, the same combing head identifier, and the same process cycle identifier into a combing head cycle object. When multiple sampling points exist within the same process cycle identifier, the window encapsulation unit extracts at least one of the following as a cycle statistic: mean, extreme value, duration, and number of state changes.

5. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 1, characterized in that, The temporal heterogeneous graph generation module includes a node generation unit and an edge generation unit; The node generation unit generates combing head nodes, shared mechanism nodes, merged combing strip quality nodes, and process event nodes. The combing head node carries the combing head cycle object. The edge generation unit generates material flow edges, shared mechanism edges, event constraint edges, and time continuity edges. Material flow edges connect combing head nodes and merged combing strip quality nodes according to material transfer delays. Shared mechanism edges connect shared mechanism nodes and combing head nodes affected by the shared mechanism. Event constraint edges connect process event nodes and nodes affected by the process event. Time continuity edges connect combing head nodes of the same combing head under adjacent process cycle identifiers.

6. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 3, characterized in that, The multi-source data access module also includes a single-head feature generation unit; The single-head feature generation unit performs combing head region cropping on the cotton web image to obtain the cotton web region image corresponding to the combing head, and generates at least one of the following values ​​based on the cotton web region image: edge breakage value, overlap continuity value, local stacking value, or texture fluctuation value. The single-head feature generation unit splices the data generated from the cotton web area image with the cotton roll unwinding value, local air suction value, web breakage status or cotton gushing status to form a single-head monitoring feature.

7. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 5, characterized in that, The deep learning early warning module includes a visual encoder, a temporal encoder, a heterogeneous graph encoding network, and an anomaly decoding network; the visual encoder receives single-head monitoring data carried by the combing head node or single-head monitoring features generated from the single-head monitoring data, and outputs a single-head cotton web representation. The timing encoder receives the single-head web representation of the same combing head under continuous process cycle identifiers and outputs the single-head timing representation; the heterogeneous graph coding network receives the combing head node representation, shared mechanism node representation, merged combing sliver quality node representation, process event node representation and edge type identifier, and outputs the fused node representation. The anomaly decoding network receives the fusion node representations and outputs the anomaly contribution value and anomaly source category.

8. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 7, characterized in that, The edge type identifiers include material flow edge identifiers, shared mechanism edge identifiers, event constraint edge identifiers, and time-continuous edge identifiers. The heterogeneous graph coding network converts edge type identifiers into edge type embeddings, generates edge type weights based on source node representations, target node representations, and edge type embeddings, and aggregates adjacent node representations according to edge type weights to obtain fused node representations that output the contribution value of participating anomalies and the anomaly source category. The deep learning early warning module also includes a process event masking unit. The process event masking unit adds event tags to the corresponding process cycle identifier based on the process event data and converts the event tags into process event node representations. The anomaly decoding network outputs process transition source category, single-head cotton web source category, shared mechanism source category, merged combed sliver source category, or sensor status source category based on the process event node representation. The process transition source category is used to record the early warning trigger source caused by the process event.

9. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 1, characterized in that, The anomaly warning result includes an anomaly location field, an anomaly source field, an anomaly link field, and a handling field; the anomaly location field records the machine identifier, combing head identifier, and process cycle identifier; The anomaly source field includes anomaly source category and anomaly source details. The anomaly source category records at least one of the following: single-head cotton web source category, shared mechanism source category, process transition source category, combined combed sliver source category, or sensor status source category. The anomaly source details record at least one of the following: cotton web edge breakage, discontinuous cotton web overlap, local accumulation, single-head suction anomaly, cotton roll unwinding anomaly, separation and joining anomaly, shared mechanism anomaly, combined combed sliver anomaly, or sensor status anomaly. The Abnormal Link field records the node and edge type that participated in generating the abnormal contribution value; the Handling field records the inspection object, the sending object, the review status, and the maintenance status; when the Abnormal Link field contains quality missing test association, conservative association, or low confidence flag, the Abnormal Source field retains the candidate abnormal source category, and the Handling field records the pending review status.

10. The deep learning-based multi-source data fusion anomaly early warning system for cotton combing machines according to claim 7, characterized in that, The deep learning early warning module also includes a sample construction unit and a model update unit; The sample construction unit extracts training samples from historical data according to machine identifier, combing head identifier and process cycle identifier. The training samples include normal samples, single-head abnormal samples, shared mechanism abnormal samples, process transition samples and merged combing strip abnormal samples. The sample construction unit generates sample labels based on maintenance records, manual review records, quality traceability records, or shutdown alarm records; The model update unit updates the model parameters of the visual encoder, temporal encoder, heterogeneous graph encoding network and anomaly decoding network with training samples and sample labels, and uses the anomaly source category as the training classification target, and the machine identifier, combing head identifier and process cycle identifier as localization supervision information.

Images