Flour equipment fault diagnosis method and system

Abstract

The invention relates to the field of flour equipment fault diagnosis, in particular to a flour equipment fault diagnosis method and system. The method comprises the following steps: constructing a session configuration packet containing a direction constraint edge set, a propagation time delay candidate set and a sensing channel set based on contents such as a component list, a component connection relation and process flow direction information; performing synchronous acquisition, time anchor point alignment and quality marking on multi-source time sequence data to form an aligned session data packet; an evidence chain fragment packet and a stage label sequence are generated through consistency damage detection and change point detection, a multi-domain feature and causal alignment view is constructed, and a causal input tensor packet is formed; and obtaining an abnormal representation packet through a time sequence causal convolutional network, completing traceability analysis in combination with the propagation time delay candidate set on the direction constraint component propagation graph, and generating a life evaluation packet through survival analysis. Unified modeling, causal traceability and life evaluation of flour equipment faults are realized, and continuity and consistency of a diagnosis process are improved.

Description

Technical Field

[0001] This invention relates to the field of flour milling equipment fault diagnosis, and in particular to a method and system for diagnosing flour milling equipment faults. Background Technology

[0002] In the field of flour milling equipment fault diagnosis, existing solutions typically focus on collecting and judging time-series data from the sensor channel set. They combine component lists, component connection relationships, process flow information, power transmission chain information, and sensor capability lists to arrange monitoring points, and complete access and operation management under installation constraints. However, these solutions suffer from limitations such as the session configuration package failing to fully encompass the directional constraint edge set, the propagation delay candidate set failing to unify the association with the sensor channel set, the difficulty in maintaining consistent time anchor alignment and quality labeling in aligned session data packets, and the inability of evidence chain fragment packets to stably cover cross-channel linkage for consistency violation detection and change point detection. Existing methods often rely on the judgment of local fragments of the time-series data and independent channel analysis paths, and lack a comprehensive constraint expression for the coupling relationship between process flow information and power transmission chain information. In scenarios where installation constraints and the sensor channel set are limited, situations arise where, after time anchor alignment, the quality label remains untraceable in aligned session data packets, and candidate fault event fragments and stage label sequences fail to form consistent fragment boundaries between channels. This makes it difficult to meet the stable link requirement of generating a causal aligned view from the evidence chain fragment packet and further forming a causal input tensor packet. Regarding the joint generation of the source tracing result package and the lifetime assessment package driven by the anomaly characterization package, existing technologies generally have common shortcomings in aspects such as the expression of propagation delay candidate set constraints, the maintenance of attribution consistency on the propagation graph of directional constraint components, the organization of survival analysis covariates, and the solidification of the caliber of the lifetime assessment package. It is difficult to form a continuous and consistent process from the session configuration package, the aligned session data package, the evidence chain fragment package, the causal input tensor package, the anomaly characterization package, the source tracing result package to the lifetime assessment package and the disposal strategy package during the operation of flour milling equipment. This leads to deviations in the generation link of the source tracing result package and the lifetime assessment package in terms of cross-stage linkage and cross-channel consistency, which affects the output consistency of the disposal strategy package. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for diagnosing flour milling equipment faults, comprising: Obtain the component list, component connection relationship, process flow information, power transmission chain information, sensor capability list and installation constraint information, and generate a session configuration package containing a set of directional constraint edges, a set of propagation delay candidates and a set of sensor channels; Based on the session configuration package, the time-series data of the sensor channel set is collected synchronously, and time anchor point alignment and quality marking are performed to generate an aligned session data packet; Based on the aligned session data packets, consistency violation detection and change point detection are performed, candidate fault event fragments are extracted and stage label sequences are generated to construct evidence chain fragment packets; Based on the evidence chain fragment package, multi-domain features are constructed and combined with the propagation delay candidate set to generate a causal alignment view, and a causal input tensor package is constructed. Based on the causal input tensor packet, an anomaly representation packet is generated through a temporal causal convolutional network; Based on the anomaly representation package, graph attention reverse attribution of the propagation delay candidate set constraint is performed on the propagation graph of the directional constraint component to generate a source tracing result package. Based on the source tracing result package, a survival analysis is performed to obtain a lifespan assessment package; Based on the lifetime assessment package, a treatment strategy package is generated and output.

[0004] Furthermore, the set of directional constraint edges is determined by the process flow information and the power transmission chain information to establish the directed connection between upstream and downstream components; the set of propagation delay candidates is obtained by calculating the cross-channel response sequence on the normal operating segment of the historical operating data, and a corresponding delay candidate value and its confidence level are recorded for each directional constraint edge.

[0005] Furthermore, the session configuration package includes synchronization trigger parameters and sampling rate stratification parameters; the sensor channel set covers key branching nodes and convergence nodes in the process flow, as well as monitoring points corresponding to key rotating components in the power transmission chain, and configures different sampling rate stratifications for different monitoring points.

[0006] Furthermore, the time anchor alignment includes unified time axis generation, sampling clock drift compensation, and trigger jitter correction; the quality markers include channel health self-check markers, missing test markers, and drift markers, and the alignment session data packet carries the quality markers for subsequent detection and modeling.

[0007] Furthermore, the consistency violation detection includes determining whether the order of upstream and downstream channels deviates based on the set of directional constraint edges, and determining whether the cross-channel delay drifts abnormally based on the set of propagation delay candidates; the stage label sequence includes one or more of the precursor stage, development stage, impact stage and recovery stage.

[0008] Furthermore, the multi-domain features include time-domain statistical features, frequency-domain statistical features, and time-frequency-domain statistical features, and the stage difference features are obtained by differentiating the multi-domain features of adjacent stages; the causal alignment view is obtained by aligning the aggregated features of upstream channels or upstream components according to the propagation delay candidate set and stacking multiple views.

[0009] Furthermore, the temporal causal convolutional network includes causal convolutional layers and dilated convolutional layers, and uses residual connections to output multi-layer temporal abstract features; the anomaly representation package includes anomaly embedding sequences and early warning scores generated from the multi-layer temporal abstract features.

[0010] Furthermore, the graph attention reverse attribution includes calculating edge attention weights on the propagation graph of the directional constraint component and suppressing edges that do not satisfy the propagation delay candidate set; the source tracing result package includes root cause component ranking results, propagation chain confidence, and root cause stability index based on multi-event segment consistency.

[0011] Furthermore, the survival analysis adopts a lifetime modeling approach based on a risk function, using the abnormal embedding sequence in the abnormal characterization package and the root cause component, propagation chain confidence, and operating condition label in the source tracing result package as covariates, outputting a risk curve of future failure risk probability changing over time and a distribution or range of remaining useful lifetime; the lifetime assessment package also includes a closed-loop verification result generated based on the consistency between the risk curve and the root cause component, and is used to drive the generation or update of the treatment strategy package.

[0012] Furthermore, a flour milling equipment fault diagnosis system is characterized by comprising: a multi-sensor synchronous acquisition unit, a data alignment and quality management unit, a propagation prior construction unit, an event evidence chain generation unit, a feature construction and causal alignment unit, a temporal causal convolution representation unit, a directional constraint graph tracing unit, a survival risk and remaining lifetime prediction unit, and a disposal strategy generation unit; the units are connected in sequence to implement the method described in any one of claims 1-9.

[0013] The key innovations of this invention include: (1) The session configuration package is generated based on the component list, component connection relationship, process flow information, power transmission chain information, sensor capability list and installation constraint information. The directional constraint edge set, the propagation delay candidate set and the sensor channel set are simultaneously solidified in the session configuration package in a unified data organization manner, so that the synchronous acquisition of the sensor channel set, the constraint reference of the propagation delay candidate set and the graph structure mapping of the directional constraint edge set are kept consistent and associated in the same session context.

[0014] (2) Based on the alignment session data packet, perform the consistency violation detection and the change point detection, and generate the stage label sequence while extracting candidate fault event fragments, thereby constructing the evidence chain fragment package that can be reused across channels, so that the evidence chain fragment package has the same fragment boundary and the same stage caliber when constructing multi-domain features, generating causal alignment view and forming causal input tensor package in the future.

[0015] (3) After generating the anomaly representation packet based on the causal input tensor packet through a temporal causal convolutional network, the propagation delay candidate set constraint is introduced on the propagation graph of the directional constraint component and graph attention reverse attribution is performed to generate the source tracing result packet. The source tracing result packet is used as the direct input for survival analysis to obtain the lifetime assessment packet, forming a continuous link where the anomaly representation packet drives the source tracing result packet and further drives the lifetime assessment packet.

[0016] The following are its main beneficial effects: (1) In view of the fact that the session configuration package in the background technology is difficult to fully carry the unified association of the direction constraint edge set, the propagation delay candidate set and the sensor channel set and the resulting cross-step reference inconsistency problem, the session configuration package binds the direction constraint edge set, the propagation delay candidate set and the sensor channel set to the same session context, so that the synchronous acquisition based on the session configuration package and the subsequent call to the propagation delay candidate set and the direction constraint edge set have traceable same source configuration basis, thereby reducing the perturbation of the caliber of the alignment session data packet, the evidence chain fragment packet and the causal input tensor packet by cross-step configuration drift.

[0017] (2) In view of the problems in the background technology that it is difficult to maintain the consistency of time anchor alignment and quality label in the alignment session data packets, the difficulty of forming consistent fragment boundaries between candidate fault event fragments and stage label sequences, and the difficulty of the evidence chain fragment package covering cross-channel linkage, by unifying the output of the consistency violation detection and the change point detection into the evidence chain fragment package and carrying the stage label sequence, the candidate fault event fragments have verifiable same-stage expressions between channels, so that the multi-domain feature construction based on the evidence chain fragment package and the causal alignment view generation are consistent in fragment granularity and stage scope, and reduce the causal input tensor packet organization deviation caused by fragment boundary inconsistency.

[0018] (3) To address the issues of insufficient expression of propagation delay candidate set constraints, insufficient maintenance of attribution consistency on the propagation graph of directional constraint components, and discontinuous link in the joint generation of the source tracing result package and the lifetime assessment package in the background technology, the propagation delay candidate set constraints are introduced on the propagation graph of the directional constraint components and graph attention reverse attribution is performed to generate the source tracing result package. Then, the source tracing result package is directly used for survival analysis to obtain the lifetime assessment package. This ensures that the input-output relationship from the anomaly representation package to the source tracing result package and then to the lifetime assessment package is closed within the same link, thereby ensuring that the generation of the lifetime assessment package maintains the same component propagation semantics as the source tracing result package and facilitates continuous docking with the generation of the disposal strategy package. Attached Figure Description

[0019] Figure 1A flowchart illustrating a flour milling equipment fault diagnosis method provided in an embodiment of this application; Figure 2 This is a structural block diagram of a flour milling equipment fault diagnosis system provided in an embodiment of this application. Detailed Implementation

[0020] Example 1: Refer to Figure 1 This is a flowchart illustrating a flour milling equipment fault diagnosis method provided in an embodiment of the present invention. The process may include at least steps S100-S800: S100: Obtain the component list, component connection relationship, process flow information, power transmission chain information, sensor capability list and installation constraint information, and generate a session configuration package containing a set of directional constraint edges, a set of propagation delay candidates and a set of sensing channels; S200. Based on the session configuration package, synchronously collect the time-series data of the sensor channel set, perform time anchor alignment and quality marking, and generate an aligned session data packet; S300. Based on the aligned session data packet, perform consistency violation detection and change point detection, extract candidate fault event fragments and generate stage label sequences to construct evidence chain fragment packets. S400. Based on the evidence chain fragment package, construct multi-domain features and combine them with the propagation delay candidate set to generate a causal alignment view, and construct a causal input tensor package; S500. Based on the causal input tensor packet, an anomaly representation packet is generated through a temporal causal convolutional network. S600. Based on the anomaly representation package, perform graph attention reverse attribution of the propagation delay candidate set constraint on the propagation delay component propagation graph to generate a source tracing result package. S700. Based on the source tracing result package, perform survival analysis to obtain a life assessment package; S800: Based on the lifetime assessment package, generate a treatment strategy package and output it.

[0021] Specifically, in S100, the component list, component connection relationships, process flow information, power transmission chain information, sensor capability list and installation constraint information are obtained, and a session configuration package containing a set of directional constraint edges, a set of propagation delay candidates and a set of sensing channels is generated. Specifically, this step is triggered by the session configuration module before the flour processing equipment is put into operation, after the process is changed, or after the parts are inspected and repaired. The session configuration module reads the parts list, parts connection relationship, process flow information and power transmission chain information from the equipment asset ledger, mechanical assembly drawings, process flow documents, drive transmission diagrams and on-site inspection records, and reads the sensor capability list and installation constraint information from the sensor selection library, installation work orders and on-site spatial mapping records. The component list is defined as a set of diagnosable objects of the flour processing equipment, including at least the fields of component identifier, component type, section to which it belongs, installation location, rated speed or rated power, allowable operating temperature range, and maintenance window; the component connection relationship is defined as a record of structural and media connections between components, including at least the fields of component identifiers at both ends of the connection, connection type, connection orientation, connection strength level, and detachable marking; the process flow information is defined as a record of the flow relationship between materials and airflow in the sections, including at least the fields of upstream section, downstream section, media type, channel identifier, and bypass condition; the power transmission chain information is defined as a record of the transmission path from the drive source to the driven component, including at least the fields of drive source component identifier, transmission component identifier, driven component identifier, number of transmission stages, coupling method, and lubrication method. The sensor capability list is defined as a collection of candidate sensors and their acquisition capabilities, including at least the fields of sensor type, measurement range, upper limit of sampling rate, resolution, interface protocol, power supply method and protection level; the installation constraint information is defined as a record of the space and environmental limitations for sensor installation, including at least the fields of installable surface location, mounting hole or fixture specifications, wiring channel, dust resistance level, vibration resistance level, accessibility limit and safety isolation distance.

[0022] Furthermore, the session configuration module performs a consistency check on the component list and component connection relationships. This consistency check includes checking the uniqueness of component identifiers, the existence of both ends of the connection, the determinability of the connection direction, and the identification of isolated components. When a missing or conflicting component is found, the conflict record is written to the configuration change record and marked as pending manual review. After completing the consistency check, a set of directional constraint edges is generated based on the process flow information and the power transmission chain information. The set of directional constraint edges is defined as a set of directed edges used for propagation inference. The direction is jointly determined by the upstream-to-downstream relationship described by the process flow information and the drive-to-follower relationship described by the power transmission chain information. When the same pair of components has both structural and media connections, the session configuration module generates directional constraint edges according to the connection type priority rule and writes them into the edge type field. In one embodiment, the connection type priority rule is fixed in the order of power transmission chain first, process flow second, and structural support third. When a directional conflict occurs, the conflicting edge is written into the conflict edge buffer along with a conflict reason field for review in subsequent version iterations. The minimum field set of the directional constraint edge set includes edge number, upstream component identifier, downstream component identifier, edge type, edge priority, observability flag, and associated channel candidate field. The observability flag is determined by the installation constraint information and the sensor capability list. Specifically, it is marked as observable when a sensing channel can be deployed on the upstream side, downstream side, or edge body; otherwise, it is marked as weakly observable and the reason for the blind zone is recorded.

[0023] Furthermore, after the set of directional constraint edges is formed, the session configuration module reads historical operating data and extracts normal operating condition segments. This historical operating data originates from the field data server, PLC operating records, and historical alarm archives of the equipment maintenance system. The normal operating condition segments are obtained by jointly filtering operating condition tags and operation logs. Operating condition tags at least include fields for production level, main motor load level, airflow setting level, and feeding status. Within the normal operating condition segments, the session configuration module performs cross-channel response order statistical processing and delay drift statistical processing on the associated channel candidates to obtain a propagation delay candidate set. This propagation delay candidate set is defined as the set of delay candidate values ​​corresponding to each directional constraint edge and its confidence level label. The confidence level label is calculated jointly by sample coverage, channel health self-check records, and delay stability. The channel health self-check records are derived from the quality label field within the historical operating data or the equipment-side self-check reporting field. To adapt to online update scenarios, the session configuration module writes a version number, generation timestamp, and source segment identifier to the propagation delay candidate set. The source segment identifier is associated with the filtering conditions of normal operating condition segments, which facilitates subsequent configuration backtracking. When an event record of component replacement, process line change, or sensor modification occurs, the session configuration module is automatically triggered to recalculate the propagation delay candidate set of the affected edge and write the change difference to the configuration change record.

[0024] Furthermore, the session configuration module determines the sensor channel set based on the directional constraint edge set and the propagation delay candidate set. The sensor channel set is defined as a set of channels used for synchronous acquisition and includes at least the channel number, the corresponding component identifier, the sensor type, the installation location description, the sampling rate level, the synchronous trigger group identifier, and the interface protocol field. The generation process of the sensor channel set includes candidate point mapping, installability determination, and sampling capability matching. Candidate point mapping maps key branching nodes and convergence nodes in the directional constraint edge set as priority placement points, and maps the drive source components, transmission components, and driven components in the power transmission chain information as priority placement points for rotating components. Installability determination filters out unplaceable points based on installation constraint information and generates an installation method field for the remaining points. In one embodiment, the installation method field includes one or more of bolt fixing, clamp fixing, and magnetic base fixing. Sampling capability matching selects sensor types that meet the range and sampling rate upper limit constraints based on the sensor capability list, and binds the interface protocol field to the field data access method. In one embodiment, the field data access method includes a message queue telemetry transport channel and an industrial fieldbus channel. To meet the automation requirements for continuous operation, the session configuration module generates a synchronization trigger group identifier for the sensor channel set. The synchronization trigger group identifier groups channels on the same propagation chain into the same synchronization trigger group and records the trigger source channel and trigger delay tolerance field in the session configuration package. When the trigger source channel is abnormal or the interface is interrupted, the session configuration module switches the synchronization trigger group identifier to a backup trigger source and writes it into the configuration change record. The backup trigger source comes from a channel with a stable channel health self-check record within the same synchronization trigger group.

[0025] In one embodiment, the flour processing equipment includes a feeding mechanism, a mill, a sieving machine, a purifier, a fan, an elevator, and pneumatic piping. The session configuration module reads component identifiers such as the main motor of the mill, the grinding roller bearing housing, the sieving machine base, and the elevator head from the asset ledger and generates a component list. It reads the structural connection between the grinding roller bearing housing and the machine base from the assembly drawing and generates component connection relationships. It reads the process flow information from the feeding mechanism to the mill, from the mill to the sieving machine, and from the sieving machine to the purifier from the process flow document. It reads the power transmission chain information from the main motor to the coupling, the reducer to the grinding roller shaft system from the transmission diagram. Based on this, the session configuration module generates a set of directional constraint edges. Within the normal operating conditions segment of historical operating data, the vibration channels of the grinding roller bearing housing, the main motor current channel, the differential pressure channel of the air transport pipeline, and the vibration channel of the screening machine base are statistically analyzed in sequence to form a candidate set of propagation delays. Subsequently, based on the installation constraint information, suitable installation surfaces are selected near the reserved holes of the grinding roller bearing housing, the reinforcing beam of the screening machine base, and the maintenance port of the air transport pipeline. Based on the sensor capability list, vibration, temperature, current, and differential pressure type sensors are selected and a set of sensor channels is generated. The sensor channel set records the synchronous trigger group identifier and associates it with the field data access method. The field data access method is consistent with the field gateway configuration and is recorded in the session configuration package.

[0026] This step outputs a session configuration package, which includes at least a set of directional constraint edges, a set of propagation delay candidates, and a set of sensor channels, and carries a version number, configuration change records, and a synchronization trigger group identifier. The session configuration package is invoked as input to the S200, allowing the S200 to perform synchronous acquisition based on the set of sensor channels and complete time anchor alignment and quality marking according to the synchronization trigger group identifier. In summary, this step achieves the structured integration of component topology, process flow, power transmission, and acquisition constraints, generates a session configuration package, and establishes a version tracking mechanism, facilitating subsequent steps to conduct synchronous acquisition and evidence chain construction under consistent standards.

[0027] S200. Based on the session configuration package, synchronously collect the time-series data of the sensor channel set, perform time anchor alignment and quality marking, and generate an aligned session data packet; Specifically, this step takes the session configuration package generated in S100 as input. The session configuration package carries at least the sensor channel set, synchronization trigger parameters, sampling rate hierarchical parameters, channel index mapping relationship, assembly rule fragments related to installation constraint information, and configuration version identifier. The multi-sensor synchronous acquisition unit reads the session configuration package to complete the session initialization registration. The multi-sensor synchronous acquisition unit consists of sensor-end acquisition nodes, acquisition gateways, and fieldbus access interfaces. The sensor-end acquisition nodes correspond to each monitoring point in the sensor channel set. The acquisition gateway is used to complete multi-channel synchronous triggering, sampling clock management, data frame encapsulation, and uploading buffering. The fieldbus access interface is used to access the operating condition and operation log signals of the Programmable Logic Controller (PLC) and generate homogeneous timestamps. The synchronous acquisition refers to registering a unified trigger for data acquisition or a unified start time for the sensor channel set under the same session number, according to the synchronous trigger parameters. At the acquisition gateway, the same trigger batch number and the same sampling window number are appended to the multi-channel data frames, thereby forming a traceable, consistent sampling sequence within the session. The time-series data refers to multi-channel sequence data continuously output according to the sampling clock, including one or more of the vibration sequence, current sequence, rotational speed sequence, temperature sequence, pressure sequence, or acoustic sequence corresponding to the sensor channel set. Each sequence is bound to a channel index, sampling rate level, and sampling window number. The relationship between the session configuration package and the sensor channel set has been given in the draft claims. This step reuses this relationship at runtime to complete the acquisition task loading and channel instantiation.

[0028] Furthermore, after session initialization, the acquisition gateway enters a synchronous trigger waiting state. The triggering conditions are limited by the synchronous triggering parameters, which include two implementation paths: hard triggering and soft triggering. In one embodiment, hard triggering involves the acquisition gateway sending a trigger line level protocol and trigger jitter tolerance to each sensor-end acquisition node, and completing batch triggering through a unified trigger line or synchronous pulse distributor. After the trigger arrives, the sensor-end acquisition nodes perform layered sampling on their respective channels according to the sampling rate layering parameters. High sampling rate layer channels generate original high-frequency sequence segments, and low sampling rate layer channels generate down-frequency sequence segments. The acquisition gateway aggregates and encapsulates the segments of different layers according to the sampling window number. In another embodiment, soft triggering involves the acquisition gateway receiving a working condition start-up command from the fieldbus access interface and completing a unified start-up time registration within the gateway. The working condition start-up command comes from the programmable logic controller's running status bit, start / stop edge, or process segment switching signal. After receiving the working condition start-up command, the acquisition gateway performs parallel start-up on the sensor channel set and records the source channel, command sequence number, and session number association relationship of the working condition start-up command for subsequent auditing and traceability. Both types of triggering paths require the acquisition gateway to complete sampling clock synchronization preparation before triggering. Sampling clock synchronization preparation includes aligning and registering the local clock of the acquisition gateway with the external time source, registering the local timing baseline of the sensor-end acquisition node, and writing the time synchronization status flag. The external time source can be provided by one of Precision Time Protocol (PTP), Network Time Protocol (NTP), or fieldbus time. When the time synchronization status flag is abnormal, the acquisition gateway can still enter the acquisition process, but will write the time synchronization abnormality into the quality flag described later, for the S300's consistency violation detection and change point detection to remove or downgrade its use.

[0029] Furthermore, the time anchor alignment is performed on the acquisition gateway side, and includes three processing links: unified timeline generation, sampling clock drift compensation, and trigger jitter correction. Unified timeline generation refers to the acquisition gateway generating a session-level timeline baseline according to the session number, and mapping the timestamps of data frames uploaded by each sensor acquisition node to the session-level timeline baseline. The mapping process uses a double-key alignment of the trigger batch number, sampling window number, and node local timing baseline, thereby forming a multi-channel alignment index on the session-level timeline. Sampling clock drift compensation refers to the acquisition gateway statistically analyzing the time interval stability of adjacent sampling windows within the sliding sampling window range, and writing drift markers for channels that deviate from the registered value of the sampling rate stratification parameter. Simultaneously, interpolation resampling or window boundary truncation and rearrangement are performed on compensable channels. The compensation path is controlled by the compensation strategy field in the session configuration package. Trigger jitter correction refers to the acquisition gateway correcting and registering the arrival time difference of the same batch of triggers according to the trigger jitter tolerance. When the arrival difference exceeds the trigger jitter tolerance, the batch of sampling windows is marked as jitter abnormal, and the original arrival time difference record is retained for subsequent traceability. After the time anchor alignment is completed, the acquisition gateway generates an alignment index table for each sampling window. The alignment index table contains at least the session number, sampling window number, channel index, alignment timestamp, and sampling rate level fields, and is written as the index header of the alignment session data packet.

[0030] Furthermore, the quality markers are executed collaboratively with the data alignment and quality governance unit on the acquisition gateway side. These quality markers include at least a channel health self-check marker, a missing detection marker, and a drift marker, and are carried in the same packet as the alignment session data packet. The channel health self-check markers are generated by the acquisition gateway in two phases: session initialization and sampling window operation. During initialization, connectivity detection, range boundary detection, and noise floor assessment are performed on each of the sensor channels, and the results are written into the channel health self-check markers. During operation, saturation detection, silence detection, and abrupt spike detection are performed on each sampling window, and the channel health self-check markers are updated. When a sensor-end acquisition node is detected to be offline or the upload sequence number is broken, the acquisition gateway writes a missing detection marker to the missing sampling window and retains a missing placeholder record in the alignment index table, enabling the S300 to distinguish between real changes and acquisition gaps when extracting candidate fault event fragments. The drift marker is written by the aforementioned sampling clock drift compensation link. In addition to indicating the drift state, the drift marker also records a drift processing action type field. This action type field indicates whether the channel has undergone interpolation resampling, window boundary truncation and rearrangement, or only registered an uncompensated state within the sampling window. Besides the aforementioned minimum set, the quality marker can also extend to carry timing anomaly markers, jitter anomaly markers, and communication congestion markers. These extended markers are written as preferred fields into the quality header of the aligned session data packet for subsequent training data filtering and online inference weighting strategy invocation.

[0031] In one embodiment, after the flour mill, purifier, and elevator in the flour processing workshop complete the sensor channel set arrangement according to the session configuration package, the acquisition gateway receives the start / stop signals and process section switching signals from the programmable logic controller, and receives the working condition sampling command during the shift start-up phase to enter the synchronous acquisition process. The vibration channel at the bearing housing of the mill mill and the main motor current channel are configured as high sampling rate levels, while the negative pressure channel and ambient temperature channel of the purifier are configured as low sampling rate levels. The acquisition gateway aggregates multi-level sequence segments under the same trigger batch number and generates an alignment index table for each sampling window. At the same time, it writes a missing measurement mark for vibration channels that experience short-term disconnections, writes a drift mark for current channels that experience sampling interval fluctuations, and records the interpolation resampling action type field. When a process section switching occurs during shift operation, the fieldbus access interface writes the process section switching signal into the working condition and operation log and adds a common source timestamp. The acquisition gateway maps this timestamp to a unified time axis and writes it into the alignment session data packet, so that the S300 can subsequently correlate the working condition changes with candidate fault event segments during consistency violation detection. When the session configuration package is updated, the acquisition gateway generates a new session number according to the configuration version identifier and archives the old session closure record. At the same time, the version switching boundary is written into the session boundary field of the alignment index table to form an auditable version evolution trajectory.

[0032] The alignment session data packet is output by the acquisition gateway and falls into the session-level cache and persistent storage. The alignment session data packet contains at least an alignment index table, multi-channel time-series data corresponding to the sensor channel set, and the quality tag. It is specified in the text that it will be used as the input of the alignment session data packet of S300 to perform consistency violation detection and change point detection and extract candidate fault event fragments.

[0033] In summary, the technical effects of this step are as follows: This step completes the alignment registration of multi-channel time-series data under the framework of session-level synchronous triggering and unified timeline mapping, and adds quality markers and version boundary records at the data frame layer, thereby forming the aligned session data packet that can be directly called by subsequent steps.

[0034] S300. Based on the aligned session data packet, perform consistency violation detection and change point detection, extract candidate fault event fragments and generate stage label sequences to construct evidence chain fragment packets. This step follows the alignment session data packet generated by S200. The alignment session data packet is associated with the session configuration packet through a session number. The alignment session data packet contains at least a unified timeline, time-series data of the sensor channel set, and quality marker fields for operating condition and operation log fragments. The event evidence chain generation unit is deployed on the edge industrial computing node or workshop server of the flour processing equipment. After receiving the alignment session data packet, it looks up the session configuration packet by session number, loads the directional constraint edge set, the propagation delay candidate set, the synchronization trigger parameter and the sampling rate hierarchical parameter, and generates a detection context structure. Subsequently, under the constraints of the detection context structure, it completes consistency violation detection, change point detection, candidate fault event fragment extraction, stage label sequence generation, and encapsulates and generates an evidence chain fragment package for the "based on the evidence chain fragment package" step of S400 to call.

[0035] Specifically, the consistency violation detection is carried out around the set of directional constraint edges, which describes the directed connections between components from upstream to downstream and has a component index binding relationship with the set of sensing channels. After the event evidence chain generation unit extracts the upstream channel segment and the downstream channel segment from the alignment session data packet, it first performs quality gating processing. The quality gating processing registers the missing test dense segment, drift continuous segment, or channel self-test abnormal segment as undecidable segment based on the quality tag and writes it into the consistency evidence substructure. The undecidable segment is skipped in subsequent verification, and at the same time, the corresponding channel index, time segment, triggered quality tag type and original sampling segment reference position are registered in the consistency evidence substructure. Subsequently, cross-channel sequence relationship verification is performed on the identifiable segments. The cross-channel sequence relationship verification process calls the propagation delay candidate set, maps the candidate delay values ​​with confidence labels to the current directed connection, and calculates the response start point difference, peak arrival difference, and energy mutation arrival difference within the sliding time window to form the sequence deviation. When the sequence deviation and the candidate delay value deviate continuously and the deviation direction contradicts the direction of the directed connection, the consistency violation trigger point is registered and written into the consistency violation evidence sequence. The consistency violation evidence sequence records the trigger edge identifier, trigger channel pair, deviation type, deviation level, and quality label summary.

[0036] Change point detection revolves around single-channel evolution and multi-channel linkage, with input from the aligned session data packets. In technical solution one, the event evidence chain generation unit constructs statistical streams and generates change scores for each sampling rate level according to the sampling rate hierarchical parameters. The statistical streams include at least mean drift statistics, variance drift statistics, spectral energy shift statistics, and impact component density statistics. Then, the operating condition and operation log segments are mapped to disturbance segments, and suppression mapping is performed on the change scores within the disturbance segments, outputting a change point candidate sequence. At the same time, the "disturbance mapping source" field is registered in the candidate sequence, pointing to the corresponding log segment index, which facilitates subsequent segment interpretation link reference. In technical solution two, the event evidence chain generation unit generates trend inflection point candidates, periodic mutation candidates, and impact mutation candidates at three scales: long window, medium window, and short window, respectively, and merges them according to the time neighborhood to obtain a change point candidate sequence. During the merging stage, quality markers are referenced to avoid outputting candidates in low-confidence segments, and a "confidence level" field is registered in the change point candidate sequence to reflect the missing measurement density and drift density of the time neighborhood where the candidate point is located. In technical solution three, the event evidence chain generation unit synthesizes the component change score of multiple sensor channels under the same component index according to the component aggregation rule, and generates a neighborhood consensus change point candidate sequence by combining the adjacency relationship of the directional constraint edge set. The neighborhood consensus change point candidate sequence records the consensus strength mark of the triggering component and the neighborhood component set, and registers the corresponding channel index mapping for subsequent fragment binding and calling.

[0037] Candidate fault event fragment extraction is performed based on the consistency violation evidence sequence and the change point candidate sequence. The event evidence chain generation unit first performs trigger fusion processing, matching the consistency violation trigger point with the change point candidate point in terms of time neighborhood, and introducing a propagation delay candidate set constraint, requiring that the time difference between the upstream trigger point and the downstream trigger point along the direction constraint edge set falls within the tolerance bandwidth range of the candidate delay value; trigger pairs that do not meet the constraint are registered as conflict trigger pairs and written into the conflict record substructure. The conflict record substructure at least registers the conflict edge identifier, the conflict trigger point pair, the deviation candidate delay value, and the corresponding confidence level flag. Subsequently, the event evidence chain generation unit generates the event start point and event end point according to the start and end times of the trigger point cluster, and expands the pre-buffer window and the post-buffer window according to the fragment expansion parameter field in the session configuration package; when the interval between adjacent event fragments is less than the merging interval parameter, event fragment merging processing is performed and the trigger cluster index is retained to obtain the candidate fault event fragment set. Each candidate fault event fragment is bound to a subset of the trigger channel set, a subset of the trigger direction constraint edges, a subset of the propagation delay candidate set, and a subset of the operating condition and operation log fragments, and carries a quality tag summary, thereby forming a traceable evidence binding relationship under the same fragment number.

[0038] The generation of stage label sequences is carried out within each candidate fault event segment. These stage label sequences include one or more of the following: precursor stage, development stage, impact stage, and recovery stage. The event evidence chain generation unit first generates stage boundaries based on the change score trend, then corrects the stage boundaries by combining the first trigger time of the consistency violation evidence sequence with the peak trigger density time. During boundary correction, the propagation delay candidate set is invoked, and the order in which the stage start point satisfies the directional constraint edge set is verified. When stage boundaries conflict, the conflict is written into the conflict record substructure, triggering a recalculation process. Under the same session number, the recalculation process switches to Technical Solution Two to regenerate the change point candidate sequence and then updates the stage label sequence. For densely missing test segments, stage label sequence generation introduces a missing test compensation rule. This rule infers the stage start and end times based on the neighborhood consensus change point candidate sequence and writes the compensation source into the source marker field of the stage label sequence, thus ensuring the stage label sequence maintains an executable running link even when quality markers are abnormal.

[0039] In one embodiment, the flour processing equipment includes a mill, a sieving machine, and a pneumatic conveying pipeline. The sensor channel set covers the main motor current channel, the mill roller bearing vibration channel, the sieving machine eccentric mechanism vibration channel, and the pipeline pressure difference channel. After the aligned session data packets continuously arrive at the edge industrial computing node, the event evidence chain generation unit synchronously performs consistency violation detection and change point detection within each sliding time window. When a candidate impact mutation occurs in the mill roller bearing vibration channel and a variance drift mutation occurs in the main motor current channel, and at the same time, a consistency violation trigger point occurs in the directed connection corresponding to the power transmission chain and falls within the tolerance bandwidth range of the propagation delay candidate set, the fusion link is triggered to generate candidate fault event fragments and generate a stage label sequence containing the precursor stage and the development stage. After extracting the candidate fault event fragment set and generating the same-stage label sequence, the event evidence chain generation unit encapsulates the candidate fault event fragment set, stage label sequence, consistency violation evidence sequence, change point candidate sequence, conflict record substructure, quality mark summary, session number, fragment number, channel index mapping, and component index mapping into an evidence chain fragment package. Within the evidence chain fragment package, the subsequent input position S400 is registered as the step of "based on the evidence chain fragment package, constructing multi-domain features and combining them with the propagation delay candidate set to generate a causal alignment view, and constructing a causal input tensor package." The technical effect of this step is that the aligned session data packet, after consistency violation detection and change point detection, is structured into an evidence chain fragment package containing candidate fault event fragments and stage label sequences, forming traceable evidence fields and fragment boundaries.

[0040] S400. Based on the evidence chain fragment package, construct multi-domain features and combine them with the propagation delay candidate set to generate a causal alignment view, and construct a causal input tensor package; Specifically, S400 is executed by the feature construction and causal alignment unit, and its input sources include: the evidence chain fragment package output by S300, and the propagation delay candidate set generated by S100 and carried through S200; wherein, the evidence chain fragment package at least encapsulates the session identifier, fault event fragment boundary, the stage label sequence, multi-channel aligned timing data fragments within the fragment, and the corresponding quality markers, and the propagation delay candidate set at least encapsulates the delay candidate values ​​and confidence markers indexed by the direction constraint edge identifier, thereby providing a callable propagation prior for cross-channel sequence relationships and delay drift. During runtime, after receiving the evidence chain fragment package, the feature construction and causal alignment unit first segments the fragment into stages based on the stage label sequence and establishes a stage context record for each stage. The stage context record includes at least the stage start and end time anchors, the sampling rate hierarchy mapping within the stage, the channel validity status, and a summary of the missing test distribution. The channel validity status is obtained by parsing the quality markers. For channels with abnormal channel health self-check markers or drift markers, one of three strategies—channel deweighting, window removal, or fragment interpolation—is used for processing. The processing strategy and triggering conditions are written into the stage audit field for subsequent cross-step traceability. After the stage segmentation and quality governance are completed, the feature construction and causal alignment unit generates a stage index table to constrain the window range of subsequent multi-domain feature calculations. The stage index table is bound to the evidence chain fragment package as part of the minimum required input set within this step.

[0041] Furthermore, the multi-domain features include at least time-domain statistical features, frequency-domain statistical features, and time-frequency-domain statistical features, defined as a set of reproducible statistical descriptions for each sensing channel and channel pair within the time window defined by the stage index table; wherein, the time-domain statistical features include amplitude distribution summary, trend summary, impulse summary, and steady-state summary; the amplitude distribution summary is obtained from the quantile statistics and amplitude dispersion statistics of samples within the window; the trend summary is obtained from the residual statistics and monotonicity statistics of piecewise linear fitting; the impulse summary is obtained from the peak density, peak spacing distribution, and kurtosis class statistics; and the steady-state summary is obtained from the proportion of variance stable intervals and autocorrelation decay summary within the window; the frequency-domain statistical features are obtained by performing spectral estimation on the data within the stage window, and the spectral estimation adopts one of two implementations: Discrete Fourier Transform (DFT) or power spectral density estimation, and the output includes at least the position of the main frequency peak, the energy proportion of the main frequency peak, the harmonic cluster energy summary, and the bandwidth expansion summary; the time-frequency-domain statistical features adopt Short-Time Fourier Transform (SFT). The output includes at least a time-frequency energy ridge summary, an energy mutation zone location summary, and a time-frequency sparsity summary. For multi-domain features at the channel pair level, the feature construction and causal alignment unit, combined with the propagation delay candidate set, constructs a cross-channel response sequence description for the upstream and downstream channels corresponding to the direction constraint edges. This description includes at least a phase difference mutation summary, a cross-channel cross-correlation peak location summary, and a delay drift summary. The delay drift summary is obtained by aligning and comparing the cross-correlation peak location within the stage with the candidate values ​​in the propagation delay candidate set, and recording the comparison residual as a cross-channel delay deviation field. After completing the multi-domain feature calculation within the stage, the feature construction and causal alignment unit generates stage difference features between adjacent stages. These stage difference features are obtained by differentiating or ratioming the corresponding multi-domain features of adjacent stages, and a change rate summary is appended at the stage transition point, thus forming an inter-stage dynamic description that can be directly consumed by the subsequent characterization network. Subsequently, the feature construction and causal alignment unit aggregates the multi-channel features according to the component index. The aggregation follows the binding relationship between the component and the sensing channel, and performs one of three strategies on the multi-channel features within the same component: weighted aggregation, maximum response aggregation, or robust statistical aggregation. The strategy selection is constrained by both the sampling rate level and the quality label, and the strategy identifier is written into the component aggregation metadata field to form a component aggregation feature package.

[0042] Understandably, the causal alignment view is a set of multi-view representations obtained by aligning the component aggregated feature package along the propagation prior with time delay. Its generation process is executed by the feature construction and causal alignment unit at the stage granularity: for each directional constraint edge, the candidate time delay value and confidence mark of the propagation time delay candidate set are read, and the upstream component aggregated features are translated and aligned according to the candidate time delay. The translation alignment includes one of two implementations: integer sampling point translation and fractional sampling point interpolation translation. When the confidence mark is lower than the threshold or the stage window length is insufficient, the feature construction and causal alignment unit triggers a candidate degradation rule, marking the directional constraint edge as a low-confidence alignment view and replacing it with zero translation or neighborhood mean translation, while writing the triggering reason in the alignment anomaly record field. After alignment, the feature construction and causal alignment unit outputs at least two types of views for each stage: an upstream alignment view and a downstream in-situ view, and forms a multi-view stack for different candidate time delays, thereby obtaining the causal alignment view. Based on the causal alignment view, the feature construction and causal alignment unit constructs the causal input tensor package. The causal input tensor package includes at least three parts: tensor data body, tensor axis semantic mapping, and version record. The tensor data body is jointly indexed by the stage dimension, component dimension, feature dimension, and view dimension. The tensor axis semantic mapping records the alignment relationship between stage labels, component identifiers, feature names, and view numbers. The version record includes at least multi-domain feature caliber versions and propagation delay candidate set versions. The causal input tensor package serves as the direct input to the S500's causal input tensor package, and the tensor axis semantic mapping is synchronously written to the audit log for the propagation graph tracing results of the S600 to be back-referenced with the feature source of this step. In summary, this step achieves the following technical effect: by performing staged feature construction, stage difference generation, component index aggregation, and combining the propagation delay candidate set on the evidence chain fragment package to complete multi-view alignment, a structured causal input tensor package with auditable version records is formed and continuously invoked by subsequent steps.

[0043] S500. Based on the causal input tensor packet, an anomaly representation packet is generated through a temporal causal convolutional network. Specifically, the causal input tensor packet output by S400 is used as the direct input for this step. The causal input tensor packet is formed by aligning the multi-domain features within the evidence chain fragment packet after being constrained by the propagation delay candidate set. The packet also solidifies the mapping relationship between stage boundaries and component indices, thereby ensuring that the feature fragments corresponding to each sensor channel set have computable adjacency and traceability within the same time anchor framework. Further, the temporal causal convolutional representation unit reads the tensor body data from the causal input tensor packet and simultaneously reads the stage label sequence index, component index mapping, quality marker mapping, and causal alignment view index. The stage label sequence index is used as a segmentation gating condition to perform intra-stage slicing and cross-stage splicing order verification on the tensor body data. When a missing or drifting marker is detected in the quality marker mapping, the corresponding channel fragment is written into a quality mask and weight decay or fragment skipping is performed in subsequent convolution calculations. Simultaneously, this process is recorded as an inference audit record and bound to the current session identifier and network version identifier. The inference audit record is output along with the anomaly representation packet for consistency tracing in subsequent main steps.

[0044] Furthermore, the temporal causal convolutional network, as the core computational carrier of this step, includes at least causal convolutional layers and dilated convolutional layers, forming multiple temporal abstract feature channels through residual connections. The causal convolutional layer refers to a convolutional operation structure where the output at any given time point depends only on the current and historical time segments. When the temporal causal convolutional representation unit performs computation on this layer, it first locks the available time windows within a stage based on the stage label sequence index, and then performs hole-filling or hole propagation on unavailable sampling points according to the quality mask to avoid introducing future segments across stage boundaries. The dilated convolutional layer refers to a convolutional structure that inserts intervals between convolutional sampling points. The temporal causal convolutional representation unit expands the receptive field within this layer according to a preset dilation interval sequence, and uses the causal alignment view index as a multi-view channel routing condition, enabling multi-channel alignment features under the same component index mapping to enter subsequent abstract layers in parallel at several receptive field scales. Residual connections refer to the connection method that performs same-dimensional superposition or mapping superposition of the input features of the previous layer and the output features of the current layer. The temporal causal convolutional representation unit writes residual path markers and records dimension mapping strategies at the residual connections to facilitate the location of dimension inconsistencies or gradient anomalies that occur during network operation. Understandably, the minimum set of parameters essential for realizing the core improvement of this invention includes at least the configuration of convolutional kernel width, dilation interval sequence, network layer depth, feature channel dimension, residual mapping strategy, and stage gating strategy. On this basis, the extended functions can preferably include multi-scale branch convergence, stage attention convergence, and anomaly threshold post-processing modules, but these extended functions are disabled by default and do not affect the generation chain of the anomaly representation package. The temporal causal convolutional representation unit adopts a combination of event-driven and period-driven strategies at the operation scheduling level: when the evidence chain fragment package constructed by S300 enters the "stage closure" state or reaches the session cycle boundary, the inference task is triggered, and the snapshot of the causal input tensor package and the network version identifier are frozen at the trigger. After the inference is completed, the inference audit record is written to the session log to form an auditable version evolution trajectory.

[0045] Furthermore, after the temporal causal convolutional representation unit completes the extraction of multi-layer temporal abstract features, it generates the anomaly representation package as the sole output of this step. The anomaly representation package contains at least an anomaly embedding sequence field and an early warning score field. The anomaly embedding sequence field refers to the embedding sequence obtained by encoding the multi-layer temporal abstract features in chronological order within a stage. The early warning score field refers to the stage-level score sequence formed after performing stage convergence and confidence mapping on the anomaly embedding sequence field. Simultaneously, the anomaly representation package further encapsulates an anomaly evidence summary field. The anomaly evidence summary field contains an anomaly component candidate set, stage label sequence references, and quality mask references, which are used by S600 to execute graph attention reverse attribution and generate a source tracing result package. The anomaly embedding sequence field, along with the operating condition labels, is then used by S700 to execute survival analysis to obtain a life assessment package. In one engineering embodiment, the flour processing equipment includes several key rotating and conveying components. The sensor channel set covers motor current channels, bearing vibration channels, shell temperature channels, and wind pressure channels. When the evidence chain fragment package output by S300 shows signs of consistency disruption in both the precursor and development stages, the temporal causal convolutional representation unit on the edge computing node reads the snapshot of the causal input tensor package and triggers an inference task. First, it generates a local fragment of the abnormal embedding sequence field within the precursor stage window, and then completes the cross-scale abstraction of the dilated convolution layer within the development stage window, writing the stage-level early warning score field into the abnormal representation package. If a missing measurement marker is detected in a vibration channel during inference, the quality mask reference is synchronously written into the abnormal evidence summary field, enabling S600 to perform edge weight suppression in conjunction with quality information during graph attention reverse attribution. Finally, the abnormal representation package carries the inference audit record, network version identifier, and stage label sequence reference and is consumed as a common input of S600 and S700, forming a cross-main step connection link. In summary, the technical effect of this step is as follows: the temporal causal convolutional network completes the multi-layer temporal abstract representation of the causal input tensor package, and the stage-level anomaly information is solidified in the anomaly representation package for continuous use in subsequent main steps.

[0046] S600. Based on the anomaly representation package, perform graph attention reverse attribution of the propagation delay candidate set constraint on the propagation delay component propagation graph to generate a source tracing result package. Specifically, in this step, the directional constraint graph tracing unit calls the anomaly representation package output from the previous main step as input. The anomaly representation package is a structured payload formed after processing by the temporal causal convolutional network, which at least includes anomaly embedding sequences and early warning scores bound to the fault event segments and their stage label sequences. During the input loading stage, the directional constraint graph tracing unit reads the set of directional constraint edges and the propagation delay candidate set that are consistent with the component index binding relationship from the session configuration package, and further reads the quality tag reference carried in the aligned session data packet, which is used to perform masked loading of the effective interval, missing interval, and drift interval of the anomaly embedding sequence, thereby forming an anomaly payload input view for graph-oriented computation. The anomaly payload input view retains the fault event segment identifier, stage label sequence reference, and channel-to-component mapping reference within the segment, so that the subsequent edge attention calculation can perform consistent index switching between component granularity and channel granularity, and maintain the same time axis caliber as the constraint logic of the propagation delay candidate set.

[0047] Furthermore, the directional constraint component propagation graph is instantiated and constructed from the directional constraint edge set in this step. The construction process uses the component list as the complete set of nodes and solidifies each directed connection in the directional constraint edge set as a propagation edge record. The propagation edge record carries at least an upstream component identifier, a downstream component identifier, an edge source type flag, and an edge version flag. The edge source type flag is used to distinguish between material direction constraints from the process flow information and power direction constraints from the power transmission chain information, thereby differentiating the propagability of edges from different sources during graph computation. For cases with conflicting direction descriptions or closed loops, this step performs loop detection and conflict registration processing during the graph construction phase, writing conflicting edges into conflict registration records and writing loop flags for related edges, thereby reducing the weight or isolating edges with loop flags during subsequent attention computation. The corresponding graph version flag and conflict registration record are written together into the traceability metadata area of ​​this step for subsequent cross-segment recalculation and audit reproduction.

[0048] After completing the construction of the directional constraint component propagation graph, this step proceeds to the graph attention calculation stage. The graph attention reverse attribution uses edge attention weights as the core intermediate product. These edge attention weights represent the contribution of each propagation edge to the current fault event segment and its stage. They are generated by the component anomaly representation obtained by aggregating the anomaly embedding sequence through component indexing, and the structural relationship between the two ends of the propagation edge. Specifically, the directional constraint graph tracing unit first aggregates the anomaly embedding sequence according to the component index based on the sensor channel set and component mapping relationship in the session configuration package, obtaining a component anomaly sequence. Then, it segments the component anomaly sequence into intra-stage sequence segments according to the stage label sequence. Subsequently, within each stage, the component anomaly sequence segments at both ends of the propagation edge are relatively matched to generate candidate edge attention weights. The calculation of the candidate edge attention weights incorporates the propagation delay candidate set for consistency constraints. Specifically, for each propagation edge, its corresponding delay candidate value and confidence level are read, and the observation delay offset is extracted based on the sequential relationship of sequence segments within the stage. When the deviation between the observation delay offset and the delay candidate value exceeds the allowable range of the confidence level, this step writes the edge into an edge suppression flag and performs suppression processing on its edge attention weight, thereby forming a constrained attention weight record that satisfies the requirement of "calculating edge attention weights on the propagation graph of the direction constraint component and suppressing edges that do not satisfy the propagation delay candidate set." In one implementation, the suppression processing uses an intra-stage consistency gating method, where suppressed edges are directly set to non-propagable within the same stage, and the suppression reason code is recorded. In another implementation, the suppression processing uses a cross-stage soft suppression method, where deviating edges are strongly suppressed in the precursor and development stages, while limited propagation weights are retained and stage difference suppression flags are recorded in the impact and recovery stages, facilitating subsequent root cause stability and stage sequence verification.

[0049] After obtaining the constrained attention weight records, this step performs graph attention back attribution to generate a source tracing result package. The graph attention back attribution is an attribution process that traces back from the downstream component with significant anomalies along the set of reverse edges in the propagation graph of the constrained component in the direction to the upstream component. Its inputs are the anomaly representation package, the propagation graph of the constrained component in the direction, the propagation delay candidate set, and the constrained attention weight records. Its outputs are the root cause component ranking results and the propagation chain confidence, and further outputs a root cause stability index based on the consistency of multiple event segments. Specifically, this step first locates the anomaly initiation stage and the anomaly peak stage within the fault event segment based on the early warning score, and selects significant anomalies as the attribution starting point set within the corresponding stages. Subsequently, it traces back layer by layer along the back propagation direction, accumulating the edge attention weights of the edges traversed by each candidate backtracking path, and performing path reduction or path confidence decay processing on paths containing edge suppression labels to obtain the candidate root cause component set and its ranking score. Furthermore, to meet the stability requirements across fault event segments, this step repeats the above attribution process on multiple fault event segments and aggregates the candidate root cause component sets obtained from each segment according to the component index to form a root cause stability index. At the same time, this step uses the stage label sequence to verify the stage sequence of the root cause, that is, to statistically analyze the leading and persistent attribution scores of candidate root cause components in the precursor or development stage, and writes the verification results into the component factors of the propagation chain confidence, thereby obtaining the converged root cause components and propagation chain confidence. In an engineering embodiment, the flour processing equipment includes a crushing and sieving linkage section, a pneumatic conveying section, and a power drive section. The component list includes at least a grinding roller assembly, a bearing assembly, a belt drive assembly, a fan assembly, and a sieving assembly. The sensor channel set includes at least a bearing vibration channel, a motor current channel, a wind pressure channel, and a temperature channel. When the anomaly characterization package shows that a certain fault event segment exhibits an increase in the abnormal embedding sequence of the vibration channel during the development stage, accompanied by an increase in the early warning score of the motor current channel, this step performs reverse attribution on the directional constraint component propagation map using the sieving assembly and the fan assembly as the starting set of the significantly abnormal components. It also performs time delay consistency screening on propagation edges such as "bearing assembly to grinding roller assembly" and "grinding roller assembly to fan assembly" in conjunction with the propagation time delay candidate set. For propagation edges with large deviations between the observation sequence and the time delay candidate value, an edge suppression mark is written. The edge attention weights of the remaining propagation edges are accumulated to form a backtracking path score, thereby outputting the root cause component ranking result with the bearing assembly as the first component. The corresponding propagation chain confidence and root cause stability index are simultaneously written into the source tracing result package for subsequent life assessment.

[0050] Finally, this step encapsulates the intermediate and final products of the source tracing calculation into the source tracing result package and outputs it. The source tracing result package includes at least the root cause component ranking result, propagation chain confidence, and root cause stability index. It also writes the graph version marker of the directional constraint component propagation graph, the propagation delay candidate set version marker, the anomaly characterization package reference marker, and the fault event fragment identifier and stage label sequence reference into the traceability metadata area, for use as covariate inputs in the survival analysis stage. Specifically, the root cause component ranking result and propagation chain confidence are loaded together with the anomaly embedding sequence as covariate inputs to the survival analysis model in subsequent S700. The root cause stability index is used to gate the root cause consistency across fragments, thereby driving the life assessment package to selectively load risk modeling loads under different operating condition labels. In cases where the root cause stability index is insufficient or there is a conflict between the propagation chain confidence and the stage sequence verification, this step will write the conflict registration record into the rollback trigger record. This record will be used by subsequent processes when re-evaluating the propagation delay candidate set or recalculating the edge attention weights, thereby achieving automated recalculation and version traceability without changing the order of the main process steps.

[0051] In summary, the technical effects of this step are as follows: by introducing the propagation delay candidate set constraint onto the propagation graph of the directional constraint component and performing graph attention reverse attribution, a structured root cause component ranking result, propagation chain confidence, and root cause stability index are formed, and these are encapsulated into the source tracing result package for subsequent survival analysis and disposal strategy generation steps.

[0052] S700. Based on the source tracing result package, perform survival analysis to obtain a life assessment package; In this step, the source tracing result packet output by S600 is used as input, and is received and parsed by the survival risk and remaining life prediction unit in the flour equipment fault diagnosis system. The relevant fields constitute the step connection relationship consistent with the draft claims. The source tracing result packet includes at least the root cause component ranking result, propagation chain confidence, and root cause stability index, and may include the association record of the abnormal embedding sequence and early warning score corresponding to the same fault event segment; furthermore, the source tracing result packet and the alignment session data packet generated by S200 share the same session identifier, which is used to locate its source channel and verify the time anchor alignment status. The survival risk and remaining life prediction unit is triggered to run after detecting the source tracing result packet writing completion mark. Preferably, the first round of inference is completed on the device edge controller side using a streaming access method; when the offline training window update trigger condition is met, the relevant historical running data is included in the model update queue to form a version record for audit query.

[0053] Specifically, the survival risk and remaining lifetime prediction unit first performs field completeness and consistency checks on the source tracing result package. This includes verifying the mapping between the component identifier in the root cause component sorting results and the component list, verifying the edge direction matching between the propagation chain topology in the propagation chain confidence and the propagation graph of the directional constraint components, and recalculating and verifying the statistical caliber of the root cause stability index on multiple fault event segments. When a field is found to be missing or conflicting, the cause of the anomaly is recorded and the source tracing result package is marked as a low-confidence input for downgrade reference in the subsequent handling strategy package generation stage. Subsequently, the time anchor point of the survival event is determined by combining the candidate fault event segments and the stage label sequence. The survival event is an event marker indicating that the equipment or target component has failed to enter the impact stage or recovery stage. The survival time is obtained by slicing the unified time axis of the aligned session data packets. When no survival event occurs before the end of the observation window, the sample is marked as a censored sample and the censoring time is recorded. The censored sample and the event sample together constitute the training input set and inference input set of the survival analysis model.

[0054] Furthermore, the survival risk and remaining life prediction unit constructs covariate vectors from the root cause component ranking results, propagation chain confidence, root cause stability index, and operating condition labels in the source tracing result package, and aggregates the abnormal embedding sequences according to the stage label sequences to generate stage-level abnormal covariates. The root cause component ranking results provide discrete covariate encoding at the component level, the propagation chain confidence provides continuous covariate encoding at the propagation chain level, the root cause stability index provides robustness covariate encoding at the cross-segment consistency level, and the operating condition labels provide conditional covariate encoding at the operating condition stratification level. The survival analysis model adopts a risk function-driven life modeling approach. The model structure includes at least a baseline risk modeling submodule, a covariate mapping submodule, and a time window sampling submodule. The baseline risk modeling submodule represents the basic failure risk pattern under covariate-free conditions, the covariate mapping submodule maps the covariate vector to modulation coefficients of the risk pattern, and the time window sampling submodule extracts training and inference windows from a unified time axis and generates time window index records. In technical solution one, the baseline risk modeling submodule uses piecewise constant risk representation and combines it with the maximum likelihood criterion to complete parameter estimation; in technical solution two, the baseline risk modeling submodule uses accelerated failure time frame to complete lifetime distribution fitting; in technical solution three, the covariate mapping submodule uses tree model risk stratification to complete fitting of nonlinear covariate effects; all three solutions record the training window range, input field scope and parameter group summary when the model version is updated, and generate a new version marker and freeze the old version for backtracking when a change in data scope or a new value range of operating condition label is detected.

[0055] During the inference phase, the survival risk and remaining useful life prediction unit inputs the covariate vector corresponding to the current source tracing result package into the survival analysis model, outputs a risk curve showing the change of future failure risk probability over time, and calculates the remaining useful life distribution or interval based on the risk curve. The risk curve is a probability sequence organized by discrete time points along a unified time axis, and the remaining useful life distribution or interval is a time range record under a preset confidence level. The survival risk and remaining useful life prediction unit simultaneously generates closed-loop verification results. These results include at least component consistency verification flags for the risk curve and root cause component ranking results, and link consistency verification flags for the risk curve and propagation chain confidence levels. If verification fails, a rollback trigger flag is written, which is used by the handling strategy package generation phase to determine whether to call the recalculation process or adopt a conservative template. Finally, the risk curve, remaining useful life distribution or interval, and closed-loop verification results are encapsulated into a life assessment package, which is written to the interface buffer for S800 to read. Simultaneously, the life assessment package records snapshots of the associated root cause component ranking results and propagation chain confidence levels to maintain cross-step continuity.

[0056] In one embodiment, the flour processing equipment is in continuous production. The sensor channel set covers the vibration, temperature, and current channels of key rotating components, and the aligned session data packets form a continuous observation sequence on a unified time axis. When a fault event segment shows an increase in early warning score in the precursor stage and a continuous shift in abnormal embedding sequence in the development stage, the source tracing result package output by S600 gives the root cause component ranking result as a bearing component in the transmission chain, and the propagation chain confidence is concentrated on the direction of propagation from the bearing to the coupling. After reading the source tracing result package, the survival risk and remaining useful life prediction unit incorporates the root cause stability index and the operating condition label into the covariate vector, and outputs the risk curve and the remaining useful life distribution or interval in the inference window. When the closed-loop verification result shows that the risk curve suddenly rises in a short period of time and the root cause stability index decreases synchronously, a rollback trigger flag is written and the version record of the life assessment package is retained, so that S800 can select a maintenance handling template that includes re-inspection and load reduction actions when generating the handling strategy package.

[0057] Summary of the technical effects of this step: This step completes the survival analysis modeling, inference and encapsulation process around the traceability result package, forms a lifetime assessment package and establishes data connection with the S800 disposal strategy package generation process, and forms a traceable operation closed loop through closed-loop verification results and rollback trigger flags.

[0058] S800: Based on the life assessment package, generate and output a treatment strategy package; Specifically, the life assessment package output by S700 is used as input for this step. The disposal strategy generation unit accesses and parses the field structure of the life assessment package. The life assessment package carries at least a risk curve, the distribution or range of remaining useful life, and the closed-loop verification result. It may also carry reference identifiers of root cause components, propagation chain confidence, and root cause stability indicators associated with the traceability result package, thereby establishing a traceable association of "lifespan - traceability - operating condition" within the same session. During the access phase, the disposal strategy generation unit performs version record loading processing, writing the corresponding model version identifier, session identifier, and operating condition label from the life assessment package into the strategy generation context. It also reads the set of template primary keys that match the flour processing equipment model, component category, and process stage from the maintenance disposal template library. The maintenance disposal template library is a structured strategy knowledge base, containing template primary keys, applicable component types, applicable operating condition labels, executable action sequence fields, required spare parts fields, personnel qualification fields, shutdown window constraint fields, and feedback acceptance fields. For the minimum set of parameters required for template matching, the disposal strategy generation unit shall at least take the remaining useful lifetime distribution or interval, risk curve and closed-loop verification results from the lifetime assessment package, and filter them in combination with the applicable condition expressions in the template library; root cause components, propagation chain confidence and root cause stability index are preferred inputs, used to sort and resolve conflicts when multiple templates coexist.

[0059] During template screening and strategy instantiation, the disposal strategy generation unit performs risk window extraction on the risk curve, projects the curve into a risk window sequence according to a preset time granularity, and maps the remaining useful life distribution or interval into a maintenance timing constraint field. The maintenance timing constraint field includes at least a suggested execution window, a latest execution window, and a window confidence flag. Further, the disposal strategy generation unit performs consistency gating on the closed-loop verification results. When a conflict flag or non-convergence flag appears in the closed-loop verification result, the disposal strategy generation unit switches the strategy generation mode to "conservative mode." Without changing the main process step order, it elevates the execution conditions of high-risk shutdown actions in the template action sequence field to manual review triggering and records the review reason in the strategy generation audit field. When the closed-loop verification result is marked as passed, the disposal strategy generation unit switches the strategy generation mode to "automatic mode" and generates a directly deployable disposal action record according to the template action sequence field. Understandably, the action record adopts a work order-based structure, including action number, action type, target component reference, action preconditions, execution time estimate, required spare parts reference, acceptance feedback field, and execution responsibility role. The action preconditions are generated jointly by the risk window sequence and maintenance timing constraint field, and the required spare parts reference is generated by combining the required spare parts field of the template with the inventory status of the equipment asset file. For the optimal implementation involving propagation chain confidence and root cause stability indicators, the disposal strategy generation unit uses the root cause component as the primary target component and adjacent components on the propagation chain as linked inspection components, inserting linked inspection action records into the action sequence. When the propagation chain confidence is lower than a preset consistency threshold, the disposal strategy generation unit does not change the reference identifier of the root cause component, but converges the coverage of the linked inspection action records to key node components, and records the convergence basis as a strategy explanation field, thereby maintaining the auditable evolution of the strategy structure under different confidence levels.

[0060] Furthermore, to meet the automated operation requirements of the engineering operable implementation, this step can deploy a multi-sensor synchronous acquisition unit and a disposal strategy generation unit in the grinding section of the flour processing workshop for coordinated operation: When the life assessment package output by the upstream step shows that the remaining useful life range related to the grinding roller assembly enters the suggested execution window corresponding to the maintenance timing constraint field, and the closed-loop verification result is marked as passed, the disposal strategy generation unit automatically retrieves the "grinding roller assembly bearing maintenance template" from the maintenance disposal template library, and generates a disposal action record in conjunction with the equipment inspection system and shift plan. The action record includes issuing a load reduction operation command for the target component, arranging a shutdown inspection, disassembling and inspecting the lubrication system, replacing spare parts, and performing reinstallation acceptance and feedback; when a conflict mark appears in the closed-loop verification result, the disposal strategy generation unit marks the execution condition of the "replace spare parts" action record as a manual review trigger, and generates a review work order and binds the risk window sequence as a reference. In the above embodiments, the instantiation of template parameters by the disposal strategy generation unit does not depend on additional new inputs. The minimum input is still the risk curve, remaining useful life distribution or interval and closed-loop verification results in the lifetime assessment package. The root cause component and propagation chain confidence are preferred fields, which are used to further refine the target component reference and linkage check scope of the disposal action record, thereby supporting a consistent strategy generation interface for the same device under different fault propagation modes.

[0061] The disposal strategy generation unit encapsulates the generated results into a disposal strategy package and outputs it. This disposal strategy package includes at least a strategy version identifier, session identifier, strategy generation mode marker, maintenance timing constraint field, risk window sequence field, disposal action record set field, required spare parts reference field, acceptance feedback field, and strategy generation audit field. It may further include root cause component reference, propagation chain confidence, and root cause stability index copy fields. The disposal strategy package is written to the strategy display queue of the maintenance work order system or the equipment's human-machine interface terminal via the system output interface. Simultaneously, the disposal action record set field is filled back into the input position of the subsequent disposal execution loop, allowing the field execution unit to generate feedback acceptance data after execution, thus providing a data source for strategy version recording and audit tracing within the same session. The technical effect of this step is to transform the lifetime and risk information in the lifetime assessment package into a structured disposal strategy package, forming a policy output link that can be distributed, fed back, and audited.

[0062] Example 2: Figure 2 A structural block diagram of a flour milling equipment fault diagnosis system according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include: The multi-sensor synchronous acquisition unit 01 receives a session configuration package output from the propagation prior construction unit. The session configuration package includes a set of direction constraint edges, a set of propagation delay candidates, and a set of sensor channels. Under the access relationship defined by the sensor capability list and installation constraint information, it performs synchronous triggering and sampling rate hierarchical acquisition scheduling on the set of sensor channels. It frames and encapsulates the time-series data of each sensor channel according to the channel identifier and sampling level, and performs acquisition link connectivity verification and sampling clock status readback registration on the channel side. When a channel disconnection, sampling interruption, or trigger mismatch is detected, a corresponding quality tag is generated and bound to the time-series data in the same frame. The multi-sensor synchronous acquisition unit transmits the time-series data of the sensor channel set along with the quality tag to the data alignment and quality governance unit for time anchor alignment and quality governance. Simultaneously, it maintains a consistent mapping between the session identifier and channel index relationship of the session configuration package within this unit for subsequent backfeeding and updates.

[0063] The data alignment and quality governance unit 02 receives timing data and quality tags from the multi-sensor synchronous acquisition unit, and calls the synchronization trigger parameters and sampling rate layering parameters in the session configuration package to perform unified timeline generation, sampling clock drift compensation, and trigger jitter correction on each sampling level, aligning cross-channel data to the same time anchor point. For channels with missing measurements, it performs missing measurement tag merging; for channels with drift, it performs drift tag fixing; and for channels with health anomalies, it performs channel health self-check tag fixing. All tags, along with channel indexes and time anchor points, are written into the alignment session data packet. The data alignment and quality governance unit outputs the alignment session data packet to the propagation prior construction unit for updating the propagation delay candidate set, and simultaneously outputs the alignment session data packet to the event evidence chain generation unit for performing consistency violation detection and change point detection. During the alignment process, the start and end positions of abnormal segments are registered as time anchor points for reference in subsequent stage label sequences.

[0064] The propagation prior construction unit 03 receives the alignment session data packet from the data alignment and quality governance unit, reads the directional constraint edge set in the session configuration packet, and retrieves the response order of corresponding channel pairs within the alignment session data packet based on the upstream-to-downstream connection relationship of the directional constraint edge set. It then completes the candidate statistics and confidence mark update for cross-channel order relationships and writes the updated propagation delay candidate set back to the session configuration packet. Within the segment range where the quality mark indicates missing or drifting channels, it performs masking or weight reduction registration on the candidate statistics of the corresponding channel pairs and records the binding relationship between the weight reduction state and the time anchor range within the propagation delay candidate set. The propagation prior construction unit sends the updated session configuration packet back to the multi-sensor synchronous acquisition unit for its next round of synchronous acquisition scheduling and invocation. Simultaneously, it provides the session configuration packet to the event evidence chain generation unit for reference in consistency violation detection, using the directional constraint edge set and the propagation delay candidate set.

[0065] The event evidence chain generation unit 04 receives alignment session data packets from the data alignment and quality governance unit and session configuration packets from the propagation prior construction unit. It performs upstream and downstream channel sequence deviation determination around the directional constraint edge set and cross-channel delay anomaly drift determination around the propagation delay candidate set. Upon triggering a determination, it extracts candidate fault event fragments from the alignment session data packet and performs change point detection on each candidate fault event fragment to determine the stage segmentation boundary within the fragment. It generates a stage label sequence and binds it to the fragment's start and end time anchors. The candidate fault event fragments, stage label sequence, trigger channel index, and quality tag summary are encapsulated together into an evidence chain fragment packet. In the event of multiple trigger source conflicts or overlapping stage boundaries, it performs consistency arbitration based on time anchors and writes the arbitration result and conflict record into the evidence chain fragment packet. The event evidence chain generation unit outputs the evidence chain fragment packet to the feature construction and causal alignment unit for constructing a multi-domain feature and causal alignment view. It also retains the fragment index relationship of the evidence chain fragment packet and associates it with the session identifier of the alignment session data packet for subsequent backlinking reference in the tracing result packet.

[0066] The feature construction and causal alignment unit 05 receives evidence chain fragment packets from the event evidence chain generation unit and reads back the session configuration packet output by the propagation prior construction unit. Within the fragment range defined by the evidence chain fragment packets and the stage boundaries defined by the stage label sequence, it performs time-domain statistical features, frequency-domain statistical features, and time-frequency-domain statistical features calculations on the corresponding fragments of the aligned session data packets. It also performs stage difference feature generation between adjacent stages and aggregates multi-domain features by component index. Subsequently, it references the propagation delay candidate set to perform translational alignment on the aggregated features of the upstream channel or upstream component, forming a causal alignment view. It then performs multi-view stacking on the alignment views of different propagation delay candidate set entries to generate a causal input tensor packet. The feature construction and causal alignment unit outputs the causal input tensor packet to the temporal causal convolution representation unit for generating anomaly representation packets. It retains the association between the stage label sequence mapping relationship and the delay candidate reference identifier within the causal input tensor packet for subsequent directional constraint graph tracing unit to look back during graph attention reverse attribution.

[0067] The temporal causal convolutional representation unit 06 receives causal input tensor packets from the feature construction and causal alignment unit. It feeds the multi-view stacked input into causal convolutional and dilated convolutional layers according to the stage index order of the stage label sequence, and completes the cascaded output of multi-layer temporal abstract features within the residual connection structure. When the input contains missing or drifting segments indicated by quality markers, it performs mask registration on the feature channels within the corresponding stage and maintains consistency with inter-layer propagation. The temporal causal convolutional representation unit encapsulates the abnormal embedding sequence and early warning score generated by the multi-layer temporal abstract features into an abnormal representation package, and outputs the abnormal representation package to the directional constraint graph tracing unit for graph attention reverse attribution. Simultaneously, it retains the segment index relationship between the abnormal representation package and the evidence chain segment package for use by the survival risk and remaining lifespan prediction unit.

[0068] The directional constraint graph tracing unit 07 receives anomaly representation packets from the temporal causal convolution representation unit and session configuration packets from the propagation prior construction unit. It constructs a directional constraint component propagation graph based on the directional constraint edge set, maps and assembles the anomaly embedding sequence at the component index granularity, calculates edge attention weights on the directional constraint component propagation graph, and performs suppression registration on edges that do not satisfy the consistency constraint of the propagation delay candidate set. After completing the edge attention weight calculation, it performs graph attention reverse attribution on the anomaly representation packets, generates root cause component ranking results, and concatenates the attribution paths according to the directional constraint edge set to form propagation chain confidence. It performs statistical registration of root cause stability indices on the attribution results of multiple evidence chain fragment packets, performs stage sequence verification based on the sequence relationship of stage label sequences, and writes the verification records and conflict fragment indices into the tracing result packet. The directional constraint graph tracing unit outputs the tracing result package to the survival risk and remaining life prediction unit for it to form a life assessment package, and writes back the root cause component sorting result of the tracing result package to the corresponding evidence chain fragment package index for the disposal strategy generation unit to use in the back chain.

[0069] The survival risk and remaining lifetime prediction unit 08 receives the source tracing result package from the directional constraint graph source tracing unit and the anomaly representation package from the temporal causal convolution representation unit. It assembles the anomaly embedding sequence, the root cause component ranking result, the propagation chain confidence, and the operating condition label into covariate inputs for the survival analysis model. It performs time anchor binding on the stage index corresponding to the evidence chain fragment package, outputs a risk curve showing the future failure risk probability over time, and calculates the remaining useful lifetime distribution or interval. When there is a conflict triggering condition between the risk curve and the root cause component ranking result, it generates a closed-loop verification result and binds the conflict fragment index and the root cause stability index of the source tracing result package into the lifetime assessment package. The survival risk and remaining lifetime prediction unit outputs the lifetime assessment package to the disposal strategy generation unit for generating a disposal strategy package, and maintains a traceable association between the risk curve index relationship and the time anchor range within the lifetime assessment package for subsequent retraining or re-evaluation link calls.

[0070] The disposal strategy generation unit 09 receives a lifetime assessment package from the survival risk and remaining lifetime prediction unit and a tracing result package from the directional constraint graph tracing unit. Based on the root cause component sorting result, the propagation chain confidence, the risk curve, and the remaining useful lifetime distribution or interval, it performs structured assembly of the disposal strategy package, binding and registering disposal actions, timing windows, execution conditions, and time anchor ranges. It also performs rollback registration and recalculation trigger registration for the conflict fragment index indicated by the closed-loop verification result. After the disposal strategy package is generated, it outputs the disposal strategy package and establishes an association index with the session identifier of the aligned session data package, the evidence chain fragment package, the anomaly characterization package, the tracing result package, and the lifetime assessment package, completing the cross-unit tracing link closure. Simultaneously, it writes the conflict fragment index related to the propagation delay candidate set to the propagation prior construction unit for use in the next round of session configuration package updates.

Claims

1. A method for diagnosing faults in flour milling equipment, characterized in that, include: Obtain the component list, component connection relationship, process flow information, power transmission chain information, sensor capability list and installation constraint information, and generate a session configuration package containing a set of directional constraint edges, a set of propagation delay candidates and a set of sensor channels; Based on the session configuration package, the time-series data of the sensor channel set is collected synchronously, and time anchor point alignment and quality marking are performed to generate an aligned session data packet; Based on the aligned session data packets, consistency violation detection and change point detection are performed, candidate fault event fragments are extracted and stage label sequences are generated to construct evidence chain fragment packets; Based on the evidence chain fragment package, multi-domain features are constructed and combined with the propagation delay candidate set to generate a causal alignment view, and a causal input tensor package is constructed. Based on the causal input tensor packet, an anomaly representation packet is generated through a temporal causal convolutional network; Based on the anomaly representation package, graph attention reverse attribution of the propagation delay candidate set constraint is performed on the propagation graph of the directional constraint component to generate a source tracing result package. Based on the source tracing result package, a survival analysis is performed to obtain a lifespan assessment package; Based on the lifetime assessment package, a treatment strategy package is generated and output.

2. The method according to claim 1, characterized in that, The set of directional constraint edges is determined by the process flow information and the power transmission chain information to establish the directed connection between upstream and downstream components; the set of propagation delay candidates is obtained by calculating the cross-channel response sequence on the normal operating condition segment of the historical operating data, and a corresponding delay candidate value and its confidence level are recorded for each directional constraint edge.

3. The method according to claim 1, characterized in that, The session configuration package includes synchronization trigger parameters and sampling rate stratification parameters; the sensor channel set covers key branching nodes and convergence nodes in the process flow, as well as monitoring points corresponding to key rotating components in the power transmission chain, and configures different sampling rate stratifications for different monitoring points.

4. The method according to claim 1, characterized in that, The time anchor alignment includes unified time axis generation, sampling clock drift compensation, and trigger jitter correction; the quality markers include channel health self-check markers, missing test markers, and drift markers, and the alignment session data packet carries the quality markers for subsequent detection and modeling.

5. The method according to claim 1, characterized in that, The consistency violation detection includes determining whether the order of upstream and downstream channels deviates based on the set of directional constraint edges, and determining whether cross-channel delay drifts abnormally based on the set of propagation delay candidates; the stage label sequence includes one or more of the precursor stage, development stage, impact stage and recovery stage.

6. The method according to claim 1, characterized in that, The multi-domain features include time-domain statistical features, frequency-domain statistical features, and time-frequency-domain statistical features, and the stage difference features are obtained by differentiating the multi-domain features of adjacent stages; the causal alignment view is obtained by aligning the aggregated features of upstream channels or upstream components according to the propagation delay candidate set and stacking multiple views.

7. The method according to claim 1, characterized in that, The temporal causal convolutional network includes causal convolutional layers and dilated convolutional layers, and uses residual connections to output multi-layer temporal abstract features; the anomaly representation package includes anomaly embedding sequences and early warning scores generated from the multi-layer temporal abstract features.

8. The method according to claim 1, characterized in that, The graph attention reverse attribution includes calculating edge attention weights on the propagation graph of the directional constraint component and suppressing edges that do not satisfy the propagation delay candidate set; the source tracing result package includes root cause component ranking results, propagation chain confidence, and root cause stability index based on multi-event segment consistency.

9. The method according to claim 1, characterized in that, The survival analysis employs a lifetime modeling approach based on a risk function, using the abnormal embedding sequence in the anomaly characterization package and the root cause component, propagation chain confidence, and operating condition label in the source tracing result package as covariates. It outputs a risk curve showing the change of future failure risk probability over time and the distribution or range of remaining useful life. The lifetime assessment package also includes a closed-loop verification result generated based on the consistency between the risk curve and the root cause component, which is used to drive the generation or update of the treatment strategy package.

10. A flour milling equipment fault diagnosis system, characterized in that, include: The system comprises a multi-sensor synchronous acquisition unit, a data alignment and quality governance unit, a propagation prior construction unit, an event evidence chain generation unit, a feature construction and causal alignment unit, a temporal causal convolution representation unit, a directional constraint graph tracing unit, a survival risk and remaining lifetime prediction unit, and a disposal strategy generation unit; the units are connected in sequence to implement the method described in any one of claims 1-9.

Images