A machine learning based fuel nozzle cavitation prediction method, apparatus, and medium

Abstract

The application discloses a fuel nozzle cavitation prediction method and device based on machine learning and a medium, relates to the technical field of machine learning, and comprises the following steps: based on a double-time-scale cavitation event acceptance diagram, multi-source event features are extracted and transmitted along node acceptance edges, event propagation order checking, propagation time delay deviation evaluation and cross-channel support weight are performed, physical constraint coding is implemented in combination with nozzle structure and operating conditions, and a cavitation event fusion representation is generated; parallel reasoning of cavitation event fusion representation is performed on cavitation influence reservation and exclusion, cavitation state explanation information and counterfactual state explanation information are acquired, state explanation difference comparison is performed, cavitation counterfactual contribution is determined, continuous time window credible decision is made in combination with the double-time-scale cavitation event acceptance diagram, and fuel nozzle cavitation prediction information is generated. The application suppresses isolated jumps and channel conflicts, and improves the accuracy and stability of cavitation identification.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method, device and medium for predicting cavitation in fuel nozzles based on machine learning. Background Technology

[0002] With the development of high-pressure common rail injection, precision fuel supply, and intelligent condition monitoring technologies, cavitation inside fuel nozzles has become a crucial research subject affecting injection stability, flow consistency, and structural lifespan. Current research is gradually shifting from single-pressure signal discrimination to multi-source sensing integrating pressure, vibration, acoustics, and operating conditions, and incorporating machine learning methods for feature extraction, state classification, and trend prediction. The development of graph neural networks, attention mechanisms, and physically constrained learning has provided new technical pathways for representing fluid propagation, structural response, and cross-channel correlations within nozzles.

[0003] Existing machine learning-based fuel nozzle cavitation prediction technologies still have two shortcomings: First, they often treat each signal channel as an independent time-series input, lacking a unified expression of fluid propagation order, structural transmission order, and cross-running window relationships, making it easy to mix local jumps, propagation delays, and real cavitation events for identification; second, the prediction process mainly relies on feature correlation, lacking constraints on nozzle structure, operating conditions, and physical propagation relationships, as well as lacking differential reasoning under conditions of cavitation influence retention and exclusion, making it difficult to verify the contribution source and continuous reliability of the predicted state. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a machine learning-based fuel nozzle cavitation prediction method to solve the problems of incomplete expression of multi-source signal propagation relationship and insufficient cavitation prediction reliability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a machine learning-based method for predicting cavitation in fuel nozzles, comprising: collecting multi-source operating information, nozzle structure information, and sensor arrangement information of fuel nozzles; performing time-series alignment, dividing operating windows, and calibrating signal propagation ranges; acquiring signal change segments and corresponding operating conditions within the operating windows; mapping signal channel positions and aggregating operating segments to generate a multi-source operating window sequence; locating signal changes and marking candidate events in the multi-source operating window sequence; obtaining cavitation candidate event nodes; performing dual-time-scale relay arrangement of fluid propagation and structural propagation; establishing node relay edges; and simultaneously relaying the cavitation candidate events sequentially along the connection direction of the node relay edges. Selected event nodes are grouped into cavitation event chains to generate a dual-timescale cavitation event continuation graph. Based on the dual-timescale cavitation event continuation graph, multi-source event features are extracted and transmitted along the node continuation edges. Event propagation order verification, propagation delay deviation assessment, and cross-channel support weighting are performed. Physical constraint coding is implemented in conjunction with nozzle structure and operating conditions to generate a cavitation event fusion representation. Parallel reasoning for cavitation impact retention and exclusion is performed on the cavitation event fusion representation to obtain cavitation state explanation information and counterfactual state explanation information. State explanation differences are compared to determine the cavitation counterfactual contribution. At the same time, continuous time window credible adjudication is performed in conjunction with the dual-timescale cavitation event continuation graph to generate fuel nozzle cavitation prediction information.

[0007] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the step of generating a multi-source runtime window sequence is as follows: Based on the unified time reference of the collection time of multi-source operation information, duplicate records and time-series conflict records are eliminated, and the signal propagation relationship of nozzle structure information and sensor layout information is calibrated to determine the signal propagation range. At the same time, it is attached to the corresponding collection time period to form a time-aligned operation segment. Based on time-aligned running segments, the continuous acquisition is checked and the running status is segmented. Running windows are divided, and the effective response segment of the running window is limited by the signal propagation range. Signal change segments and corresponding operating conditions are extracted from the effective response segment to form running window sample segments. Based on the sensor layout information, the signal channel position correspondence is analyzed to determine the acquisition position corresponding to each signal channel. The acquisition position is then mapped to the nozzle structure information. The runtime window sample segments are marked with positions and time-series aggregated according to a unified channel order to form a multi-source runtime window sequence.

[0008] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps for obtaining cavitation candidate event nodes are as follows: Extract continuous change trajectories from signal change segments of multi-source runtime window sequences. Around the corresponding operating conditions, check the increase / decrease direction and continuity of adjacent signal changes according to the acquisition time. Locate the starting position and recovery position of the signal change from the original increase / decrease trend to abnormal deviation. Eliminate isolated jumps that do not continue to the future, and collect change segments with the same deviation direction and continuous extension along time as candidate event segments. Based on candidate event fragments, signal channel mapping and acquisition location are performed, and timing continuity verification and location consistency checks are conducted. Candidate event fragments with channel mapping conflicts and inconsistent acquisition locations are eliminated, and the remaining candidate event fragments are marked as cavitation candidate event nodes.

[0009] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps for generating a dual-time-scale cavitation event sequence diagram are as follows: Based on cavitation candidate event nodes, the node response order in the fuel flow direction and nozzle structure transmission direction is checked according to the corresponding signal channel and signal propagation range. Node connection edges are established for cavitation candidate event nodes whose response order and signal propagation range are mutually connected, forming dual-time-scale node connection information. Based on the dual-time-scale node connection information, the cavitation candidate event nodes are sequentially arranged along the connection direction of the node receiving edge. The cavitation candidate event nodes that maintain propagation and acceptance across runtime windows are connected to the same cavitation event chain. The cavitation event chain and the corresponding node receiving edge are aggregated to generate a dual-time-scale cavitation event receiving graph.

[0010] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps of extracting and transmitting multi-source event features are as follows: The dual-timescale cavitation event reception graph is input into the graph structure machine learning network. According to the node arrangement order of the cavitation event chain, the signal change features corresponding to each cavitation candidate event node are extracted. The signal change features are then linked to nodes according to the signal channel affiliation to form multi-source event features corresponding to each cavitation candidate event node. The signal change features are converted into node feature representations through the node feature encoding layer. The node reception edges corresponding to fluid propagation and the node reception edges corresponding to structural propagation are respectively encoded by edge type and transmission time sequence through the dual-timescale edge relationship encoding layer. At the same time, the graph message passing layer transmits and updates the node feature representations level by level along the node reception edges to form a node event feature sequence. The graph-structured machine learning network includes a node feature encoding layer, a dual-time-stamped edge relation encoding layer, a graph message passing layer, a cross-channel attention fusion layer, and a dual-path inference layer. The node feature encoding layer takes on the multi-source event features of the cavitation candidate event nodes. The dual-time-stamped edge relation encoding layer takes on the propagation type and temporal relationship of the edges received by the two types of nodes. The graph message passing layer transmits node feature representations based on the edges received by the two types of nodes. The cross-channel attention fusion layer allocates fusion weights based on the degree of common support of different signal channels. The dual-path inference layer retains and excludes the cavitation effect to perform state inference. Based on the node event feature sequence, the response order of adjacent nodes is checked, the propagation delay deviation of adjacent nodes is calculated, and the response direction and duration of each signal channel in the same cavitation event chain are compared to determine the degree of cross-channel common support. The fusion weight of node event features is adjusted through the cross-channel attention fusion layer to generate weighted event propagation features.

[0011] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps for generating a cavitation event fusion representation are as follows: The weighted event propagation features are reattached to the dual-timescale cavitation event reception diagram. Based on the nozzle structure and operating conditions associated with the cavitation event chain, the physical consistency of the structure and operating conditions is checked to determine the physical reception relationship of the multi-source event features in the corresponding propagation process. The dual-timescale edge relationship coding layer then performs constraint marking and relationship coding on the physical reception relationship to generate physical constraint coding information. By combining physical constraint encoding information with weighted event propagation features, and through a cross-channel attention fusion layer, the fusion weight of event features that do not conform to physical continuity is reduced, while event features that have the same propagation direction and support each other across channels are aggregated to generate a cavitation event fusion representation.

[0012] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps for obtaining cavitation state interpretation information and counterfactual state interpretation information are as follows: The dual-path inference layer of the input graph structure machine learning network is used to fuse cavitation events into a representation. The cavitation state inference branch retains the weighted event propagation features and physical constraint encoding information for state decoding, while the counterfactual state inference branch keeps the physical constraint encoding information, node connection relationships and operating condition associations unchanged. The weighted event propagation features are replaced with the baseline event propagation features corresponding to normal samples under the same operating conditions, and counterfactual state decoding is performed to form cavitation state explanation information and counterfactual state explanation information, respectively. The cavitation state interpretation information and the counterfactual state interpretation information are attached back to the dual-time-scaled cavitation event succession graph according to the runtime window. The state interpretation differences are compared, and the differences are marked to the corresponding cavitation candidate event nodes to generate state interpretation difference information.

[0013] As a preferred embodiment of the machine learning-based fuel nozzle cavitation prediction method of the present invention, the steps for generating fuel nozzle cavitation prediction information are as follows: Based on the state interpretation difference information, node relationship encoding and temporal association reasoning are performed along the node bearing edge to trace the propagation path of the state interpretation difference information between adjacent runtime windows and form a cavitation counterfactual contribution sequence. The cavitation counterfactual contribution sequence is reattached to the dual-time-scale cavitation event succession diagram. Contribution continuity verification and node succession integrity verification are performed according to the runtime window sequence. Contribution credibility weighting and prediction state decoding are also performed to generate fuel nozzle cavitation prediction information.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the machine learning-based fuel nozzle cavitation prediction method as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based fuel nozzle cavitation prediction method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by using dual-timescale inheritance arrangement and graph structure machine learning feature transfer, a unified fusion of fluid propagation, structural response, and cross-channel support relationships is achieved, which is used to suppress isolated jumps and channel conflicts, and improve the accuracy and stability of cavitation identification; by using parallel reasoning to retain and exclude cavitation effects, continuous adjudication of state interpretation differences and cross-time window counterfactual contributions is achieved, which is used to distinguish between real cavitation contributions and occasional anomalies, and improve the interpretability of machine learning predictions and the reliability of early warnings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a machine learning-based fuel nozzle cavitation prediction method.

[0019] Figure 2 This is a comparison chart of the reference propagation delay and the actual transmission interval of the node receiving edge.

[0020] Figure 3 A graph showing the changes in cavitation prediction probability and cavitation counterfactual contribution under continuous running windows.

[0021] Figure 4 A comparative chart showing the overall performance of fuel nozzle cavitation prediction for different schemes. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a machine learning-based fuel nozzle cavitation prediction method, comprising the following steps: S1: Collect multi-source operating information, nozzle structure information and sensor layout information of fuel nozzles, perform time sequence alignment, divide operating windows and calibrate signal propagation range, obtain signal change segments and corresponding operating conditions within the operating window, perform signal channel position mapping and operating segment aggregation, and generate multi-source operating window sequence.

[0026] S1.1: Based on the unified time reference of the collection time of multi-source operation information, duplicate records and time-series conflict records are eliminated, and the signal propagation relationship of nozzle structure information and sensor layout information is calibrated to determine the signal propagation range. At the same time, it is attached to the corresponding collection time period to form a time-aligned operation segment.

[0027] Furthermore, multi-source information acquisition is initiated during the continuous operation of the fuel nozzle. Sensors positioned at the nozzle inlet, outlet, and corresponding locations on the housing synchronously generate operating signals, which are then recorded according to the acquisition time and source. Nozzle structural information is extracted from the nozzle design documents and structural configuration records, retaining the flow path, key structural locations, and structural connection relationships. Sensor placement information is compiled based on installation records, specifying the acquisition location, installation direction, and signal channel affiliation for each sensor. The multi-source operating information, nozzle structural information, and sensor placement information are linked together according to the same fuel nozzle and the same operating section, eliminating records with inconsistent affiliations and missing acquisition sources. The acquisition time from the multi-source operating information is then used as a unified time reference to eliminate duplicate and time-series conflicting records, forming a continuous time-series record.

[0028] Combining the flow path and structural connection relationships in the nozzle structure information, and the acquisition positions and installation directions in the sensor layout information, propagation relationship calibration is performed on each signal channel. The response sequence of the preceding and following acquisition positions is checked along the fuel flow direction, and the transmission positions of vibration and acoustic responses are checked along the nozzle structural connection direction. Channel positions that can form continuous responses are written into the corresponding propagation relationships to determine the applicable signal propagation range for each acquisition period. The signal propagation range is then linked to a time-aligned continuous time sequence record, preserving the corresponding states between acquisition time, signal channel, acquisition position, and propagation relationship to form a time-aligned running segment.

[0029] It should be noted that the signal propagation range refers to the spatial distance and time delay range during which fluid pressure changes or structural vibrations generated at the acquisition location are transmitted along a predetermined direction and can form a continuous response at other acquisition locations under the corresponding operating conditions. Example ranges: fluid propagation distance is 5-80 mm and response delay is 0.05-8 ms; structural propagation distance is 5-120 mm and response delay is 0.01-1 ms. The specific range is calibrated based on nozzle size, fuel flow rate, structural material, and sensor spacing.

[0030] S1.2: Based on the time-aligned running segments, the continuous acquisition is checked and the running status is segmented. The running windows are divided, and the effective response segments of the running windows are limited in combination with the signal propagation range. The signal change segments and corresponding operating conditions are extracted from the effective response segments to form the running window sample segments.

[0031] Furthermore, the continuity of acquisition time is checked channel by channel around the time-aligned operation segment, verifying the acquisition interval, repetition status, and changes in operating conditions between adjacent records. Records whose acquisition interval and changes in operating conditions are all within the acquisition continuity threshold range are grouped into the same continuous segment. Segment boundaries are set at locations where acquisition is interrupted, records are repeated, nozzles start / stop switch, or where the acquisition continuity threshold range is exceeded, forming multiple operating state transition segments. Running windows are divided according to the start and end times of the transition segments, and the temporal relationship between adjacent running windows is preserved.

[0032] Based on the signal propagation range of each operating window, the acquisition location, response delay, and duration of the channel response are checked along the fuel flow propagation direction and the nozzle structure propagation direction. Periods exceeding the signal propagation range or failing to reach the response acceptance threshold are excluded, and the portions meeting the propagation location and timing acceptance conditions are defined as valid response segments. The start, continuation, and recovery positions of signal changes are marked according to the acquisition time, and the changes with consistent and continuous directions are collected to form signal change segments. These segments are then time-sequentially coupled with the corresponding operating conditions to form operating window sample segments.

[0033] It should be noted that the continuous sampling threshold (example range: the deviation between adjacent sampling intervals does not exceed 10% to 20% of the nominal sampling period, and the change in operating parameters does not exceed 1.2 to 1.5 times the upper limit of normal fluctuation) is specifically calibrated based on the sampling jitter and operating condition fluctuation distribution in normal continuous operation segments.

[0034] The response acceptance threshold (the example range is that the response delay of adjacent channels does not exceed 1.1 to 1.3 times the upper limit of the corresponding signal propagation delay, and the response duration is not less than 2 to 4 sampling periods) is specifically determined based on the acceptance differentiation effect between normal propagation samples and labeled cavitation samples.

[0035] S1.3: Based on the sensor layout information, perform signal channel position correspondence analysis to determine the acquisition position corresponding to each signal channel, and map the acquisition position to the nozzle structure information. Mark the position and time sequence of the runtime window sample segments according to the unified channel order to form a multi-source runtime window sequence.

[0036] Furthermore, using the signal channel identifiers in the sensor deployment information as the starting point for analysis, the affiliation between each signal channel and the sensor installation location is verified, identifying situations where the same signal channel corresponds to multiple acquisition locations, multiple signal channels occupy the same acquisition location, and channel locations are missing. Conflicts are corrected based on the actual sensor installation relationship, and a unique acquisition location marker is assigned to each verified signal channel. According to the nozzle flow channel extension relationship and structural connection relationship, the acquisition locations are mapped to the corresponding structural locations in the nozzle structural information, establishing a positional correspondence between signal channels, acquisition locations, and nozzle structural locations.

[0037] A unified channel order is established based on the location correspondence. Signal change segments within each runtime window sample segment are sorted by channel, and acquisition location markers are written to the corresponding segments. The channel affiliation, location order, and acquisition time of each segment within the same runtime window are verified. Segments with conflicting location markers are repositioned, segments repeatedly assigned to the same channel location are merged, and segments whose acquisition location cannot be determined are temporarily excluded from timing aggregation. Runtime window sample segments with completed location markers are aggregated according to the order of the runtime windows, preserving the timing continuity between adjacent runtime windows to form a multi-source runtime window sequence.

[0038] S2: Perform signal change localization and candidate event marking on the multi-source runtime window sequence, obtain cavitation candidate event nodes, and perform dual-timescale inheritance and arrangement of fluid propagation and structure propagation, establish node inheritance edges, and at the same time, merge the successively inherited cavitation candidate event nodes into cavitation event chains along the connection direction of the node inheritance edges to generate a dual-timescale cavitation event inheritance graph.

[0039] S2.1: Extract continuous change trajectories from signal change segments of multi-source runtime window sequences, check the increase / decrease direction and continuity of adjacent signal changes according to the corresponding operating conditions and the acquisition time, locate the starting position and recovery position where the signal change changes from the original increase / decrease trend to abnormal deviation, eliminate isolated jumps that do not continue, and collect change segments with consistent deviation direction and continuous extension along time as candidate event segments.

[0040] Furthermore, signal change segments are retrieved sequentially along the multi-source runtime window sequence. Changes belonging to the same signal channel at adjacent acquisition times are sequentially connected to verify the direction, amplitude, and duration of change, eliminating trajectory breaks caused by acquisition interruptions. Stable change phases are organized around the corresponding operating conditions, with continuously increasing, continuously decreasing, and stable change states considered as the original increasing / decreasing trends. The directional differences and continuity between subsequent signal changes and the original increasing / decreasing trends are compared moment-by-moment. Locations where the direction of change reverses, the degree of change deviates continuously, or the original increasing / decreasing trend is interrupted are marked as the starting positions of abnormal deviations. Locations where the signal returns to the original increasing / decreasing trend under the corresponding operating conditions are marked as recovery positions.

[0041] The abnormal deviation segment is defined by the starting and recovery positions of the abnormal deviation. The amplitude, duration, and continuity of the changes during the abnormal deviation period are extracted to calculate the signal abnormal deviation intensity. Changes that occur only at a single acquisition moment, do not maintain the same deviation direction, or whose abnormal deviation intensity does not reach the deviation intensity threshold are removed. For change segments whose abnormal deviation intensity reaches the deviation intensity threshold, continuity is checked according to the deviation direction and time sequence. Contents with consistent deviation directions and adjacent changes that can be continuously connected are grouped into the same segment, retaining the corresponding signal channel, operating condition, abnormal deviation starting position, and recovery position to form candidate event segments.

[0042] The formula for calculating the intensity of abnormal signal deviation is: ; in, This indicates the signal abnormality intensity of the current abnormal deviation segment. This indicates the average deviation of the signal change from its original increasing or decreasing trend. This represents the baseline fluctuation amplitude of normal signal changes under corresponding operating conditions. Indicates the duration of the abnormal deviation segment. This indicates the baseline duration of normal changes under the corresponding operating conditions. This indicates the percentage of sampling points within the abnormal deviation segment that maintain the same direction of deviation. It represents a very small positive number.

[0043] The formula multiplies the ratio of the average deviation amplitude to the baseline fluctuation amplitude, the ratio of the duration of the abnormal deviation segment to the baseline duration, and the proportion of sampling points in the same deviation direction. This ensures that the signal's abnormal deviation strength must be supported by the continuity of deviation amplitude, duration, and direction simultaneously. Using a multiplicative relationship can suppress the interference of large jumps at a single sampling moment and slight fluctuations over a long period on anomaly detection. Furthermore, exponential mapping limits the calculated value to between 0 and 1, facilitating a unified comparison with the deviation strength threshold.

[0044] In the formula, the average deviation amplitude is obtained by statistically analyzing the difference between the actual signal change at each sampling time within the abnormal deviation segment and the original increasing / decreasing trend; the reference fluctuation amplitude is obtained by statistically analyzing the signal change fluctuation of a normal operating segment under the same operating conditions, and is used to normalize the average deviation amplitude. The duration of the abnormal deviation segment is determined by the time interval between the start position of the abnormal deviation and the recovery position, and the reference duration is obtained by statistically analyzing the duration of the normal change segment under the corresponding operating conditions; the sampling point ratio is determined by the proportion of the number of sampling points maintaining the same deviation direction to the total number of sampling points in the abnormal deviation segment. The minimum positive number is set according to the sensor sampling accuracy and participates in the denominator calculation to avoid calculation abnormalities caused by a reference value of 0.

[0045] It should be noted that the deviation intensity threshold (example range: 0.55 to 0.75) is specifically determined based on the upper limit of the deviation intensity distribution of the normal operating segment under the same operating conditions, as well as the recognition accuracy and false negative rate of the labeled cavitation samples.

[0046] S2.2: Based on candidate event fragments, perform signal channel mapping and acquisition location attachment, and perform timing continuity verification and location consistency check. Eliminate candidate event fragments with channel mapping conflicts and inconsistent acquisition locations, and mark the retained candidate event fragments as cavitation candidate event nodes.

[0047] Furthermore, starting with the signal channel identifier carried by the candidate event fragment, the candidate event fragment is mapped to the corresponding signal channel in the multi-source runtime window sequence, and the consistency between the channel identifier, signal type, and runtime window affiliation is verified. According to the correspondence between signal channel and acquisition location, the candidate event fragment is attached to the corresponding acquisition location in the nozzle structure information, and the start and end times, deviation direction, and duration of the candidate event fragment are retained. The temporal sequence of adjacent candidate event fragments is checked along the runtime window sequence to confirm that the candidate event fragments can form a continuous sequence within the same signal channel, avoiding the incorrect merging of changes from different acquisition periods.

[0048] For candidate event segments that have completed location attachment, verify whether the acquisition position corresponding to the signal channel is consistent with the sensor layout information, and check whether the acquisition positions of adjacent candidate event segments conform to the propagation order in the nozzle structure. Set conflict flags for cases where the same signal channel is mapped to different acquisition positions, different signal channels repeatedly occupy the same position, or the segment position exceeds the signal propagation range, and remove these from the candidate event range. Candidate event segments that pass the timing continuity check and location consistency check are marked as nodes, and the correspondence between the signal channel, acquisition position, runtime window, and changing segment is written, forming cavitation candidate event nodes.

[0049] S2.3: Based on cavitation candidate event nodes, according to the corresponding signal channels and signal propagation ranges, the node response order in the fuel flow direction and nozzle structure transmission direction is checked respectively. For cavitation candidate event nodes whose response order and signal propagation range are mutually connected, node connection edges are established to form dual-time-scale node connection information.

[0050] Furthermore, based on the signal channels, acquisition locations, runtime windows, and change sections carried by cavitation candidate event nodes, node combinations capable of forming forward and backward responses are retrieved according to the signal propagation range. Along the fuel flow direction, corresponding nodes are arranged according to their positional relationship in the nozzle channel, and the response times and change directions of upstream and downstream nodes are checked to determine whether the signal change is transmitted backward in the order of fuel flow. Along the nozzle structure transmission direction, corresponding nodes are arranged according to the connection relationship between the sensor acquisition location and the nozzle structure, and the order of occurrence of vibration and acoustic responses between adjacent structural locations is checked to form two types of node response sequences.

[0051] For node combinations arranged in the order of completion, the response time intervals of adjacent nodes are compared with the corresponding signal propagation ranges to exclude node combinations with reversed response directions, jumps in transmission positions, or inconsistent time intervals. Fluid propagation node connection edges are established for cavitation candidate event nodes with continuous response order in the fuel flow direction, and structural propagation node connection edges are established for cavitation candidate event nodes with continuous response order in the nozzle structure transmission direction. The connection direction, node order, and transmission timing are written into the node connection edges. The two types of node connection edges and their corresponding cavitation candidate event nodes are collected to form dual-time-stamped node connection information.

[0052] S2.4: Based on the connection information of the dual-time-scaled nodes, arrange the cavitation candidate event nodes sequentially along the connection direction of the node receiving edge, connect the cavitation candidate event nodes that maintain propagation and receiving across runtime windows to the same cavitation event chain, and collect the cavitation event chain and the corresponding node receiving edge to generate a dual-time-scaled cavitation event receiving graph.

[0053] Furthermore, based on the cavitation candidate event nodes and node receiving edges in the dual-time-stamped node connection information, the node relationships are traced along the connection directions of the fluid propagation node receiving edges and the structural propagation node receiving edges, respectively. Cavitation candidate event nodes in the same propagation direction are arranged according to their response times, and it is checked whether the propagation endpoint of the previous node and the response start point of the next node form a continuous connection. Nodes within the same runtime window are directly arranged according to their connection order. For nodes spanning adjacent runtime windows, the propagation direction, acquisition location, and response time sequence are checked. Nodes with continuous connection relationships are connected to the same node sequence to avoid cross-connection of nodes in different propagation processes.

[0054] Around the completed sequence of nodes, starting from the first cavitation candidate event node, the process extends level by level along the node receiving edges, grouping nodes with consistent connection directions and maintaining propagation continuity across runtime windows into the same cavitation event chain. When a node receiving edge is interrupted, the connection direction changes, or runtime windows cannot continuously receive connections, the current cavitation event chain ends at the corresponding position, and the orchestration restarts with nodes not yet included in the event chain. A correspondence is established between each cavitation event chain and the fluid propagation node receiving edges and the structural propagation node receiving edges, preserving the association between node order, propagation type, and runtime window, and a cavitation event graph structure is generated, producing a dual-timescale cavitation event receiving graph.

[0055] It should be noted that the dual-timescale cavitation event succession diagram is a unified collection of cavitation candidate event nodes, node succession edges, and cavitation event chains formed during the fluid propagation process and the nozzle structure propagation process. It is used to characterize the response sequence of cavitation events, their continuity across runtime windows, and the node association status during the two types of propagation processes.

[0056] A cavitation event chain refers to a continuous sequence of events formed by connecting multiple cavitation candidate event nodes with temporal succession relationships in the direction of fluid propagation or structural propagation, according to the propagation order. It is used to characterize the propagation process of cavitation anomaly response at different locations inside the nozzle and between different operating windows.

[0057] S3: Based on the dual-timescale cavitation event transmission graph, multi-source event features are extracted and transmitted along the node transmission edges. Event propagation order verification, propagation delay deviation assessment, and cross-channel support weighting are performed. Physical constraint coding is implemented in combination with nozzle structure and operating conditions to generate a fusion representation of cavitation events.

[0058] S3.1: The graph-structured machine learning network includes a node feature encoding layer, a dual-time-stamped edge relation encoding layer, a graph message passing layer, a cross-channel attention fusion layer, and a dual-path inference layer. The node feature encoding layer receives multi-source event features from cavitation candidate event nodes. The dual-time-stamped edge relation encoding layer receives the propagation type and temporal relationship of the edges received by the two types of nodes. The graph message passing layer transmits node feature representations based on the edges received by the two types of nodes. The cross-channel attention fusion layer allocates fusion weights based on the degree of common support of different signal channels. The dual-path inference layer retains and excludes the influence of cavitation for state inference.

[0059] Furthermore, the graph-structured machine learning network is a machine learning processing network designed for the dual-timescale cavitation event succession relationship. It incorporates the multi-source event features carried by cavitation candidate event nodes, the propagation type carried by the node succession edges, and the transmission timing into the computation. The multi-source event features are formed by aggregating signal change features corresponding to different signal channels according to the channel affiliation and acquisition location of the cavitation candidate event nodes, including the signal change direction, duration, and response timing. The information of each node is processed according to the node arrangement order of the cavitation event chain. While preserving the differences between the fluid propagation process and the nozzle structure propagation process, it completes node feature transmission, cross-channel feature fusion, and cavitation state inference, enabling multi-source operating signals to form a unified event representation according to the actual propagation relationship.

[0060] The graph-structured machine learning network sequentially comprises a node feature encoding layer, a dual-time-scale edge relation encoding layer, a graph message passing layer, a cross-channel attention fusion layer, and a dual-path inference layer. The node feature encoding layer converts multi-source event features into node feature representations. The dual-time-scale edge relation encoding layer writes propagation type and transmission timing codes into the nodes' receiving edges. The graph message passing layer combines the node feature representations and edge relation codes, propagating and updating the node feature representations level by level along the connection direction of the nodes' receiving edges. The cross-channel attention fusion layer receives the updated node event features and adjusts the fusion weights based on the common support of each signal channel to form weighted event propagation features. After introducing physical constraint encoding information, the cross-channel attention fusion layer constrains and aggregates the weighted event propagation features to form a cavitation event fusion representation. The dual-path inference layer receives the cavitation event fusion representation and performs cavitation effect retention inference and cavitation effect exclusion inference, respectively.

[0061] It should be noted that during the network training phase, the historical multi-source runtime window sequences are constructed into a dual-timescale cavitation event connection diagram for training, based on cavitation candidate event nodes, fluid propagation node connection edges, and structural propagation node connection edges. Based on the cavitation calibration information from fuel nozzle flow field simulation and visualization experiments, cavitation occurrence state, cavitation degree, and corresponding propagation position labels are configured for each runtime window. The node feature encoding layer and the dual-timescale edge relationship encoding layer encode the multi-source event features and node connection edge relationships in the training samples, respectively. The cavitation state inference branch retains the weighted event propagation features and physical constraint encoding information to complete state decoding. The counterfactual state inference branch replaces the weighted event propagation features with the baseline event propagation features corresponding to normal samples under the same operating conditions, while maintaining the physical constraint encoding information, node connection relationships, and operating conditions unchanged. During training, the network parameters are updated jointly based on the state deviation between the output of the cavitation state inference branch and the cavitation state label, the state deviation between the output of the counterfactual state inference branch and the normal state label, and the contribution deviation between the difference between the outputs of the two branches and the cavitation contribution calibration value. By verifying the synchronous check of prediction bias and state continuity across runtime windows, training is stopped when the prediction bias no longer decreases and the state continuity across runtime windows no longer improves, and the corresponding network parameters are saved as the cavitation prediction model.

[0062] To balance cavitation state recognition, counterfactual contribution verification, and state continuity across runtime windows, the state bias of the two inference branches, cavitation contribution bias, and state change bias of adjacent runtime windows are all included in the loss calculation during the training phase. The biases are weighted and aggregated according to their impact on prediction performance to form a joint loss function, which is used to update network parameters in reverse to reduce cavitation state misjudgment and abnormal jumps across runtime windows.

[0063] The joint loss function is: ; in, This represents the total loss value during network training; The output of the cavitation state inference branch is the probability of cavitation occurrence. The probability of cavitation occurring is output by the counterfactual state reasoning branch. This is the cavitation state label corresponding to the current runtime window. The label for the normal state corresponding to the counterfactual state reasoning branch. The cavitation contribution calibration value is assigned to the cavitation state reasoning branch and the counterfactual state reasoning branch. Predict the change in cavitation probability for adjacent runtime windows. This represents the change in the emptying state label of adjacent runtime windows. For branch state deviation weights, Contribute bias weights to cavitation. This is the weight for continuity deviation across runtime windows.

[0064] Branch state deviation weights (example range 0.4 to 0.6) are used to highlight the degree of fit between the outputs of the two inference branches and the corresponding state labels. The specific weights are set based on the accuracy of cavitation state recognition and the degree of recovery of counterfactual normal states in the validation samples.

[0065] The cavitation contribution deviation weight (example range is 0.2 to 0.35) is used to constrain the deviation between the output difference of the two branches and the cavitation contribution calibration value. The specific setting is based on the cavitation contribution identification deviation and contribution ranking consistency in the verification samples.

[0066] The cross-run window continuity bias weight (example range 0.1 to 0.25) is used to constrain the deviation between the predicted state change and the calibrated state change of adjacent run windows. The specific weight is set according to the cross-run window state continuity and the degree of abnormal jump suppression in the validation sample.

[0067] The initial learning rate for network training is set to 0.0003, which is determined based on the rate of decrease of the joint loss and the magnitude of parameter updates in the early stages of training. When the joint loss of the validation samples does not decrease for 3 to 5 consecutive training rounds, the learning rate is adjusted to 0.5 times the current value, and the minimum learning rate is set to 0.00001 to reduce parameter fluctuations in the later stages of training.

[0068] The cavitation occurrence probability output by the cavitation state inference branch is the cavitation prediction probability obtained by inputting the cavitation event fusion representation of the current runtime window into the cavitation state inference branch and performing state decoding while retaining the weighted event propagation characteristics and physical constraint encoding information.

[0069] The cavitation probability output by the counterfactual state inference branch is the cavitation prediction probability obtained by replacing the weighted event propagation features with the baseline event propagation features corresponding to normal samples under the same operating conditions, while keeping the physical constraint encoding information, node connection relationships and operating conditions unchanged, and then decoding the counterfactual state.

[0070] S3.2: Input the dual-timescale cavitation event receiving graph into the graph structure machine learning network. According to the node arrangement order of the cavitation event chain, extract the signal change features corresponding to each cavitation candidate event node. According to the signal channel affiliation, attach the signal change features to nodes to form multi-source event features corresponding to each cavitation candidate event node. Convert the signal change features into node feature representations through the node feature encoding layer. Then, through the dual-timescale edge relationship encoding layer, encode the edge type and transmission time sequence of the node receiving edges corresponding to fluid propagation and the node receiving edges corresponding to structural propagation, respectively. At the same time, the graph message passing layer transmits and updates the node feature representations level by level along the node receiving edges to form a node event feature sequence.

[0071] Furthermore, the dual-timescale cavitation event succession graph is fed into a graph-structured machine learning network. Cavitation candidate event nodes are located sequentially according to the node arrangement order in each cavitation event chain. The direction of change, duration, and response timing are extracted from the signal change segments corresponding to each node, while retaining the signal channel, acquisition location, and runtime window markers. Using the cavitation candidate event nodes as attachment objects, the signal change characteristics of the current node itself, as well as the signal change characteristics of other signal channels that form a propagation connection with the current node, are attached to the corresponding nodes according to signal channel affiliation, acquisition location, and node succession relationships, forming multi-source event features corresponding to each cavitation candidate event node.

[0072] The node feature encoding layer unifies the dimensions and compresses the features of multi-source events received by each cavitation candidate event node, converting the change direction, duration, and response timing of different signal channels into node feature representations that can participate in network computation. The dual-time-scale edge relationship encoding layer identifies the propagation type corresponding to the edge received by each node, writes it into fluid propagation markers and structure propagation markers respectively, and forms a transmission timing encoding based on the response sequence and transmission interval between connected nodes, so that the node feature representation is associated with the corresponding propagation type, connection direction, and transmission timing.

[0073] The graph message passing layer uses the starting node of the cavitation event chain as the starting point for transmission, and transmits node feature representations level by level along the connection direction of the node receiving edges. It aggregates the feature representations of the preceding nodes, the current node, and the corresponding edge relationship codes. Feature transmission is performed separately for fluid propagation node receiving edges and structural propagation node receiving edges, preserving the differences in node responses during the two types of propagation processes. After each feature transmission, the current node feature representation is updated based on the node feature representation transmitted by the preceding node, the edge relationship code of the corresponding node receiving edge, and the multi-source event features received by the current node. The updated node feature representation is then transmitted to the next cavitation candidate event node. The updated feature content of each node is collected according to the node arrangement order, forming a node event feature sequence reflecting the level-by-level transmission process of multi-source signals along the dual-time-scale propagation relationship.

[0074] S3.3: Based on the node event feature sequence, check the response order of adjacent nodes, calculate the propagation delay deviation of adjacent nodes, and compare the response direction and duration of each signal channel in the same cavitation event chain to determine the degree of cross-channel common support. Adjust the fusion weight of node event features through the cross-channel attention fusion layer to generate weighted event propagation features.

[0075] Furthermore, according to the node arrangement order in the node event feature sequence, the response time of adjacent cavitation candidate event nodes is checked one by one to determine the actual transmission interval from the preceding node to the subsequent node. The actual transmission interval is compared with the signal propagation delay associated with the corresponding node's receiving edge to calculate the degree of propagation delay deviation. When the degree of propagation delay deviation does not exceed the propagation delay deviation threshold, the corresponding node event feature is retained; when it exceeds the propagation delay deviation threshold, a delay deviation flag is set for the corresponding node event feature. The response sequence is further checked along the node receiving edge, and node event features with response order opposite to the connection direction or without subsequent responses are eliminated.

[0076] The node event characteristics of each signal channel are aligned according to the runtime window. The direction of signal change, overlap of change segments, and response duration are checked. Signal channels with consistent change directions and sequential timing are marked as supporting channels. The degree of cross-channel common support is determined based on the number of supporting channels and the degree of response duration. When the degree of cross-channel common support reaches the common support threshold, the fusion weight of the corresponding node event characteristics is increased; when it does not reach the common support threshold, the corresponding fusion weight is decreased. The node event characteristics with adjusted weights are aggregated according to the node order of the cavitation event chain to generate weighted event propagation characteristics.

[0077] It should be noted that the propagation delay deviation threshold (example range: 0.1 to 0.3) applies to situations where the degree of propagation delay deviation is expressed as the relative deviation between the actual transmission interval and the reference propagation delay. The specific threshold is determined based on the fluctuation range of normal propagation delay under different operating conditions and the cavitation event calibration sample.

[0078] The common support threshold (example range: 0.6 to 0.8) indicates that at least 60% to 80% of the effective signal channels in the same cavitation event chain provide consistent support for the same change process. The specific threshold is determined based on the number of multi-source sensors, channel reliability, and the consistency of channel responses of labeled cavitation samples.

[0079] The formula for calculating the deviation in propagation delay between adjacent nodes is: ; in, Represents a node To node The degree of deviation in propagation delay of the corresponding node's connected edge. Represents a node To node The actual transmission interval, Represents a node To node The reference propagation delay, Represents a node To node The actual transmission interval, This represents the correction coefficient for abrupt changes in adjacent time delays. It represents a very small positive number.

[0080] It should be noted that the node ,node and nodes These are three cavitation candidate event nodes arranged sequentially according to the propagation direction within the same cavitation event chain, where node... For nodes Preorder node, node For the current front-end node performing propagation latency verification, the node... For nodes The subsequent nodes connected along the node's receiving edge.

[0081] The formula uses the relative difference between the actual transmission interval from node i to node j and the corresponding reference propagation delay as its main component. This serves to uniformly characterize propagation advance and propagation delay, avoiding the impact of different reference propagation delays on comparisons between edges received by different nodes. Furthermore, it combines the node... To node With nodes To node The time delay changes between propagation deviations are locally corrected by abrupt changes, which enhances the identification of abnormal time delay changes in continuous propagation. The degree of propagation time delay deviation is limited to between 0 and 1 by exponential mapping, which facilitates subsequent comparison with the propagation time delay deviation threshold.

[0082] In the formula, the actual transmission interval from node i to node j is obtained from the difference in response times between the two nodes, and the corresponding reference propagation delay is calibrated based on normal samples under the same propagation type, acquisition location, and operating conditions. The actual transmission interval from node h to node i is used to measure the degree of change in the current propagation delay relative to the previous propagation delay. The adjacent delay abrupt change correction coefficient is determined by the differentiation effect between normal samples and cavitation samples, and the minimum positive number is set according to the sensor sampling time resolution to avoid division by zero and logarithmic aberrations.

[0083] like Figure 2As shown, the propagation timing continuity in the dual-timescale cavitation event continuity diagram is displayed with the baseline propagation delay corresponding to each node's connection edge as the horizontal axis and the actual transmission interval between adjacent cavitation candidate event nodes as the vertical axis. Normal connection edges are mainly distributed near the ideal congruence line, indicating that the actual transmission interval can maintain stable continuity with the baseline propagation delay calibrated under the same propagation type, acquisition location, and operating conditions. Abnormal connection edges deviate significantly from the ideal congruence line, reflecting that the corresponding nodes have premature responses, delayed responses, and local delay abrupt changes. By comparing the actual transmission interval with the baseline propagation delay, the degree of propagation delay deviation can be determined, and delay deviation markers can be set for node event features with large delay deviations, restricting erroneous node connection edges from entering subsequent feature aggregation. The figure shows a high degree of concentration of normal connection edges, and abnormal connection edges can be clearly distinguished, indicating that propagation delay deviation assessment helps reduce node connections that do not conform to the actual propagation order, improving the propagation integrity of the cavitation event chain and the accuracy of cavitation identification.

[0084] It should be noted that a normal receiving edge is a node receiving edge where the actual transmission interval is close to the corresponding reference propagation delay, and the node response order conforms to the signal propagation relationship.

[0085] Delay anomaly receiving edge: Delay anomaly receiving edge indicates that the actual transmission interval deviates significantly from the reference propagation delay, and there are node receiving edges with early response, delayed response, and local delay abrupt changes.

[0086] Ideal congruence line: The ideal congruence line represents the reference relationship formed when the actual transmission interval is exactly equal to the reference propagation delay. It is used to intuitively judge the degree of delay deviation of the node receiving edge.

[0087] S3.4: The weighted event propagation features are reattached to the dual-timescale cavitation event connection diagram. Based on the nozzle structure and operating conditions associated with the cavitation event chain, the physical consistency of the structure and operating conditions is checked. The physical connection relationship of the multi-source event features in the corresponding propagation process is determined. The dual-timescale edge relationship coding layer performs constraint marking and relationship coding on the physical connection relationship to generate physical constraint coding information.

[0088] Furthermore, based on the correspondence between cavitation event chains and node receiving edges, the weighted event propagation features are associated with the corresponding cavitation candidate event nodes, and the runtime window, acquisition location, and propagation type of each node are retained. For each cavitation event chain, the corresponding nozzle structure information and operating conditions are extracted. The arrangement order of node acquisition locations is checked along the fuel flow direction, and the structural connection relationship between nodes is checked along the nozzle structure transmission direction. Then, the node response direction, actual transmission interval, and signal propagation relationship under the corresponding operating conditions are aligned, and node receiving edges with opposite propagation directions, conflicting position orders, and transmission timings exceeding the signal propagation range are marked.

[0089] Based on the verified node acceptance edges, weighted event propagation characteristics that conform to the nozzle structure information and operating conditions in terms of propagation direction, acquisition position order, and transmission timing are established as physical acceptance relationships. Weighted event propagation characteristics with verification conflicts retain corresponding constraint tags. The dual-timescale edge relationship coding layer writes relationship types according to two categories: fluid propagation and structural propagation node acceptance edges. The physical acceptance relationships are converted into consistency codes, constraint tags are converted into deviation codes, and associated with the corresponding cavitation candidate event nodes and weighted event propagation characteristics. The coding is then collected according to the node arrangement order of the cavitation event chain to generate physical constraint coding information.

[0090] S3.5: Combine physical constraint encoding information with weighted event propagation features. Through a cross-channel attention fusion layer, reduce the fusion weight of event features that do not conform to the physical succession relationship, and aggregate event features that have the same propagation direction and support each other across channels to generate a cavitation event fusion representation.

[0091] Furthermore, the physical constraint encoding information is linked to the weighted event propagation features item by item according to the correspondence between the cavitation event chain, cavitation candidate event nodes, and node receiving edges, identifying the propagation type, physical receiving state, and cross-channel support state associated with each event feature. The cross-channel attention fusion layer, based on the original fusion weights of the weighted event propagation features, verifies whether the corresponding physical constraint encoding indicates that the propagation direction, acquisition position order, and transmission timing conform to the physical receiving relationship. When there are event features with opposite propagation directions, conflicting position orders, or transmission timings exceeding the signal propagation range, the fusion weights are adjusted according to the deviation state corresponding to the physical constraint encoding, restricting conflicting content from participating in subsequent aggregation.

[0092] For event features that conform to physical continuity, the response direction and temporal continuity status of different signal channels are checked along the node arrangement order of the cavitation event chain. Event features with consistent propagation directions and cross-channel common support are grouped into the same fusion node. The cross-channel attention fusion layer reallocates channel weights based on the degree of common support, performs weighted aggregation on event features pointing to the same cavitation propagation process, and preserves the changes that can be continuously inherited between fluid propagation and structural propagation. Event features that have completed physical constraints and cross-channel fusion according to the aggregation of each cavitation event chain are used to form a cavitation event fusion representation.

[0093] It should be noted that the cavitation event fusion representation is a unified feature representation formed by weighted aggregation of event propagation characteristics from different signal channels, combined with nozzle structure, operating conditions, and physical connections. It retains cavitation changes with consistent propagation direction, continuous timing, and cross-channel support, while weakening interference features such as positional conflicts, opposite propagation directions, and abnormal transmission timing. This information is used to support subsequent cavitation state reasoning and counterfactual contribution judgment.

[0094] S4: Parallel reasoning is performed on the fusion representation of cavitation events to retain and exclude cavitation effects. Cavitation state interpretation information and counterfactual state interpretation information are obtained, and the differences in state interpretation are compared to determine the counterfactual contribution of cavitation. At the same time, a continuous time window credible decision is made by combining the dual-timescale cavitation event transition diagram to generate fuel nozzle cavitation prediction information.

[0095] S4.1: The dual-path inference layer of the input graph structure machine learning network is used to fuse cavitation events. The cavitation state inference branch retains the weighted event propagation features and physical constraint encoding information for state decoding. The counterfactual state inference branch keeps the physical constraint encoding information, node connection relationship and operating condition association unchanged. The weighted event propagation features are replaced with the baseline event propagation features corresponding to normal samples under the same operating conditions. Counterfactual state decoding is then performed to form cavitation state explanation information and counterfactual state explanation information, respectively.

[0096] Furthermore, the cavitation event fusion representation is fed into the dual-path inference layer. Weighted event propagation features and physical constraint encoding information are extracted according to the cavitation event chain, and a correspondence between the two types of information is established based on the order of the cavitation candidate event nodes. The cavitation state inference branch retains the weighted event propagation features corresponding to each cavitation candidate event node, and combines them with the physical constraint encoding information to perform state decoding level by level along the node's connecting edges. This gathers the support direction and contribution degree of each node to the cavitation state, forming cavitation state explanation information.

[0097] The counterfactual state reasoning branch extracts operational segments from normal samples that fall within the same operational range as the current operating condition. These segments are mapped according to signal channel, acquisition location, propagation type, and node arrangement. The average state of the event propagation characteristics for each node's normal operation segment is statistically analyzed to form a baseline event propagation characteristic. Following the order of cavitation candidate event nodes, the current weighted event propagation characteristic is replaced node by node with the baseline event propagation characteristic. Simultaneously, the node connection edges, physical constraint encoding information, operational condition associations, and state decoding parameters in the dual-timescale cavitation event connection graph remain unchanged. Feature transfer and state decoding are then re-executed along the original node connection edges to form counterfactual state explanation information.

[0098] S4.2: The cavitation state interpretation information and the counterfactual state interpretation information are reattached to the dual-time-scaled cavitation event succession diagram according to the runtime window, the state interpretation difference is compared, and the difference comparison content is marked to the corresponding cavitation candidate event node to generate state interpretation difference information.

[0099] Furthermore, the cavitation state inference branch and the counterfactual state inference branch use the same state decoding parameters to output the cavitation occurrence probability of each cavitation candidate event node and its corresponding runtime window. The difference in cavitation occurrence probability of the same node under the two inference branches is determined as the node's counterfactual contribution, characterizing the actual influence of weighted event propagation features on the corresponding cavitation prediction state. According to the start and end positions of the runtime windows, the cavitation state explanation information and the counterfactual state explanation information are associated with the corresponding cavitation event chains in the dual-timescale cavitation event succession diagram, and node-by-node alignment is completed according to the arrangement order of the cavitation candidate event nodes. The runtime windows, node positions, and propagation relationships corresponding to the two types of state explanation information are checked to exclude explanations with cross-time period misalignment and inconsistent node affiliation. Using the aligned cavitation candidate event nodes as the comparison unit, the state explanation formed when retaining weighted event propagation features is compared item by item with the counterfactual state explanation formed when suppressing weighted event propagation features, determining the changes in node state, propagation support direction, and explanation contribution.

[0100] The difference comparison content of each cavitation candidate event node is organized along the connection direction of the cavitation event chain. Nodes with consistent state interpretation are marked as indifferent nodes, and nodes with changed state interpretation are marked as contributing difference nodes. The runtime window and node connection edge corresponding to the difference are recorded. Contributing difference nodes pointing to the same cavitation propagation process within consecutive runtime windows are sequentially aggregated, preserving the connection state of the difference from the starting node to subsequent nodes. The marked difference comparison content is then attached back to the corresponding cavitation candidate event node to form state interpretation difference information.

[0101] S4.3: Based on the state interpretation difference information, perform node relationship encoding and temporal correlation reasoning along the node receiving edge, track the propagation path of the state interpretation difference information between adjacent runtime windows, and form a cavitation counterfactual contribution sequence.

[0102] Furthermore, starting from the cavitation candidate event node that has completed the difference marking in the state interpretation difference information, the preceding and subsequent nodes are retrieved according to the connection direction of the node's receiving edge. The receiving direction, propagation type, and difference state between nodes are converted into node relationship codes. Along the same cavitation event chain, the runtime window and response sequence of adjacent nodes are checked. The difference ending position of the preceding node is sequentially connected to the difference starting position of the subsequent node to confirm whether the state interpretation difference continues along the fluid propagation relationship or the structural propagation relationship. Propagation tracking is terminated for difference content with opposite connection directions, misaligned runtime windows, or lacking node receiving edge support.

[0103] The temporal correlation reasoning proceeds in the order of runtime windows from earliest to latest. Differences in state interpretation between adjacent runtime windows with continuous node connections are sequentially connected, recording the process from the originating node, propagation along node connections to subsequent nodes, and eventual decay of the difference. When multiple node connections exist, each propagation branch is traced based on edge type and response order. When different branches converge on the same node, their corresponding contributions are merged according to arrival time and shared supporting states. The encoded node relationships, runtime window order, and difference continuation states are then aggregated according to the cavitation event chain to form a cavitation counterfactual contribution sequence.

[0104] S4.4: Reattach the cavitation counterfactual contribution sequence to the dual-time-scale cavitation event succession graph, perform contribution continuity verification and node succession integrity verification according to the runtime window sequence, and perform contribution credibility weighting and prediction state decoding to generate fuel nozzle cavitation prediction information.

[0105] Furthermore, according to the cavitation event chain, cavitation candidate event nodes, and runtime windows corresponding to the cavitation counterfactual contribution sequence, the contribution content of each node is reattached to the corresponding position in the dual-timescale cavitation event succession diagram. The contribution direction and continuity status of adjacent nodes are checked sequentially along the runtime windows. If the contribution termination position of the previous runtime window can be connected with the contribution start position of the next runtime window, the continuous contribution relationship is preserved; if a temporal succession cannot be formed, a contribution interruption flag is set at the corresponding node. Then, the contribution sequence is checked along the node succession edges to see if it covers the complete propagation process. Succession anomaly flags are set for positions with missing nodes, broken succession edges, and conflicting node orders.

[0106] Based on the markers formed by the continuity and integrity checks of contributions, the credibility weights of the counterfactual contributions of each cavitation candidate event node are reassigned. Contributions that continuously span runtime windows and have complete node integrity retain their original weights, while contributions with interruptions or abnormal integrity are weighted down according to the number of interruptions and the proportion of missing nodes. The weighted node contributions are aggregated level by level along the cavitation event chain, and the cavitation state of the corresponding runtime window is decoded by the dual-path inference layer to determine the cavitation occurrence state, propagation location, and continuation trend. The decoded content of each runtime window is then collected according to the fuel nozzle runtime sequence to generate fuel nozzle cavitation prediction information.

[0107] like Figure 3As shown, the actual cavitation state within the continuous operating window gradually increases from a low level, entering a concentrated cavitation phase between operating windows 7 and 9, and then gradually decreases as the cavitation process weakens. The predicted probability of the cavitation state inference branch maintains a similar direction of change as the actual cavitation state, forming continuous responses in the cavitation initiation, enhancement, stabilization, and dissipation stages, indicating that the weighted event propagation characteristics and physical constraint encoding information can jointly support cavitation state decoding. The counterfactual state inference branch, while maintaining the physical constraint encoding information, node connection relationships, and operating condition correlations unchanged, suppresses the weighted event propagation characteristics, and its predicted probability is significantly lower than that of the cavitation state inference branch. The difference between the two inference branches forms the cavitation counterfactual contribution, which remains continuously high during the actual cavitation persistence stage and returns to a lower level during the normal operation stage and the cavitation dissipation stage. This change indicates that the cavitation counterfactual contribution can distinguish between actual cavitation with continuous propagation support and occasional anomalies lacking node connection relationships, and reduces misjudgments in a single operating window through cross-operating window contribution continuity verification, improving the interpretability and early warning reliability of fuel nozzle cavitation prediction information.

[0108] It should be noted that the actual cavitation state refers to the actual cavitation degree obtained through fuel nozzle flow field simulation and cavitation state calibration for each operating window.

[0109] Cavitation state inference branch prediction probability: The cavitation state inference branch prediction probability represents the cavitation occurrence probability obtained by decoding while retaining the weighted event propagation characteristics and physical constraint encoding information.

[0110] Counterfactual state reasoning branch prediction probability: The counterfactual state reasoning branch prediction probability represents the cavitation occurrence probability obtained after keeping the physical constraint encoding information and the correlation of operating conditions unchanged and suppressing the weighted event propagation characteristics.

[0111] Cavitation counterfactual contribution: The cavitation counterfactual contribution represents the difference between the predicted probability of the cavitation state inference branch and the predicted probability of the counterfactual state inference branch, and is used to measure the actual contribution of event propagation characteristics to the cavitation predicted state.

[0112] like Figure 4As shown, the benchmark group, the dual-timescale graph structure group, and the complete scheme group are comprehensively compared from six dimensions: accuracy, stability, noise resistance, robustness, interpretability, and early warning reliability. The benchmark group mainly relies on ordinary signal features to complete cavitation judgment, and its performance in terms of noise resistance, robustness, and interpretability is relatively low. The dual-timescale graph structure group, through dual-timescale cavitation event transmission graph and graph structure machine learning feature transfer, incorporates fluid propagation node transmission edges, structural propagation node transmission edges, and cavitation candidate event nodes into a unified representation, and its accuracy, stability, and noise resistance are significantly improved. The complete scheme group further incorporates cross-channel common support weighting, physical constraint encoding, parallel reasoning for cavitation influence retention and exclusion, and continuous time window credible adjudication, maintaining a high level in all six evaluation dimensions. Among them, the improvement in interpretability and early warning reliability indicates that the cavitation state explanation information, counterfactual state explanation information, and cavitation counterfactual contribution sequence can clearly identify the contribution source of the predicted state; the improvement in accuracy, stability, and robustness indicates that the present invention can suppress isolated jumps and channel conflicts, and achieve stable fusion of multi-source cavitation propagation features.

[0113] It should be noted that the benchmark group represents a conventional machine learning prediction scheme that does not incorporate dual-timescaled emptying event succession graphs, counterfactual parallel reasoning, and continuous time-window credible adjudication.

[0114] Dual-timescale graph structure group: The dual-timescale graph structure group represents a scheme that uses fluid propagation nodes to accept edges and structure propagation nodes to accept edges for graph structure feature transfer, but does not include a complete counterfactual adjudication mechanism.

[0115] Complete Solution Group: The solution group represents the present invention solution that simultaneously employs dual-timescale succession orchestration, cross-channel support weighting, physical constraint encoding, counterfactual parallel reasoning, and continuous time window trusted decision-making.

[0116] This embodiment also provides a computer device applicable to the machine learning-based fuel nozzle cavitation prediction method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based fuel nozzle cavitation prediction method proposed in the above embodiment.

[0117] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0118] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based fuel nozzle cavitation prediction method as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0119] In summary, this invention achieves a unified fusion of fluid propagation, structural response, and cross-channel support relationships through dual-timescale inheritance orchestration and graph structure machine learning feature transfer, which is used to suppress isolated jumps and channel conflicts, and improve the accuracy and stability of cavitation identification; through parallel reasoning of cavitation influence retention and exclusion, it achieves continuous adjudication of state interpretation differences and cross-time window counterfactual contributions, which is used to distinguish between true cavitation contributions and occasional anomalies, and improve the interpretability of machine learning predictions and the reliability of early warnings.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A machine learning-based method for predicting cavitation in fuel nozzles, characterized in that, include: Collect multi-source operation information, nozzle structure information and sensor layout information of fuel nozzles, perform time sequence alignment, divide operation windows and calibrate signal propagation range, obtain signal change segments and corresponding operating conditions within the operation window, perform signal channel position mapping and operation segment aggregation, and generate multi-source operation window sequence; Signal change localization and candidate event marking are performed on the multi-source runtime window sequence to obtain cavitation candidate event nodes. Then, the fluid propagation and structure propagation dual-timescale inheritance arrangement is performed to establish node inheritance edges. At the same time, the cavitation candidate event nodes that are successively inherited are merged into cavitation event chains along the connection direction of the node inheritance edges to generate a dual-timescale cavitation event inheritance graph. Based on the dual-timescale cavitation event transmission graph, multi-source event features are extracted and transmitted along the node transmission edges. Event propagation order verification, propagation delay deviation assessment, and cross-channel support weighting are performed. Physical constraint coding is implemented in combination with nozzle structure and operating conditions to generate a fusion representation of cavitation events. Parallel reasoning is used to retain and exclude the cavitation impact of the fusion representation of cavitation events. Cavitation state interpretation information and counterfactual state interpretation information are obtained, and the differences in state interpretation are compared to determine the counterfactual contribution of cavitation. At the same time, continuous time window credible adjudication is performed by combining the dual-timescale cavitation event transition diagram to generate fuel nozzle cavitation prediction information.

2. The machine learning-based fuel nozzle cavitation prediction method as described in claim 1, characterized in that, The steps for generating the multi-source runtime window sequence are as follows: Based on the unified time reference of the collection time of multi-source operation information, duplicate records and time-series conflict records are eliminated, and the signal propagation relationship of nozzle structure information and sensor layout information is calibrated to determine the signal propagation range. At the same time, it is attached to the corresponding collection time period to form a time-aligned operation segment. Based on time-aligned running segments, the continuous acquisition is checked and the running status is segmented. Running windows are divided, and the effective response segment of the running window is limited by the signal propagation range. Signal change segments and corresponding operating conditions are extracted from the effective response segment to form running window sample segments. Based on the sensor layout information, the signal channel position correspondence is analyzed to determine the acquisition position corresponding to each signal channel. The acquisition position is then mapped to the nozzle structure information. The runtime window sample segments are marked with positions and time-series aggregated according to a unified channel order to form a multi-source runtime window sequence.

3. The machine learning-based fuel nozzle cavitation prediction method as described in claim 2, characterized in that, The steps for obtaining cavitation candidate event nodes are as follows: Extract continuous change trajectories from signal change segments of multi-source runtime window sequences. Around the corresponding operating conditions, check the increase / decrease direction and continuity of adjacent signal changes according to the acquisition time. Locate the starting position and recovery position of the signal change from the original increase / decrease trend to abnormal deviation. Eliminate isolated jumps that do not continue to the future, and collect change segments with the same deviation direction and continuous extension along time as candidate event segments. Based on candidate event fragments, signal channel mapping and acquisition location are performed, and timing continuity verification and location consistency checks are conducted. Candidate event fragments with channel mapping conflicts and inconsistent acquisition locations are eliminated, and the remaining candidate event fragments are marked as cavitation candidate event nodes.

4. The machine learning-based fuel nozzle cavitation prediction method as described in claim 1 or 3, characterized in that, The steps for generating the dual-time-scaled emptying event succession graph are as follows: Based on cavitation candidate event nodes, the node response order in the fuel flow direction and nozzle structure transmission direction is checked according to the corresponding signal channel and signal propagation range. Node connection edges are established for cavitation candidate event nodes whose response order and signal propagation range are mutually connected, forming dual-time-scale node connection information. Based on the dual-time-scale node connection information, the cavitation candidate event nodes are sequentially arranged along the connection direction of the node receiving edge. The cavitation candidate event nodes that maintain propagation and acceptance across runtime windows are connected to the same cavitation event chain. The cavitation event chain and the corresponding node receiving edge are aggregated to generate a dual-time-scale cavitation event receiving graph.

5. The machine learning-based fuel nozzle cavitation prediction method as described in claim 4, characterized in that, The steps for extracting and transmitting multi-source event features are as follows: The dual-timescale cavitation event reception graph is input into the graph structure machine learning network. According to the node arrangement order of the cavitation event chain, the signal change features corresponding to each cavitation candidate event node are extracted. The signal change features are then linked to nodes according to the signal channel affiliation to form multi-source event features corresponding to each cavitation candidate event node. The signal change features are converted into node feature representations through the node feature encoding layer. The node reception edges corresponding to fluid propagation and the node reception edges corresponding to structural propagation are respectively encoded by edge type and transmission time sequence through the dual-timescale edge relationship encoding layer. At the same time, the graph message passing layer transmits and updates the node feature representations level by level along the node reception edges to form a node event feature sequence. The graph-structured machine learning network includes a node feature encoding layer, a dual-time-stamped edge relation encoding layer, a graph message passing layer, a cross-channel attention fusion layer, and a dual-path inference layer. The node feature encoding layer takes on the multi-source event features of the cavitation candidate event nodes. The dual-time-stamped edge relation encoding layer takes on the propagation type and temporal relationship of the edges received by the two types of nodes. The graph message passing layer transmits node feature representations based on the edges received by the two types of nodes. The cross-channel attention fusion layer allocates fusion weights based on the degree of common support of different signal channels. The dual-path inference layer retains and excludes the cavitation effect to perform state inference. Based on the node event feature sequence, the response order of adjacent nodes is checked, the propagation delay deviation of adjacent nodes is calculated, and the response direction and duration of each signal channel in the same cavitation event chain are compared to determine the degree of cross-channel common support. The fusion weight of node event features is adjusted through the cross-channel attention fusion layer to generate weighted event propagation features.

6. The machine learning-based fuel nozzle cavitation prediction method as described in claim 5, characterized in that, The steps for generating the cavitation event fusion representation are as follows: The weighted event propagation features are reattached to the dual-timescale cavitation event reception diagram. Based on the nozzle structure and operating conditions associated with the cavitation event chain, the physical consistency of the structure and operating conditions is checked to determine the physical reception relationship of the multi-source event features in the corresponding propagation process. The dual-timescale edge relationship coding layer then performs constraint marking and relationship coding on the physical reception relationship to generate physical constraint coding information. By combining physical constraint encoding information with weighted event propagation features, and through a cross-channel attention fusion layer, the fusion weight of event features that do not conform to physical continuity is reduced, while event features that have the same propagation direction and support each other across channels are aggregated to generate a cavitation event fusion representation.

7. The machine learning-based fuel nozzle cavitation prediction method as described in claim 6, characterized in that, The steps for obtaining cavitation state interpretation information and counterfactual state interpretation information are as follows: The dual-path inference layer of the input graph structure machine learning network is used to fuse cavitation events into a representation. The cavitation state inference branch retains the weighted event propagation features and physical constraint encoding information for state decoding, while the counterfactual state inference branch keeps the physical constraint encoding information, node connection relationships and operating condition associations unchanged. The weighted event propagation features are replaced with the baseline event propagation features corresponding to normal samples under the same operating conditions, and counterfactual state decoding is performed to form cavitation state explanation information and counterfactual state explanation information, respectively. The cavitation state interpretation information and the counterfactual state interpretation information are attached back to the dual-time-scaled cavitation event succession graph according to the runtime window. The state interpretation differences are compared, and the differences are marked to the corresponding cavitation candidate event nodes to generate state interpretation difference information.

8. The machine learning-based fuel nozzle cavitation prediction method as described in claim 7, characterized in that, The steps for generating fuel nozzle cavitation prediction information are as follows: Based on the state interpretation difference information, node relationship encoding and temporal association reasoning are performed along the node bearing edge to trace the propagation path of the state interpretation difference information between adjacent runtime windows and form a cavitation counterfactual contribution sequence. The cavitation counterfactual contribution sequence is reattached to the dual-time-scale cavitation event succession diagram. Contribution continuity verification and node succession integrity verification are performed according to the runtime window sequence. Contribution credibility weighting and prediction state decoding are also performed to generate fuel nozzle cavitation prediction information.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the machine learning-based fuel nozzle cavitation prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the machine learning-based fuel nozzle cavitation prediction method according to any one of claims 1 to 8.

Images