Forklift attachment fatigue life prediction method and system adopting transfer learning

Abstract

The invention relates to the technical field of mechanical life prediction, in particular to a forklift attachment fatigue life prediction method and system adopting transfer learning. Obtaining a fatigue frequent point list and a frequent stress direction list; taking the accessory structure topology, the accessory material parameters, the fatigue frequent point list and the frequent stress direction list as constraints, acquiring a plurality of first accessory operation condition time sequence data and a first label for identifying the fatigue life which are in one-to-one correspondence, and training a fatigue life prediction basic model; and acquiring a plurality of second accessory operation condition time sequence data and a second label for identifying the fatigue life in one-to-one correspondence according to the model of the target forklift accessory, the fatigue frequent point list and the frequent stress direction list, and carrying out transfer learning on the fatigue life prediction basic model to obtain a fatigue life prediction model of the target forklift accessory. And executing a forklift attachment fatigue life prediction task. The problems of insufficient fatigue life prediction precision and low efficiency caused by insufficient fatigue monitoring samples are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical life prediction, in particular to a forklift accessory fatigue life prediction method and system using transfer learning. BACKGROUND

[0002] In forklift operation, the accessory is a key component to realize its multifunctionality, which is divided into carrying type, clamping type, adjusting type, etc., and is suitable for various special working conditions. Its fatigue life is crucial to the safety and stability of operation, so accurate prediction of the fatigue life of the forklift accessory is an important demand in the industry. In the current fatigue life prediction of forklift accessories, one means is to collect historical fatigue data, train a fatigue life prediction model for prediction. However, due to uncertain accessory scheduling and large load fluctuations, it is difficult to obtain sufficient and effective fatigue monitoring samples, resulting in insufficient precision and low efficiency of the trained prediction model in the prior art for modeling and predicting the fatigue life of the forklift accessory, which is difficult to meet the demand for fatigue life prediction of the forklift accessory. SUMMARY

[0003] The present application provides a forklift accessory fatigue life prediction method and system using transfer learning to solve the technical problems of insufficient precision and low efficiency of the trained prediction model due to insufficient fatigue monitoring samples in the prior art.

[0004] The technical solution of the present application to solve the above technical problems is as follows: In a first aspect, the present application provides a forklift accessory fatigue life prediction method using transfer learning, comprising: loading operation monitoring data of a target forklift accessory, performing fatigue position frequent pattern analysis to obtain a fatigue frequent point list; based on the operation monitoring data, traversing the fatigue frequent point list, performing stress direction frequent pattern analysis to obtain a frequent stress direction list; taking the accessory structure topology, accessory material parameters, the fatigue frequent point list and the frequent stress direction list as constraints, collecting a plurality of one-to-one first accessory operation working condition time series data and first identification fatigue life labels, training a fatigue life prediction base model; taking the target forklift accessory model, the fatigue frequent point list and the frequent stress direction list, collecting a plurality of one-to-one second accessory operation working condition time series data and second identification fatigue life labels, performing transfer learning on the fatigue life prediction base model to obtain a target forklift accessory fatigue life prediction model, and performing a forklift accessory fatigue life prediction task.

[0005] Optionally, the work monitoring data of the target forklift accessory is loaded, a fatigue position frequent pattern analysis is performed, and a fatigue frequent point list is obtained, including: collecting fatigue detection record data with accessory material parameters as constraints, performing fatigue symptom type frequent pattern analysis, and obtaining a frequent fatigue symptom type set; extracting forklift accessory defect monitoring data from the work monitoring data, and counting a forklift accessory position set and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom type set appears; and adding a forklift accessory position in which the frequent fatigue type trigger frequency in the frequent fatigue type trigger frequency set is greater than or equal to a trigger frequency threshold to the fatigue frequent point list.

[0006] In the method, the forklift accessory defect monitoring data is extracted from the work monitoring data, and a forklift accessory position set and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom type set appears are counted, including: extracting first group defect monitoring data to Qth group defect monitoring data from the forklift accessory defect monitoring data, where Q represents the number of groups of forklift accessory defect monitoring data; integrating the first group defect monitoring data to the Qth group defect monitoring data according to positions to obtain first position defect monitoring data set to Nth position defect monitoring data set, where N represents the number of positions; deleting sets in which the number of defect monitoring data is less than or equal to a data number threshold to obtain a plurality of remaining position defect monitoring data sets; and traversing the plurality of remaining position defect monitoring data sets, counting a forklift accessory position set and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom type set appears based on the frequent fatigue symptom type set.

[0007] In the method, the forklift accessory defect monitoring data is extracted from the work monitoring data, and a forklift accessory position set and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom type set appears are counted, including: extracting first group defect monitoring data to Qth group defect monitoring data from the forklift accessory defect monitoring data, where Q represents the number of groups of forklift accessory defect monitoring data; integrating the first group defect monitoring data to the Qth group defect monitoring data according to positions to obtain first position defect monitoring data set to Nth position defect monitoring data set, where N represents the number of positions; deleting sets in which the number of defect monitoring data is less than or equal to a data number threshold to obtain a plurality of remaining position defect monitoring data sets; and traversing the plurality of remaining position defect monitoring data sets, counting a forklift accessory position set and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom type set appears based on the frequent fatigue symptom type set.

[0008] Optionally, based on the job monitoring data, the list of fatigue frequent points is traversed, stress direction frequent pattern analysis is performed, and a list of frequent stress directions is obtained, including: extracting a first fatigue frequent position from the list of fatigue frequent points; based on the first fatigue frequent position, extracting a first set of fatigue frequent position stress directions from the job monitoring data; based on an angle deviation threshold, performing clustering on the first set of fatigue frequent position stress directions to obtain a plurality of clusters of first fatigue frequent position stress directions; taking a plurality of first fatigue frequent position centroid stress directions of the plurality of clusters of first fatigue frequent position stress directions as a first set of reconstructed fatigue frequent position stress directions, and simultaneously taking the number of clusters of the plurality of clusters of first fatigue frequent position stress directions as a plurality of initial trigger frequencies of the plurality of first fatigue frequent position centroid stress directions, and adding the plurality of initial trigger frequencies to the first set of reconstructed fatigue frequent position stress direction initial trigger frequencies; until the Lth set of reconstructed fatigue frequent position stress directions and the Lth set of reconstructed fatigue frequent position stress direction initial trigger frequencies are obtained; based on the first set of reconstructed fatigue frequent position stress directions and the first set of reconstructed fatigue frequent position stress direction initial trigger frequencies, until the Lth set of reconstructed fatigue frequent position stress directions and the Lth set of reconstructed fatigue frequent position stress direction initial trigger frequencies, stress direction frequent pattern analysis is performed on the first fatigue frequent position; until stress direction frequent pattern analysis is completed on all fatigue frequent points, and a list of frequent stress directions is output.

[0009] Optionally, under the constraints of the tool structure topology, tool material parameters, the fatigue frequent point list and the frequent force direction list, a plurality of one-to-one corresponding first tool operation working condition time series data and first identification fatigue life labels are collected, a fatigue life prediction base model is trained, including: based on the tool structure topology and the tool material parameters, a first level constraint is constructed; based on the fatigue frequent point list and the frequent force direction list, a second level constraint is constructed; a set of samples of tools to be analyzed is loaded, wherein the set of samples of tools to be analyzed has the same to-be-analyzed structure topology, to-be-analyzed material parameters, to-be-analyzed fatigue frequent point list, to-be-analyzed frequent force direction list, tool operation fatigue frequent point force time series data and tool operation environment monitoring time series information; when the to-be-analyzed structure topology has a structure topology similarity with the tool structure topology greater than or equal to a structure topology similarity threshold, and the to-be-analyzed material parameters are the same as the to-be-analyzed material parameters, it is considered that the first level constraint is met; when the first level constraint is met, and the point position distribution similarity of the fatigue frequent point list and the to-be-analyzed fatigue frequent point list is greater than or equal to a point position distribution similarity threshold, and the force direction distribution similarity of the frequent force direction list and the to-be-analyzed frequent force direction list is greater than or equal to a force direction distribution similarity threshold, it is considered that the second level constraint is met; when the second level constraint is met, the tool operation fatigue frequent point force time series data and the tool operation environment monitoring time series information are set as the first tool operation working condition time series data, and the set of samples of tools to be analyzed is subjected to fatigue symptom duration mode analysis to obtain the first identification fatigue life label.

[0010] Optionally, under the constraints of the target forklift tool model, the fatigue frequent point list and the frequent force direction list, a plurality of one-to-one corresponding second tool operation working condition time series data and second identification fatigue life labels are collected, and transfer learning is performed on the fatigue life prediction base model to obtain a target forklift tool fatigue life prediction model, the fatigue life prediction base model being a model integrating the outputs of a plurality of topologically different sub-models, including: obtaining a plurality of fatigue life prediction base sub-models of the fatigue life prediction base model; constructing a fully connected neural network, and connecting the plurality of fatigue life prediction base sub-models in parallel to the input layer of the fully connected neural network; taking the second tool operation working condition time series data as the input of the plurality of fatigue life prediction base sub-models, and taking the second identification fatigue life label as the output of the fully connected neural network, performing transfer learning on the fatigue life prediction base model to obtain the target forklift tool fatigue life prediction model.

[0011] In a second aspect, the present application provides a forklift tool fatigue life prediction system using transfer learning, comprising: a frequent mode analysis module, configured to load work monitoring data of a target forklift accessory, perform fatigue position frequent mode analysis, and obtain a fatigue frequent point list; a stress direction analysis module, configured to traverse the fatigue frequent point list based on the work monitoring data, perform stress direction frequent mode analysis, and obtain a frequent stress direction list; a basic model training module, configured to collect a plurality of first accessory work condition time sequence data and first label fatigue life corresponding to the fatigue frequent point list and the frequent stress direction list as constraints, and train a fatigue life prediction basic model; a target model training module, configured to collect a plurality of second accessory work condition time sequence data and second label fatigue life corresponding to the fatigue frequent point list and the frequent stress direction list, and perform transfer learning on the fatigue life prediction basic model to obtain a target forklift accessory fatigue life prediction model and perform a forklift accessory fatigue life prediction task.

[0012] By implementing the present application, the work monitoring data of the target forklift accessory can be loaded, fatigue position frequent mode analysis can be performed, and a fatigue frequent point list can be obtained, so that the key positions prone to fatigue can be accurately located, avoiding indiscriminate analysis of all positions of the accessory and reducing the redundant workload of subsequent data processing and model training, thereby improving efficiency. By implementing the present application, the work monitoring data can be traversed based on the fatigue frequent point list, stress direction frequent mode analysis can be performed, and a frequent stress direction list can be obtained, so that the main stress state of each fatigue-prone position can be determined, the correlation between fatigue and stress direction can be revealed, and key stress feature data can be provided for subsequent model training, so that the model can more accurately capture factors affecting fatigue life. By implementing the present application, the fatigue life prediction basic model can be trained by collecting a plurality of first accessory work condition time sequence data and first label fatigue life corresponding to the fatigue frequent point list and the frequent stress direction list as constraints, so that the data used to train the basic model is highly related to the target accessory in core characteristics by double constraint sample screening, thereby improving the generalization ability and accuracy of the basic model. By implementing the present application, the target forklift accessory model, the fatigue frequent point list and the frequent stress direction list, a plurality of one-to-one second accessory operation working condition time sequence data and a second identification fatigue life label are collected, the fatigue life prediction base model is migrated learning, the target forklift accessory fatigue life prediction model is obtained, the forklift accessory fatigue life prediction task is executed, the problem that the monitoring sample is few due to the uncertain scheduling and the large load fluctuation of the target accessory is solved, the existing knowledge of the base model is utilized by means of the migration learning, the dependence on a large number of samples of the target accessory is reduced, the modeling difficulty and cost are reduced, and finally the target model is completely adapted to the specific model of the forklift accessory, can accurately predict the fatigue life, provides a scientific basis for the maintenance and replacement of the accessory, reduces the operation interruption and safety accidents caused by fatigue failure.

[0013] In summary, by implementing the present application, the problem that the monitoring sample is few due to the uncertain scheduling and the large load fluctuation of the forklift accessory is solved by means of migration learning, a precise and efficient forklift accessory fatigue life prediction model is trained to realize the rapid and accurate prediction of the fatigue life of the forklift accessory, further reliable technical support is provided for the whole life cycle management of the forklift accessory, which helps to plan the maintenance strategy in advance, reduces the equipment failure risk, and improves the safety and stability of the forklift operation. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A flowchart of a forklift accessory fatigue life prediction method using migration learning provided by the present application is shown in the figure. Figure 2 A structure diagram of a forklift accessory fatigue life prediction system using migration learning provided by the present application is shown in the figure.

[0015] In the figure, the components represented by each number are as follows: Frequent pattern analysis module 11, stress direction analysis module 12, base model training module 13, target model training module 14. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] In the description of the present application, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified.

[0018] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope consistent with the principles and characteristics disclosed.

[0019] Embodiment one, as Figure 1 shown, the present application provides a forklift accessory fatigue life prediction method and system using transfer learning, comprising: S100: loading the operation monitoring data of the target forklift accessory, performing fatigue position frequent pattern analysis, and obtaining a fatigue frequent point list; S200: based on the operation monitoring data, traversing the fatigue frequent point list, performing stress direction frequent pattern analysis, and obtaining a frequent stress direction list; S300: taking the accessory structure topology, accessory material parameters, the fatigue frequent point list and the frequent stress direction list as constraints, collecting a plurality of one-to-one corresponding first accessory operation working condition time series data and first identification fatigue life labels, and training a fatigue life prediction base model; S400: taking the target forklift accessory model, the fatigue frequent point list and the frequent stress direction list, collecting a plurality of one-to-one corresponding second accessory operation working condition time series data and second identification fatigue life labels, performing transfer learning on the fatigue life prediction base model, obtaining a target forklift accessory fatigue life prediction model, and performing a forklift accessory fatigue life prediction task.

[0020] In step S100 of the embodiments of the present application, the operation monitoring data of the target forklift accessory is loaded, fatigue position frequent pattern analysis is performed, and a fatigue frequent point list is obtained, comprising: Collect fatigue detection record data with the tool material parameters as constraints, perform fatigue symptom type frequent pattern analysis, and obtain a frequent fatigue symptom type set; From the job monitoring data, extract forklift tool defect monitoring data, and count the forklift tool position set of at least any one of the frequent fatigue symptom types in the frequent fatigue symptom type set and the frequent fatigue type trigger frequency set; Add the forklift tool position with a frequent fatigue type trigger frequency greater than or equal to the trigger frequency threshold in the frequent fatigue type trigger frequency set to the fatigue frequent point list.

[0021] In the embodiments of the present application, the core purpose of step S100 is to accurately screen out the key positions with high fatigue problems, i.e., fatigue frequent points, from the massive job data of the target forklift tool, to provide focused core research objects for subsequent stress direction analysis and model training, to avoid low efficiency and resource waste caused by indiscriminate analysis, and to ensure the pertinence and accuracy of subsequent fatigue life prediction from the source.

[0022] First, fatigue detection record data needs to be collected with the tool material parameters as constraints, and fatigue symptom type frequent pattern analysis needs to be performed to obtain a frequent fatigue symptom type set.

[0023] First, the key material attributes of the target forklift tool, i.e., material parameters, need to be extracted. These material parameters that represent material performance directly determine the fatigue failure mode of the tool. The material parameters that need to be determined include but are not limited to basic material categories, mechanical property parameters, and process characteristic parameters.

[0024] The basic material categories include high-strength steel, aluminum alloy, engineering plastic, and composite material. The mechanical property parameters include fatigue limit, tensile strength, yield strength, elastic modulus, and fracture toughness. The process characteristic parameters include heat treatment state, surface treatment method, and forming process.

[0025] Then, the above parameters are converted into quantifiable screening rules to form a "constraint condition list" to ensure that only fatigue detection records matching the material characteristics of the target tool are collected. For example, if the target tool material is "Q355 high-strength steel, fatigue limit 320 MPa, and welded forming", the constraint rule is set as: only collect fatigue detection records of tools with Q355 steel, fatigue limit within 300-340 MPa (allowing ±5% error), and forming process containing welding.

[0026] Then, fatigue position frequent pattern analysis is performed on the fatigue detection record data that meets the constraint conditions The fatigue symptom description in the data is described, and is standardized and classified according to the “failure mode + failure feature”, so as to avoid statistical deviation caused by description difference. The failure mode (fatigue symptom type) is, for example, fatigue crack, fatigue wear, fatigue fracture and the like; and the failure feature (fatigue symptom description) is, for example, linear crack at a welding position, excessive wear at a connecting pin shaft of a tool, sudden fracture of a tool bearing beam and the like.

[0027] A corresponding standardized symptom type label is matched for each fatigue detection record, that is, one record corresponds to one core symptom type, such as “welding position crack” mapped to “fatigue crack”. The total number and frequency of occurrence of each standardized symptom type label in the data set are counted. Then, the standardized symptom type labels are sorted in descending order of frequency of occurrence, and a “frequency threshold” is set. The frequency threshold can be set according to the data volume, for example, the frequency of occurrence ≥ 15%, to screen out the symptom types meeting the definition of “frequency”.

[0028] Then, the fatigue detection record data collected is statistically analyzed, and the fatigue symptom types meeting the frequency threshold are identified, and the high-frequency types are integrated to form a “frequent fatigue symptom type set”.

[0029] In step S100 of the embodiment of the present application, the forklift tool defect monitoring data is extracted from the operation monitoring data, and the forklift tool position set and the frequent fatigue type triggering frequency set of at least any one frequent fatigue symptom type in the frequent fatigue symptom type set are counted, including: The first group of defect monitoring data to the Qth group of defect monitoring data is extracted from the forklift tool defect monitoring data, where Q represents the number of groups of forklift tool defect monitoring data; The first group of defect monitoring data to the Qth group of defect monitoring data is integrated according to the position, to obtain the first position defect monitoring data set to the Nth position defect monitoring data set, where N represents the number of positions; The sets of defect monitoring data with a number less than or equal to a data number threshold are deleted, to obtain a plurality of remaining position defect monitoring data sets; The plurality of remaining position defect monitoring data sets are traversed, and the forklift tool position set and the frequent fatigue type triggering frequency set of at least any one frequent fatigue symptom type in the frequent fatigue symptom type set are counted based on the frequent fatigue symptom type set.

[0030] In the embodiments of the present application, the purpose of the above subdivision step in step S100 is to accurately associate "frequent fatigue symptoms" with "specific occurrence positions" from the defect monitoring data of the forklift accessory, and to quantify the frequency of high-frequency fatigue symptoms at different positions, to provide "position + frequency" double data support for subsequent screening of fatigue frequent points, and to remove invalid information through data screening to ensure the reliability and pertinence of the statistical results.

[0031] Firstly, a first group of defect monitoring data to a Qth group of defect monitoring data need to be extracted from the forklift accessory defect monitoring data, wherein Q represents the number of groups of forklift accessory defect monitoring data. Specifically, a suitable grouping basis needs to be selected according to the collection logic of the defect monitoring data. For example, if the monitoring data is collected according to the daily operation cycle, the data is grouped by "day", and 1 day of monitoring data is 1 group, a total of Q groups (Q is the total number of days); if it is divided according to "sensor collection batch", the single collection data of each monitoring sensor is 1 group, and Q is the total number of sensor collection batches.

[0032] Then, the "forklift accessory defect monitoring data" separated from the operation monitoring data is split into independent data sets according to the above grouping dimensions, and is sequentially identified as "first group of defect monitoring data", "second group of defect monitoring data",..., "Qth group of defect monitoring data", to ensure that the boundaries of each group of data are clear and non-repetitive. The forklift accessory defect monitoring data includes defect type, defect position, occurrence time, and other information.

[0033] Further, the first group of defect monitoring data to the Qth group of defect monitoring data need to be integrated according to the position to obtain a first position defect monitoring data set to an Nth position defect monitoring data set, wherein N represents the number of positions.

[0034] Firstly, the key structure positions of the target forklift accessory need to be standardizedly labeled, covering all possible defect positions. For example, for a carrying accessory such as a push-pull device, possible defect positions include the slide plate connection point, the push-pull arm root, and the guide wheel support; for a clamping accessory such as a paper roll clamp, possible defect positions include the middle part of the clamping arm, the connection between the clamping arm and the drive shaft, and the buffer pad mounting position; for an adjusting accessory such as a distance-adjusting fork, possible defect positions include the fork body telescopic guide rail, the hydraulic cylinder connection flange, and the fork distance adjusting gear box.

[0035] The first group to the Qth group of defect monitoring data are traversed, and for each defect record, the corresponding standardized position label is matched according to the "occurrence position" information. All defect records under the same label, regardless of which group they come from, are aggregated to form a "first position defect monitoring data set" (corresponding to the first position label) to an "Nth position defect monitoring data set" (corresponding to the Nth position label), N being the total number of standardized position labels.

[0036] Further, the set of defect monitoring data whose number is less than or equal to the data number threshold value is deleted, and a plurality of remaining position defect monitoring data sets are obtained.

[0037] The data number threshold value can be set according to the total amount of data, the number of positions and industry experience, for example: if the total number of defect records is 1000, and there are 20 key positions of the tool (N=20), the data number threshold value can be set as "data number≥20", to ensure that the sample size of each remaining data set is sufficient to reflect the defect law of the position; if some positions have simple structure and low probability of defect, the data number threshold value can be appropriately reduced, such as≥10.

[0038] The number of defect records contained in each data set in the "first position to Nth position defect monitoring data set" is counted, the data set whose data number is less than or equal to the data number threshold value is deleted, and the remaining data set is the "remaining position defect monitoring data set".

[0039] In step S100 of the embodiment of the present application, the plurality of remaining position defect monitoring data sets are traversed, and based on the set of frequent fatigue symptom types, a set of forklift tool positions and a frequent fatigue type trigger frequency set in which at least any one of the frequent fatigue symptom types appears are counted, including: The first frequent fatigue symptom type to the Qth frequent fatigue symptom type is extracted from the set of frequent fatigue symptom types, wherein Q represents the number of frequent fatigue symptom types; The first remaining position defect type set is extracted from the plurality of remaining position defect monitoring data sets; The trigger frequency of the first frequent fatigue symptom type in the first remaining position defect type set is counted and set as the first trigger frequency; The trigger frequency of the Qth frequent fatigue symptom type in the first remaining position defect type set is counted and set as the Qth trigger frequency; The first trigger frequency to the Qth trigger frequency is summed and set as the first remaining position frequent fatigue type trigger frequency; When the first remaining position frequent fatigue type trigger frequency is not equal to 0, the first remaining position frequent fatigue type trigger frequency is added to the frequent fatigue type trigger frequency set, and the first remaining position is added to the set of forklift tool positions.

[0040] In the embodiments of the present application, the purpose of the above-mentioned subdivision step in step S100 is to perform refined “frequent fatigue symptom-position-frequency” correlation statistics on the screened effective position defect data, i.e., the retained position defect monitoring data set, accurately identify the tool positions that have appeared at least one high-frequency fatigue symptom, and quantify the total frequency of the positions appearing high-frequency fatigue symptoms, thereby providing accurate and quantifiable core basis for finally determining the “fatigue frequent point” and avoiding omissions or deviations caused by single symptom statistics.

[0041] First, each frequent fatigue symptom type needs to be extracted from the set of frequent fatigue symptom types in order, and sequentially named as the first frequent fatigue symptom type to the Qth frequent fatigue symptom type, wherein Q represents the number of frequent fatigue symptom types.

[0042] Then, a specific “retained position” is selected from the retained position defect monitoring data set, i.e., the first retained position, which is an effective position screened by data volume, such as the middle part of the clamp arm. The “defect types” corresponding to all defect records of the position, such as cracks, deformations, wear, looseness, etc., are extracted, and these defect types are integrated to form a “first retained position defect type set”, which provides an analysis object for subsequent statistics of the occurrence of high-frequency symptoms at the position.

[0043] Further, the triggering frequency of the first frequent fatigue symptom type in the first retained position defect type set needs to be calculated and set as the first triggering frequency. Optionally, the calculation method can be first triggering frequency = (the number of times of the occurrence of a certain frequent fatigue symptom type in the position defect type set) ÷ (the total number of defect records of the position defect type set) × 100%.

[0044] Taking “the first frequent fatigue symptom type” fatigue crack as an example, the number of times of its occurrence in the “first retained position defect type set”, such as the defect type set of the middle part of the clamp arm, is counted. Assuming that the total number of defect records is 50 and fatigue cracks occur 25 times, then the first triggering frequency = 25 / 50 × 100% = 50%.

[0045] In the same way, the triggering frequencies of the second frequent fatigue symptom type to the Qth frequent fatigue symptom type at the position are counted in turn, and the second triggering frequency to the Qth triggering frequency are obtained.

[0046] Then, the first triggering frequency to the Qth triggering frequency obtained in the above step are summed up. For example, the three triggering frequencies of the above-mentioned middle part of the clamp arm are 50%, 20%, and 10% respectively, and then the first retained position frequent fatigue type triggering frequency = 50% + 20% + 10% = 80%.

[0047] Finally, when the frequency of frequent fatigue type triggering at the first retention position is not equal to 0, the frequency of frequent fatigue type triggering at the first retention position is added to the set of frequent fatigue type triggering frequencies, and the first retention position is added to the set of forklift attachment positions.

[0048] Furthermore, it is necessary to add the forklift attachments whose frequent fatigue type trigger frequencies are greater than or equal to the trigger frequency threshold to the list of frequent fatigue points.

[0049] The trigger frequency threshold is determined by combining industry standards, attachment usage scenarios, and historical fault data, such as 30%.

[0050] Organize the forklift attachment location set and the frequent fatigue type trigger frequency set obtained in the early stage, and establish a one-to-one correspondence, such as storing "middle of clamp arm, 80%" and "fork body telescopic guide rail, 25%" in a table or dictionary.

[0051] The corresponding relationships are iterated through, and the trigger frequency of each position is compared with the trigger frequency threshold. Positions with trigger frequencies ≥ the trigger frequency threshold are included in the list of frequent fatigue points, such as the middle of the clamping arm with a trigger frequency of 80%. Positions with trigger frequencies below the trigger frequency threshold are excluded, such as the telescopic guide rail of the fork with a trigger frequency of 25%. The list of frequent fatigue points is thus formed.

[0052] In step S200 of this application embodiment, based on the work monitoring data, the list of frequent fatigue points is traversed, and a frequent force direction pattern analysis is performed to obtain a list of frequent force directions, including: Extract the first frequent fatigue position from the list of frequent fatigue points; Based on the first frequently fatigued locations, the force directions of the first set of frequently fatigued locations are extracted from the work monitoring data; Based on the angle deviation threshold, clustering is performed on the force direction of the first group of frequent fatigue locations to obtain multiple clusters of the force direction of the first frequent fatigue locations. Take the multiple first fatigue frequent position centroid force directions of the multiple cluster first fatigue frequent position force directions and set them as the first group of reconstructed fatigue frequent position force directions. At the same time, set the number of clusters of the multiple cluster first fatigue frequent position force directions as multiple initial trigger frequencies of the multiple first fatigue frequent position centroid force directions and add them to the first group of reconstructed fatigue frequent position force directions initial trigger frequencies. Until the force direction of the frequently reconstructed fatigue location in the Lth group and the initial triggering frequency of the force direction of the frequently reconstructed fatigue location in the Lth group are obtained; Based on the force direction of the first group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the first group of reconstructed fatigue frequent locations, until the force direction of the Lth group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the Lth group of reconstructed fatigue frequent locations, a frequent force direction pattern analysis is performed on the first fatigue frequent location. Continue until the frequent stress direction pattern analysis is completed for all fatigue-prone points, and output a list of frequent stress directions.

[0053] In this embodiment of the application, the purpose of step S200 is to accurately extract and analyze the force direction characteristics of each location in operation for the "frequent fatigue points" determined in S100, screen out the high-frequency force direction of each frequent fatigue point, provide key "frequent fatigue location - force direction" correlation data for the subsequent construction of fatigue life prediction model, reveal the intrinsic relationship between force direction and fatigue failure, and improve the accuracy of fatigue life prediction model prediction.

[0054] First, the first position needs to be extracted from the list of frequently fatigued points in sequence as the first frequently fatigued position. For example, the "middle of the clamping arm" is selected as the first frequently fatigued position, which is the first analysis object.

[0055] Then, based on the first frequently fatigued location, the force directions of the first set of frequently fatigued locations need to be extracted from the operation monitoring data. The operation monitoring data contains force monitoring information for each position of the attachment, typically collected by force sensors and angle sensors installed at key positions of the attachment, recording the magnitude, direction, and angle of the force at different times. Here, a filtering condition needs to be set as "monitoring location = first frequently fatigued location," such as "monitoring location = middle of the clamp arm." All force direction data meeting this condition are extracted from the operation monitoring data. This force direction data includes force direction angle values ​​for different operation scenarios and different time points, such as 0°, 45°, 90°, etc. These data are then integrated into a dataset of "force directions of the first set of frequently fatigued locations."

[0056] Furthermore, based on the attachment's operating scenario and mechanical analysis requirements, reasonable angular deviation thresholds, such as ±5° or ±10°, need to be set. That is, when the angular difference between two force directions is less than or equal to the angular deviation threshold, they are considered "similar directions" and can be grouped into the same cluster. For example, if the angular deviation threshold is set to ±5°, then force directions at 30°, 33°, and 35° can be grouped into one cluster.

[0057] Clustering algorithms suitable for angular data, such as K-means clustering and density clustering (DBSCAN), are used to cluster the data on "force directions at frequent fatigue locations in the first group". Taking K-means as an example, the algorithm automatically divides force directions with similar angles into multiple clusters, ultimately obtaining "multi-cluster force directions at frequent fatigue locations in the first group", such as cluster 1: 25°-35°, cluster 2: 85°-95°, and cluster 3: 175°-185°.

[0058] Furthermore, it is necessary to take the multiple first fatigue frequent position centroid force directions of the multiple clusters of first fatigue frequent position force directions and set them as the first group of reconstructed fatigue frequent position force directions. At the same time, the number of clusters of the multiple clusters of first fatigue frequent position force directions is set as multiple initial trigger frequencies of the multiple first fatigue frequent position centroid force directions and added to the first group of reconstructed fatigue frequent position force directions initial trigger frequencies.

[0059] First, it is necessary to calculate the direction of force on the center of mass. For each cluster of force directions, calculate the average value of the angles of all force directions within that cluster, i.e., the "center of mass direction", which serves as the representative force direction for that cluster. For example, cluster 1 (25°-35°) contains angle data [25°, 28°, 30°, 32°, 35°], and its center of mass direction = (25+28+30+32+35) / 5 = 30°. This 30° is one of the "center of mass force directions at the first frequently fatigued location".

[0060] Next, it is necessary to count the number within each cluster, that is, to count the number of force direction data entries contained in each cluster. This number directly reflects the frequency of occurrence of the corresponding centroid direction and serves as the "initial trigger frequency". For example, if cluster 1 contains 5 data entries, then its corresponding initial trigger frequency = 5.

[0061] Finally, the "centroid force direction" of all clusters is integrated into "force direction of the first group of reconstructed fatigue frequent position", such as [30°, 90°, 180°], and the corresponding "initial trigger frequency" is integrated into "initial trigger frequency of the first group of reconstructed fatigue frequent position force direction", such as [5, 8, 3], to ensure that the elements of the two datasets correspond one-to-one, such as 30° corresponding to 5, 90° corresponding to 8, and 180° corresponding to 3.

[0062] Repeat the above operation sequentially for "force direction of frequent fatigue positions in the second group" and "force direction of frequent fatigue positions in the third group" to obtain "force direction and initial trigger frequency of frequent fatigue positions in the second group" and "force direction and initial trigger frequency of frequent fatigue positions in the third group" respectively, until finally obtaining the force direction of frequent fatigue positions in the Lth group and the initial trigger frequency of the force direction of frequent fatigue positions in the Lth group.

[0063] Furthermore, based on the force direction of the first group of reconstructed frequently fatigued locations and the initial trigger frequency of the force direction of the first group of reconstructed frequently fatigued locations, up to the force direction of the Lth group of reconstructed frequently fatigued locations and the initial trigger frequency of the force direction of the Lth group of reconstructed frequently fatigued locations, a frequent force direction pattern analysis needs to be performed on the first frequently fatigued location. The purpose of this step is to integrate the reconstruction data of all batches at this location, count the total trigger frequency of the force direction of each centroid, filter out the frequently occurring force directions, and determine the core force direction characteristics of this frequently fatigued point.

[0064] The analysis logic for the frequent pattern analysis of force direction here is the same as that for the frequent pattern analysis of fatigue location and the frequent pattern analysis of fatigue symptom type mentioned above. That is, summarizing the total trigger frequency, setting a frequency threshold, and filtering the high-frequency force direction at the location based on the frequency threshold. The specific analysis process will not be elaborated here.

[0065] Finally, the selected high-frequency force directions, such as 30° and 90°, are used as the frequent force directions for the first frequent fatigue locations. Then, the frequent force direction pattern analysis is performed for all locations in the "Fatigue Frequent Point List," and the high-frequency force directions for all locations are integrated to form the final "Frequent Force Direction List." This list contains two types of data: one-to-one corresponding fatigue frequent locations and high-frequency force directions. For example, the clamping arm middle section is at -30° and 90°; the fork telescopic guide rail is at -45° and 135°; and the hydraulic cylinder connecting flange is at -0° and 180°.

[0066] In step S300 of this embodiment, constrained by the attachment structure topology, attachment material parameters, the list of frequent fatigue points, and the list of frequent force directions, several one-to-one corresponding first attachment operating condition time-series data and first labels identifying fatigue life are collected to train a fatigue life prediction basic model, including: Based on the attachment structure topology and the attachment material parameters, a first-level constraint is constructed; Based on the list of frequent fatigue points and the list of frequent force directions, construct a secondary constraint. Load the sample set of attachments to be analyzed, wherein the sample set of attachments to be analyzed has the same topology of the structure to be analyzed, parameters of the material to be analyzed, list of frequent fatigue points to be analyzed, list of frequent force directions to be analyzed, time series data of force on frequent fatigue points of attachments, and time series information of attachment operation environment monitoring. When the topological similarity between the structure topology to be analyzed and the topological similarity between the attachment structure topology is greater than or equal to the topological similarity threshold, and the material parameters to be analyzed are the same as the material parameters to be analyzed, it is considered that the first-level constraint is satisfied. When the first-level constraint is satisfied, and the similarity of the point distribution between the list of frequent fatigue points and the list of frequent fatigue points to be analyzed is greater than or equal to the point distribution similarity threshold, and the similarity of the force direction distribution between the list of frequent force directions and the list of frequent force directions to be analyzed is greater than or equal to the force direction distribution similarity threshold, it is considered that the second-level constraint is satisfied. When the secondary constraint is satisfied, the stress time series data of the frequent fatigue points of the attachment and the time series information of the attachment working environment monitoring are set as the first attachment working condition time series data. The mode analysis of the duration of fatigue symptoms is performed on the attachment sample set to be analyzed to obtain the first label of fatigue life.

[0067] In this embodiment of the application, the purpose of step S300 is to select sample data that highly matches the core characteristics of the target forklift attachment through dual constraints, and to train a fatigue life prediction basic model with strong generalization ability and high prediction accuracy based on these high-quality data, so as to lay a reliable model foundation for subsequent transfer learning to adapt to the target attachment and solve the prediction bias problem caused by poor sample fit in traditional modeling.

[0068] First, based on the attachment structure topology and the attachment material parameters, a first-level constraint needs to be constructed. The attachment structure topology is the key structural features of the extracted target attachment, including structural type, such as the "slide-push-pull arm" connection structure of a handling push-pull device, and the "double clamp arm-drive shaft" symmetrical structure of a clamping paper roll clamp; key component dimensions, such as clamp arm length and connecting shaft diameter; load transmission path, such as the force transmission route from the contact end to the frame when force is applied, etc., forming a "target structure topology parameter set".

[0069] The attachment material parameters are the material parameters determined in S100, including basic material category, mechanical property parameters, process characteristic parameters, etc., forming a "target material parameter set".

[0070] Then, the "target structural topology parameter set" and the "target material parameter set" are used as hard conditions for screening samples. The rule is stated as: "The structural topology of the attachment sample to be analyzed must be highly similar to the structural topology of the target attachment, and the material parameters must be completely consistent with the material parameters of the target attachment."

[0071] Furthermore, secondary constraints need to be constructed based on the list of frequent fatigue points and the list of frequent force directions. These two lists serve as supplementary filtering conditions, and the rule can be expressed as: "The distribution of frequent fatigue point locations in the sample to be analyzed must be highly similar to that of the target attachment, and the distribution of frequent force directions at each frequent fatigue point must be highly similar to that of the target attachment."

[0072] Furthermore, a sample set of attachments to be analyzed needs to be loaded. This sample set can be derived from various sources, including historical operational data from enterprises, publicly available industry databases, and laboratory simulation test data. Each attachment sample must possess the following six core data categories to ensure it matches the constraints and training requirements: the topology of the structure to be analyzed, the parameters of the materials to be analyzed, a list of frequently occurring fatigue points to be analyzed, a list of frequently occurring force directions to be analyzed, the time-series data of the force at the frequently occurring fatigue points of the attachment, and the time-series information of the attachment's operating environment monitoring.

[0073] The topology of the structure to be analyzed refers to the structural characteristic parameters of the sample attachment, which correspond to the target structure topology parameter set. The material parameters to be analyzed refer to the material property parameters of the sample attachment, which correspond to the target material parameter set. The stress time-series data of frequent fatigue points during attachment operation refers to the stress data of each frequent fatigue point changing over time during the operation of the sample attachment, such as the magnitude and direction of the force per second. The time-series information of the attachment's operating environment monitoring includes data on the changes of environmental parameters during the operation of the sample attachment over time, such as temperature, humidity, and the fluctuation range of the operating load.

[0074] Then, the topology of the structure to be analyzed in the attachment sample is compared with the "target structure topology parameter set" of the target attachment, including structural type matching degree, key component size deviation rate, load transfer path overlap degree, etc., and the "structural topology similarity" is calculated by weighting. A structural topology similarity threshold is set, such as 85%. When the structural topology similarity of the attachment sample to be analyzed is greater than or equal to the structural topology similarity threshold, it is determined to be a structural topology match.

[0075] Next, the "material parameters to be analyzed" of the sample of the attachment to be analyzed are compared item by item with the "target material parameter set" of the target attachment. For example, it is determined whether the materials are the same, whether the fatigue limit deviation is ≤3%, and whether the heat treatment method is consistent. When all parameters are completely matched, or the deviation of key parameters is within the allowable range, it is determined that the material parameters are consistent.

[0076] Only when the sample of the attachment to be analyzed simultaneously satisfies "structural topological similarity ≥ structural topological similarity threshold" and "material parameters are consistent" is it considered to satisfy the first-level constraint; otherwise, the sample of the attachment to be analyzed is removed.

[0077] Furthermore, when the sample of the attachment to be analyzed satisfies the first-level constraint, and the similarity of the point distribution between the list of frequent fatigue points and the list of frequent fatigue points to be analyzed is greater than or equal to the point distribution similarity threshold, and the similarity of the force direction distribution between the list of frequent force directions and the list of frequent force directions to be analyzed is greater than or equal to the force direction distribution similarity threshold, it is considered to satisfy the second-level constraint.

[0078] Specifically, the determination of the similarity of the distribution of points involves calculating the overlap rate of the points.

[0079] Specifically, the "Fatigue Frequent Point List" of the attachment sample to be analyzed is compared with the "Fatigue Frequent Point List" of the target attachment. The proportion of overlapping positions to the total number of positions in the target list is counted. For example, if the fatigue frequent point list has 3 positions and the fatigue frequent point list to be analyzed has 2 overlapping positions, the overlap rate is approximately 66.7%, and this proportion is the "point distribution similarity". A point distribution similarity threshold is set, such as ≥80%. When the point distribution similarity of the attachment sample to be analyzed is greater than or equal to the point distribution similarity threshold, it is determined to be a fatigue frequent point distribution match.

[0080] Specifically, the determination of the similarity of the force direction distribution is to calculate the direction matching degree one by one according to the position.

[0081] Specifically, for each overlapping fatigue-frequent point, the "frequently applied force direction to be analyzed" of the attachment sample at that location is compared with the "frequently applied force direction" of the target attachment at that location. The proportion of the number of directions with an angle deviation ≤ an angle threshold is counted out of the total number of target directions. As mentioned, the angle threshold can be ±5°. If there are two force directions at the target location and the attachment sample to be analyzed matches one, then the direction matching degree at that location is 50%. The average of the direction matching degrees for all overlapping locations is taken as the "force direction distribution similarity" of the attachment samples being analyzed.

[0082] Set a similarity threshold, such as 75%: when the similarity of the force direction distribution of the analyzed attachment samples is greater than or equal to the force direction distribution similarity threshold, it is determined that the force direction distribution matches.

[0083] A sample is considered to satisfy the second-level constraint only if it satisfies the first-level constraint and simultaneously satisfies both the "similarity of point distribution ≥ the similarity threshold of point distribution" and the "similarity of force direction distribution ≥ the similarity threshold of force direction distribution"; otherwise, the sample is discarded.

[0084] Furthermore, when the secondary constraint is satisfied, the stress time series data of the frequent fatigue points of the attachment and the time series information of the attachment working environment monitoring are set as the first attachment working condition time series data. The mode analysis of the duration of fatigue symptoms is performed on the attachment sample set to be analyzed to obtain the first label of fatigue life.

[0085] For samples that meet the second-level constraints, their "force timing data of frequent fatigue points of attachments" and "time series information of attachment working environment monitoring" are integrated to form the input feature data of the fatigue life prediction basic model, namely the time series data of the first attachment working condition - this data contains the core dynamic factors affecting the fatigue life of attachments.

[0086] Then, a "mode analysis of the duration of fatigue symptoms" is performed on the "sample set of attachments to be analyzed". This involves statistically analyzing the duration from when the sample attachment is put into use until the first occurrence of any symptom in the "set of frequent fatigue symptom types", which is the fatigue life duration. The mode of the duration data under multiple identical working conditions is taken, which is the duration value with the highest frequency. This mode is the "label of the first fatigue life". Using the mode here can reduce the impact of a single abnormal data on the accuracy of the label.

[0087] Furthermore, it is necessary to train a basic model for predicting fatigue life.

[0088] Considering the intended use of the fatigue life prediction model, it can be constructed using a method of integrating multiple topological sub-models and fusing output means. The fatigue life prediction model comprises three sub-models: an LSTM sub-model, a CNN-LSTM sub-model, and a GRU sub-model.

[0089] The LSTM sub-model is adapted to the stress time series data of frequent fatigue points of the attachment, and captures the long-term dependencies of the time series data, such as the influence of stress on fatigue accumulation in different operation stages. The structure consists of an input layer, an LSTM layer, a fully connected layer, and an output layer.

[0090] The CNN-LSTM sub-model is adapted and fused with frequent fatigue stress and temporal environmental data. Local features are extracted by CNN and long-term patterns are captured by LSTM. The structure is "input layer, CNN layer, LSTM layer, fully connected layer, output layer".

[0091] The GRU sub-model is adapted to the integrated job condition time series data, simplifies the network structure and improves training efficiency, and complements LSTM. The structure is "input layer, GRU layer, fully connected layer, output layer".

[0092] After the three sub-models output their predicted values, the arithmetic mean is taken as the final output of the base model.

[0093] In the parameter settings of the fatigue life prediction basic model, the optimizer is Adam, the learning rate is set to 0.001, the loss function is mean squared error (MSE), the batch size is 32, L2 regularization is used, the coefficient is 0.0001, and training stops if the loss on the validation set does not decrease for 5 consecutive rounds.

[0094] The training samples for the fatigue life prediction model are time-series data of the first attachment's operating conditions and labels indicating fatigue life. The sample size is no less than 1200 sets, with 80% for training and 20% for validation. The maximum number of training rounds is set to 100.

[0095] The convergence criterion for the basic model for fatigue life prediction is MSE ≤ 0.001 for the training set and MSE ≤ 0.0015 for the validation set.

[0096] In step S400 of this application embodiment, using the target forklift attachment model, the list of frequent fatigue points, and the list of frequent force directions, several corresponding second attachment operating condition time-series data and second fatigue life labels are collected. Transfer learning is then performed on the fatigue life prediction basic model to obtain the target forklift attachment fatigue life prediction model. The fatigue life prediction basic model is an integrated model of the output averages of multiple sub-models with different topologies, including: Multiple fatigue life prediction sub-models are obtained from the fatigue life prediction basic model. Construct a fully connected neural network and connect multiple basic sub-models for fatigue life prediction in parallel to the input layer of the fully connected neural network; Using the second attachment's operating condition time series data as input to multiple fatigue life prediction sub-models, and using the second label identifying fatigue life as the output of the fully connected neural network, transfer learning is performed on the fatigue life prediction base model to obtain the target forklift attachment fatigue life prediction model.

[0097] In this embodiment of the application, the purpose of step S400 is to adapt the fatigue life prediction basic model trained in step S300 to a specific target forklift attachment through transfer learning, thereby obtaining a more targeted and accurate target forklift attachment fatigue life prediction model, solving the problem of prediction deviation of the basic model for specific attachment models, while retaining the general fatigue laws learned by the basic model.

[0098] First, it is necessary to obtain multiple basic sub-models for fatigue life prediction from the basic model for fatigue life prediction. That is, from the basic model for fatigue life prediction trained in step S300, all sub-models that constitute the integrated model are extracted, such as LSTM sub-model, CNN-LSTM sub-model, and GRU sub-model. These sub-models are the "multiple basic sub-models for fatigue life prediction", and the pre-training parameters of each sub-model are retained.

[0099] Next, a fully connected neural network needs to be constructed, and multiple basic sub-models for fatigue life prediction are connected in parallel to the input layer of the fully connected neural network.

[0100] Specifically, a fully connected network with 2-3 hidden layers needs to be designed. For example, the number of neurons in the input layer equals the number of basic sub-models. If there are 3 basic sub-models, the input layer has 3 neurons, the number of neurons in hidden layer 1 is 16, the number of neurons in hidden layer 2 is 8, the number of neurons in the output layer is 1, and the output is the predicted fatigue life value of the target attribute. The activation function used is ReLU.

[0101] Then, the outputs of the extracted basic sub-models for fatigue life prediction are connected to the input layer of the newly constructed fully connected neural network, forming a transfer learning architecture of "parallel connection of basic sub-models → integration of fully connected network". This retains the feature extraction capability of the basic sub-models and adapts the target attachment characteristics through the fully connected network.

[0102] Furthermore, the second attachment's operating condition time series data is used as the input to multiple fatigue life prediction sub-models, and the second label identifying fatigue life is used as the output of the fully connected neural network. Transfer learning is then performed on the fatigue life prediction base model to obtain the target forklift attachment fatigue life prediction model.

[0103] The "second attachment operating condition time series data" is collected as input, constrained by the target forklift attachment model, a list of frequent fatigue points, and a list of frequent force directions. It includes the force time series data of frequent fatigue points of that attachment model and the time series information of the operating environment. The "second label identifying fatigue life" is the life label obtained by performing a mode analysis of the duration of fatigue symptoms on the target attachment model sample.

[0104] During implementation, the "second attachment working condition time series data" needs to be input into multiple fatigue life prediction basic sub-models simultaneously. Each sub-model outputs preliminary prediction values, which are used as inputs to the fully connected neural network.

[0105] Then, targeting the "second label for fatigue life", the parameters of the fully connected neural network are optimized through backpropagation to minimize the mean square error between the predicted value and the label. After training, the overall model consisting of "parallel basic sub-models + fully connected network" is the "target forklift attachment fatigue life prediction model", which can be directly used to predict the fatigue life of this type of attachment.

[0106] In summary, by implementing the forklift attachment fatigue life prediction method using transfer learning provided in this embodiment, at least the following technical effects can be achieved: By using transfer learning, an architecture integrating parallel basic sub-models and a fully connected network is constructed. This reuses the general knowledge already learned by the basic sub-models, significantly reducing the reliance on a large number of monitoring samples for the target attachments. This solves the problem of insufficient monitoring samples for forklift attachments due to scheduling uncertainties and large load fluctuations, thus enabling the training of an accurate and efficient fatigue life prediction model for forklift attachments.

[0107] By performing "frequent fatigue location pattern analysis" on the target forklift attachment operation monitoring data, key locations with high fatigue incidence are accurately screened, forming a list of frequent fatigue points. This avoids indiscriminate analysis of all attachment locations, reduces data processing time and manpower costs, and improves the efficiency of the overall prediction process.

[0108] 3. Based on the list of frequent fatigue points, further "frequent force direction pattern analysis" is performed. Combined with the operation monitoring data, the main force directions of each high-fatigue location are extracted to form a list of frequent force directions. This avoids the prediction bias caused by neglecting key force characteristics in traditional methods and improves the accuracy of the model.

[0109] Example 2, as Figure 2 As shown, based on the same inventive concept as the forklift attachment fatigue life prediction method using transfer learning provided in Embodiment 1, this embodiment of the invention also provides a forklift attachment fatigue life prediction system using transfer learning, comprising: The frequent pattern analysis module 11 is used to load the operation monitoring data of the target forklift attachment, perform fatigue position frequent pattern analysis, and obtain a list of frequent fatigue points. Force direction analysis module 12 is used to traverse the list of frequent fatigue points based on the operation monitoring data, perform frequent force direction pattern analysis, and obtain a list of frequent force directions. The basic model training module 13 is used to collect several one-to-one corresponding first attachment working condition time series data and first label identifying fatigue life, with attachment structure topology, attachment material parameters, the list of frequent fatigue points and the list of frequent force directions as constraints, and to train the fatigue life prediction basic model. The target model training module 14 is used to collect several corresponding second attachment working condition time series data and second fatigue life labels based on the target forklift attachment model, the list of frequent fatigue points and the list of frequent force directions, and to perform transfer learning on the fatigue life prediction basic model to obtain the target forklift attachment fatigue life prediction model and execute the forklift attachment fatigue life prediction task.

[0110] Furthermore, the frequent pattern analysis module 11 includes the following execution steps: Constrained by the material parameters of the attachments, fatigue detection data is collected, and frequent pattern analysis of fatigue symptom types is performed to obtain a set of frequent fatigue symptom types. Extract forklift attachment defect monitoring data from the operation monitoring data, and statistically analyze the forklift attachment location set and the frequent fatigue symptom trigger frequency set that show at least one frequent fatigue symptom type in the frequent fatigue symptom type set. The forklift attachments whose frequent fatigue type trigger frequencies are greater than or equal to the trigger frequency threshold are added to the list of frequent fatigue points.

[0111] Specifically, from the operation monitoring data, forklift attachment defect monitoring data is extracted, and the set of forklift attachment locations and the set of frequent fatigue symptom trigger frequencies that occur at least one frequent fatigue symptom type in the set of frequent fatigue symptom types are statistically analyzed, including: From the forklift attachment defect monitoring data, extract the first set of defect monitoring data up to the Qth set of defect monitoring data, where Q represents the number of sets of forklift attachment defect monitoring data; The first set of defect monitoring data is integrated according to location up to the Qth set of defect monitoring data to obtain the first location defect monitoring dataset up to the Nth location defect monitoring dataset, where N represents the number of locations; Delete sets whose defect monitoring data are less than or equal to the data quantity threshold to obtain several remaining defect monitoring datasets. Traverse the several retained location defect monitoring datasets, and based on the frequent fatigue symptom type set, statistically analyze the forklift attachment location set and the frequent fatigue type trigger frequency set that occur at least one frequent fatigue symptom type in the frequent fatigue symptom type set.

[0112] Specifically, the process involves traversing the plurality of retained location defect monitoring datasets, and based on the set of frequent fatigue symptom types, statistically analyzing the forklift attachment location set and the frequent fatigue type trigger frequency set that exhibit at least one frequent fatigue symptom type from the set of frequent fatigue symptom types, including: From the set of frequent fatigue symptom types, extract the first frequent fatigue symptom type up to the Qth frequent fatigue symptom type, where Q represents the number of frequent fatigue symptom types; From the aforementioned datasets of defect monitoring at several remaining locations, extract the first set of defect types at remaining locations; The trigger frequency of the first frequent fatigue symptom type in the first retention location defect type set is statistically analyzed and set as the first trigger frequency; The frequency of the Qth frequent fatigue symptom type in the first retention location defect type set is statistically analyzed and set as the Qth trigger frequency. The summation of the first trigger frequency up to the Qth trigger frequency is set as the first retention position frequent fatigue type trigger frequency; When the frequency of frequent fatigue type triggering at the first retention position is not equal to 0, the frequency of frequent fatigue type triggering at the first retention position is added to the set of frequent fatigue type triggering frequencies, and the first retention position is added to the set of forklift attachment positions.

[0113] Furthermore, the force direction analysis module 12 includes the following execution steps: Extract the first frequent fatigue position from the list of frequent fatigue points; Based on the first frequently fatigued locations, the force directions of the first set of frequently fatigued locations are extracted from the work monitoring data; Based on the angle deviation threshold, clustering is performed on the force direction of the first group of frequent fatigue locations to obtain multiple clusters of the force direction of the first frequent fatigue locations. Take the multiple first fatigue frequent position centroid force directions of the multiple cluster first fatigue frequent position force directions and set them as the first group of reconstructed fatigue frequent position force directions. At the same time, set the number of clusters of the multiple cluster first fatigue frequent position force directions as multiple initial trigger frequencies of the multiple first fatigue frequent position centroid force directions and add them to the first group of reconstructed fatigue frequent position force directions initial trigger frequencies. Until the force direction of the frequently reconstructed fatigue location in the Lth group and the initial triggering frequency of the force direction of the frequently reconstructed fatigue location in the Lth group are obtained; Based on the force direction of the first group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the first group of reconstructed fatigue frequent locations, until the force direction of the Lth group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the Lth group of reconstructed fatigue frequent locations, a frequent force direction pattern analysis is performed on the first fatigue frequent location. Continue until the frequent stress direction pattern analysis is completed for all fatigue-prone points, and output a list of frequent stress directions.

[0114] Furthermore, the basic model training module 13 includes the following execution steps: Based on the attachment structure topology and the attachment material parameters, a first-level constraint is constructed; Based on the list of frequent fatigue points and the list of frequent force directions, construct a secondary constraint. Load the sample set of attachments to be analyzed, wherein the sample set of attachments to be analyzed has the same topology of the structure to be analyzed, parameters of the material to be analyzed, list of frequent fatigue points to be analyzed, list of frequent force directions to be analyzed, time series data of force on frequent fatigue points of attachments, and time series information of attachment operation environment monitoring. When the topological similarity between the structure topology to be analyzed and the topological similarity between the attachment structure topology is greater than or equal to the topological similarity threshold, and the material parameters to be analyzed are the same as the material parameters to be analyzed, it is considered that the first-level constraint is satisfied. When the first-level constraint is satisfied, and the similarity of the point distribution between the list of frequent fatigue points and the list of frequent fatigue points to be analyzed is greater than or equal to the point distribution similarity threshold, and the similarity of the force direction distribution between the list of frequent force directions and the list of frequent force directions to be analyzed is greater than or equal to the force direction distribution similarity threshold, it is considered that the second-level constraint is satisfied. When the secondary constraint is satisfied, the stress time series data of the frequent fatigue points of the attachment and the time series information of the attachment working environment monitoring are set as the first attachment working condition time series data. The mode analysis of the duration of fatigue symptoms is performed on the attachment sample set to be analyzed to obtain the first label of fatigue life.

[0115] Furthermore, the target model training module 14 includes the following execution steps: Multiple fatigue life prediction sub-models are obtained from the fatigue life prediction basic model. Construct a fully connected neural network and connect multiple basic sub-models for fatigue life prediction in parallel to the input layer of the fully connected neural network; Using the second attachment's operating condition time series data as input to multiple fatigue life prediction sub-models, and using the second label identifying fatigue life as the output of the fully connected neural network, transfer learning is performed on the fatigue life prediction base model to obtain the target forklift attachment fatigue life prediction model.

[0116] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0117] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0122] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the fatigue life of forklift attachments using transfer learning, characterized in that, include: Load the operation monitoring data of the target forklift attachments, perform fatigue position frequent pattern analysis, and obtain a list of fatigue frequent points; Based on the work monitoring data, the list of frequent fatigue points is traversed, and the frequent force direction pattern analysis is performed to obtain the list of frequent force directions. Using the attachment structure topology, attachment material parameters, the list of frequent fatigue points, and the list of frequent force directions as constraints, collect several one-to-one corresponding time-series data of the first attachment's working conditions and labels that identify fatigue life, and train a basic model for fatigue life prediction. Using the target forklift attachment model, the list of frequent fatigue points, and the list of frequent force directions, collect several corresponding second attachment operating condition time-series data and second fatigue life labels, perform transfer learning on the fatigue life prediction basic model, obtain the target forklift attachment fatigue life prediction model, and execute the forklift attachment fatigue life prediction task.

2. The method as described in claim 1, characterized in that, Load the target forklift attachment's operational monitoring data, perform frequent fatigue location pattern analysis, and obtain a list of frequent fatigue points, including: Constrained by the material parameters of the attachments, fatigue detection data is collected, and frequent pattern analysis of fatigue symptom types is performed to obtain a set of frequent fatigue symptom types. Extract forklift attachment defect monitoring data from the operation monitoring data, and statistically analyze the forklift attachment location set and the frequent fatigue symptom trigger frequency set that show at least one frequent fatigue symptom type in the frequent fatigue symptom type set. The forklift attachments whose frequent fatigue type trigger frequencies are greater than or equal to the trigger frequency threshold are added to the list of frequent fatigue points.

3. The method as described in claim 2, characterized in that, From the operation monitoring data, forklift attachment defect monitoring data is extracted, and the set of forklift attachment locations and the set of frequent fatigue symptom trigger frequencies that show at least one frequent fatigue symptom type in the frequent fatigue symptom type set are statistically analyzed, including: From the forklift attachment defect monitoring data, extract the first set of defect monitoring data up to the Qth set of defect monitoring data, where Q represents the number of sets of forklift attachment defect monitoring data; The first set of defect monitoring data is integrated according to location up to the Qth set of defect monitoring data to obtain the first location defect monitoring dataset up to the Nth location defect monitoring dataset, where N represents the number of locations; Delete sets whose defect monitoring data are less than or equal to the data quantity threshold to obtain several remaining defect monitoring datasets. Traverse the several retained location defect monitoring datasets, and based on the frequent fatigue symptom type set, statistically analyze the forklift attachment location set and the frequent fatigue type trigger frequency set that occur at least one frequent fatigue symptom type in the frequent fatigue symptom type set.

4. The method as described in claim 3, characterized in that, Traverse the aforementioned datasets of defect monitoring at various retention locations, and based on the set of frequent fatigue symptom types, statistically analyze the set of forklift attachment locations where at least one frequent fatigue symptom type appears in the set of frequent fatigue symptom types and the set of frequent fatigue type trigger frequencies, including: From the set of frequent fatigue symptom types, extract the first frequent fatigue symptom type up to the Qth frequent fatigue symptom type, where Q represents the number of frequent fatigue symptom types; From the aforementioned datasets of defect monitoring at several remaining locations, extract the first set of defect types at remaining locations; The trigger frequency of the first frequent fatigue symptom type in the first retention location defect type set is statistically analyzed and set as the first trigger frequency; The frequency of the Qth frequent fatigue symptom type in the first retention location defect type set is statistically analyzed and set as the Qth trigger frequency. The summation of the first trigger frequency up to the Qth trigger frequency is set as the first retention position frequent fatigue type trigger frequency; When the frequency of frequent fatigue type triggering at the first retention position is not equal to 0, the frequency of frequent fatigue type triggering at the first retention position is added to the set of frequent fatigue type triggering frequencies, and the first retention position is added to the set of forklift attachment positions.

5. The method as described in claim 1, characterized in that, Based on the aforementioned work monitoring data, the list of frequent fatigue points is traversed, and a frequent force direction pattern analysis is performed to obtain a list of frequent force directions, including: Extract the first frequent fatigue position from the list of frequent fatigue points; Based on the first frequently fatigued locations, the force directions of the first set of frequently fatigued locations are extracted from the work monitoring data; Based on the angle deviation threshold, clustering is performed on the force direction of the first group of frequent fatigue locations to obtain multiple clusters of the force direction of the first frequent fatigue locations. Take the multiple first fatigue frequent position centroid force directions of the multiple cluster first fatigue frequent position force directions and set them as the first group of reconstructed fatigue frequent position force directions. At the same time, set the number of clusters of the multiple cluster first fatigue frequent position force directions as multiple initial trigger frequencies of the multiple first fatigue frequent position centroid force directions and add them to the first group of reconstructed fatigue frequent position force directions initial trigger frequencies. Until the force direction of the frequently reconstructed fatigue location in the Lth group and the initial triggering frequency of the force direction of the frequently reconstructed fatigue location in the Lth group are obtained; Based on the force direction of the first group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the first group of reconstructed fatigue frequent locations, until the force direction of the Lth group of reconstructed fatigue frequent locations and the initial trigger frequency of the force direction of the Lth group of reconstructed fatigue frequent locations, a frequent force direction pattern analysis is performed on the first fatigue frequent location. Continue until the frequent stress direction pattern analysis is completed for all fatigue-prone points, and output a list of frequent stress directions.

6. The method as described in claim 1, characterized in that, Constrained by the attachment structure topology, attachment material parameters, the list of frequent fatigue points, and the list of frequent force directions, several one-to-one corresponding time-series data of the first attachment's operating conditions and labels identifying fatigue life are collected to train a basic fatigue life prediction model, including: Based on the attachment structure topology and the attachment material parameters, a first-level constraint is constructed; Based on the list of frequent fatigue points and the list of frequent force directions, construct a secondary constraint. Load the sample set of attachments to be analyzed, wherein the sample set of attachments to be analyzed has the same topology of the structure to be analyzed, parameters of the material to be analyzed, list of frequent fatigue points to be analyzed, list of frequent force directions to be analyzed, time series data of force on frequent fatigue points of attachments, and time series information of attachment operation environment monitoring. When the topological similarity between the structure topology to be analyzed and the topological similarity between the attachment structure topology is greater than or equal to the topological similarity threshold, and the material parameters to be analyzed are the same as the material parameters to be analyzed, it is considered that the first-level constraint is satisfied. When the first-level constraint is satisfied, and the similarity of the point distribution between the list of frequent fatigue points and the list of frequent fatigue points to be analyzed is greater than or equal to the point distribution similarity threshold, and the similarity of the force direction distribution between the list of frequent force directions and the list of frequent force directions to be analyzed is greater than or equal to the force direction distribution similarity threshold, it is considered that the second-level constraint is satisfied. When the secondary constraint is satisfied, the stress time series data of the frequent fatigue points of the attachment and the time series information of the attachment working environment monitoring are set as the first attachment working condition time series data. The mode analysis of the duration of fatigue symptoms is performed on the attachment sample set to be analyzed to obtain the first label of fatigue life.

7. The method as described in claim 1, characterized in that, Using the target forklift attachment model, the list of frequent fatigue points, and the list of frequent force directions, several corresponding time-series data of the second attachment's operating conditions and labels indicating fatigue life are collected. Transfer learning is then performed on the fatigue life prediction base model to obtain a fatigue life prediction model for the target forklift attachment. This fatigue life prediction base model is an integrated model that takes the average output of multiple sub-models with different topologies, including: Multiple fatigue life prediction sub-models are obtained from the fatigue life prediction basic model. Construct a fully connected neural network and connect multiple basic sub-models for fatigue life prediction in parallel to the input layer of the fully connected neural network; Using the second attachment's operating condition time series data as input to multiple fatigue life prediction sub-models, and using the second label identifying fatigue life as the output of the fully connected neural network, transfer learning is performed on the fatigue life prediction base model to obtain the target forklift attachment fatigue life prediction model.

8. A forklift attachment fatigue life prediction system employing transfer learning, characterized in that, The system is used to implement the forklift attachment fatigue life prediction method using transfer learning as described in any one of claims 1-7, including: The frequent pattern analysis module is used to load the operation monitoring data of the target forklift attachments, perform fatigue position frequent pattern analysis, and obtain a list of frequent fatigue points; The force direction analysis module is used to traverse the list of frequent fatigue points based on the operation monitoring data, perform frequent force direction pattern analysis, and obtain a list of frequent force directions. The basic model training module is used to collect several one-to-one corresponding time-series data of the first attachment's working conditions and labels that identify fatigue life, constrained by the attachment structure topology, attachment material parameters, the list of frequent fatigue points and the list of frequent force directions, and to train the basic model for fatigue life prediction. The target model training module is used to collect several corresponding second attachment operating condition time series data and second fatigue life labels based on the target forklift attachment model, the list of frequent fatigue points and the list of frequent force directions, and to perform transfer learning on the fatigue life prediction basic model to obtain the target forklift attachment fatigue life prediction model and execute the forklift attachment fatigue life prediction task.

Images