A cabinet structure assembly post-assembly push performance testing system and method

Abstract

The application provides a cabinet structure assembly post-pressing performance test system and method. It belongs to the technical field of mechanical and electrical product quality inspection. The method comprises the following steps: analyzing and processing CAD model data of the cabinet structure after assembly, automatically identifying and positioning key bearing parts in the model, marking the mechanical properties and extracting the geometric parameters of each test point, and generating a multi-dimensional test point feature set; through the cooperation of adaptive loading path planning and multi-modal data fusion, the test points such as the top beam, the side door hinge, the bottom bearing foot and the like can be automatically identified in the pressing test according to the CAD model, and the loading direction and displacement trajectory can be dynamically corrected in real time with the help of a six-dimensional force sensor, so that the data dispersion problem caused by unstable force direction and uneven loading speed in manual operation is fundamentally avoided, and the repeated test deviation is narrowed from plus or minus five to plus or minus zero point five.

Description

Technical Field

[0001] This invention proposes a system and method for testing the push-press performance of cabinet structural components after assembly, belonging to the field of electromechanical product quality inspection technology. Background Technology

[0002] With the continuous development of the communication equipment manufacturing industry, the accuracy and efficiency of quality inspection of cabinet structural components have become one of the key factors affecting product reliability and delivery capability. Currently, after the cabinet structural components are assembled, the testing of their push-pressure performance is generally carried out manually using handheld thrust gauges or pressure gauges. Operators need to manually control the loading speed and direction of force. Due to human intervention, the data deviation of the same product in repeated tests can reach ±5 Newtons, showing significant dispersion. At the same time, traditional testing methods usually only apply load to single locations such as the top thrust, which is insufficiently covered for parts such as side door hinges and bottom load-bearing feet. The overall structural integrity is not fully considered, and there is a risk of missing local weak points. In the pass / fail judgment stage, operators directly read the instrument values ​​and compare them with fixed thresholds. There is a lack of quantitative evaluation methods for products in a critical state, resulting in poor consistency of judgment results.

[0003] In the field of automation and intelligent testing of rack structural components' compressive strength, the industry has made some technological explorations. Regarding intelligent rack strength testing, Chinese patent publication number CN112557202A, entitled "Modular Data Center Rack Intelligent Strength Testing Device," utilizes a feeding conveyor, turntable, and gantry structure. Testing units are installed on the gantry and supports, enabling the application of lateral, longitudinal, and vertical pressure to the rack, achieving strength testing in three directions. This solution facilitates seamless testing and intelligent control; however, the placement of the testing units is relatively fixed, primarily addressing standardized testing needs for a single rack model. When dealing with mixed-model rack production lines, the testing unit positions are difficult to flexibly adapt to the critical stress points of racks of different sizes, often failing to effectively cover local structural components such as side door hinges and bottom load-bearing feet. Regarding pressure resistance testing for specific cabinets, Chinese Patent Publication No. CN114544368A, entitled "A Device for Testing the Pressure Resistance Performance of Low-Voltage Cabinets," utilizes a movable connecting plate to move a pressure sensor, enabling testing at different locations within the cabinet. However, the pressure sensor's freedom of movement is limited to a single plane, making it difficult to adapt to the multi-dimensional and multi-poseal pressure performance testing requirements of the overall cabinet structure. Its positioning and adjustment capabilities also face adaptation bottlenecks when dealing with diverse test points across different cabinet models, resulting in a long tooling adjustment cycle. In the area of ​​flexible testing tooling, Chinese Patent Publication No. CN120760789A, entitled "A Flexible Testing Tooling and Method for Sheet Metal Parts," employs a lifting and lowering testing mechanism above a workpiece placement platform. Multiple testing components simultaneously descend and contact the sheet metal part, acquiring positional data of the reference plane and the testing surface, thus achieving precise testing of the surface quality of bent sheet metal parts. However, the core purpose of this solution is static inspection of the dimensions and surface indentations of sheet metal parts, rather than dynamic testing of the pushing performance of assembled rack structural components under stress. Its detection components are based on contact displacement sensors, which cannot simulate the pushing loads experienced by the rack in actual use, nor can they provide closed-loop feedback control of the force value. Furthermore, while some production lines have introduced automated testing solutions using fixed tooling and limit switches, which have improved testing efficiency to some extent, these solutions suffer from poor tooling adaptability, making it difficult to meet the actual needs of mixed-model rack production lines. They also lack real-time force feedback and self-calibration functions, and a closed-loop testing system has not yet been established.

[0004] While existing technologies have achieved preliminary results in areas such as rack strength testing, pressure sensor displacement adjustment, and sheet metal flexibility testing, none of these solutions effectively address the numerous shortcomings of traditional push-pressure performance testing methods that rely primarily on manual inspection. Firstly, in manual testing, operators must manually control the loading speed and force direction, leading to significant data deviations of up to ±5 Newtons in repeated tests for the same product, indicating substantial dispersion. Secondly, traditional testing schemes typically apply loads only to single locations such as the top push point, neglecting areas like side door hinges and bottom load-bearing feet, failing to adequately consider the overall structural integrity, and creating a risk of missing weak points. Furthermore, in the pass / fail judgment stage, operators directly read instrument values ​​and compare them to fixed thresholds, lacking quantitative evaluation methods for products in critical states, resulting in poor consistency in judgment results. Additionally, while some production lines have introduced automated testing schemes using fixed tooling and limit switches, the tooling has poor adaptability, making it difficult to meet the actual needs of mixed-model rack production lines, and lacks real-time force feedback and self-calibration functions, thus failing to establish a closed-loop testing system. Therefore, a method is urgently needed to provide precise force loading control, real-time force feedback, and self-calibration capabilities while achieving multi-point adaptive coverage and rapid model changeover, thus forming a complete testing closed loop. This closed-loop control function effectively improves the intelligence level and production line adaptability of cabinet structural component quality inspection. Summary of the Invention

[0005] This invention provides a system and method for testing the push-press performance of cabinet structural components after assembly, in order to solve the problems mentioned in the background art above: This invention proposes a method for testing the push-press performance of cabinet structural components after assembly, the method comprising: S1. Analyze and process the CAD model data of the rack structure after assembly, automatically identify and locate the key load-bearing parts in the model, mark the mechanical properties and extract the geometric parameters of each identified test point, and generate a multi-dimensional test point feature set; predict the loading direction and classify the force mode of the multi-dimensional test point feature set to generate candidate data for the initial loading path. S2. Perform real-time feedback fusion processing on the initial loading path candidate data using six-dimensional force sensors, and dynamically correct the loading direction and displacement trajectory by combining the force signals collected by the sensors to generate adaptive loading path data; generate multi-modal synchronous acquisition instructions from the adaptive loading path data to generate multi-source synchronous acquisition instruction data. S3. Perform convolutional neural network feature extraction processing on the multi-source acquisition synchronous command data, input the acquired force and displacement curves, strain distribution data and structural vibration mode data into the pre-trained CNN model to generate push-press performance feature vector data; perform health index construction processing on the push-press performance feature vector data to generate push-press performance health index data. S4. Perform dynamic threshold comparison processing on the pushing performance health index data, quantitatively compare the health index with the preset dynamic qualified threshold, and generate qualified judgment result data; perform anomaly location analysis on the qualified judgment result data, trace the weak parts of the unqualified products, and generate anomaly location data. S5. Perform closed-loop testing and calibration on the abnormal location data, trigger local fine-tuning commands for the located weak points, execute the locking bolt torque compensation operation, restart the pushing performance test process, and generate a closed-loop test completion signal; archive the test report on the closed-loop test completion signal to generate the final state data of the pushing performance test.

[0006] This invention proposes a system for testing the push-press performance of rack structural components after assembly, the system comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0007] The beneficial effects of this invention are as follows: Through the synergy of adaptive loading path planning and multimodal data fusion, multiple test points, such as the top crossbeam, side door hinges, and bottom load-bearing feet, can be automatically identified based on the CAD model during the push-press test. Furthermore, the six-dimensional force sensor provides real-time feedback to dynamically correct the loading direction and displacement trajectory, fundamentally avoiding the data dispersion problems caused by unstable force direction and uneven loading speed in manual operation. This reduces the deviation of repeated tests from ±5 Newtons to ±0.5 Newtons. Simultaneously, multimodal data fusion, combined with CNN feature extraction, maps multi-source information such as force-displacement curves, strain distribution, and structural vibration modes into a unified push-press performance health index, replacing the previous single-threshold judgment logic. This enhances the consistency of critical value product judgment, covering both overall structural assessment and precise tracing of local weak points. The introduction of a test-calibration closed loop allows non-conforming products to automatically trigger locking bolt torque compensation and retesting, reducing manual intervention, lowering the risk of missing local weak points, and avoiding batch misjudgments caused by subjective judgment differences. Attached Figure Description

[0008] Figure 1 This is a diagram illustrating the steps of the method described in this invention; Figure 2 This is a detailed flowchart of step S5 in this invention. Detailed Implementation

[0009] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0010] Example 1 One embodiment of the present invention, such as Figure 1 As shown, a method for testing the push-press performance of a cabinet structural component after assembly, the method comprising: S1. The CAD model data of the rack structure components after assembly is parsed and processed. Key load-bearing parts such as the top beam, side door hinges, and bottom load-bearing feet in the model are automatically identified and located. Mechanical properties are labeled and geometric parameters are extracted for each identified test point to generate a multi-dimensional test point feature set. Loading direction is predicted and force mode is classified for the multi-dimensional test point feature set to generate candidate data for the initial loading path. S2. Perform real-time feedback fusion processing on the initial loading path candidate data using six-dimensional force sensors, and dynamically correct the loading direction and displacement trajectory by combining the force signals collected by the sensors to avoid the risk of rigid impact and generate adaptive loading path data; generate multi-modal synchronous acquisition commands for the adaptive loading path data, synchronously trigger the force and displacement curve acquisition unit, strain distribution acquisition unit and structural vibration mode acquisition unit, and generate multi-source acquisition synchronous command data. S3. Perform convolutional neural network feature extraction processing on the multi-source acquisition synchronous command data. Input the acquired force and displacement curves, strain distribution data and structural vibration mode data into the pre-trained CNN model to extract multi-dimensional feature vectors of the pushing performance and generate pushing performance feature vector data. Perform health index construction processing on the pushing performance feature vector data and map the multi-dimensional feature vectors to the pushing performance health index evaluation model to generate pushing performance health index data. S4. Perform dynamic threshold comparison processing on the pushing performance health index data, quantitatively compare the health index with the preset dynamic qualified threshold, eliminate subjective judgment deviation of critical value products, and generate qualified judgment result data; perform anomaly location analysis on qualified judgment result data, trace the weak parts of unqualified products, and generate anomaly location data. S5. Perform closed-loop testing and calibration on the abnormal location data, trigger local fine-tuning commands for the located weak points, execute the torque compensation operation of the locking bolts, and restart the pushing performance test process to generate a closed-loop test completion signal; archive the test report on the closed-loop test completion signal, and store the full process test data, health index, pass / fail judgment results and calibration records in a structured manner to generate the final state data of the pushing performance test.

[0011] The working principle and effects of the above technical solution are as follows: By analyzing the CAD model data of the cabinet structural components to locate key test points, geometric and mechanical features are accurately extracted, effectively improving the pertinence and overall accuracy of the cabinet push-press performance test. Through real-time data fusion from sensors, the loading trajectory is dynamically corrected, weakening rigid impacts during testing and reducing structural deformation interference and equipment wear. Relying on convolutional neural networks to complete feature extraction and analysis of multi-source test data significantly improves the analysis accuracy of structural stress, deformation, and vibration data, enhancing the comprehensiveness of the overall push-press performance evaluation. Dynamic threshold comparison is used for qualification judgment, reducing the judgment deviation of critical products and avoiding test result distortion caused by subjective judgment factors. Simultaneously, through weak point traceability calibration and closed-loop retesting mechanisms, both accurate correction of structural assembly defects and stability of batch cabinet product test results are achieved, improving the standardization and reliability of the overall testing process.

[0012] In one embodiment of the present invention, S1 includes: S11. Read the original data stream of the CAD model corresponding to the assembled cabinet structural components, traverse the model entity layer to divide the component units, distinguish the three types of load-bearing component entities: top beam, side door hinge, and bottom load-bearing foot, scan the coordinate values ​​of the entity boundary nodes, mark the spatial position coordinate parameters of each node, and output the original dataset of component entity coordinate points; traverse the model material thickness, cross-sectional dimensions, and assembly gap values ​​associated with each point in the original dataset of coordinate points, enter the corresponding geometric parameter entries for the points, and output the complete dataset of point geometric parameters. S12. Traverse the complete dataset of geometric parameters of the points, match the mechanical constants of material yield strength, elastic modulus, and Poisson's ratio, bind them to the corresponding coordinate points to generate point mechanical attribute annotation entries, integrate all annotation entries to form a multi-dimensional test point feature set containing coordinate, geometric, and mechanical information; retrieve the component support constraint distribution state, deduce the force transmission direction under external force for each test point, classify the three basic stress modes of compression, shear, and bending, assign corresponding stress mode codes, and output the initial stress mode classification dataset. S13. Read the force codes of each test point in the initial force pattern classification dataset, arrange multiple external force application trajectories along the force transmission direction, record the starting and ending point travel angle displacement increment values ​​of each trajectory, and summarize all trajectory entries to form initial loading path candidate data; traverse each trajectory node of the initial loading path candidate data, calculate the stress difference deformation amplitude values ​​between nodes, reduce the displacement step size of trajectory nodes whose deformation amplitude exceeds the safe range, and output the preprocessed loading path candidate data. S14. Traverse the candidate data of the preprocessed loading path to match the overall assembly support boundary constraint range of the component, trim the invalid travel segments that extend the trajectory outside the constraint boundary, correct the coordinate values ​​of the trajectory endpoint, and output the boundary constraint adapted loading path dataset; compare the overall stress fluctuation amplitude of each path after adaptation, sort the path entries from low to high according to the stress fluctuation amplitude, and output the path priority sorted dataset. S15. Read the first priority loading path entry in the path priority sorting dataset, enter the initial running configuration cache of the loading execution unit, fill in the configuration values ​​of the upper limit of loading speed, displacement, upper limit of loading force, and running time, and output the initial configuration cache data of the loading unit; verify that there are no numerical overflow conflicts of the three types of parameters of speed, force, and displacement in the cache data, push the cache data to the loading driver hardware register, and generate the initial loading start preparation state signal.

[0013] The working principle and effects of the above technical solution are as follows: By analyzing the cabinet CAD model data layer by layer and dividing it into component units, the coordinates and geometric parameters of key load-bearing parts can be accurately collected, improving the precision and coverage of test point selection and reducing positional deviations and parameter omissions caused by manual point selection. By matching mechanical parameters to each test point and completing the force mode division, the test logic can be planned to fit the actual force state, enhancing the fit of the loading path design. By performing deformation screening and step size optimization on multiple initial loading trajectories, the abnormal deformation amplitude during the test can be reduced, lowering the probability of component pre-damage. By combining assembly constraint boundaries to correct invalid trajectories and performing path priority sorting, test failures caused by overloading can be avoided. This ensures the safety and stability of the loading operation and makes the basic working conditions of subsequent push-press tests more closely match the actual assembly and use conditions, effectively improving the accuracy and operational stability of the overall testing operation.

[0014] In one embodiment of the present invention, S2 includes: S21. Receive the initial loading start-up preparation state signal to start the continuous sampling process of the six-dimensional force sensor, collect the real-time values ​​of XYZ three-axis thrust and tension and three-axis torque at equal time intervals, package the force sensing sampling frame sequence per unit time, and output the continuous force sensing signal time sequence sampling dataset; read the real-time execution displacement position value of the loading unit, synchronously package the displacement time sequence sampling frame sequence, and output the displacement trajectory time sequence sampling dataset. S22. Synchronously align the timestamp nodes of the force sensor timing sampling dataset and the displacement trajectory timing sampling dataset. Under the same timestamp, match the three sets of values ​​of force, torque and displacement to construct a spatiotemporal matching sampling unit. Summarize all units to form a force-displacement synchronous fusion sampling dataset. Traverse each spatiotemporal matching unit of the fusion sampling dataset, calculate the angle difference between the real-time force vector direction and the preset loading and travel direction, and output the direction deviation difference sequence data. S23. Read the direction deviation difference sequence data one by one, adjust the loading angle correction amount in the next sampling period according to the positive and negative magnitude of the difference, simultaneously fine-tune the displacement step amplitude per unit period, iterate and update the coordinates of the entire trajectory points frame by frame, and output the adaptive loading path data after frame-by-frame correction; traverse all trajectory segments of the adaptive loading path data, calculate the instantaneous force change gradient values ​​of adjacent points, and mark the trajectory segments whose gradient values ​​exceed the impact safety threshold. S24. Reduce the displacement step increment and force rise slope values ​​of the trajectory segment marked with impact risk, recalculate the coordinates and loading rate values ​​of all points within the segment, and output the optimized adaptive loading path data after the impact risk is resolved; split the time axis nodes within the optimized adaptive loading path data, generate a unified time-series trigger mark code for each time node, and output the time-series synchronous trigger mark sequence data. S25. Distribute the timing synchronization trigger mark sequence data to the registers of the three sets of acquisition hardware units. The mark code is synchronously pushed to the force and displacement curve acquisition unit, strain distribution acquisition unit, and structural vibration mode acquisition unit. The three sets of units synchronously latch their own acquisition start time and generate multi-source acquisition synchronization instruction data. Latch the system timing reference value at the time of the synchronization instruction issuance, enter the acquisition timing reference log entry, and output the acquisition timing reference log data.

[0015] The working principle and effects of the above technical solution are as follows: By continuously acquiring force and displacement time-series data through a six-dimensional force sensor, the dynamic operating state during the loading process can be completely recorded, improving the integrity and continuity of test data acquisition. The fusion and matching of multiple data types is achieved through timestamp alignment, eliminating time-series misalignment issues during multi-device acquisition and improving the accuracy of loading deviation calculation. Real-time deviation value iterative correction of the loading trajectory and step parameters allows the loading action to adapt to the real-time stress state of the component, significantly reducing stress deviation issues during testing. By identifying and optimizing high-gradient impact trajectory segments, the structural disturbance caused by instantaneous loading impact is weakened, reducing additional deformation damage to the cabinet structural components caused by rigid contact. A unified timing triggering mechanism allows multiple acquisition units to simultaneously record data, eliminating analysis errors caused by asynchronous multi-source data acquisition and ensuring a more unified and reliable data foundation for subsequent performance analysis.

[0016] In one embodiment of the present invention, step S22 includes: Retrieve the continuous force sensing signal timing sampling dataset and the displacement trajectory timing sampling dataset, batch analyze the timing encoding parameters of the timing frames within the two datasets, unify the sampling timing accuracy and frame synchronization interval of the two datasets, and generate a timing reference unified sampling dataset; perform timing node calibration processing on the timing reference unified sampling dataset to eliminate frame misalignment deviation caused by data acquisition time difference, and generate a timing-accurate aligned dataset. The time-series precisely aligned dataset is reorganized into frame units. Using a unified time-series node as the matching benchmark, the triaxial force value, triaxial torque, and real-time displacement value corresponding to the same time-series node are bound and integrated to construct a single-frame integrated spatiotemporal matching sampling unit. All time-series frame units are integrated to generate a complete force-displacement synchronous fusion sampling dataset. Traverse all spatiotemporal matching sampling units within the force-displacement synchronous fusion sampling dataset, extract the three-dimensional coordinate parameters of the real-time force vector of each unit, combine the basic orientation parameters of the loaded trajectory to perform vector calculations, solve the spatial angle offset between the real-time force vector and the standard orientation, and generate a set of single-unit angle offset values. The set of angle offset values ​​of a single unit is processed in the whole domain. The angle offset values ​​of each sampling unit are connected in the time sequence. Abnormal and sudden values ​​caused by acquisition disturbances are removed. The time series data is smoothed and regularized to generate continuous and stable angle offset time series data. The continuous and stable angular offset time series data is encapsulated in a data structure, and the numerical storage format and time series arrangement rules are unified to fully record the angular offset change information throughout the time period and generate the direction deviation difference sequence data for trajectory correction.

[0017] The working principle and effects of the above technical solution are as follows: By unifying the timing accuracy of multiple types of acquired data and calibrating timing nodes, frame misalignment problems caused by inconsistent acquisition rates of different sensors can be eliminated, improving the fit of multi-source test data fusion. By integrating force, torque, and displacement parameters into a unified sampling unit with unified timing nodes, a precise correspondence between the force and displacement states during the loading process can be established, enhancing the rigor of subsequent trajectory deviation analysis. By solving the real-time loading angle offset through vector operations, subtle directional deviations during the loading process can be accurately captured, reducing the inaccuracy of traditional deviation judgment methods. By removing abnormal data mutations and performing timing smoothing, signal disturbances caused by the on-site acquisition environment can be filtered out, improving the stability of deviation data. This ensures the continuity of loading direction deviation analysis and provides accurate and reliable data support for subsequent dynamic correction of the loading trajectory, effectively improving the overall test control accuracy.

[0018] In one embodiment of the present invention, step S23 includes: Read the full-time difference values ​​of the direction deviation difference sequence data, distinguish positive and negative attributes and divide the amplitude levels of the time-series difference values ​​to generate graded and classified deviation value data; perform periodic correction parameter conversion on the graded and classified deviation value data, match the angle adjustment amplitude corresponding to different deviation intervals, and generate loading angle correction parameter dataset. The loading angle correction parameter dataset is retrieved, and the travel angle parameter of each sampling period is finely adjusted cycle by cycle, and the displacement step scale parameter of the corresponding period is adapted synchronously to generate single-cycle trajectory correction parameter data; the correction parameter data of multiple consecutive cycles are iteratively superimposed to update the trajectory coordinate point information point by point to generate trajectory point iterative update data. Integrate the iterative update data of the entire trajectory points, complete the iterative optimization of the parameters of the entire loading trajectory, and generate adaptive loading path data after frame-by-frame correction; perform global trajectory segmentation processing on the adaptive loading path data, divide the continuous independent trajectory segments according to a fixed displacement step size, and generate an equidistant trajectory segment dataset. Traverse all trajectory segments in the equidistant trajectory segment dataset, extract the real-time force values ​​and displacement parameters of adjacent trajectory points, perform instantaneous force change rate calculation, obtain the instantaneous force change gradient values ​​corresponding to each segment, and generate a full-segment force gradient statistical data set. The system retrieves the preset impact safety threshold parameters, compares each gradient value in the full-segment force gradient statistics dataset with the threshold parameters in batches, filters out abnormal trajectory segments with excessive gradient values, completes the risk segment marking process, and generates trajectory segment data with impact risk marking.

[0019] The working principle and effects of the above technical solution are as follows: By classifying and converting the loading deviation values, the adjustment range corresponding to different deviation states can be matched, improving the precision and adaptability of trajectory correction. By fine-tuning the travel angle and displacement step scale cycle by cycle, and updating the trajectory coordinates through iterative superposition, the loading path can be made to fit the real-time stress state of the component, reducing the matching deviation caused by fixed trajectory testing. By equally segmenting the optimized loading path, the detection granularity of the entire trajectory can be refined, improving the completeness of capturing stress changes. By calculating the instantaneous stress change gradient segment by segment, the location of stress mutation during loading can be accurately identified, reducing the situation where sudden changes in local stress are ignored. This not only allows for the early screening of trajectory segments with potential impact risks, avoiding hidden damage to the cabinet structure caused by instantaneous strong loading, but also makes subsequent path optimization work more targeted, effectively improving the stability and safety of the overall push-pressure test.

[0020] In one embodiment of the present invention, S3 includes: S31. Receive the multi-source acquisition synchronization command data to start the parallel acquisition operation of three sets of units. The force and displacement curve acquisition unit outputs the full-time force and displacement paired sampling sequence, the strain distribution acquisition unit outputs the strain amplitude distribution matrix of each test point, and the structural vibration mode acquisition unit outputs the multi-order vibration frequency amplitude phase matrix. Output three types of independent raw acquisition datasets; unify the sampling interval and timestamp format of the three types of raw acquisition datasets, align the time sequence nodes of each sampling frame, and output the time sequence aligned multi-source raw acquisition fusion dataset. S32. Divide the time-aligned multi-source raw acquisition and fusion dataset into fixed-length time-series sample slices. Perform normalized numerical scaling on each sample slice to scale it to the zero-to-one value range and output the normalized time-series sample slice dataset. Load the weight bias parameters of the completed and converged convolutional neural network and input the normalized time-series sample slice dataset into the first convolutional operation layer of the network. Perform sliding multiplication and addition operations of the convolutional kernel to extract shallow temporal spatial features. S33. Shallow temporal and spatial features are compressed through pooling layers to reduce feature dimensions. The compressed features are then fed into multiple fully connected layers for feature fusion and mapping operations, reducing redundant feature components and amplifying effective push-compression related feature components layer by layer. The network outputs a fixed-length one-dimensional feature array, and all sample arrays are aggregated to form push-compression performance feature vector data. The push-compression performance feature vector data is split into three sub-feature vector subsets: force, deformation, and vibration. The mean, variance, extreme values, and dispersion statistics of each subset are statistically analyzed, and the sub-feature statistics dataset is output. S34. Retrieve the built-in mapping weight matrix of the pushing performance health index assessment model, multiply and sum the three types of sub-feature vectors of force deformation vibration with the corresponding weight matrix element by element to obtain the basic calculated value of a single health index, traverse all time series samples to generate the full time series health index basic value sequence data; perform moving mean smoothing filtering on the full time series health index basic value sequence data to eliminate numerical jitter caused by instantaneous sampling noise, and output the smoothed pushing performance health index data; S35. Extract the four statistical indicators of the entire process of the push-press performance health index data: maximum value, minimum value, average value, and fluctuation range. Bind the corresponding rack structural component assembly number and sampling start and end time to three sets of identification information. Output the complete health index archived pre-data with product identification. Verify that the archived pre-data has no empty values, garbled characters, or out-of-bounds values. Write it to a temporary storage buffer and output the health index temporary cache storage data.

[0021] The working principle and effects of the above technical solution are as follows: By acquiring various test data in parallel across multiple units and completing time-series alignment and fusion, the acquisition standards of multi-source data can be unified, reducing analysis errors caused by time-series misalignment of data from different dimensions and improving the regularity of test data. Through sample slicing and numerical normalization, analysis interference caused by differences in the dimensions of different parameters can be eliminated, improving the adaptability of model feature extraction. Utilizing convolutional neural networks to complete feature extraction and fusion layer by layer can filter out a large number of invalid data components, strengthen core features related to cabinet push-press performance, and improve the precision of feature recognition. By classifying and statistically analyzing the changing patterns of various feature data, the performance status of structural stress, deformation, and vibration can be comprehensively covered. By smoothing and filtering the health index time-series data, numerical fluctuations caused by sampling noise can be weakened, avoiding the problem of instantaneous abnormal data interfering with the overall evaluation results. This not only makes the performance evaluation results more consistent with the actual working conditions of the components but also ensures the integrity and effectiveness of test data archiving.

[0022] In one embodiment of the present invention, S32 includes: The system receives a time-aligned multi-source raw acquisition and fusion dataset, and performs equal segmentation of the continuously acquired data according to a fixed time length. The complete time-series data is divided into multiple independent time-series sample units with unified dimensions. All sample units are aggregated to generate a fixed-length time-series sample slice dataset. The system then performs anomaly screening on the time-series sample slice dataset to remove aberration abnormal samples generated during the acquisition process, thereby generating clean time-series sample slice data. The pure time series sample slice data is uniformly scaled across the entire domain to adjust the overall data value range to the standard zero-to-one range, eliminating data bias caused by differences in the units of different acquisition parameters, and generating standardized and normalized time series sample data; the normalized time series sample data is then dimensionally normalized to unify the matrix size of all samples, generating a dimensionally unified normalized sample dataset. Read the network parameter file stored in the convolutional neural network, import the weight parameters and bias parameters of each layer in batches, complete the batch loading and configuration of network running parameters, and generate complete network parameter configuration data; verify the validity of the complete network parameter configuration data, check for missing parameters and abnormal values, and generate usable network parameter deployment data; The sample dataset with unified and normalized dimensions is fed into the first operation layer of the convolutional neural network in batches. The convolutional kernel is driven to complete the sliding traversal operation on the surface of the sample data matrix with a fixed stride, and matrix multiplication and addition operations are performed to complete the preliminary extraction of basic data features and generate shallow feature matrix data of the samples. The shallow feature matrix data corresponding to each sample is subjected to full-domain feature filtering, retaining the temporal variation features and spatial distribution features within the data, eliminating invalid and duplicate data components, summarizing the extraction results of all samples, and generating the overall shallow temporal and spatial feature data of the dataset.

[0023] The working principle and effects of the above technical solution are as follows: By equally slicing continuous time-series test data and screening for abnormal samples, invalid data generated during the collection process can be cleaned up, reducing the interference of abnormal data on the model calculation process and improving the overall quality of sample data. Through global numerical scaling and dimension normalization operations, the numerical ranges and data specifications of different physical parameters can be unified, eliminating analytical biases caused by inconsistencies in the dimensions of various test parameters and improving the adaptability of model calculations. By batch importing and verifying network operating parameters, parameter anomalies can be identified, ensuring the stability of the convolutional neural network's operating state and avoiding feature extraction failures caused by parameter anomalies. By completing shallow feature extraction through standardized matrix operations, the temporal changes and spatial distribution characteristics of test data can be comprehensively captured. This not only strengthens the data foundation for intelligent feature analysis but also improves the accuracy and stability of subsequent push-pressure performance feature recognition, making the model analysis results more consistent with the actual test conditions of rack structural components.

[0024] In one embodiment of the present invention, step S4 includes: S41. Read the complete push performance health index sequence from the temporary cache storage data of the health index, retrieve the dynamic qualified threshold upper and lower limit values ​​of the batch of cabinet structural parts of the same specification, enter the threshold comparison operation cache area, and output the threshold comparison benchmark configuration data; take out the health index value of each time node, compare the value with the upper and lower limits of the threshold respectively, mark each time node with three categories of judgment labels: above the upper limit, within the interval, and below the lower limit, and output the preliminary label dataset of a single node. S42. Calculate the percentage of nodes that are within the range throughout the entire process in the initial label dataset of a single node, the percentage of nodes with a cumulative duration below the lower limit, and the percentage of nodes with a peak duration above the upper limit. Output the statistical dataset of the percentage of nodes. Set the numerical range of the critical interval. Nodes whose health index falls into the critical interval are not directly marked as unqualified. Separately divide the critical sample subset and output the critical sample isolated dataset and the regular qualified and unqualified sample dataset. S43. The regular qualified and unqualified sample dataset directly generates the corresponding overall product preliminary qualified judgment mark. The critical sample isolated dataset retrieves the slope value of the health index change between adjacent time periods. If the slope is stable and fluctuates slightly, it is classified as qualified. If the slope rises or falls sharply, it is classified as unqualified. All marks are integrated to form qualified judgment result data. The qualified judgment result data is traversed to the product items marked as unqualified. The original force, strain and vibration sampling values ​​of the product are retrieved from the multi-source acquisition and fusion dataset of the whole process. S44. Backtrack the original sampled values ​​and match the coordinates of each point in the feature set of multi-dimensional test points one by one. Compare the strain peak value, stress extreme value, and vibration amplitude of each point. Sort the point entries according to the extreme value values ​​from large to small and output the original sequence data of weak point sorting. Filter the points in the sorted sequence whose extreme values ​​do not exceed the material bearing limit of the component and retain the points that exceed the limit to form a simplified weak point sequence data. S45. Bind the component name, assembly position, geometric parameters, and mechanical property information corresponding to each point in the simplified weak point sequence data, assemble them into a complete set of entries containing component information of point coordinates exceeding limits, and generate abnormal part location data; enter the excess value and percentage value of each weak point in the abnormal part location data, and output the quantitative auxiliary data of the degree of point exceeding limits.

[0025] The working principle and effects of the above technical solution are as follows: By comparing point-by-point values ​​with dynamic thresholds corresponding to each batch, it can adapt to the production tolerance range of different batches of cabinets, improving the degree to which the acceptance criteria align with actual production conditions. By statistically analyzing the duration ratio of various nodes and dividing critical sample intervals, it avoids directly judging products with critical values, reducing misjudgments caused by crude classification under critical operating conditions. Refining the critical sample judgment criteria based on the slope of index fluctuations before and after time periods eliminates the result deviation caused by subjective human judgment, improving the fairness of the overall machine judgment. By retrospectively comparing the extreme values ​​of stress deformation at each test point with the complete set of original collected data, it can accurately screen out locations where the load exceeds the standard, reducing the time spent investigating weak areas. Simultaneously recording the parameters and over-limit ratios of the supporting components at the points, it fully restores the details of assembly defects, which can quickly locate the assembly positions that need adjustment and provide accurate quantitative basis for subsequent bolt torque compensation calibration, improving the execution efficiency of the entire defect correction process.

[0026] One embodiment of the present invention, such as Figure 2 As shown, S5 includes: S51. Read the abnormal location data and the quantitative auxiliary data of the degree of exceeding the limit of the point, match the standard torque reference value of the assembly locking bolt corresponding to each weak point, adjust the reference torque compensation increment value according to the percentage of exceeding the limit, generate the target locking torque value of the bolt for each point, and output the local fine-tuning torque command dataset; split the local fine-tuning torque command dataset into independent command frames for each point, push the command frames to the torque tightening actuator controller register, and generate the weak point fine-tuning execution command signal. S52. The torque tightening actuator receives the fine-tuning execution command signal, applies the target locking torque point by point to complete the bolt compensation locking operation, and after the locking operation is completed, it feeds back the torque execution in place response signal and outputs torque locking completion response timing data; the response timing data is pushed to the test main control unit, the main control unit clears all the previously collected cached feature vector health index judgment result storage data, releases the temporary storage space occupied, and outputs test cache clearing completion status data; S53. After the test cache is cleared and the status data is cleared, the main control unit restarts the entire push-pressure performance test process. The loading path is repeatedly loaded to generate multi-source data, collect CNN features, extract health index, calculate and judge the entire process, and output the original data set of the second test process. The feature vector health index calculation steps are repeated on the original data set of the second test process to obtain the second push-pressure performance health index and the second pass judgment result data. S54. Compare the secondary pass / fail judgment result data with the initial judgment result data. If the secondary judgment marks the result as pass, the loop test process is terminated. If the secondary judgment still marks the result as fail, the abnormal part location data is regenerated and the fine-tuning instruction is issued again. The loop iterates until the judgment mark is pass. The closed-loop test completion signal is output at the moment the iteration ends. The closed-loop test completion signal triggers the full-process data packaging process. The packaged content includes the initial, secondary and multiple rounds of raw data collection, the pushing performance feature vector of each round, the health index value of each round, the pass / fail judgment result of each round, and the bolt torque calibration record of each round. S55. After packaging, the entire process data is uniformly converted into a structured row-column storage format, and the product assembly number, test start and end time, equipment number, operator number, and auxiliary identifier fields are filled in. The structured packaging final state dataset is then output. The structured packaging final state dataset is written to the persistent storage disk partition to create independent index entries, and the data archiving and storage operation is completed to generate the push-pressure performance test final state data. The read and write permissions of the final state data index entries are verified to be complete and undamaged. Log entries are entered into the storage completion log record entries, and the archived storage log auxiliary data is output.

[0027] The working principle and effects of the above technical solution are as follows: By matching the corresponding bolt torque compensation value according to the degree of exceedance at the point, precise adjustment of the defect location can be achieved, improving the correction accuracy of cabinet assembly defects and reducing assembly deviations caused by blind adjustment. Automatic clearing and resetting of test cache data avoids interference from residual historical test data on the new round of testing, ensuring the purity of retest data. Through a closed-loop mode of multiple retests and iterative corrections, the problem of weak component assembly can be completely solved, reducing the probability of product test failures and improving the stability of finished product assembly. By summarizing complete data from multiple rounds of testing and calibration, information from the entire testing process can be completely retained, preventing the loss of critical test data. The structured data archiving method standardizes the storage format of test data, facilitating subsequent product quality traceability and process verification, and continuously accumulating test data to provide reliable data support for optimizing cabinet structure assembly processes, thus improving the overall standardization of testing and calibration operations.

[0028] In one embodiment of the present invention, S53 includes: S531: Receive test buffer clearing completion status data, activate the process restart trigger mechanism of the test main control unit, reset the loading parameters, acquisition parameters and calculation parameters of the test system, complete the system initialization configuration for a new round of testing, and generate secondary test system initialization data; perform a self-check of the whole machine parameters on the secondary test system initialization data to ensure that each test unit is in standby running state, and generate secondary test ready status data; S532. Based on the secondary test ready state data, restart the loading path generation process, replicate the standard adaptive loading path generation process, complete the test trajectory iterative correction and risk section optimization, and generate secondary test adaptive loading path data; rely on the secondary test adaptive loading path data to start the hardware loading mechanism, complete the standard push-load action of the cabinet structural components, and generate secondary loading execution process data. S533. Synchronously trigger the parallel operation of multi-source acquisition units to complete the full-time acquisition and storage of force displacement data, strain distribution data, and structural vibration mode data according to a unified time series reference. Summarize all acquired information to complete time series alignment and data fusion processing, and output the original acquisition dataset of the entire secondary test process. Perform sample slicing and normalization scaling processing on the original acquisition dataset of the entire secondary test process to generate secondary standardized test sample data. S534. Input the secondary standardized test sample data into the pre-set convolutional neural network model, and complete the layer-by-layer extraction and fusion of shallow and deep features through multi-layer convolution and pooling operations. Reduce invalid data components and retain the pushing performance related features to generate secondary pushing performance feature vector data. Classify and decompose the secondary pushing performance feature vector data, divide it into three feature subsets: force, deformation, and vibration, and generate secondary multi-dimensional feature subset data. S535. Import the secondary multi-dimensional feature subset data into the pushing performance health index evaluation model, complete the multi-dimensional feature weighting operation and numerical smoothing filtering, and output the secondary pushing performance health index data; connect the secondary pushing performance health index data to the dynamic threshold comparison process, complete the quantitative rating and qualification judgment, and generate secondary qualification judgment result data.

[0029] The working principle and effects of the above technical solution are as follows: By resetting all system parameters and performing a self-test before restarting, interference from residual parameters from the previous test can be eliminated, reducing the probability of operational failures during the retest process. The mature adaptive loading path generation steps are replicated to perform a second loading operation, ensuring consistency between the retest conditions and the first test, thus improving the comparative reference value of the two test results. Multiple acquisition units simultaneously collect and organize raw data, unifying data processing standards and reducing analytical biases caused by differences in data specifications between different rounds. Convolutional networks extract and classify multi-dimensional performance characteristics in layers, comprehensively capturing the stress, deformation, and vibration states of the cabinet structural components after calibration, enhancing the detail of the retest performance analysis. A unified application of health index assessment and threshold judgment processes ensures consistent judgment criteria between the two tests, accurately measuring the performance improvement brought about by bolt adjustment while stably controlling the retest pass / fail judgment standard, guaranteeing the authenticity and reliability of the closed-loop calibration test results.

[0030] In one embodiment of the present invention, S534 includes: The system receives secondary standardized test sample data, and then sends the standardized sample data in batches to the input layer of the convolutional neural network to complete the model loading and parameter adaptation processing of the full-domain sample data, generating model input layer adaptation data. The model input layer adaptation data is then distributed hierarchically and allocated to each convolutional operation unit according to the network operation logic, generating hierarchical operation arrangement data. By utilizing each convolutional operation unit to perform sliding convolution multiplication and addition operations on the split data, the temporal variation characteristics and spatial distribution characteristics within the sample data are mined, and the shallow features are fully extracted and integrated to generate secondary test shallow feature mapping data. The secondary test shallow feature mapping data is then subjected to pooling dimensionality reduction processing to compress redundant feature dimensions, retain core feature information, and generate dimensionality-reduced compressed feature data. The dimensionality-reduced and compressed feature data is sent to the fully connected network layer to carry out cross-dimensional feature fusion iterative operations. Irrelevant noise feature components are suppressed and removed step by step to strengthen the core related features corresponding to the rack push-up performance and generate deep fusion feature data. The deep fusion feature data is then subjected to dimensional regularization and sequence sorting to generate secondary push-up performance feature vector data with unified dimensions. Read the global feature components of the secondary push performance feature vector data, classify and split them according to the test parameter attribute features, separate the independent feature component groups corresponding to mechanical force, structural deformation and vibration mode, and generate multi-class feature split original data; group, normalize and sequence integrate the multi-class feature split original data to form secondary multi-dimensional feature subset data with clear classification and regular dimensions.

[0031] The working principle and effects of the above technical solution are as follows: By adapting the model and arranging the retest sample data in layers, the processing load of each computing unit can be balanced, improving the overall data computing efficiency of the convolutional network and reducing the analysis error caused by local computational backlog. By mining test features in the temporal and spatial dimensions through multi-layer convolutional operations, subtle performance changes after rack calibration can be fully captured, improving the completeness of retest feature extraction. By simplifying feature information through pooling dimensionality reduction, a large amount of redundant data content is eliminated, reducing the computational pressure of subsequent feature fusion. By filtering noise features step by step and strengthening core performance features, the purity of retest feature data can be significantly improved, making the performance characterization results more consistent with the actual state of the components. By classifying and organizing the final feature vectors, three independent feature subsets of force, deformation, and vibration are formed, which can not only meet the analysis needs of multi-dimensional performance review, but also make the comparison of test features between the two rounds more intuitive and accurate, effectively improving the quality verification effect of closed-loop retest.

[0032] One embodiment of the present invention provides a system for testing the push-press performance of rack structural components after assembly, the system comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0033] Example 2 This embodiment further determines the compensation priority, single torque compensation amount, graded compensation method, timing of retest and closed-loop termination conditions for each abnormal weak point based on the degree of exceeding the limit of the abnormal weak point, the remaining allowable torque of the locking bolt, the structural force transmission relationship between abnormal weak points, the response coupling degree of adjacent test points, the degree of improvement of the health index after each torque compensation, the migration of abnormal positions, and the convergence trend of retest results.

[0034] The method described in this embodiment includes: S1. Generate candidate data for the initial loading path; According to step S1 of Embodiment 1, the CAD model data of the assembled cabinet structural components is analyzed to identify key load-bearing parts such as the top beam, side door hinges and bottom load-bearing feet. The coordinate information, geometric parameters and mechanical properties of each test point are extracted to generate a multi-dimensional test point feature set. Initial loading path candidate data are generated based on the force mode, support constraint state and force transmission direction of each test point.

[0035] S2. Generate adaptive loading path data and multi-source acquisition synchronization instruction data; According to step S2 of Embodiment 1, a six-dimensional force sensor is used to collect triaxial force, triaxial torque and real-time displacement data during the loading process. The loading travel angle and displacement step amplitude are corrected according to the deviation between the real-time force vector direction and the preset loading direction to generate adaptive loading path data. The force and displacement curve acquisition unit, strain distribution acquisition unit and structural vibration mode acquisition unit are triggered simultaneously to generate multi-source acquisition synchronization command data.

[0036] S3. Generate push-compression performance health index data; According to step S3 in Example 1, the force and displacement curves, strain distribution data and structural vibration mode data are time-series aligned and fused. The processed multi-source data is input into a pre-trained convolutional neural network model to extract force features, deformation features and vibration features. The pushing performance health index data corresponding to each test point is generated through the pushing performance health index evaluation model.

[0037] S4. Generate abnormal location data; According to step S4 of Embodiment 1, the pushing performance health index is compared with the dynamic qualified threshold. For the test points that are judged to be unqualified, the corresponding extreme values ​​of force, strain peaks and vibration amplitudes are traced back to determine the comprehensive degree of exceeding the limit of each test point. The test point coordinates, component name, assembly position, corresponding locking bolt and comprehensive degree of exceeding the limit are bound together to generate abnormal part location data and point degree of exceeding the limit quantitative auxiliary data.

[0038] S5. Perform local fine-tuning and retest the pushing performance; According to step S5 of Embodiment 1, the abnormal location data and the quantitative auxiliary data of the degree of exceeding the limit of the point are read, the specifications and standard torque reference values ​​of the locking bolts corresponding to each abnormal weak point are matched, a local fine-tuning torque command is generated, and the torque compensation of the corresponding locking bolt is completed through the torque tightening actuator.

[0039] After torque compensation is completed, the test main control unit restarts the loading path generation, multi-source data acquisition, pushing performance feature extraction, pushing performance health index calculation and qualification judgment process to obtain the compensated pushing performance health index, abnormal part location data and point over-limit quantitative auxiliary data.

[0040] If there is only one abnormal weak point in the retest results, and there is no obvious response coupling between the abnormal weak point and the adjacent test point, continue to perform local fine-tuning in the manner described in Example 1; if there are two or more abnormal weak points at the same time, or if any abnormal weak point causes a significant change in the stress state of the adjacent test point after torque compensation, proceed to step S6.

[0041] S6. Determine the compensation sequence, compensation amount, and closed-loop termination conditions for multiple abnormal weak points. Step S6 includes steps S61 to S63.

[0042] S61. Determine the compensation priority for each abnormal weak point; Read the abnormal location data and the quantitative auxiliary data of the degree of point exceeding the limit generated in step S4, and read the actual torque of each locking bolt fed back by the torque tightening actuator in step S5.

[0043] For each abnormally weak point, the upper limit of the allowable torque for the corresponding locking bolt is retrieved, and the remaining allowable torque is determined by the difference between the upper limit of the allowable torque and the current actual torque. The remaining allowable torque is then compared with the adjustable torque range of the locking bolt from the standard torque reference value to the upper limit of the allowable torque to determine the proportion of the remaining allowable torque.

[0044] A larger remaining allowable torque ratio indicates that the corresponding locking bolt still has a large safe adjustment range; a smaller remaining allowable torque ratio indicates that continuing to increase the torque may cause the locking bolt to be overtightened, the threads to be damaged, or the connection to be subjected to excessive local pressure.

[0045] Based on the overall degree of exceeding limits at each abnormal weak point, the severity of the abnormality at each weak point is uniformly assessed. The greater the overall degree of exceeding limits, the higher the degree to which the corresponding abnormal weak point deviates from the allowable stress state, and the greater the urgency for it to receive torque compensation.

[0046] The test control unit further reads the multi-dimensional test point feature set, determines the top beam, side door hinge or bottom load-bearing foot to which the abnormal weak point belongs according to the coordinates of the abnormal weak point, and determines the structural force transmission relationship between the current abnormal weak point and other abnormal weak points based on the assembly connection relationship, support constraint relationship and force transmission direction between different components.

[0047] When torque compensation is applied to the locking bolt corresponding to the current abnormal weak point, the preload can be transmitted to the adjacent abnormal position along the connection path of the cabinet structural component. If it is expected to reduce the stress extreme value, strain peak value or vibration amplitude of more than one adjacent abnormal weak point at the same time, the current abnormal weak point is considered to have a high positive force transmission benefit.

[0048] When the effect of torque compensation is mainly limited to the current abnormal weak point and has little impact on the stress state of other abnormal weak points, the current abnormal weak point is considered to have a moderate positive force transmission benefit.

[0049] When torque compensation is applied to the current abnormal weak point, the load may concentrate at the adjacent connection position, and may increase the stress extreme value, strain peak value or vibration amplitude of other abnormal weak points. In such cases, the positive force transmission benefit of the current abnormal weak point is considered to be low.

[0050] Further reading of force and displacement curves, strain distribution data, structural vibration modal data, and push-compression performance health index data before and after each torque compensation was conducted to compare the response change direction of the current abnormal weak point with that of the adjacent test point in each round of testing.

[0051] When the health index of the current abnormal weak point is improved, the extreme value of the force, the peak value of the strain and the vibration amplitude of the adjacent test points decrease synchronously, indicating that the load change between the current abnormal weak point and the adjacent test points has a positive synergistic relationship and the response coupling degree of the adjacent test points is low.

[0052] When the health index of the current abnormal weak point is improved, the extreme value of force, peak value of strain or vibration amplitude of the adjacent test point increases in the opposite direction, indicating that there is a load redistribution between the current abnormal weak point and the adjacent test point, and the response coupling degree of the adjacent test point is high.

[0053] The test control unit combines the overall degree of exceeding the limit of each abnormal weak point with the proportion of the remaining allowable torque of the corresponding locking bolt to form the basic compensation priority of each abnormal weak point.

[0054] When the overall degree of exceeding the limit is high and the remaining allowable torque ratio is large, the priority of basic compensation is increased; when the overall degree of exceeding the limit is low, the priority of basic compensation is decreased; when the overall degree of exceeding the limit is high but the remaining allowable torque ratio is small, the priority of basic compensation is restricted to avoid prioritizing the increase of the locking bolt torque that is already close to the upper limit of the allowable torque simply because the degree of abnormality is high.

[0055] Based on the basic compensation priority, adjustments are made according to the positive force transmission benefit, and adjustments are made according to the response coupling degree of adjacent test points, thus forming the final compensation priority for each abnormal weak point.

[0056] When the overall exceedance of a certain abnormal weak point is high, the corresponding locking bolt has sufficient remaining allowable torque, it is expected that compensation can simultaneously improve other abnormal weak points, and the possibility of causing load transfer to adjacent test points is low, the abnormal weak point is ranked in a priority position.

[0057] When a certain abnormal weak point has a high degree of overall over-limit, but the remaining allowable torque of the corresponding locking bolt is small, or compensation may significantly increase the stress response of adjacent test points, the compensation sequence of the abnormal weak point should be reduced.

[0058] For locking bolts whose actual torque has reached the upper limit of the allowable torque, automatic torque compensation instructions will no longer be generated, and the corresponding abnormal weak points will be marked as manual verification points.

[0059] For locking bolts whose remaining allowable torque ratio is lower than the preset safety margin, the corresponding abnormal weak point will be marked as a limited compensation point, and only small-scale graded compensation is allowed.

[0060] The test control unit sorts the abnormal weak points from high to low according to the final compensation priority, and sends the abnormal weak point with the highest compensation priority that has not reached the allowable torque limit to step S62.

[0061] Through the above processing, the comprehensive degree of exceeding the limit of abnormal weak points, the remaining allowable torque of the locking bolts, the structural force transmission relationship between abnormal weak points, and the degree of response coupling between adjacent test points are all involved in determining the compensation sequence, thereby avoiding simply increasing the torque in descending order according to the percentage of exceeding the limit.

[0062] S62. Determine the single torque compensation amount, the graded compensation method, and the timing of retesting; Read the abnormal weak point with the highest compensation priority determined in step S61, and retrieve the pushing performance health index of the abnormal weak point before and after each torque compensation.

[0063] The improvement rate of the health index of the current abnormal weak point is determined by the change between the deviation of the thrust performance health index from the dynamic qualified threshold range before torque compensation and the deviation of the thrust performance health index from the dynamic qualified threshold range after torque compensation.

[0064] When the thrust performance health index after torque compensation is closer to the dynamic qualified threshold range, and no extreme stress value, strain peak value or vibration amplitude increase occurs at adjacent test points, the torque compensation of the wheel is considered to have a high health index improvement efficiency.

[0065] When the health index of the current abnormal weak point is improved, but the deviation of the health index of adjacent test points increases, or a new abnormal weak point is formed, the efficiency of the health index improvement of the wheel torque compensation is reduced.

[0066] When the improvement of the current abnormal weak point is less than the deterioration of the adjacent test point, it is determined that the torque compensation of the wheel has not improved the overall pushing performance of the cabinet structural components, and the torque compensation of the wheel is marked as invalid compensation.

[0067] The test control unit further compares the number of abnormal weak points, the maximum comprehensive over-limit degree, and the deviation of the pushing performance health index of each abnormal weak point in the results of two adjacent rounds of retesting to determine the convergence trend of the retest results.

[0068] When the number of abnormal weak points decreases round by round, the maximum comprehensive over-limit degree decreases round by round, and the pushing performance health index of each abnormal weak point gradually approaches the dynamic qualified threshold range, the retest results are considered to have a high degree of convergence.

[0069] When the number of abnormal weak points remains unchanged, the rate of decrease in the maximum comprehensive over-limit decreases round by round, or the abnormal location changes repeatedly between different test points, it is determined that the convergence of the retest results has decreased.

[0070] The test control unit uses the remaining allowable torque of the corresponding locking bolt as the basis for determining the maximum torque compensation amount for a single operation, ensuring that the actual torque after a single torque compensation is completed does not exceed the upper limit of the allowable torque.

[0071] Within the remaining allowable torque range, the first compensation range is determined based on the overall degree of exceeding the limit of the current abnormal weak point and the proportion of the remaining allowable torque.

[0072] When the overall exceedance level is high and the remaining allowable torque ratio is large, the first compensation range is increased; when the overall exceedance level is low, the first compensation range is decreased; when the remaining allowable torque ratio is small, even if the overall exceedance level is high, the first compensation range is limited.

[0073] Subsequently, the second compensation magnitude was determined based on the health index improvement efficiency corresponding to each torque compensation and the convergence degree of the retest results.

[0074] When the efficiency of health index improvement is high and the retest results continue to converge, the second compensation range can be maintained or appropriately increased within the torque safety range; when the efficiency of health index improvement decreases or the convergence of retest results decreases, the second compensation range is reduced; when the previous round of torque compensation is marked as invalid compensation, the previous round of compensation range is no longer used.

[0075] The first compensation range and the second compensation range are combined and processed together. The remaining allowable torque of the corresponding locking bolt is used to limit the result of the combined processing, so as to form the single target torque compensation amount of the current abnormal weak point.

[0076] By simultaneously incorporating the overall degree of exceeding limits and the proportion of remaining permissible torque into the compensation priority determination in step S61 and the single target torque compensation amount determination in step S62, the severity of the anomaly and the torque safety margin can be continuously transferred from the compensation order determination stage to the actual compensation magnitude determination stage.

[0077] When the response coupling between the current abnormal weak point and the adjacent test point is low, and the single target torque compensation amount accounts for a small proportion of the remaining allowable torque, the torque tightening actuator completes the torque compensation in one go according to the single target torque compensation amount.

[0078] After the torque tightening actuator sends a feedback signal indicating that the torque has been executed, the test control unit re-executes the loading path generation, multi-source data acquisition, pushing performance feature extraction, pushing performance health index calculation, and abnormal part location according to the test process described in Example 1.

[0079] When the current abnormal weak point has a moderate degree of response coupling with the adjacent test point, or when the proportion of the single target torque compensation to the remaining allowable torque exceeds the preset classification ratio, the single target torque compensation is divided into two compensation stages.

[0080] In the first compensation stage, a large portion of the single target torque compensation amount is applied. After completion, the current abnormal weak point and adjacent test points that have a direct force transmission relationship with it are immediately retested.

[0081] When the retest results show that the health index of the pushing performance of the current abnormal weak point has improved and no new over-limit conditions have appeared at the adjacent test points, the second compensation stage is executed; when the extreme value of force, peak value of strain or vibration amplitude of the adjacent test points increases significantly, the second compensation stage is cancelled.

[0082] When the current abnormal weak point has a high degree of response coupling with the adjacent test point, or when abnormal position migration has occurred after previous torque compensations, the single target torque compensation amount is divided into three compensation stages.

[0083] The three compensation phases are executed in the order of applying the larger portion first, followed by the two smaller portions. After each compensation phase is completed, the push-pressure performance test process corresponding to the current abnormal weak point and adjacent test points is restarted, and the health index improvement efficiency, the response coupling degree of adjacent test points, and the convergence degree of the retest results are re-determined.

[0084] After any compensation stage is completed, if the increase in the extreme value of force, peak value of strain, or vibration amplitude of adjacent test points relative to the value before compensation exceeds the preset allowable increase limit, the subsequent compensation stage will be stopped.

[0085] When the deviation between the actual torque increment fed back by the torque tightening actuator and the single target torque compensation exceeds the allowable execution deviation, the test control unit suspends the current compensation process, does not use the test data of this round for updating the efficiency of health index improvement and the convergence of retest results, and generates a torque execution anomaly flag.

[0086] By adaptively employing single-stage, two-stage, or three-stage compensation based on the degree of coupling between adjacent test points and the amount of single-stage target torque compensation, the load redistribution state at highly coupled connection locations can be observed step by step through small torque increments, reducing the risk of local stress concentration at adjacent connection locations caused by applying a large torque at once.

[0087] S63. Generate a closed-loop optimization instruction set and determine the closed-loop termination condition; After each torque compensation or compensation stage is completed, the abnormal weak points before and after torque compensation are matched point by point according to the test point coordinates and component names in the abnormal part location data.

[0088] For abnormal states that remain at or near the original test point after torque compensation, they are marked as in-situ residual abnormalities.

[0089] If, after torque compensation, the original abnormal weak point recovers to the dynamic qualified threshold range, but other test points newly show an out-of-limit state, it is marked as a newly added migration abnormality.

[0090] If the original weak point has not been recovered and new out-of-limit states appear at adjacent test points at the same time, it is marked as an extended migration anomaly.

[0091] The system reads the number of newly added abnormal weak points, the spatial migration distance of the abnormal location relative to the original abnormal location, the degree of increase in the extreme values ​​of the stress at adjacent test points, the degree of increase in the peak strain, the degree of increase in the vibration amplitude, and whether the component to which the abnormal weak point belongs has changed, in order to determine the degree of migration of the abnormal location.

[0092] If the location of the anomaly remains unchanged and the current weak point improves synchronously with the adjacent test point, the degree of migration of the anomaly location will be determined to be within a low range.

[0093] When the number of newly added abnormal weak points increases, the abnormal location shifts from the top beam to the side door hinge or the bottom load-bearing foot, or the response of adjacent test points deteriorates significantly, the degree of abnormal location migration should be increased.

[0094] The test control unit combines the compensation priority formed in step S61, the single target torque compensation amount formed in step S62, the remaining allowable torque of the corresponding locking bolt, the degree of abnormal position migration, and the degree of convergence of the retest results to form a closed-loop execution judgment result.

[0095] A high compensation priority indicates that the corresponding abnormal weak point is suitable for continued compensation; the proportion of the single target torque compensation amount to the remaining allowable torque is within a safe range, indicating that the compensation action has a low risk of over-locking; the degree of abnormal position migration is low, indicating that the risk of continued compensation causing abnormal transfer is low; the degree of convergence of the retest results is high, indicating that continuous compensation is gradually bringing the cabinet structural components closer to a qualified state.

[0096] When the compensation priority is high, the single target torque compensation amount is within a safe range, the degree of abnormal position migration is low, and the retest results continue to converge, the closed-loop execution judgment result is improved.

[0097] When the degree of abnormal location migration increases, the convergence of retest results decreases, or the remaining allowable torque of the corresponding locking bolt is insufficient, the closed-loop execution judgment result is reduced.

[0098] The test control unit rearranges the remaining abnormal weak points from high to low according to the closed-loop execution judgment results, and binds the coordinates of the abnormal weak points, the corresponding locking bolts, the single target torque compensation amount, the number of compensation stages, the retest range, the compensation execution order and the termination mark to generate a closed-loop optimization instruction set.

[0099] When the closed-loop execution judgment result meets the conditions for continued compensation, the degree of abnormal position migration is lower than the migration safety threshold, the convergence degree of the retest result remains stable or increases, and the corresponding locking bolt has sufficient remaining allowable torque, the test main control unit selects the abnormal weak point with the highest closed-loop execution judgment result and issues the next compensation command to the torque tightening actuator.

[0100] When the response coupling degree between adjacent test points between two abnormal weak points is lower than the preset independent threshold, the test control unit can complete the torque compensation of the two abnormal weak points in sequence according to the closed-loop execution judgment result, and then restart the push-press performance test process in a unified manner.

[0101] When there is a clear structural force transmission relationship between two abnormal weak points or when there is a high degree of response coupling between adjacent test points, torque compensation is performed only on the abnormal weak point with the highest closed-loop execution judgment result each time. After the compensation is completed, a retest is performed immediately, and torque compensation commands are not issued to multiple highly coupled abnormal weak points at the same time.

[0102] When the push-pressure performance health index of all test points enters the dynamic qualified threshold range, and no new migration anomalies or extended migration anomalies are found in two consecutive retests, the cabinet structural components are deemed to have completed calibration, and a closed-loop test completion signal is generated.

[0103] When the actual torque of the corresponding locking bolt reaches the upper limit of the allowable torque, the automatic torque compensation for that locking bolt will stop.

[0104] If, after two consecutive rounds of torque compensation, the deviation of the pushing performance health index of the original abnormal weak point from the dynamic qualified threshold does not decrease significantly, it is determined that continuing to increase the torque of the corresponding locking bolt will not improve the abnormal state, and the automatic compensation for the abnormal weak point will be stopped.

[0105] When the convergence of the results of two consecutive retests decreases, it is determined that the continuous torque compensation has not brought the cabinet structural components to a stable and qualified state, and automatic compensation is stopped.

[0106] When the migration degree of the abnormal location exceeds the migration safety threshold, or when the overall excess degree of the new abnormal weak point after torque compensation is higher than the overall excess degree of the original abnormal weak point before compensation, it is determined that the current torque compensation has caused an unfavorable load redistribution, and automatic compensation is stopped.

[0107] When an abnormal state repeatedly migrates between two or more test points, or when the number of automatic torque compensation cycles reaches the preset maximum number of cycles, automatic compensation is stopped to avoid continuous looping of testing and calibration.

[0108] After automatic compensation stops, the test control unit no longer increases the target locking torque of the corresponding locking bolt. It binds the current abnormal part location data, the data of each torque compensation process, the abnormal position migration record, the actual torque in place, and the reason for automatic compensation termination, writes them into the final state data of the pushing performance test, and generates a closed-loop test completion signal with a manual verification mark.

[0109] Step S61 determines the compensation sequence for multiple abnormal weak points, step S62 determines the single target torque compensation amount constrained by the torque safety space and the improvement effect of each iteration, and step S63 determines whether to continue compensation based on the abnormal position migration and the convergence trend of the retest results. This makes the torque compensation process for multiple abnormal weak points a process in which the compensation sequence, compensation magnitude, compensation feedback, and closed-loop exit are interconnected, reducing the risks of overtightening of locking bolts, damage to connecting parts, abnormal position migration, and failure to terminate the test and calibration closed loop.

[0110] 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 the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for testing the push-press performance of cabinet structural components after assembly, characterized in that, The method includes: S1. Analyze and process the CAD model data of the rack structure after assembly, automatically identify and locate the key load-bearing parts in the model, mark the mechanical properties and extract the geometric parameters of each identified test point, and generate a multi-dimensional test point feature set; predict the loading direction and classify the force mode of the multi-dimensional test point feature set to generate candidate data for the initial loading path. S2. Perform real-time feedback fusion processing on the initial loading path candidate data using six-dimensional force sensors, and dynamically correct the loading direction and displacement trajectory by combining the force signals collected by the sensors to generate adaptive loading path data; generate multi-modal synchronous acquisition instructions from the adaptive loading path data to generate multi-source synchronous acquisition instruction data. S3. Perform convolutional neural network feature extraction processing on the multi-source acquisition synchronous command data, input the acquired force and displacement curves, strain distribution data and structural vibration mode data into the pre-trained CNN model to generate push-press performance feature vector data; perform health index construction processing on the push-press performance feature vector data to generate push-press performance health index data. S4. Perform dynamic threshold comparison processing on the pushing performance health index data, quantitatively compare the health index with the preset dynamic qualified threshold, and generate qualified judgment result data; perform anomaly location analysis on the qualified judgment result data, trace the weak parts of the unqualified products, and generate anomaly location data. S5. Perform closed-loop testing and calibration on the abnormal location data, trigger local fine-tuning commands for the located weak points, execute the locking bolt torque compensation operation, restart the pushing performance test process, and generate a closed-loop test completion signal; archive the test report on the closed-loop test completion signal to generate the final state data of the pushing performance test.

2. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 1, characterized in that, S1 includes: S11. Read the original data stream of the CAD model corresponding to the assembled cabinet structural components, traverse the model entity layer to divide the component units, distinguish the three types of load-bearing component entities: top beam, side door hinge, and bottom load-bearing foot, scan the coordinate values ​​of the entity boundary nodes, mark the spatial position coordinate parameters of each node, and output the original dataset of component entity coordinate points; traverse the model material thickness, cross-sectional dimensions, and assembly gap values ​​associated with each point in the original dataset of coordinate points, enter the corresponding geometric parameter entries for the points, and output the complete dataset of point geometric parameters. S12. Traverse the complete dataset of geometric parameters of the points, match the mechanical constants of material yield strength, elastic modulus, and Poisson's ratio, bind them to the corresponding coordinate points to generate point mechanical attribute annotation entries, integrate all annotation entries to form a multi-dimensional test point feature set containing coordinate, geometric, and mechanical information; retrieve the component support constraint distribution state, deduce the force transmission direction under external force for each test point, classify the three basic stress modes of compression, shear, and bending, assign corresponding stress mode codes, and output the initial stress mode classification dataset. S13. Read the force codes of each test point in the initial force pattern classification dataset, arrange multiple external force application trajectories along the force transmission direction, record the starting and ending point travel angle displacement increment values ​​of each trajectory, and summarize all trajectory entries to form initial loading path candidate data; traverse each trajectory node of the initial loading path candidate data, calculate the stress difference deformation amplitude values ​​between nodes, reduce the displacement step size of trajectory nodes whose deformation amplitude exceeds the safe range, and output the preprocessed loading path candidate data. S14. Traverse the candidate data of the preprocessed loading path to match the overall assembly support boundary constraint range of the component, trim the invalid travel segments that extend the trajectory outside the constraint boundary, correct the coordinate values ​​of the trajectory endpoint, and output the boundary constraint adapted loading path dataset; compare the overall stress fluctuation amplitude of each path after adaptation, sort the path entries from low to high according to the stress fluctuation amplitude, and output the path priority sorted dataset. S15. Read the first priority loading path entry in the path priority sorting dataset, enter the initial running configuration cache of the loading execution unit, fill in the configuration values ​​of the upper limit of loading speed, displacement, upper limit of loading force, and running time, and output the initial configuration cache data of the loading unit; verify that there are no numerical overflow conflicts of the three types of parameters of speed, force, and displacement in the cache data, push the cache data to the loading driver hardware register, and generate the initial loading start preparation state signal.

3. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 1, characterized in that, The S2 includes: S21. Receive the initial loading start-up preparation state signal to start the continuous sampling process of the six-dimensional force sensor, collect the real-time values ​​of XYZ three-axis thrust and tension and three-axis torque at equal time intervals, package the force sensing sampling frame sequence per unit time, and output the continuous force sensing signal time sequence sampling dataset; read the real-time execution displacement position value of the loading unit, synchronously package the displacement time sequence sampling frame sequence, and output the displacement trajectory time sequence sampling dataset. S22. Synchronously align the timestamp nodes of the force sensor timing sampling dataset and the displacement trajectory timing sampling dataset. Under the same timestamp, match the three sets of values ​​of force, torque and displacement to construct a spatiotemporal matching sampling unit. Summarize all units to form a force-displacement synchronous fusion sampling dataset. Traverse each spatiotemporal matching unit of the fusion sampling dataset, calculate the angle difference between the real-time force vector direction and the preset loading and travel direction, and output the direction deviation difference sequence data. S23. Read the direction deviation difference sequence data one by one, adjust the loading angle correction amount in the next sampling period according to the positive and negative magnitude of the difference, simultaneously fine-tune the displacement step amplitude per unit period, iterate and update the coordinates of the entire trajectory points frame by frame, and output the adaptive loading path data after frame-by-frame correction; traverse all trajectory segments of the adaptive loading path data, calculate the instantaneous force change gradient values ​​of adjacent points, and mark the trajectory segments whose gradient values ​​exceed the impact safety threshold. S24. Reduce the displacement step increment and force rise slope values ​​of the trajectory segment marked with impact risk, recalculate the coordinates and loading rate values ​​of all points within the segment, and output the optimized adaptive loading path data after the impact risk is resolved; split the time axis nodes within the optimized adaptive loading path data, generate a unified time-series trigger mark code for each time node, and output the time-series synchronous trigger mark sequence data. S25. Distribute the timing synchronization trigger mark sequence data to the registers of the three sets of acquisition hardware units. The mark code is synchronously pushed to the force and displacement curve acquisition unit, strain distribution acquisition unit, and structural vibration mode acquisition unit. The three sets of units synchronously latch their own acquisition start time and generate multi-source acquisition synchronization instruction data. Latch the system timing reference value at the time of the synchronization instruction issuance, enter the acquisition timing reference log entry, and output the acquisition timing reference log data.

4. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 3, characterized in that, S22 includes: Retrieve the continuous force sensing signal timing sampling dataset and the displacement trajectory timing sampling dataset, batch analyze the timing coding parameters of the timing frames within the two datasets, unify the sampling timing accuracy and frame synchronization interval of the two datasets, and generate a timing reference unified sampling dataset; perform timing node calibration processing on the timing reference unified sampling dataset to generate a timing precise alignment dataset; The time-series precisely aligned dataset is reorganized into frame units. Using a unified time-series node as the matching benchmark, the triaxial force value, triaxial torque, and real-time displacement value corresponding to the same time-series node are bound and integrated to construct a single-frame integrated spatiotemporal matching sampling unit. All time-series frame units are integrated to generate a complete force-displacement synchronous fusion sampling dataset. Traverse all spatiotemporal matching sampling units within the force-displacement synchronous fusion sampling dataset, extract the three-dimensional coordinate parameters of the real-time force vector of each unit, combine the basic orientation parameters of the loaded trajectory to perform vector calculations, solve the spatial angle offset between the real-time force vector and the standard orientation, and generate a set of single-unit angle offset values. The set of angle offset values ​​of a single unit is processed across the entire domain, and the angle offset values ​​of each sampling unit are concatenated in the time sequence to generate continuous and stable angle offset time series data. The continuous and stable angular offset time series data is encapsulated in a data structure, and the numerical storage format and time series arrangement rules are unified to fully record the angular offset change information throughout the time period and generate the direction deviation difference sequence data for trajectory correction.

5. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 1, characterized in that, The S3 includes: S31. Receive the multi-source acquisition synchronization command data to start the parallel acquisition operation of three sets of units. The force and displacement curve acquisition unit outputs the full-time force and displacement paired sampling sequence, the strain distribution acquisition unit outputs the strain amplitude distribution matrix of each test point, and the structural vibration mode acquisition unit outputs the multi-order vibration frequency amplitude phase matrix. Output three types of independent raw acquisition datasets; unify the sampling interval and timestamp format of the three types of raw acquisition datasets, align the time sequence nodes of each sampling frame, and output the time sequence aligned multi-source raw acquisition fusion dataset. S32. Divide the time-aligned multi-source raw acquisition and fusion dataset into fixed-length time-series sample slices. Perform normalized numerical scaling on each sample slice to scale it to the zero-to-one value range and output the normalized time-series sample slice dataset. Load the weight bias parameters of the completed and converged convolutional neural network and input the normalized time-series sample slice dataset into the first convolutional operation layer of the network. Perform sliding multiplication and addition operations of the convolutional kernel to extract shallow temporal spatial features. S33. Shallow temporal and spatial features are compressed through pooling layers to reduce feature dimensions. The compressed features are then fed into multiple fully connected layers for feature fusion and mapping operations, reducing redundant feature components and amplifying effective push-compression related feature components layer by layer. The network outputs a fixed-length one-dimensional feature array, and all sample arrays are aggregated to form push-compression performance feature vector data. The push-compression performance feature vector data is split into three sub-feature vector subsets: force, deformation, and vibration. The mean, variance, extreme values, and dispersion statistics of each subset are statistically analyzed, and the sub-feature statistics dataset is output. S34. Retrieve the built-in mapping weight matrix of the pushing performance health index assessment model, multiply and sum the three types of sub-feature vectors of force deformation vibration with the corresponding weight matrix element by element to obtain the basic calculated value of a single health index, traverse all time series samples to generate the full time series health index basic value sequence data; perform moving mean smoothing filtering on the full time series health index basic value sequence data, and output the smoothed pushing performance health index data. S35. Extract the four statistical indicators of the entire process of the push-press performance health index data: maximum value, minimum value, average value, and fluctuation range. Bind the corresponding rack structural component assembly number and sampling start and end time to three sets of identification information. Output the complete health index archived pre-data with product identification. Verify that the archived pre-data has no empty values, garbled characters, or out-of-bounds values. Write it to a temporary storage buffer and output the health index temporary cache storage data.

6. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 1, characterized in that, The S4 includes: S41. Read the complete push performance health index sequence from the temporary cache storage data of the health index, retrieve the dynamic qualified threshold upper and lower limit values ​​of the batch of cabinet structural parts of the same specification, enter the threshold comparison operation cache area, and output the threshold comparison benchmark configuration data; take out the health index value of each time node, compare the value with the upper and lower limits of the threshold respectively, mark each time node with three categories of judgment labels: above the upper limit, within the interval, and below the lower limit, and output the preliminary label dataset of a single node. S42. Calculate the percentage of nodes that are within the range throughout the entire process in the initial label dataset of a single node, the percentage of nodes with a cumulative duration below the lower limit, and the percentage of nodes with a peak duration above the upper limit. Output the statistical dataset of the percentage of nodes. Set the numerical range of the critical interval. Nodes whose health index falls into the critical interval are not directly marked as unqualified. Separately divide the critical sample subset and output the critical sample isolated dataset and the regular qualified and unqualified sample dataset. S43. Generate a preliminary qualification mark for the overall product. Retrieve the slope values ​​of the health index changes in adjacent time periods. If the slope is stable with small fluctuations, it is classified as qualified. If the slope rises or falls sharply, it is classified as unqualified. Integrate all marks to form qualification result data. Traverse the qualified result data and mark the product items as unqualified. Retrieve the original force, strain and vibration sampling values ​​of the product from the multi-source acquisition and fusion dataset throughout the process. S44. Backtrack the original sampled values ​​and match the coordinates of each point in the feature set of multi-dimensional test points one by one. Compare the strain peak value, stress extreme value, and vibration amplitude of each point. Sort the point entries according to the extreme value values ​​from large to small and output the original sequence data of weak point sorting. S45. Bind the component name, assembly position, geometric parameters, and mechanical properties of each point in the simplified weak point sequence data, assemble them into a complete set of entries, and generate abnormal part location data; enter the over-limit value and the percentage of the over-limit value for each weak point in the abnormal part location data, and output the quantitative auxiliary data of the degree of over-limit of the point.

7. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 1, characterized in that, The S5 includes: S51. Read the abnormal location data and the quantitative auxiliary data of the degree of exceeding the limit of the point, match the standard torque reference value of the assembly locking bolt corresponding to each weak point, adjust the reference torque compensation increment value according to the percentage of exceeding the limit, generate the target locking torque value of the bolt for each point, and output the local fine-tuning torque command dataset; split the local fine-tuning torque command dataset into independent command frames for each point, push the command frames to the torque tightening actuator controller register, and generate the weak point fine-tuning execution command signal. S52. The torque tightening actuator receives the fine-tuning execution command signal, applies the target locking torque point by point to complete the bolt compensation locking operation, and after the locking operation is completed, it feeds back the torque execution in place response signal and outputs torque locking completion response timing data; the response timing data is pushed to the test main control unit, the main control unit clears all the previously collected cached feature vector health index judgment result storage data, releases the temporary storage space occupied, and outputs test cache clearing completion status data; S53. After the test cache is cleared and the status data is cleared, the main control unit restarts the entire push-pressure performance test process. The loading path is repeatedly loaded to generate multi-source data, collect CNN features, extract health index, calculate and judge the entire process, and output the original data set of the second test process. The feature vector health index calculation steps are repeated on the original data set of the second test process to obtain the second push-pressure performance health index and the second pass judgment result data. S54. Compare the secondary pass / fail judgment result data with the initial judgment result data. If the secondary judgment marks the pass / fail, the loop test process is terminated. If the secondary judgment still marks the fail / fail, the abnormal part location data is regenerated and the fine-tuning instruction is issued again. The loop iterates until the judgment mark is pass / fail. The closed-loop test completion signal is output at the moment the iteration ends. The closed-loop test completion signal triggers the full-process data packaging process. S55. After packaging, the entire process data is uniformly converted into a structured row-column storage format, and the product assembly number, test start and end time, equipment number, operator number, and auxiliary identifier fields are filled in. The structured packaging final state dataset is output. The structured packaging final state dataset is written to the persistent storage disk partition to create independent index entries, and push-pressure performance test final state data is generated. The read and write permissions of the final state data index entries are verified to be complete and undamaged. The storage completion log record entries are entered, and the archived storage log auxiliary data is output.

8. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 7, characterized in that, S53 includes: S531: Receive test buffer clearing completion status data, activate the process restart trigger mechanism of the test main control unit, reset the loading parameters, acquisition parameters and calculation parameters of the test system, and generate secondary test system initialization data; perform whole machine parameter self-check on the secondary test system initialization data, and generate secondary test ready status data; S532. Restart the loading path generation process based on the secondary test ready status data, replicate the standard adaptive loading path generation process, and generate secondary test adaptive loading path data; start the hardware loading mechanism based on the secondary test adaptive loading path data, complete the standard push loading action of the cabinet structural components, and generate secondary loading execution process data. S533. Synchronously trigger the parallel operation of multi-source acquisition units to complete the full-time acquisition and storage of force displacement data, strain distribution data, and structural vibration mode data according to a unified time series reference. Summarize all acquired information to complete time series alignment and data fusion processing, and output the original acquisition dataset of the entire secondary test process. Perform sample slicing and normalization scaling processing on the original acquisition dataset of the entire secondary test process to generate secondary standardized test sample data. S534. Input the secondary standardized test sample data into the pre-set convolutional neural network model, and complete the layer-by-layer extraction and fusion of shallow and deep features through multi-layer convolution and pooling operations to generate secondary pushing performance feature vector data; classify and decompose the secondary pushing performance feature vector data, divide it into three feature subsets of force, deformation and vibration, and generate secondary multi-dimensional feature subset data. S535. Import the secondary multi-dimensional feature subset data into the pushing performance health index evaluation model, complete the multi-dimensional feature weighting operation and numerical smoothing filtering, and output the secondary pushing performance health index data; connect the secondary pushing performance health index data to the dynamic threshold comparison process, complete the quantitative rating and qualification judgment, and generate secondary qualification judgment result data.

9. The method for testing the push-press performance of a cabinet structural component after assembly according to claim 8, characterized in that, S534 includes: The system receives secondary standardized test sample data, and then sends the standardized sample data in batches to the input layer of the convolutional neural network to complete the model loading and parameter adaptation processing of the full-domain sample data, generating model input layer adaptation data. The model input layer adaptation data is then distributed hierarchically and allocated to each convolutional operation unit according to the network operation logic, generating hierarchical operation arrangement data. By utilizing each convolutional operation unit to perform sliding convolution multiplication and addition operations on the split data, the temporal variation characteristics and spatial distribution characteristics within the sample data are mined, and the shallow features are fully extracted and integrated to generate secondary test shallow feature mapping data. The secondary test shallow feature mapping data is then subjected to pooling dimensionality reduction processing to generate dimensionality-reduced compressed feature data. The dimensionality-reduced and compressed feature data is sent to the fully connected network layer to carry out cross-dimensional feature fusion iterative operations. Irrelevant noise feature components are suppressed and removed step by step to strengthen the core related features corresponding to the rack push-up performance and generate deep fusion feature data. The deep fusion feature data is then subjected to dimensional regularization and sequence sorting to generate secondary push-up performance feature vector data with unified dimensions. Read the global feature components of the secondary push performance feature vector data, classify and split them according to the test parameter attribute features, separate independent feature component groups, and generate multi-class feature split original data; group, normalize and sequence integrate the multi-class feature split original data to form secondary multi-dimensional feature subset data.

10. A system for testing the push-press performance of cabinet structural components after assembly, characterized in that, The system includes: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

Images