Rail transit interface assembly reassembling state data processing method and system

Abstract

The application relates to the technical field of data processing, and discloses a rail transit interface assembly reassembling state data processing method and system, wherein the method comprises the following steps: obtaining an equal displacement interval pressure sequence; confirming a physical contact starting point; segmenting a contact deformation sequence through saliency extreme value tracking and physical consistency constraint; extracting a shape feature vector; and outputting an assembly state category label by using cost-sensitive LightGBM integrated learning. Compared with the prior art which only checks a final pressure value and ignores waveform features in an assembly process, especially in the case that pressure curve forms are diverse due to component tolerances and operation differences, the technical problem that precise classification of process abnormalities such as seal cutting and assembly skewing cannot be realized. Due to the fact that the application extracts curve shape features which are robust to individual differences and combines cost-sensitive integrated learning, precise classification of various assembly abnormalities is realized, and the automatic detection capability of assembly process defects is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, in particular to a method and system for processing reassembly status data of rail transit interface components. Background Art

[0002] At present, in the reassembly process of rail transit interface components, presses are widely used to press-fit seals, and pressure-displacement data sequences are collected synchronously, so as to ensure assembly quality by monitoring the assembly process. Existing processes generally only use the pressure value at the end of press-fitting as the evaluation basis, and compare it with preset upper and lower threshold: as long as the final value falls within the allowable range, it is determined as qualified, otherwise it is determined as abnormal. This detection method based on a single end-point completely ignores the rich dynamic information contained in the pressure curve in the whole process from physical contact, elastic pressing to final fitting, and its data processing path usually only involves simple extraction and threshold comparison of the end-point, lacking an analysis link for the complete curve shape.

[0003] For example, in actual reassembly scenarios, process defects such as trimmed sealing edges, inclined installation or inadequate assembly often occur. These defects will cause local shape distortion in the waveform formed by press-fitting displacement and pressure, such as abnormal local pressure drop, slope break or high-frequency fluctuation; however, due to individual manufacturing tolerances of the sealing components themselves, superimposed with human differences in press-fitting speed and force from operators, the pressure curves of the same type of defects have diverse shapes, but the final pressure may still remain within the normal range, resulting in a large number of hidden defects being missed. Such missed detection allows components with process defects to flow into subsequent operation links, which directly buries potential safety hazards to the long-term air tightness, pressure resistance and electrical contact stability of rail transit interfaces. The existing single final value judgment technology cannot fully meet the demand for accurate anomaly detection in scenarios where assembly process waveforms are variable.

[0004] Therefore, there is an urgent need for a processing method that can automatically extract curve shape features robust to individual differences from the complete pressure-displacement curve, and perform reliable automatic classification of various assembly anomaly types on this basis, so that even under complex conditions where component tolerances and operation differences lead to variable pressure curve waveforms, process anomalies such as sealing edge trimming, inclined installation and inadequate assembly can still be accurately identified, thereby greatly improving the detection effectiveness of reassembly quality monitoring, providing clear anomaly category guidance for targeted rework, and improving overall assembly consistency and long-term service reliability. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to propose a method for processing data on the reassembly status of rail transit interface components. This method addresses the problem in existing technologies that only check the final pressure value while ignoring waveform characteristics during the assembly process. In particular, when the pressure curve shape varies due to component tolerances and operational differences, it is impossible to accurately classify process anomalies such as edge cutting and skewed installation of seals.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for processing the reinstallation status data of rail transit interface components.

[0007] The method for processing reinstallation status data of rail transit interface components includes:

[0008] Step S10: Obtain the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component. Based on the pressure data sequence and the displacement data sequence, perform waveform preprocessing tasks using an adaptive median filtering and linear interpolation resampling mechanism to output an equal displacement interval pressure sequence.

[0009] Step S20: Based on the equal displacement interval pressure sequence, a dynamic baseline adaptive threshold and stiffness confirmation mechanism is used to perform the physical contact initiation detection task, and the contact initiation index is output;

[0010] Step S30: Based on the contact start point index and the equal displacement interval pressure sequence, a significant extreme value tracking and physical consistency constraint mechanism is used to perform key inflection point detection and curve segmentation tasks, and output the contact deformation segment sequence;

[0011] Step S40: Based on the contact deformation segment sequence, a quantile robust normalization and hybrid morphological feature quantization mechanism are used to perform shape feature extraction, and a hybrid shape feature vector is output.

[0012] Step S50: Based on the hybrid shape feature vector, the cost-sensitive LightGBM ensemble learning mechanism is used to perform the assembly anomaly classification task and output the assembly state category label.

[0013] Preferably, step S10, which involves acquiring the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component, and performing waveform preprocessing based on the pressure data sequence and the displacement data sequence using an adaptive median filtering and linear interpolation resampling mechanism to output an equal displacement interval pressure sequence, specifically includes:

[0014] Step S101: Real-time synchronous acquisition of the cumulative displacement and corresponding pressure value of the press head to form a raw data sequence set;

[0015] Step S102: Perform adaptive median filtering of the guard edge on the pressure signal in the original data sequence set to obtain the filtered sequence;

[0016] Step S103: Perform linear interpolation on the filtered sequence based on the shift, and resample to equal intervals. On the grid, pressure sequences with equal displacement intervals are obtained. ,in , This represents the length of the resampled sequence.

[0017] Preferably, step S20, which involves performing a physical contact initiation detection task based on the equal displacement interval pressure sequence using a dynamic baseline adaptive threshold and stiffness confirmation mechanism, and outputting the contact initiation index, specifically includes:

[0018] Step S201: Calculate the equal displacement interval pressure sequence First-order difference and second-order difference and to and Savitzky-Golay smoothing is performed to obtain the smoothed first-order difference. and smooth second-order difference ;

[0019] Step S202: Set a length of sliding window Continuously calculate the average pressure value within the window. and standard deviation Generate adaptive threshold :

[0020]

[0021] in, For the first Adaptive threshold at each sampling point This represents the average pressure value within the sliding window. The standard deviation of the pressure values ​​inside the sliding window. The sensitivity coefficient, This is the amount of compensation for sensor quantization error;

[0022] Step S203: The pressure value must be greater than the adaptive threshold and the smoothed first-order difference must be greater than the preset initial stiffness threshold. The first index is determined to be the contact start index. .

[0023] Preferably, in step S202, the sensitivity coefficient Take 3.5.

[0024] Preferably, step S30, which involves performing key inflection point detection and curve segmentation tasks based on the contact start point index and the equal displacement interval pressure sequence using a significant extreme value tracking and physical consistency constraint mechanism, and outputting a contact deformation segment sequence, specifically includes:

[0025] Step S301: Calculate significance index based on smoothed second-order difference And search for indicators that make significance Maximize candidate inflection points :

[0026]

[0027] in, It is a symbolic function;

[0028] Step S302: For the candidate inflection points A physical consistency constraint is applied, requiring that the smooth first-order difference at the point is greater than zero and that the pressure value at the equal displacement interval is within the range of 15% to 75% of the entire equal displacement interval pressure sequence. If the physical consistency constraint is met, it is identified as a critical inflection point. ;

[0029] Step S303: Utilize the contact origin index The aforementioned key turning point and preset endpoint From the equal displacement interval pressure sequence, segment from +1 to The subsequence is used as the contact deformation segment sequence.

[0030] Preferably, step S40, which involves performing shape feature extraction based on the contact deformation segment sequence using quantile robust normalization and a hybrid morphological feature quantization mechanism, and outputting a hybrid shape feature vector, specifically includes:

[0031] Step S401: Resample the contact deformation segment sequence to a fixed length using cubic spline interpolation. This yields a fixed-length sequence;

[0032] Step S402: Perform quantile robust normalization on the fixed-length sequence to obtain a normalized contact deformation sequence. ;

[0033] Step S403: From the normalized contact deformation sequence Extract a hybrid shape feature vector containing detrended volatility features and morphological envelope entropy features; wherein, the detrended volatility features include... The root mean square and zero-crossing count of the detrended sequence obtained after straight-line fitting, the morphological envelope entropy feature is obtained by... Multi-scale morphological filtering is used to extract the upper envelope sequence. Then, the envelope sequence The subband energy entropy is obtained by performing three-level wavelet packet decomposition.

[0034] Preferably, step S50, which involves performing an assembly anomaly classification task based on the hybrid shape feature vector using a cost-sensitive LightGBM ensemble learning mechanism and outputting assembly state category labels, specifically includes:

[0035] Step S501: Input the hybrid shape feature vector into the cost-sensitive LightGBM classifier that has been pre-optimized and trained using Bayesian parameters;

[0036] Step S502: The cost-sensitive LightGBM classifier uses unequal class cost weights in its loss function. The unequal class cost weights are used to set the weight values ​​of abnormal classes such as seal edge cutting, seal skew installation, and improper assembly to be 5 to 10 times the weight values ​​of normal classes.

[0037] Step S503: Obtain the posterior probability vector output by the cost-sensitive LightGBM classifier, and output the category corresponding to the highest probability in the posterior probability vector as the assembly state category label. The assembly state category label includes normal, seal edge cut, seal skewed, and incomplete assembly.

[0038] The present invention also provides a rail transit interface component reassembly status data processing system, comprising:

[0039] The waveform preprocessing module is used to acquire the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component. Based on the pressure data sequence and the displacement data sequence, an adaptive median filtering and linear interpolation resampling mechanism is used to perform waveform preprocessing tasks and output a pressure sequence with equal displacement intervals.

[0040] The contact initiation detection module is used to perform physical contact initiation detection tasks based on the equal displacement interval pressure sequence using a dynamic baseline adaptive threshold and stiffness confirmation mechanism, and output the contact initiation index.

[0041] The curve segmentation module is used to perform key inflection point detection and curve segmentation tasks based on the contact start point index and the equal displacement interval pressure sequence, using a significant extreme value tracking and physical consistency constraint mechanism, and outputs a contact deformation segment sequence.

[0042] The shape feature extraction module is used to perform shape feature extraction based on the contact deformation segment sequence using quantile robust normalization and hybrid morphological feature quantization mechanism, and output a hybrid shape feature vector.

[0043] The anomaly classification module is used to perform assembly anomaly classification tasks based on the hybrid shape feature vector using a cost-sensitive LightGBM ensemble learning mechanism, and output assembly state category labels.

[0044] The present invention also provides a rail transit interface component reassembly status data processing device, the rail transit interface component reassembly status data processing device comprising: a memory, a processor, and a rail transit interface component reassembly status data processing program stored in the memory and executable on the processor, wherein the rail transit interface component reassembly status data processing program implements the above method when executed by the processor.

[0045] The present invention also provides a computer program product, the computer program product including a rail transit interface component reinstallation status data processing program, which implements the above method when executed by a processor.

[0046] The beneficial effects of this invention are as follows:

[0047] 1. This invention completes waveform preprocessing through adaptive median filtering and equal displacement resampling mechanism. The width of the filtering window is dynamically adjusted according to the pressure change gradient, which effectively suppresses measurement noise while protecting the details of curve transition edges. Resampling to equal displacement interval grid eliminates the problem of inconsistent displacement step size caused by non-uniform pressing speed, providing a stable and standardized signal foundation for subsequent displacement-based differential operations and curve segmentation, and indirectly improving the accuracy of contact start point detection and transition point positioning.

[0048] 2. This invention introduces a dual judgment of dynamic baseline adaptive threshold and stiffness confirmation in the detection of physical contact start point. It uses a sliding window to statistically analyze the pressure baseline level in real time, so that the judgment threshold changes adaptively with the current pressure background. This avoids false triggering caused by differences in component preload or sensor zero drift due to fixed threshold. At the same time, it uses the stiffness condition of smooth first-order difference to eliminate random fluctuations, ensuring that the contact start point index falls accurately at the position where the workpiece and the seal actually begin to be loaded. This isolates the interference of empty stroke and allows subsequent analysis to focus on the effective contact deformation segment.

[0049] 3. This invention uses significant extreme value tracking and physical consistency constraints to locate key inflection points. By maximizing the significance index, it automatically captures the characteristic position of the curve transitioning from a steep rise to a gradual rise. It also uses pressure range constraints and positive difference constraints to filter out false extreme values ​​caused by noise in the initial segment or compaction fluctuations at the end. This can reliably segment the contact deformation segment that reflects the deformation process of the seal, providing semantically consistent and accurately segmented subsequences for shape feature extraction.

[0050] 4. This invention implements quantile robust normalization and hybrid morphological feature quantization on the contact deformation segment, and uses the 2nd and 98th percentiles to replace the extreme values ​​for scaling, which greatly reduces the impact of occasional burrs or abnormal sampling points on the normalization results. At the same time, the extracted detrended volatility feature and morphological envelope entropy feature respectively characterize the local fluctuation characteristics and shape texture complexity of the curve. The resulting hybrid shape feature vector has strong distinguishability for defects such as edge cutting and skew during the assembly process, and is not sensitive to changes in the absolute value of pressure amplitude, ensuring feature stability under individual differences.

[0051] 5. This invention uses the cost-sensitive LightGBM ensemble learning mechanism to perform classification tasks. By assigning significantly higher misclassification cost weights to rare abnormal classes such as edge cutting and skewed assembly in the loss function than to normal classes, the model is guided to focus on learning abnormal morphologies during training. Then, Bayesian optimization is used to automatically select the optimal hyperparameters, which effectively overcomes the model bias problem caused by the scarcity of abnormal samples in actual production, and improves the recall ability and overall classification reliability of various assembly abnormalities. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the first embodiment of a method for processing reinstallation status data of a rail transit interface component according to the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0054] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the method for processing reinstallation status data of rail transit interface components according to the present invention. The first embodiment of the method for processing reinstallation status data of rail transit interface components according to the present invention is presented.

[0056] In the first embodiment, the method for processing reinstallation status data of a rail transit interface component includes:

[0057] Step S10: Obtain the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component. Based on the pressure data sequence and the displacement data sequence, perform waveform preprocessing tasks using an adaptive median filtering and linear interpolation resampling mechanism to output an equal displacement interval pressure sequence.

[0058] This step processes the pressure and displacement data sequences synchronously collected by the assembly equipment during the pressing of rail transit interface components. The pressure data sequence reflects the continuous output of the pressure sensor during the application of force by the pressure head, while the displacement data sequence reflects the cumulative pressing position of the pressure head relative to the interface component, seal, or mating part. Although the two are related during acquisition, the raw data is usually affected by fluctuations in pressing speed, sampling clock jitter, sensor quantization errors, and mechanical micro-vibrations. Directly using it for differential, inflection point detection, and classification can easily amplify non-process factors. Adaptive median filtering is used to first process local glitch in the pressure signal. Its window width is gradually adjusted according to the local pressure change: a wider window is used during the idle stroke or gentle pressing stage to suppress noise, and a narrower window is used when the pressure rises rapidly or the curve shows a bend to preserve the edge shape. Linear interpolation resampling uses displacement as a reference to map the filtered pressure signal onto an equally spaced displacement grid, forming an equally spaced pressure sequence.

[0059] After this processing, subsequent steps no longer deal with the original time series with uneven sampling density, but with a standardized sequence that is evenly spaced on the displacement axis, suppresses pressure spikes, and retains key rising edges. The equal-displacement-interval pressure sequence ensures that adjacent sampling points have the same displacement meaning. Therefore, when calculating the first-order difference, second-order difference, and sliding window statistics in step S20, it is no longer necessary to compensate for the point spacing differences caused by changes in pressing speed. Simultaneously, the filtering stage retains the curve edges near the contact start and pressure inflection points, avoiding the mistaken removal of slope changes truly caused by pressure on the seal as noise. Therefore, this step establishes a unified data entry point for physical contact start detection, key inflection point detection, and subsequent shape feature quantification.

[0060] In conventional reassembly testing, pressure curves are often smoothed only through a fixed window, or even the original sampled values ​​are used directly for final value judgment. Fixed window smoothing may flatten the contact rising edge when there is strong noise, and may leave sharp peaks when there are large local changes in the curve; direct processing according to time series can also transform the fluctuations in manual pressing speed into pseudo-differential features. For example, within the same actual displacement, the sampling points are sparse when the pressing speed is fast and dense when the speed is slow. If it is not converted to equal displacement intervals, the pressure change rate calculated later will include both the actual force change and the sampling density change. This step correlates the filtering intensity with the local curve state and shifts the analysis benchmark from the time axis to the displacement axis, which can reduce the interference of speed fluctuations and isolated spikes on the subsequent detection boundaries.

[0061] Taking a specific interface seal reassembly as an example, the average pressing speed of the press is approximately 5 mm / s. However, the operator's speed fluctuates between 3 mm / s and 7 mm / s near the end of the assembly, causing the original interval between adjacent displacement points to change from 0.02 mm to 0.15 mm. Furthermore, the pressure sensor exhibits approximately 1.5 Newtons of random noise during the non-contact phase. In this step, during the idle stroke phase, due to the low pressure gradient, a wider filtering window can be used to smooth the noise. After the pressure head contacts the seal, the pressure rises rapidly, and the window narrows as the gradient increases, preventing excessive smoothing of the contour at the pressure jump. Subsequently, linear interpolation is performed with a displacement interval of 0.01 mm to obtain a pressure sequence with equal displacement intervals. In this sequence, the pressure rise starting point is clear, and the point spacing is consistent, allowing subsequent steps to be judged based on the actual pressure process rather than on fluctuations in operating speed.

[0062] Step S20: Based on the equal displacement interval pressure sequence, a dynamic baseline adaptive threshold and stiffness confirmation mechanism is used to perform the physical contact initiation detection task, and the contact initiation index is output;

[0063] The detection target in this step is the equal displacement interval pressure sequence output in step S10. The goal is to find the location where the seals, connectors, or mating surfaces in the pressure head and interface assembly begin to be actually loaded. The physical contact start point is not the same as the first non-zero reading of the pressure sensor, because there may be zero drift, pressure head self-weight disturbance, or hydraulic system pulsation during the no-stroke phase. The dynamic baseline adaptive threshold mechanism continuously calculates the mean and standard deviation of the current pressure baseline through a sliding window, and generates a judgment threshold that changes with the working conditions by combining the sensitivity coefficient and the sensor quantization error compensation amount. The stiffness confirmation mechanism uses the first-order difference after Savitzky-Golay smoothing to determine whether the pressure not only exceeds the threshold, but also has formed a continuous upward trend consistent with the pressure process. The first sampling index that meets the conditions of pressure value higher than the adaptive threshold and smoothed first-order difference higher than the preset initial stiffness threshold is determined as the contact start point index.

[0064] This step separates the effective pressure analysis range from the complete sampling sequence. Before outputting the contact start index, the system first incorporates the current empty stroke baseline level into the threshold calculation using a sliding window, and then uses stiffness conditions to exclude accidental disturbances that only show an increase in pressure value but lack a true pressure trend. The resulting contact start index can serve as the left boundary for searching key inflection points in step S30, ensuring that subsequent significant extreme value tracking is no longer affected by noise in the empty stroke segment and drift in the non-contact stage of the pressure head. For the reassembly quality assessment, the accuracy of this boundary directly affects the length of the contact deformation segment, the curve shape, and the stability of the final classification characteristics.

[0065] Traditional methods often use a fixed pressure threshold to determine the start of contact; for example, a pressure exceeding a certain Newton value is considered to have entered the contact phase. This method is usable for single products and stable equipment conditions, but in the reassembly of rail transit interface components, there are often differences in parts batches, seal pre-tightening conditions, and sensor zero-point drift. If the threshold is set too low, slight fluctuations at the end of the idle stroke may be misjudged as contact; if the threshold is set too high, the low-pressure deformation phase at the initial contact stage may be missed. This step adjusts the judgment threshold according to the sliding window baseline and requires that the pressure rise trend simultaneously meet stiffness conditions, so that contact point judgment no longer completely relies on fixed values, and can better adapt to different components and equipment conditions in different shifts.

[0066] Taking a reassembly with positive sensor drift as an example, during the idle stroke phase, the pressure value slowly drifts from near 0 Newtons to 1.2 Newtons. Sliding window statistics show that the mean value during this phase is approximately 0.8 Newtons, and the standard deviation is approximately 0.3 Newtons. When the sensitivity coefficient is set to 3.5 and a 2-Newton quantization error compensation is added, the adaptive threshold is approximately 3.85 Newtons, and the drift value will not trigger a contact judgment. When the pressure head actually contacts the sealing ring, the pressure increases from 2 Newtons to 6 Newtons within a displacement of approximately 0.2 mm. The smoothed first-order difference exceeds the preset initial stiffness threshold, and the pressure value also exceeds the adaptive threshold. The system marks the sampling point that first meets both conditions as the contact start index, thereby truncating the preceding invalid idle stroke data.

[0067] Step S30: Based on the contact start point index and the equal displacement interval pressure sequence, a significant extreme value tracking and physical consistency constraint mechanism is used to perform key inflection point detection and curve segmentation tasks, and output the contact deformation segment sequence;

[0068] This step uses the contact start index and equal displacement interval pressure sequence output from the previous step as input to identify the key transition points in the reassembly pressure curve from rapid compression to slow compaction, and outputs the contact deformation segment sequence accordingly. The significant extremum tracking mechanism focuses on the curvature change reflected by the smoothed second-order difference. When the curve gradually changes from a steep rise to a gentle rise, the second-order difference will show a significant negative change at the corresponding position. This position usually corresponds to the node where the seal transitions from the elastic compression stage to the bonding compaction stage. To avoid mistaking noise extrema for process inflection points, this step also applies physical consistency constraints: the smoothed first-order difference at the candidate point needs to remain positive, indicating that the pressure is still in the rising process; the equal displacement interval pressure value at this point should also be within the range of 15% to 75% of the maximum pressure of the entire curve, avoiding the selection of small fluctuations at the initial contact or local disturbances in the final compaction zone.

[0069] Through the above processing, the contact deformation segment sequence has a clear physical meaning, namely, the stage from the start of actual contact to when the main elastic deformation characteristics have become apparent. This sequence eliminates the empty travel before contact and also reduces the dilution of shape features by the final rigid compaction segment. For step S40, the input is not the entire pressure curve, but a key subsequence that centrally reflects the compression of the seal, the force on the lip, changes in assembly posture, and the influence of local defects. The more consistent the boundaries of the contact deformation segments, the more the subsequent quantile normalization, detrended volatility calculation, and morphological envelope entropy extraction can be carried out around the same process stage, and the more comparable the feature vectors received by the classifier will be.

[0070] Common curve segmentation methods readily identify the maximum point of the first-order difference or manually set a fixed displacement range. However, when there are material differences, local friction variations, and operating speed disturbances in the reassembly pressure curve, the extreme value of the first-order difference may appear at local noise peaks, and the fixed displacement range cannot accommodate the compression stroke of different interface components or different seals. If segmentation is too early, the contact deformation section will be truncated, making it difficult to reflect the subsequent morphology caused by skewing or trimming; if segmentation is too late, the significant pressure increase in the final compaction zone will mask subtle defects in the intermediate stages. This step identifies morphological transitions through curvature significance and uses positive difference and pressure range constraints to eliminate positions that do not conform to the physical process of press-fitting, ensuring that the segmentation results are derived from the data curve and are constrained by the assembly process rules.

[0071] In a normal seal reassembly process, the pressure curve of the equal displacement interval rises rapidly after the contact point, reaching 35 N at a displacement of approximately 2.5 mm, after which the upward trend slows significantly; the maximum pressure along the entire curve is approximately 50 N. The smoothed second-order difference shows a significant negative extreme value near 2.5 mm, where the pressure accounts for about 70% of the maximum pressure, still within the constraint range of 15% to 75%. The smoothed first-order difference also remains positive, thus this position is identified as a critical inflection point. In another reassembly sample with uneven stress, the curve shows an earlier slowdown in its rise at a displacement of approximately 1.8 mm, with a pressure of approximately 18 N, accounting for about 39% of the maximum pressure of 46 N. This position also satisfies the physical consistency constraint, and the system obtains a shorter contact deformation segment accordingly. These two contact deformation segments have different lengths, but both are truncated around their respective actual deformation processes, preserving curve differences related to the assembly state for subsequent feature extraction.

[0072] Step S40: Based on the contact deformation segment sequence, a quantile robust normalization and hybrid morphological feature quantization mechanism are used to perform shape feature extraction, and a hybrid shape feature vector is output.

[0073] This step converts the contact deformation segment sequence output from step S30 into a mixed shape feature vector that can be used by the classifier. Since the contact deformation segment lengths may differ between samples, they are first resampled into a fixed-length sequence using cubic spline interpolation, ensuring each sample has a consistent feature calculation dimension. Subsequently, quantile robust normalization is performed, using the 2nd and 98th percentiles as scaling references instead of directly using the minimum and maximum values ​​as normalization boundaries. On the normalized contact deformation sequence, the system extracts two types of morphological features: detrended volatility features to describe local fluctuations of the curve outside the overall upward trend, including the root mean square of the detrended residuals and the number of zero crossings; and morphological envelope entropy features to describe the energy complexity of the upper envelope curve after multi-scale filtering and three-level wavelet packet decomposition.

[0074] After using fixed-length resampling, regardless of the original assembly press-in stroke length or sampling frequency, the contact deformation segment is standardized to... Point sequences, with different samples aligned in the feature space, allow subsequent classifier weight learning to share a unified structure, avoiding the need for additional processing of variable-length inputs. Quantile robust normalization is utilized. and As a scaling boundary, it effectively reduces the impact of occasional punctures or abnormal sampling points on the overall scale of the sequence, preserving the subtle fluctuations within the curve, while weakening the overall pressure amplitude shift and scaling effects caused by differences in the hardness of seals made of different materials or batches. The extracted root mean square value of detrended volatility characterizes the degree of fluctuation of the curve around the overall upward trend, and the number of zero crossings reflects the frequency characteristics of the fluctuation; morphological envelope entropy extracts the texture information of the envelope shape through multi-scale morphological operators, quantifying its complexity in the form of subband energy entropy. These two features characterize the curve shape from the perspectives of local fluctuations and global contours, respectively. The resulting hybrid shape feature vector is highly sensitive to morphological differences such as irregular concavities caused by tangential edges and asymmetric tilts caused by skew, and remains relatively stable even with individual amplitude differences, providing a highly discriminative input for accurate classification.

[0075] Traditional pressure curve feature extraction often directly uses the maximum, average, or pressure values ​​corresponding to a few fixed displacement points as features. These features are highly dependent on the absolute pressure amplitude. The pressure amplitude between different batches or different parts may vary due to differences in seal hardness and lubrication conditions, leading to feature value drift even in the same normal process, easily confused with anomalies. Another approach is to use minimum-maximum normalization, but this method is highly susceptible to interference from single-point spikes. If an abnormal peak appears in a sampling, the feature scale of the entire sequence will be drastically compressed, resulting in the loss of subtle shape details. This step uses percentile normalization instead of extreme value normalization, reducing the impact of outlier sample points. Furthermore, the extracted detrended volatility and morphological envelope entropy features are refinements of curve shape and texture information, independent of absolute pressure values, thus exhibiting good transferability across different individuals. Without modifying the original acquisition hardware and pressing process, post-processing improves feature stability, facilitating the subsequent classifier's discovery of subtle but crucial morphological cues.

[0076] For example, in a contact deformation segment exhibiting abnormal behavior due to an abnormal cut edge of the seal, the initial pressure showed a serrated indentation within a range of approximately 0.5 mm in the later stages of contact, with an overall pressure peak of approximately 45 Newtons; another normal segment showed a smooth curve with a peak pressure of approximately 52 Newtons. (After a fixed length...) After 128 resampling, the two sequences have the same number of points. Quantile normalization maps the values ​​of both curves to an approximate range of 0 to 1, preserving the concave features of the trimmed piece completely and enhancing its amplitude. When calculating the root mean square value of detrended volatility, the trimmed piece, due to multiple sawtooth fluctuations, has a residual root mean square of 0.12 and 28 zero-crossings; the normal piece curve is smooth, with a root mean square of only 0.04 and 8 zero-crossings. In terms of morphological envelope entropy, the upper envelope sequence of the trimmed piece also exhibits high and low fluctuations, with uneven subband energy distribution after three-layer wavelet packet decomposition, resulting in a higher envelope entropy value; the upper envelope of the normal piece is smooth, with concentrated energy, resulting in a lower entropy value. The two features are combined into a 4-dimensional mixed shape feature vector. The corresponding component of the trimmed piece in the vector deviates significantly from the distribution of the normal piece, and this difference is learned as an effective discrimination rule in the subsequent LightGBM classifier.

[0077] Step S50: Based on the hybrid shape feature vector, the cost-sensitive LightGBM ensemble learning mechanism is used to perform the assembly anomaly classification task and output the assembly state category label.

[0078] This step receives the mixed shape feature vector output from step S40 and uses a pre-trained cost-sensitive LightGBM ensemble learning classifier to perform an assembly anomaly classification task, ultimately outputting assembly state category labels. This classifier is an ensemble of multiple gradient boosting decision trees. During training, its loss function has been modified with cost sensitivity, assigning unequal misclassification cost weights to different categories. Specifically, the cost weights for the three anomalies—cut seal edges, misaligned seals, and incomplete assembly—are set to 5 to 10 times the weight of the normal class, resulting in a higher penalty for misclassification of minority anomalies during optimization. Hyperparameters are automatically tuned on the validation set using Bayesian optimization to determine key configurations such as tree depth, learning rate, and number of leaf nodes. During actual inference, the classifier calculates the input mixed shape feature vector and outputs a posterior probability vector. Each component represents the probability estimate of the sample belonging to the corresponding category. Finally, the category corresponding to the highest probability is output as the assembly state category label, covering four states: normal, cut seal edges, misaligned seals, and incomplete assembly.

[0079] It should be noted that the assembly status category label output in step S50 is not used as a single classification result. In one optional implementation, the assembly status category label, together with the posterior probability vector output by the cost-sensitive LightGBM classifier, the feature index value of the mixed shape feature vector, the reassembly inspection timestamp, the inspection station number, and the unique identifier of the interface component, are encapsulated into an interface reassembly status object. The unique identifier of the interface component can be obtained from the QR code, RFID tag, serial number, batch number, or work order binding number on the interface component, and is used to associate and store the status data of the same interface component during disassembly, reassembly, inspection, re-inspection, and rework processes.

[0080] The interface reassembly status object includes at least the following fields: component identifier field, inspection batch field, assembly status category field, posterior probability field, reassembly status identifier field, re-inspection trigger field, rework reason field, restricted release field, and data write time field. Specifically, the assembly status category field records at least one of the following categories: normal, seal edge trimming, seal misalignment, and incomplete assembly; the posterior probability field records the maximum posterior probability corresponding to the assembly status category label; the reassembly status identifier field records one of the following status identifiers: reassembly passed, awaiting re-inspection, rework, or restricted release; the re-inspection trigger field records the reason for triggering re-inspection due to a posterior probability lower than a preset confidence threshold, incomplete pressure data sequence, or abnormal curve segmentation; the rework reason field records the rework reason code corresponding to seal edge trimming, seal misalignment, or incomplete assembly; and the restricted release field provides a readable restricted status value to the reassembly station terminal, maintenance work order system, or vehicle preparation status database.

[0081] In a specific data mapping method, when the assembly status category label is normal and the maximum a posteriori probability is not lower than a preset confidence threshold, the reassembly status identifier field is written as reassembly passed, and the restricted release field is written as unrestricted; when the assembly status category label is seal edge cutting, seal skewed installation, or incomplete assembly, and the maximum a posteriori probability is not lower than the preset confidence threshold, the reassembly status identifier field is written as rework, and the rework reason field is written with the exception code corresponding to the assembly status category label; when the maximum a posteriori probability is lower than the preset confidence threshold, or any data item in the pressure data sequence, the contact start index, or the contact deformation segment sequence is missing, the reassembly status identifier field is written as pending re-inspection; when the assembly status category label is an exception type but the rework confirmation information has not yet been written, the restricted release field is written as restricted release to prevent the interface component from being used as a reassembled complete part to enter the subsequent preparation state.

[0082] Through the above processing, the data output in step S50 is no longer confined to a single assembly anomaly category, but is transformed into an interface reassembly status object that can be read by the reassembly station terminal, maintenance work order system, preparation status database, or vehicle maintenance record system. After this interface reassembly status object is associated with and stored with the unique identifier of the interface component, it can form a continuous record of the same interface component from reassembly inspection, re-inspection confirmation, rework processing to status deregulation, thereby providing structured data for interface component reassembly quality traceability, abnormal part interception, and subsequent maintenance verification.

[0083] By introducing a cost-sensitive mechanism, the LightGBM model amplifies the gradients of misclassified anomalous samples during internal feature splitting and leaf weight calculation. This makes the decision tree structure more inclined to capture subtle feature patterns that can distinguish between normal and anomalous samples, even if these patterns only appear in a small number of samples. This directly improves the model's recall ability for anomalous classes, because even small differences in feature vectors near the actual boundary have a sufficient chance to form split nodes under the drive of high cost. At the same time, Bayesian optimization parameter tuning can automatically balance the model's bias and variance, avoiding overfitting to noisy features of individual anomalous samples in the training set due to forced high weights. Through the combination of these two mechanisms, the classifier can maintain relatively robust accuracy in outputting class labels when facing test samples with varying shapes produced by different workstations and operators on an actual production line. It also has a relatively balanced detection capability for various anomalies, ultimately mapping the morphological numerical differences extracted from step S40 into semantically clear assembly status judgments, directly guiding rework operations.

[0084] In real-world rail transit interface component reassembly scenarios, assembly anomalies are low-probability events, resulting in a far smaller number of usable anomaly samples compared to normal samples. When using a classifier with a standard loss function, the model tends to predict all samples as normal, as this yields a lower overall error in terms of quantity, but it almost completely misses process defects that require careful detection, such as edge trimming of seals and skewed assembly. Even with data balancing techniques like oversampling or undersampling, the original data distribution may still be altered or synthetic noise may be introduced. This step, without changing the original training set sample distribution, directly embeds prior knowledge into the weight parameters of the loss function, forcing the extraction of discriminative information from categories with scarce data but severe hazards. Compared to a single final value pressure threshold determination method, this invention provides a decision based on the complete curve morphology throughout the entire processing chain. Therefore, this classification step can identify anomalies where the final value pressure is normal but the process morphology has defects, compensating for the blind spots of traditional final value detection. Compared to machine learning classification without cost adjustment, the cost-sensitive mechanism is adapted to the class imbalance characteristics of this process, improving the overall detection effect of assembly process defects.

[0085] For example, in a training data collection, approximately 3000 normal assembly samples were accumulated, while only 80 samples of seal edge cutting, about 60 samples of misaligned assembly, and about 40 samples of improper assembly were collected, resulting in a severe class imbalance. A cost-sensitive setting was adopted, uniformly setting the misclassification cost weight for these three types of anomalies to 8 times that of the normal class. After model training, inference was performed on a seal edge cutting sample in the test set: In the mixed shape feature vector of this sample, the number of times the detrended volatility crossed zero was significantly higher than that of normal samples, and the envelope entropy value was higher, even though the overall pressure peak was still within the acceptable range. The LightGBM classifier calculated the posterior probability vector as [Normal: 0.21, Seal edge cutting: 0.67, Misaligned: 0.08, Improper assembly: 0.04], with the highest probability pointing to the seal edge cutting, ultimately outputting the class label for this category. However, if an unweighted ordinary classifier were used, the posterior probability for the same test sample might be [Normal: 0.53, Seal edge cutting: 0.40], with the highest probability pointing to normal, resulting in missed detections. The cost-sensitive mechanism successfully detected the edge defect, and the production line personnel then disassembled and inspected the component to confirm and replace the seal, thus avoiding the potential risk of seal failure.

[0086] Example 2: Furthermore, the rail transit interface component reassembly status data processing system provided by the present invention employs a rail transit interface component reassembly status data processing method from the above embodiments, which can solve a technical problem in rail transit interface component reassembly status data processing. The beneficial effects of the rail transit interface component reassembly status data processing system provided by the present invention are the same as those of the rail transit interface component reassembly status data processing method from the above embodiments, and other technical features of the rail transit interface component reassembly status data processing system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0087] Example 3: This invention provides a rail transit interface component reassembly status data processing device. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the rail transit interface component reassembly status data processing method described in Example 1. The rail transit interface component reassembly status data processing device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This rail transit interface component reassembly status data processing device is merely an example and should not limit the functionality or scope of the invention. A rail transit interface component reassembly status data processing device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the rail transit interface component reassembly status data processing device. The processing unit, the read-only memory, and the random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the rail transit interface component reassembly status data processing device to communicate wirelessly or wiredly with other devices to exchange data. Although a rail transit interface component reassembly status data processing device with various systems has been described, it should be understood that it is not required to implement or possess all the systems described. More or fewer systems may be implemented alternatively.

[0088] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for processing reinstallation status data of a rail transit interface component. The computer program product provided by this invention can solve a technical problem related to processing reinstallation status data of a rail transit interface component. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the method for processing reinstallation status data of a rail transit interface component provided in the above embodiments, and will not be repeated here.

[0089] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0090] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0091] 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 present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for processing reinstallation status data of rail transit interface components, characterized in that, The methods include: Step S10: Obtain the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component. Based on the pressure data sequence and the displacement data sequence, perform waveform preprocessing tasks using an adaptive median filtering and linear interpolation resampling mechanism to output an equal displacement interval pressure sequence. Step S20: Based on the equal displacement interval pressure sequence, a dynamic baseline adaptive threshold and stiffness confirmation mechanism is used to perform the physical contact initiation detection task, and the contact initiation index is output; Step S30: Based on the contact start point index and the equal displacement interval pressure sequence, a significant extreme value tracking and physical consistency constraint mechanism is used to perform key inflection point detection and curve segmentation tasks, and output the contact deformation segment sequence; Step S40: Based on the contact deformation segment sequence, a quantile robust normalization and hybrid morphological feature quantization mechanism are used to perform shape feature extraction, and a hybrid shape feature vector is output. Step S50: Based on the hybrid shape feature vector, the cost-sensitive LightGBM ensemble learning mechanism is used to perform the assembly anomaly classification task and output the assembly state category label.

2. The method for processing reinstallation status data of rail transit interface components as described in claim 1, characterized in that, Step S10 involves acquiring the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component, and performing waveform preprocessing based on the pressure data sequence and displacement data sequence using an adaptive median filtering and linear interpolation resampling mechanism to output a pressure sequence with equal displacement intervals. Specifically, this includes: Step S101: Real-time synchronous acquisition of the cumulative displacement and corresponding pressure value of the press head to form a raw data sequence set; Step S102: Perform adaptive median filtering of the guard edge on the pressure signal in the original data sequence set to obtain the filtered sequence; Step S103: Perform linear interpolation on the filtered sequence based on the shift, and resample to equal intervals. On the grid, pressure sequences with equal displacement intervals are obtained. ,in , This represents the length of the resampled sequence.

3. The method for processing reinstallation status data of rail transit interface components as described in claim 1, characterized in that, Step S20, which involves performing a physical contact initiation detection task based on the equal displacement interval pressure sequence using a dynamic baseline adaptive threshold and stiffness confirmation mechanism, and outputting the contact initiation index, specifically includes: Step S201: Calculate the equal displacement interval pressure sequence First-order difference and second-order difference and to and Savitzky-Golay smoothing is performed to obtain the smoothed first-order difference. and smooth second-order difference ; Step S202: Set a length of sliding window Continuously calculate the average pressure value within the window. and standard deviation Generate adaptive threshold : in, For the first Adaptive threshold at each sampling point This represents the average pressure value within the sliding window. The standard deviation of the pressure values ​​inside the sliding window. The sensitivity coefficient, This is the amount of compensation for sensor quantization error; Step S203: The pressure value must be greater than the adaptive threshold and the smoothed first-order difference must be greater than the preset initial stiffness threshold. The first index is determined to be the contact start index. .

4. The method for processing reinstallation status data of rail transit interface components as described in claim 3, characterized in that, In step S202, the sensitivity coefficient Take 3.

5.

5. The method for processing reinstallation status data of rail transit interface components as described in claim 3, characterized in that, Step S30, which involves performing key inflection point detection and curve segmentation based on the contact start point index and the equal displacement interval pressure sequence using a significant extreme value tracking and physical consistency constraint mechanism, and outputting the contact deformation segment sequence, specifically includes: Step S301: Calculate significance index based on smooth second-order difference And search for significance indicators Maximize candidate inflection points : in, It is a symbolic function; Step S302: For the candidate inflection points A physical consistency constraint is applied, requiring that the smooth first-order difference at the point is greater than zero and that the pressure value at the equal displacement interval is within the range of 15% to 75% of the entire equal displacement interval pressure sequence. If the physical consistency constraint is met, it is identified as a critical inflection point. ; Step S303: Utilize the contact origin index The aforementioned key turning point and preset endpoint From the equal displacement interval pressure sequence, segment from +1 to The subsequence is used as the contact deformation segment sequence.

6. The method for processing reinstallation status data of rail transit interface components as described in claim 1, characterized in that, Step S40, which involves performing shape feature extraction based on the contact deformation segment sequence using quantile robust normalization and a hybrid morphological feature quantization mechanism, and outputting a hybrid shape feature vector, specifically includes: Step S401: Resample the contact deformation segment sequence to a fixed length using cubic spline interpolation. This yields a fixed-length sequence; Step S402: Perform quantile robust normalization on the fixed-length sequence to obtain a normalized contact deformation sequence. ; Step S403: From the normalized contact deformation sequence Extract a hybrid shape feature vector containing detrended volatility features and morphological envelope entropy features; wherein, the detrended volatility features include... The root mean square and zero-crossing count of the detrended sequence obtained after straight-line fitting, the morphological envelope entropy feature is obtained by... Multi-scale morphological filtering is used to extract the upper envelope sequence. Then, the envelope sequence The subband energy entropy is obtained by performing three-level wavelet packet decomposition.

7. The method for processing reinstallation status data of rail transit interface components as described in claim 1, characterized in that, Step S50, which involves performing an assembly anomaly classification task based on the hybrid shape feature vector using a cost-sensitive LightGBM ensemble learning mechanism and outputting assembly state category labels, specifically includes: Step S501: Input the hybrid shape feature vector into the cost-sensitive LightGBM classifier that has been pre-optimized and trained using Bayesian parameters; Step S502: The cost-sensitive LightGBM classifier uses unequal class cost weights in its loss function. The unequal class cost weights are used to set the weight values ​​of abnormal classes such as seal edge cutting, seal skew installation, and improper assembly to be 5 to 10 times the weight values ​​of normal classes. Step S503: Obtain the posterior probability vector output by the cost-sensitive LightGBM classifier, and output the category corresponding to the highest probability in the posterior probability vector as the assembly state category label. The assembly state category label includes normal, seal edge cut, seal skewed, and incomplete assembly.

8. A rail transit interface component reassembly status data processing system, applied to the rail transit interface component reassembly status data processing method according to any one of claims 1 to 7, characterized in that, The system includes: The waveform preprocessing module is used to acquire the pressure data sequence and displacement data sequence during the reassembly process of the rail transit interface component. Based on the pressure data sequence and the displacement data sequence, an adaptive median filtering and linear interpolation resampling mechanism is used to perform waveform preprocessing tasks and output a pressure sequence with equal displacement intervals. The contact initiation detection module is used to perform physical contact initiation detection tasks based on the equal displacement interval pressure sequence using a dynamic baseline adaptive threshold and stiffness confirmation mechanism, and outputs the contact initiation index. The curve segmentation module is used to perform key inflection point detection and curve segmentation tasks based on the contact start point index and the equal displacement interval pressure sequence, using a significant extreme value tracking and physical consistency constraint mechanism, and outputs a contact deformation segment sequence. The shape feature extraction module is used to perform shape feature extraction based on the contact deformation segment sequence using quantile robust normalization and hybrid morphological feature quantization mechanism, and output a hybrid shape feature vector. The anomaly classification module is used to perform assembly anomaly classification tasks based on the hybrid shape feature vector using a cost-sensitive LightGBM ensemble learning mechanism, and output assembly state category labels.

9. A data processing device for the reassembly status of rail transit interface components, characterized in that, The rail transit interface component reassembly status data processing device includes: a memory, a processor, and a rail transit interface component reassembly status data processing program stored in the memory and executable on the processor. When the rail transit interface component reassembly status data processing program is executed by the processor, it implements a rail transit interface component reassembly status data processing method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a rail transit interface component reinstallation status data processing program, which, when executed by a processor, implements a rail transit interface component reinstallation status data processing method according to any one of claims 1 to 7.

Images