A Machine Learning-Based Intelligent Prediction Method and System for Quality Defects in Ring Snappers

Abstract

This invention relates to the field of intelligent prediction technology for the quality of mechanical parts, and discloses an intelligent prediction method and system for quality defects of snap rings based on machine learning. The method includes: acquiring multimodal data from snap ring production equipment; obtaining a clean dataset containing material properties through preliminary aggregation, signal smoothing, noise filtering, and iterative adjustment of filtering parameters; extracting key features and reducing dimensionality, obtaining a refined feature subset through weight optimization, and compressing and encoding edge-processed data, correcting deviations, and integrating it in real time to obtain a compressed feature package; decoding and fusing historical data in the cloud, constructing an optimized random forest model, inputting real-time data to obtain prediction results, determining the defect type, and providing feedback for parameter adjustment. This method can accurately predict snap ring defects and improve the efficiency of quality control.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology for the quality of mechanical parts, and in particular to an intelligent prediction method for quality defects of snap rings based on machine learning. Background Technology

[0002] Currently, in the field of mechanical manufacturing and intelligent quality prediction, with the continuous improvement of reliability requirements for precision components in industries such as automotive and aerospace, the quality stability of snap rings, as core connecting and fixing components, directly affects the safe operation and service life of downstream equipment. Intelligent prediction of snap ring quality defects is a key link in preventing problems such as fatigue cracks and micro-stress concentration. It requires accurate capture of multimodal data characteristics during the production process to achieve early warning of potential hazards. The integration of multimodal data filtering and machine learning models has become the core technological direction for improving the accuracy and real-time performance of defect prediction.

[0003] Current industry practices for detecting defects in snap rings primarily rely on manual sampling or simple automated equipment. This involves visually inspecting the snap rings for cracks or using a single-dimensional measuring instrument to check if geometric parameters meet standards. However, this approach is clearly inadequate in mass production environments. Manual sampling struggles to cover dynamic parameters such as temperature and pressure throughout the entire production process, resulting in low efficiency per inspection and significant susceptibility to human experience, easily missing latent defects such as microscopic stress concentrations. Simple automated equipment can only detect visible indicators of the finished product, failing to correlate production process data with the intrinsic relationship between defects and struggling to distinguish between data noise and genuine defect signals, leading to a high false positive rate. This is especially problematic in long-cycle, continuous production processes, making it difficult to quickly trace the critical points where defects originate.

[0004] In summary, existing technologies are insufficient for efficient processing of multimodal data throughout the entire production process of snap rings, and cannot meet the efficiency and safety requirements of distribution network cable operation and maintenance under mass production, resulting in insufficient accuracy in defect prediction. Summary of the Invention

[0005] This invention provides a machine learning-based intelligent prediction method and system for snap ring quality defects, which solves the problems of missed hidden defects in manual sampling and the lack of accuracy in defect prediction caused by misjudgment due to single data measurement, and meets the needs of efficient quality control and safety early warning in snap ring production.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a machine learning-based intelligent prediction method for quality defects of snap rings, comprising:

[0007] Obtain multimodal data from snap ring manufacturing equipment;

[0008] The multimodal data is initially aggregated and smoothed to obtain a raw dataset containing material properties. Noise is then filtered, and the filtering parameters are iteratively adjusted to obtain a clean dataset.

[0009] Key features of the clean dataset are extracted and then subjected to dimensionality reduction and weight optimization to obtain a refined feature subset.

[0010] The data required for edge processing within the refined feature subset is selected, and after compression encoding, signal deviation correction, and real-time data integration, a compressed feature package suitable for cloud upload is obtained.

[0011] The compressed feature package is uploaded to the cloud analysis platform for decoding, and historical data is integrated to obtain a comprehensive feature dataset. Historical patterns are matched to determine potential defect trends.

[0012] Based on the comprehensive feature dataset, a random forest model is constructed and optimized to obtain a trained defect prediction model. Real-time multimodal data is input for forward computation to obtain the model's forward computation results.

[0013] Based on the forward calculation results of the model, the potential quality defect type is determined, defect classification is triggered to obtain the final prediction result, and subsequent filtering parameters are adjusted accordingly.

[0014] Secondly, the present invention provides a machine learning-based intelligent prediction system for quality defects in snap rings, comprising:

[0015] The data acquisition module is used to acquire multimodal data from the snap ring production equipment;

[0016] The data processing module is used to perform preliminary aggregation and signal smoothing on the multimodal data to obtain a raw dataset containing material properties, and to perform noise filtering, iteratively adjusting the filtering parameters to obtain a clean dataset.

[0017] The feature optimization module is used to extract key features from the clean dataset, and after dimensionality reduction and weight optimization, a refined feature subset is obtained.

[0018] The edge processing module is used to select the data required for edge processing within the refined feature subset, and after compression encoding, signal deviation correction and real-time data integration, obtain a compressed feature package suitable for cloud upload;

[0019] The cloud fusion module is used to upload the compressed feature package to the cloud analysis platform for decoding, integrate historical data to obtain a comprehensive feature dataset, and match historical patterns to determine potential defect trends.

[0020] The model building and computation module is used to build and optimize the random forest model based on the comprehensive feature dataset to obtain the trained defect prediction model. It inputs real-time multimodal data for forward computation to obtain the model forward computation results.

[0021] The defect judgment and feedback module is used to judge the type of potential quality defects based on the forward calculation results of the model, trigger defect classification to obtain the final prediction result, and provide feedback to adjust subsequent filtering parameters.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) This invention collects real-time multimodal signals such as temperature fluctuations and pressure changes from the snap ring production equipment, generates an initial dataset through a preset acquisition module, and then uses Kalman filtering to iteratively adjust parameters to filter noise. Combined with Fourier transform frequency domain analysis, the initial noise level is determined. This invention breaks through the limitations of traditional manual sampling inspection, which only focuses on single-dimensional data and has poor noise filtering effect. It comprehensively captures the dynamic parameter characteristics of the production process, eliminates the influence of environmental interference and noise caused by equipment vibration, provides high-reliability clean dataset support for defect prediction, effectively improves the extraction rate of latent defect-related features such as micro-stress concentration, and solves the problem of missed detection of latent defects due to single data and noise interference in the existing technology.

[0024] (2) This invention extracts key features of interaction relationships in clean datasets through Pearson correlation coefficient, uses principal component analysis to reduce the dimensionality of multimodal data, and combines the gain matrix to adjust feature weights to obtain a refined feature subset. Then, the signal deviation is restored by Huffman coding compression and state estimation at the edge end. This invention breaks through the limitations of inefficient feature extraction and disconnect between edge end and cloud data processing in existing technologies, accurately selects core features with high correlation to fatigue cracks, reduces data redundancy and retains high-value information, significantly improves feature transmission efficiency and the targeting of cloud analysis, and makes up for the defects of traditional methods that cause the loss of key information due to low feature quality and poor data connection.

[0025] (3) This invention decodes and compresses feature packets and integrates historical data through a cloud platform. It uses a tree structure to traverse and match historical defect patterns, and combines a random forest model to optimize the tree structure to achieve defect prediction. At the same time, the prediction results are fed back to adjust the filtering parameters, realizing a closed loop of the whole process of "multimodal acquisition-noise filtering-feature optimization-cloud prediction-parameter feedback". It breaks through the limitations of existing technologies that are difficult to associate production parameters with defect patterns and lack dynamic optimization mechanisms. It provides maintenance personnel with accurate defect types and generation trends, solves the problems of traditional detection that cannot provide early warning and poor parameter adaptability, and takes into account the accuracy of defect prediction and dynamic optimization of the production process, meeting the needs of batch snap ring production quality control and risk avoidance. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a machine learning-based intelligent prediction method for quality defects of snap rings provided in the first embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of a machine learning-based intelligent prediction system for quality defects of snap rings, provided in the second embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0029] Reference Figure 1 The first embodiment of the present invention provides a machine learning-based intelligent prediction method for quality defects of snap rings, comprising the following steps:

[0030] S101, acquire multimodal data from the snap ring manufacturing equipment;

[0031] S102, the multimodal data is initially aggregated and signal smoothed to obtain an original dataset containing material properties, and noise is filtered. The filtering parameters are iteratively adjusted to obtain a clean dataset.

[0032] S103, extract the key features of the clean dataset, and obtain a refined feature subset after dimensionality reduction and weight optimization;

[0033] S104, Select the data required for edge processing within the refined feature subset, and obtain a compressed feature package suitable for cloud upload through compression encoding, signal deviation correction and real-time data integration;

[0034] S105, the compressed feature package is uploaded to the cloud analysis platform for decoding, historical data is integrated to obtain a comprehensive feature dataset, and historical patterns are matched to determine potential defect trends;

[0035] S106. Based on the comprehensive feature dataset, construct and optimize the random forest model to obtain the trained defect prediction model. Input real-time multimodal data for forward computation to obtain the model forward computation result.

[0036] S107. Based on the forward calculation results of the model, determine the type of potential quality defects, trigger defect classification to obtain the final prediction result, and feed back to adjust subsequent filtering parameters.

[0037] In step S101, acquiring multimodal data from the snap ring manufacturing equipment includes:

[0038] The multimodal data is collected from the snap ring production equipment, and the multimodal data includes real-time signals such as temperature fluctuations and pressure changes.

[0039] It should be noted that when collecting the multimodal data from the snap ring production equipment, it is necessary to closely follow the three key processes of the snap ring production process: stamping, heat treatment, and surface treatment. In combination with the factors affecting quality, the collection location, sensor selection, and data collection rules should be determined in a targeted manner to ensure that the data covers the core dynamic parameters that may induce defects. In the stamping process, referring to the industry-standard monitoring approach for snap ring stamping—high-precision capture of dynamic signals from the mold and cylinder—a contact-type platinum resistance temperature sensor (measurement range -20℃ to 300℃, accuracy ±0.1℃) is selected to monitor the mold cavity temperature (the mold temperature needs to be stable at 80-120℃ during snap ring stamping; small fluctuations can easily lead to material stress concentration). The sampling frequency is set to 2Hz to capture subtle temperature changes within the stamping interval. At the same time, a piezoelectric pressure sensor (measurement range 0-5MPa, accuracy ±0.02MPa) is used to collect the real-time pressure of the stamping cylinder (sudden pressure changes can cause snap ring deformation, so the pressure needs to be controlled at 1.5-2.5MPa; the sampling frequency is set to 10Hz). During the heat treatment process, vibration sensors (measurement range 0-500Hz, sensitivity 0.1mV / (m / s²)) are deployed on the equipment frame to collect high-frequency vibration signals (abnormal vibration can easily lead to uneven surface treatment), with a sampling frequency of 1000Hz; at the same time, thermocouple temperature sensors (measurement range 0-100℃, accuracy ±0.2℃) are used to monitor the temperature of the electrolyte in the electroplating tank, and a pH sensor (measurement range 0-14, accuracy ±0.01) is used to collect the pH value of the electrolyte to ensure a stable surface treatment environment.

[0040] Real-time signals collected by all sensors, such as temperature fluctuations, pressure changes, equipment vibrations, heating element current, and electrolyte pH, need to be initially aggregated by a preset acquisition module according to a unified timestamp (accurate to milliseconds) to form a raw dataset containing the characteristics of the snap ring material. Referring to the technical requirements of "data aggregation and material property correlation," the acquisition module needs to be compatible with multiple sensor interfaces and synchronously integrate various signals at a 1Hz base frequency (high-frequency data is downsampled to 1Hz using a moving average method). For example, in a batch production run, the acquisition module continuously collects data for 10 minutes, generating 600 sets of time-series raw data. Each set covers parameters such as mold temperature, stamping pressure, tempering temperature, and heating element current at the corresponding time point. The mold temperature data reflects changes in the material's thermal expansion coefficient, and the stamping pressure data reflects the material's stress-forming state, providing a complete foundation for subsequent initial noise level determination. As recorded in historical data, the mold temperature fluctuated between 85℃ and 88℃ during a certain period (exceeding the ideal range of 80-82℃), and the corresponding stamping pressure changed abruptly between 1.8MPa and 2.3MPa. Subsequent analysis confirmed that the correlation data could be traced back to a slight deformation defect in the batch of snap rings, verifying the necessity of synchronous acquisition of multimodal data.

[0041] In summary, the process of collecting multimodal data from the snap ring production equipment strictly follows the technical logic of "full process coverage, high parameter precision, and data synchronization." By selecting dedicated sensors for stamping, heat treatment, and surface treatment processes, it ensures that no key signals such as temperature, pressure, and vibration are missed. By leveraging the timestamp aggregation of the acquisition modules, a deep correlation between multidimensional data and material properties is achieved. This provides a complete data source for determining the initial noise level in subsequent steps and lays a precise and reliable data foundation for the entire snap ring quality defect prediction process.

[0042] In step S102, the multimodal data undergoes preliminary aggregation and signal smoothing to obtain an original dataset containing material properties, followed by noise filtering and iterative adjustment of the filtering parameters to obtain a clean dataset, including:

[0043] The multimodal data is initially aggregated according to the collection timestamp to generate a raw dataset containing material properties;

[0044] Kalman filtering is used to perform state vector prediction on the temperature fluctuation signal and pressure change signal in the multimodal data to obtain smoothed signal data.

[0045] If the deviation between the smoothed signal data and the preset signal threshold exceeds the allowable range, then the smoothed signal data is analyzed in the frequency domain by Fourier transform to determine the initial noise level.

[0046] The original dataset is decomposed in the frequency domain by Fourier transform to obtain a frequency domain signal dataset.

[0047] The noise level of the frequency domain signal dataset is compared with a preset noise threshold. If the noise level of the frequency domain signal dataset exceeds the preset noise threshold, the Kalman filter parameters are updated and adjusted through the covariance matrix to obtain an optimized filter parameter set.

[0048] Based on the optimized set of filtering parameters, the frequency domain signal dataset is subjected to secondary noise filtering using Kalman filtering to obtain a clean dataset with reduced noise.

[0049] It should be noted that when initially aggregating multimodal data according to the collection timestamp to generate the raw dataset containing material properties, it is necessary to rely on multi-interface adaptation and millisecond-level time synchronization to accurately integrate temperature fluctuations (such as mold cavity and tempering furnace temperatures), pressure changes (such as stamping cylinder pressure), and equipment operating parameters (such as heating tube current) of the stamping, heat treatment, and surface treatment processes in snap ring production, forming a "timestamp-multi-parameter-material property" correlation structure. Material properties are indirectly mapped through parameters; for example, a mold temperature of 80-120℃ corresponds to a material strength of 1.2×10⁻ 5The coefficient of thermal expansion at / ℃, and the reasonable stress state of the material under a yield strength of 800MPa corresponding to a stamping pressure of 1.5-2.5MPa, avoid the data being out of sync with material properties. For example, within a 15-minute acquisition cycle (1Hz frequency), 900 sets of data are generated, each set containing a timestamp, multiple parameters, and corresponding material property information, providing support for subsequent processing.

[0050] When using Kalman filtering to predict the state vectors of temperature and pressure signals in multimodal data, a targeted model needs to be constructed: the temperature signal state vector is set as [current temperature value, rate of temperature change], and the pressure signal is set as [current pressure value, rate of pressure change]. Smoothing is achieved through a two-step "prediction-update" process. In the prediction stage, the state at the next moment is estimated based on historical data. In the update stage, the weights are adjusted based on the deviation between the measured and predicted values ​​to reduce random noise. In practical applications, before filtering, the mold temperature fluctuates by ±2℃ due to equipment vibration, and the pressure contains ±0.1MPa noise due to air valve switching. After processing, the temperature fluctuation is reduced to ±0.5℃, and the pressure noise is reduced to ±0.02MPa. For example, after smoothing, the mold temperature stabilizes at 84.5-85.1℃ and the stamping pressure stabilizes at 2.05-2.15MPa, which meets the process requirements.

[0051] If the smoothed signal deviates from the preset threshold beyond the allowable range, a Fourier transform is required for frequency domain analysis to determine the initial noise level: the temperature fluctuation threshold is set to ±1℃ (corresponding to the upper limit of mold temperature fluctuation of 80-120℃), and the pressure change threshold is set to ±0.05MPa (corresponding to the upper limit of stamping pressure fluctuation of 1.5-2.5MPa). When the smoothed signal exceeds the range (e.g., mold temperature deviation of 1.2℃ for 3 consecutive frames, stamping pressure deviation of 0.06MPa), a Fourier transform is triggered to convert the time-domain signal to the frequency domain. The noise source is located through the peak frequency of the spectrum (e.g., a 2Hz temperature peak corresponds to cooling system vibration), and the noise level is quantified by the peak amplitude (e.g., a 2Hz amplitude of 0.3 corresponds to a temperature fluctuation of ±1.2℃). For example, in a certain analysis, the smoothed temperature deviation was 1.3℃. The Fourier transform revealed a peak amplitude of 0.32 at 1.8Hz, determining the initial noise level to be "temperature fluctuation noise amplitude of 0.32 (corresponding to ±1.3℃)," clearly indicating that the noise originated from the vibration of the cooling fan. The correspondence between frequency domain amplitude and physical wave amplitude is determined by fitting historical data.

[0052] When performing frequency domain decomposition on the original dataset using Fourier transform to obtain a frequency domain signal dataset, the original dataset needs to cover aggregated multimodal parameters (such as time-domain data such as mold temperature, stamping pressure, and heating tube current). Fourier transform can convert noise and effective signals that are difficult to distinguish in the time domain into amplitude characteristics of different frequencies in the frequency domain, achieving precise noise localization. During decomposition, the number of sampling points needs to match the length of the original data time series (e.g., 600 sets of data collected at 1Hz over 10 minutes, with 600 sampling points). The 0.1-5Hz frequency components should be captured—where 0.1-1Hz represents normal process fluctuation frequencies (such as slow mold heating), and 2-5Hz mainly represents noise frequencies such as equipment vibration and pressure valve switching. For example, if the mold temperature fluctuates by ±1.5℃ in the original data, a peak amplitude of 0.3 appears at 3Hz after transformation (consistent with the vibration frequency of the stamping mechanism, determined to be noise). This corresponds to the "3Hz - amplitude 0.3" record in the frequency domain dataset, which, together with other frequency components, constitutes the frequency domain dataset.

[0053] The noise level of the frequency domain signal dataset is compared with a preset threshold. If it exceeds the limit, the Kalman filter parameters are updated and adjusted through the covariance matrix to obtain an optimized parameter set. The preset noise threshold is set according to the process stability requirements. The amplitude threshold for the frequency domain noise band (2-5Hz) is 0.2. An amplitude exceeding 0.2 in this band is considered to be exceeding the limit. The comparison focuses on checking the 2-5Hz frequency band. If the amplitude at a certain frequency point exceeds the limit (e.g., 0.3 > 0.2 at 3Hz), the process noise covariance matrix (Q) and measurement noise covariance matrix (R) of the Kalman filter are updated. Q is positively correlated with the noise amplitude. When Q = 0.01, it can be adjusted to 0.015 according to the excess ratio (0.3 / 0.2 = 1.5). R is corrected according to the sensor accuracy. If the noise is not related to the measurement error, it remains unchanged at 0.02. Finally, an optimized parameter set is formed (e.g., Q = 0.015, R = 0.02). If the amplitude at 4Hz in the frequency domain data is 0.25 (exceeding the threshold of 0.2), then Q is adjusted from 0.01 to 0.016, and R is kept at 0.02 to obtain the adaptation parameters.

[0054] Based on the optimized filter parameter set, when performing secondary noise filtering on the frequency domain signal dataset using Kalman filtering to obtain a low-noise clean dataset, the iterative logic needs to be adjusted according to the frequency domain characteristics: using the optimized Q and R as inputs, the amplitude of the out-of-range frequency band (such as 2-5Hz) is suppressed, while retaining the normal process frequency characteristics of 0.1-1Hz to avoid loss of effective signals. The amplitude at each frequency point is individually iterated and corrected during filtering. For example, the initial amplitude at 3Hz is 0.3, which is reduced to 0.08 after processing, maintaining stability consistent with the normal frequency band (such as 0.12 at 0.5Hz). The final clean dataset needs to be restored to a time domain signal. For example, after filtering, the mold temperature fluctuation is reduced from ±1.5℃ to ±0.3℃, and the stamping pressure fluctuation is reduced from ±0.15MPa to ±0.03MPa. All parameter time domain curves are smooth and fit the ideal process range, with no significant noise interference.

[0055] In summary, step S102 transforms the raw collected data into a high-purity clean dataset through a coherent data processing flow. During the process, relevant technologies are used to process noise and optimize parameters, which not only solves the noise interference problem of multimodal data, but also lays a data foundation for subsequent feature extraction, model construction and improvement of defect prediction accuracy. This is in line with the core technical logic of "data preprocessing-precise analysis" for intelligent prediction of snap ring quality defects.

[0056] In step S103, the key features of the clean dataset are extracted, and after dimensionality reduction and weight optimization, a refined feature subset is obtained, including:

[0057] By calculating the Pearson correlation coefficient, key features of the interaction relationship related to the quality defects of the snap ring in the clean dataset are extracted to form a multimodal key feature set;

[0058] Principal component analysis is used to reduce the dimensionality of the multimodal key feature set. Principal components are calculated by eigenvalue decomposition to obtain the dimensionality-reduced feature vector set.

[0059] The correlation between each feature in the feature vector set and the fatigue crack is compared with a preset correlation threshold. If the correlation between a certain feature and the fatigue crack is lower than the preset correlation threshold, the importance of the feature is adjusted by matrix multiplication of the gain matrix to obtain an optimized feature weight set.

[0060] Based on the optimized feature weight set, features whose weights meet the preset weight requirements are selected from the feature vector set through threshold filtering to obtain a refined feature subset.

[0061] It should be noted that when extracting key interactive features related to circlip quality defects from the clean dataset using Pearson correlation coefficient calculation to form a multimodal key feature set, the clean dataset must cover the core parameters of the entire circlip production process (such as mold cavity temperature T, stamping cylinder pressure P, tempering furnace temperature). The core function of the Pearson correlation coefficient (including equipment vibration amplitude V and heating tube current I) is to quantify the degree of linear correlation between different parameters, focusing on capturing the interaction relationships of parameters directly related to quality defects (such as fatigue cracks and dimensional deformation). For example, circlip fatigue cracks are often caused by micro-stress concentration, which often stems from the synergistic fluctuations of temperature and pressure, and the superposition effect of vibration and pressure. Therefore, the correlation coefficients of such parameter pairs should be given priority. During calculation, the data must first be Z-score standardized to eliminate the influence of dimensions. Then, all parameters are analyzed in pairs. Parameter pairs with an absolute correlation coefficient ≥ 0.7 are usually considered strongly correlated and included in the key feature category. For example, the calculated correlation coefficient between mold temperature T and stamping pressure P is 0.83 (reflecting the synergistic increase of both leading to stress concentration), and the correlation coefficient between tempering furnace temperature and stamping pressure P is 0.83. The correlation coefficient between the temperature variation and the fatigue crack incidence rate is 0.75 (reflecting the effect of temperature unevenness on material toughness), and the correlation coefficient between the equipment vibration amplitude V and the stamping pressure P is 0.72 (reflecting that vibration exacerbates pressure fluctuations). These strongly correlated parameters, along with the correlation between individual parameters and defects, collectively constitute a multimodal key feature set, such as {[T,P], [...] - Fatigue crack correlation], [V,P], [I- ]}.

[0062] Principal component analysis (PCA) is used to reduce the dimensionality of the multimodal key feature set. When calculating principal components through eigenvalue decomposition to obtain the dimensionality-reduced feature vector set, the multimodal key feature set often exhibits dimensional redundancy (e.g., the set mentioned above contains 4 key feature groups, corresponding to 8 original parameters). PCA maps high-dimensional data to a low-dimensional space through linear transformation while preserving core data information. The operation requires first constructing the covariance matrix of the key features, then performing eigenvalue decomposition on the covariance matrix, sorting the eigenvalues ​​from largest to smallest, and retaining principal components with eigenvalues ​​> 1 (ensuring no loss of core information). For example, performing eigenvalue decomposition on a key feature set containing 8 original parameters yields eigenvalues ​​of 4.1, 2.3, 1.2, 0.8, 0.6, 0.5, 0.4, and 0.3. The first three principal components have eigenvalues ​​> 1, therefore, the first three principal components are retained as the dimensionality reduction result: First Principal Component = 0.6T + 0.5P + 0.3 The first principal component primarily reflects the synergistic effect of temperature and pressure; the second principal component = 0.4V + 0.3I (primarily reflects equipment operational stability); the third principal component = 0.5... +0.3V (mainly reflecting the superimposed effect of tempering temperature and vibration), these three principal components together constitute the set of feature vectors after dimensionality reduction, which reduces data redundancy while retaining the core information related to defects.

[0063] The correlation between each feature in the feature vector set and the fatigue crack is compared with a preset correlation threshold. If the correlation between a certain feature and the fatigue crack is lower than the preset correlation threshold, the importance of that feature is adjusted by matrix multiplication of the gain matrix. When obtaining the optimized feature weight set, the preset correlation threshold needs to be set in conjunction with the process experience of snap ring production—based on a large amount of historical data statistics, a correlation between a feature and the fatigue crack ≥ 0.6 has practical reference value for defect prediction, so the threshold is set to 0.6. Through regression analysis of historical defect data and feature vectors, the correlation between each principal component and the fatigue crack is calculated. The correlation calculation uses the Pearson correlation coefficient formula:

[0064] ;

[0065] Where: X represents the eigenvalue sequence of a certain principal component (such as the eigenvalue set of the second principal component), and Y represents the fatigue crack incidence rate sequence for the corresponding historical period. Let X be the covariance of Y. Let X be the standard deviation. The standard deviation of Y; for example, the correlation coefficient of the first principal component is calculated. r=0.81, correlation of the second principal component =0.55, Third Principal Component Correlation =0.72, where the correlation of the second principal component (0.55) is below the threshold.

[0066] At this point, a gain matrix needs to be constructed: based on the historical correlation deviation between the second principal component and the fatigue crack (the deviation value is a preset threshold minus the initial correlation degree, calculated to be 0.05), the gain coefficient is determined and a 3×3 diagonal gain matrix G is constructed. The diagonal elements corresponding to the principal component to be adjusted (the second principal component) are set to 1.09 (calculated based on the ratio of the deviation value to the initial correlation degree to ensure the correlation degree meets the standard after adjustment), and the remaining diagonal elements are set to 1 (other principal components are not adjusted). The correlation degree is adjusted by matrix multiplication: the correlation degree vector of the principal component R=[0.81,0.55,0.72] is multiplied by the gain matrix G to obtain the adjusted correlation degree vector R'=R×G=[0.81×1,0.55×1.09,0.72×1]=[0.81,0.60,0.72], where the correlation degree of the second principal component is ≈0.60 (meets the threshold requirement); at the same time, the weights are adjusted: the initial weight vector W=[0.42,0.32,0.40] is multiplied by the gain matrix G to obtain the adjusted weight vector W'=W×G=[0.42×1,0.32×1.09,0.40×1]=[0.42,0.35,0.40]. The final optimized feature weight set is {first principal component weight 0.42, second principal component weight 0.35, third principal component weight 0.40}, ensuring that all features can effectively support defect prediction.

[0067] Based on the optimized feature weight set, features with weights meeting the preset weight requirements are selected from the feature vector set through threshold filtering to obtain a refined feature subset. The preset weight requirement balances feature effectiveness and data simplicity, and is typically set to 0.35 (features with weights lower than this value contribute less to the prediction results and can be removed). The optimized feature weight set is traversed, and features with weights ≥ 0.35 are selected: the first principal component (0.42), the second principal component (0.37), and the third principal component (0.40) all meet the requirement and are retained; if a feature has a weight of 0.32 (lower than 0.35), it is removed. The final refined feature subset contains these three principal component features, each covering core information related to snap ring fatigue cracks and micro-stress concentration, with no redundant data—for example, this subset can be directly used for subsequent edge-end compression encoding, reducing data transmission volume while accurately reflecting key factors affecting snap ring quality, providing high-value input for subsequent defect trend judgment.

[0068] In summary, step S103 uses Pearson correlation coefficient to lock in interaction features strongly associated with quality defects (avoiding interference from invalid parameters), principal component analysis to reduce dimensionality and eliminate redundancy (improving subsequent efficiency), gain matrix adjustment to ensure the effectiveness of low-level associated features (avoiding loss of key information), and threshold screening to simplify features, ultimately forming a refined feature subset. The entire process aligns with the technical logic of "parameter association - information compression - feature optimization" in snap ring production.

[0069] In step S104, the data required for edge processing within the refined feature subset is selected, and after compression encoding, signal deviation correction, and real-time data integration, a compressed feature package suitable for cloud upload is obtained, including:

[0070] By using threshold filtering, the edge processing data that matches the edge processing requirements is extracted from the refined feature subset, resulting in filtered edge data.

[0071] In the local computing module of the edge device, the filtered edge data is compressed and encoded using Huffman coding to generate an initial compressed feature packet.

[0072] If the initial compressed feature packet has a potential signal deviation, the initial compressed feature packet is corrected for the signal deviation through state estimation to obtain a corrected compressed feature packet;

[0073] The real-time collected multimodal data and the corrected compressed feature package are integrated according to data type and collection timestamp to generate a compressed feature package suitable for cloud uploading.

[0074] It should be noted that when extracting edge processing data matching the edge processing requirements from the refined feature subset through threshold filtering to obtain the filtered edge data, it is necessary to first clarify the core requirements of edge processing—edge devices have limited computing power and bandwidth, so it is necessary to prioritize extracting features that are key to real-time defect early warning and have a concise data volume. Therefore, the threshold setting needs to combine the contribution of features to defect prediction and the data volume. Referring to the technical requirements for lightweight edge processing, the preset weight threshold is 0.38. This threshold is determined through two derivations: First, feature contribution statistics. Based on the analysis of historical feature data (including 1200 defect samples) from the production of snap rings in the past 3 months, features with a weight ≥ 0.38 (such as the first principal component and the third principal component strongly correlated with fatigue cracks) contribute ≥ 8% to the improvement of defect prediction accuracy, while features with a weight < 0.38 (such as features related to secondary environmental parameters) contribute ≤ 2%, which can only slightly improve prediction robustness, but require additional The first issue is data processing; the second is balancing data volume and transmission efficiency. Features with a weight <0.38, although only 4-8 bytes per data point (such as local environmental humidity-related features), accumulate to 2.4-4.8KB within a 10-minute collection period (600 data sets), accounting for 32% of the total data volume of the refined feature subset. Retaining such features would increase edge-side compression encoding time from 30ms to 65ms and cloud upload latency from 45ms to 110ms. While this doesn't exceed the 500ms latency limit, it would consume processing resources for core features and increase the risk of edge device downtime. Therefore, setting the weight threshold to 0.38 ensures that the retained features cover over 92% of the effective prediction information while keeping edge-side data processing and transmission time within 100ms, adapting to lightweight requirements. Simultaneously, the number of bytes per feature data point is required to be ≤8 bytes (adapting to edge storage and transmission capabilities). Taking the refined feature subset obtained from S104 (containing the first principal component: weight 0.42, the second principal component: weight 0.37, and the third principal component: weight 0.40) as an example, the weights are checked first during the screening process: the first principal component (0.42 ≥ 0.38) and the third principal component (0.40 ≥ 0.38) meet the requirements, while the second principal component (0.37 < 0.38) is removed; then the data volume is verified: a single feature value of the first principal component (e.g., 1.25) is stored as a floating-point number and occupies 4 bytes, while the third principal component (e.g., 0.98) occupies 4 bytes, both meeting the requirement of ≤ 8 bytes. The final filtered edge data consists of {first principal component feature sequence and third principal component feature sequence}, with each sequence containing 10 minutes of data sorted by timestamp (600 groups, totaling 600 × 8 = 4800 bytes), which retains the core defect correlation features and adapts to the edge processing capabilities.

[0075] In the local computing module of the edge device, the filtered edge data is compressed and encoded using Huffman coding to generate the initial compressed feature packet. When generating the initial compressed feature packet, the Huffman coding needs to optimize the coding length based on the frequency of occurrence of the filtered edge data values—first, statistical frequency is calculated: for example, the first principal component feature values ​​are concentrated between 1.0 and 1.5, with 1.2 appearing 120 times and 1.3 appearing 105 times (high-frequency values); the third principal component feature values ​​are concentrated between 0.8 and 1.2, with 1.0 appearing 110 times and 0.9 appearing 95 times (high-frequency values). During encoding, high-frequency values ​​are assigned 2-bit short codes (e.g., 1.2→01, 1.3→10, 1.0→11, 0.9→00), and low-frequency values ​​(e.g., 1.5, 0.8) are assigned 4-bit long codes to reduce the overall data volume. Taking 4800 bytes of filtered edge data as an example, after Huffman coding, the data volume is compressed to 1820 bytes, with a compression rate of about 62%. The generated initial compressed feature package is stored in the structure of "encoding table + compressed data block". The encoding table records the correspondence between values ​​and codes (occupying 180 bytes), and the compressed data block stores the encoded binary sequence (occupying 1640 bytes), ensuring that the data can be accurately restored when decoding in the cloud.

[0076] If the initial compressed feature packet has potential signal deviation, the signal deviation is corrected by state estimation. When the corrected compressed feature packet is obtained, the signal deviation judgment needs to compare the value after decoding the initial compressed feature packet with the historical reference data at the edge end. The preset allowable deviation range is ±0.05 (based on the stability requirements of the snap ring production parameters, deviation exceeding this range is prone to cloud misjudgment). If a certain feature value after decoding (such as the first principal component 1.35) deviates by 0.07 from the historical reference value (1.28) at the same timestamp (out of range), it is determined that there is a signal deviation. The state estimation is implemented using Kalman filtering. The historical reference value is the predicted value, and the current decoded value is the measured value. It is corrected by "prediction-update" iteration: first, the current feature value should be predicted to be 1.29 (based on the trend of the previous 3 frames of historical data), then the residual of 0.06 is calculated by combining the measured value of 1.35, and the Kalman gain is adjusted to 0.6. The final corrected value = 1.29 + 0.6 × (1.35 - 1.29) = 1.326, and the deviation is reduced to 0.046 (within the range). After correcting all out-of-bias values ​​in the initial compressed feature packet, Huffman coding is performed again to generate a corrected compressed feature packet (1830 bytes, slightly larger than the initial packet, due to slight changes in the frequency distribution of the corrected values), ensuring data accuracy.

[0077] When integrating the real-time acquired multimodal data with the corrected compressed feature package according to data type and acquisition timestamp to generate a compressed feature package suitable for cloud upload, the real-time multimodal data needs to select simplified parameters that are strongly correlated with edge processing features—focusing on two types of real-time data that are directly related to refining features: temperature fluctuation (mold cavity temperature) and pressure change (stamping cylinder pressure). The data is acquired at a frequency of 1Hz, and the data of the most recent 5 minutes is retained (300 groups, 8 bytes per group, totaling 2400 bytes). During integration, millisecond-level timestamps are used as the benchmark to match real-time data with the corrected compressed feature package data one by one. For example, real-time data with timestamp 14:30:05.123 (mold temperature 85.2℃, stamping pressure 2.12MPa) corresponds to the first principal component 1.326 and the third principal component 0.99 with the same timestamp in the corrected compressed feature package. The data is combined according to the structure of "timestamp-real-time temperature-real-time pressure-first principal component-third principal component" to form an integrated data block (300 groups, totaling 300×20=6000 bytes). Then, it is compressed to 2250 bytes through Huffman secondary encoding to generate a compressed feature package suitable for cloud upload. This package contains both refined features optimized at the edge and real-time original parameters, which is convenient for in-depth analysis of historical data in the cloud. Moreover, the 2250-byte size can be uploaded within 1 second (adapting to industrial network bandwidth).

[0078] In summary, step S104 achieves lightweight edge data through threshold filtering, adapting to edge computing power and bandwidth; Huffman coding efficiently compresses data, reducing transmission costs; state estimation correction ensures data accuracy and avoids bias interference; and real-time data integration enriches feature dimensions, providing complete input for cloud analysis. The entire process closely aligns with the collaborative logic of "efficient edge preprocessing - deep cloud analysis" in snap ring quality defect prediction.

[0079] In step S105, uploading the compressed feature package to the cloud analysis platform for decoding, integrating historical data to obtain a comprehensive feature dataset, and matching historical patterns to determine potential defect trends includes:

[0080] The compressed feature package is uploaded to a cloud analysis platform, and the compressed feature package is decoded by Huffman decoding to restore the feature vector set.

[0081] Additional historical data on snap ring production is retrieved from the cloud database. The historical data is then fused with the restored feature vector set according to data dimensions and time series to obtain an expanded comprehensive feature dataset.

[0082] The data in the comprehensive feature dataset is initially matched with the historical defect patterns in the historical data by traversing the tree structure, and the matching degree is calculated.

[0083] If the matching degree exceeds a preset matching threshold, the corresponding historical defect pattern is determined as a potential defect trend; if the matching degree does not exceed the preset matching threshold, the comprehensive feature dataset is processed by Naive Bayes classification to obtain the potential defect trend.

[0084] It should be noted that when uploading the compressed feature package to the cloud analysis platform and decoding it using Huffman decoding to restore the feature vector set, the structured data of the compressed feature package (including the Huffman coding table and binary compressed data blocks) must be read first. The decoding logic must be completely matched with the edge-end coding rules—the coding table records the correspondence between numerical values ​​and binary codes (such as "mold temperature 85℃→101" and "first principal component 1.32→011"). The cloud analysis platform first loads the coding table and then parses the binary data blocks in a "bit-by-bit matching-numerical mapping" manner to ensure that the restored data is consistent with the feature vectors filtered by the edge end. For example, the compressed features uploaded from the edge contain 2250 bytes of data (180 bytes of encoding table + 2070 bytes of compressed data block). The cloud uses the encoding table to match and map the binary sequence "101011110" to "mold temperature 85℃, first principal component 1.32, third principal component 0.99" in sequence. Finally, a set of feature vectors for 10 minutes (600 groups) is obtained. Each vector contains five dimensions of data: "timestamp - real-time mold temperature - real-time stamping pressure - first principal component - third principal component". The data volume is restored to 4800 bytes, which is completely consistent with the data filtered by the edge, with no loss or deviation.

[0085] Additional historical data on snap ring production is retrieved from the cloud database. This historical data is then fused with the reconstructed feature vector set according to data dimensions and time series to obtain an expanded comprehensive feature dataset. The additional historical data must be selected from records that perfectly match the current production process and material specifications. The cloud database stores production data for the same model of snap ring over the past three months, covering dimensions consistent with the reconstructed feature vectors, such as "mold temperature, stamping pressure, principal component characteristics, and defect records," and is divided into time series by day (retaining 10 minutes of data consistent with the current production period each day). The fusion process follows the principle of "time series alignment + dimension completion": first, the 10-minute data (timestamp 14:30-14:40) of the reconstructed feature vector is aligned with the 14:30-14:40 time period data from each day in the historical data; then, the "defect marker" dimension from the historical data is added (e.g., a historical time period data marked "fatigue crack risk"). For example, the restored feature vector contains 600 sets of data. Historical data from the past 30 days (30 × 600 = 18,000 sets) are retrieved and merged to form a comprehensive feature dataset of 18,600 sets. Each set of data adds a dimension of "historical same period defect status", such as "timestamp 14:30:05 - mold temperature 85℃ - pressure 2.12MPa - first principal component 1.326 - third principal component 0.99 - no defects in the same period in history", which not only retains real-time features, but also adds historical reference dimensions.

[0086] A preliminary matching process is performed between the data in the comprehensive feature dataset and the historical defect patterns in the historical data through tree structure traversal. When calculating the matching degree, the tree structure needs to construct hierarchical nodes based on the historical defect patterns (the node splitting threshold is set based on the 75th percentile of the feature values ​​in the historical defect patterns). For example, the root node is "whether the first principal component is >1.3" (core defect-related feature), the first-level child nodes are "whether the mold temperature is >84℃" and "whether the stamping pressure is >2.1MPa", the second-level child nodes are "whether the third principal component is >0.95", and the leaf nodes correspond to specific historical defect patterns (such as "fatigue crack", "dimensional deformation", "surface defect"). During the traversal, each group of data in the comprehensive feature dataset is matched one by one according to the node threshold. After falling into the corresponding leaf node, the matching degree between the current data and the historical defect pattern under that leaf node is calculated by cosine similarity (similarity range 0-1, the closer to 1, the higher the matching degree). For example, the current data "First principal component 1.326 > 1.3, mold temperature 85℃ > 84℃, pressure 2.12MPa > 2.1MPa, third principal component 0.99 > 0.95" falls into the "fatigue crack" leaf node. Comparing this to the historical defect patterns under this node: the historical value range of the first principal component is [0.8, 1.5], the standardized value of the current data is (1.326 - 0.8) / (1.5 - 0.8) ≈ 0.751, and the standardized value of the historical defect pattern (1.33) is (1.33 - 0.8) / (1.5 - 0.8) ≈ 0.757; the historical value range of the mold temperature is [80, 120℃], and the standardized value of the current data is (85 - 80) / (120 - 80) = 0.12. 5. The standardized value of the historical defect pattern (85.1℃) is approximately 0.128 (85.1-80) / (120-80). The historical range of stamping pressure is [1.5, 2.5MPa]. The standardized value of the current data is approximately 0.62 (2.12-1.5) / (2.5-1.5). The standardized value of the historical defect pattern (2.11MPa) is approximately 0.61 (2.11-1.5) / (2.5-1.5). The historical range of the third principal component is approximately 0.7, 1.2. The standardized value of the current data is approximately 0.58 (0.99-0.7) / (1.2-0.7). The standardized value of the historical defect pattern (1.0) is approximately 0.6 (1.0-0.7) / (1.2-0.7). The current data is then matched one by one according to the threshold of the tree nodes: after standardization, the first principal component 0.751 corresponds to the original value 1.326 > 1.3, the mold temperature 0.125 corresponds to the original value 85℃ > 84℃, the stamping pressure 0.62 corresponds to the original value 2.12MPa > 2.1MPa, and the third principal component 0.58 corresponds to the original value 0.99 > 0.95, falling into the leaf node of "fatigue crack".The matching degree (cosine similarity) is calculated based on the standardized data: Matching degree = (0.751×0.757+0.125×0.128+0.62×0.61+0.58×0.6) / (√(0.751²+0.125²+0.62²+0.58²)×√(0.757²+0.128²+0.61²+0.6²))≈(0.569+0.016+0.378+0.348) / (√1.334×√1.345)≈1.311 / (1.155×1.16)≈0.98, which is a very high matching degree.

[0087] If the matching degree exceeds the preset matching threshold, the corresponding historical defect pattern is determined as a potential defect trend; if the matching degree does not exceed the preset matching threshold, the comprehensive feature dataset is processed by Naive Bayes classification to obtain the potential defect trend. When the preset matching threshold is obtained, it needs to be set based on the historical matching effect. According to a large amount of historical data, when the matching degree is ≥0.8, the consistency rate between the historical defect pattern and the actual defect reaches 92%. Therefore, the threshold is set to 0.8. If the current data matching degree is 0.98 (exceeding the threshold), the potential defect trend is directly determined as "fatigue crack risk". If the matching degree of a certain data set is 0.75 (below the threshold), Naive Bayes classification is triggered: First, the comprehensive feature dataset is divided into three training samples according to "defect type" (fatigue crack, dimensional deformation, no defect). The conditional probability of the current data in each category is calculated (e.g., P(current data|fatigue crack)=0.42, P(current data|dimensional deformation)=0.25, P(current data|no defect)=0.33). Then, combined with the prior probability (historical fatigue crack occurrence rate of 15%), the posterior probability is calculated using the Bayes formula. Finally, "fatigue crack" with the highest posterior probability is selected as the potential defect trend, ensuring that even if the initial matching does not meet the standard, a reliable result can still be obtained through probability classification.

[0088] In summary, step S105 accurately restores the core features of the edge through Huffman decoding, providing basic data for cloud analysis; historical data fusion completes the time and defect reference dimensions, making the features more correlated; tree structure traversal enables rapid matching of defect patterns, improving judgment efficiency; and Naive Bayes classification serves as a fallback solution, ensuring that low-matching data can still obtain accurate trend judgments.

[0089] In step S106, the random forest model is constructed and optimized based on the comprehensive feature dataset to obtain a trained defect prediction model. Real-time multimodal data is then input for forward computation to obtain the model's forward computation results, including:

[0090] The data of each dimension in the comprehensive feature dataset are normalized by min-max normalization to obtain the normalized comprehensive dataset.

[0091] Multiple sample subsets are randomly selected from the normalized comprehensive dataset by bootstrap sampling, and a decision tree is constructed for each sample subset to form an initial random forest model.

[0092] Calculate the training error of the initial random forest model on the validation set. If the training error is higher than a preset error threshold, optimize the tree structure of the initial random forest model by adjusting the node splitting threshold of the decision tree and the pruning strategy, and iteratively update the model.

[0093] When the training error of the initial random forest model is lower than a preset error threshold, the iterative update is stopped, and the defect prediction model that has been trained is obtained.

[0094] After the real-time multimodal data is processed by min-max normalization, it is input into the trained defect prediction model for forward computation to obtain the model forward computation result.

[0095] It should be noted that when normalizing the data of each dimension in the comprehensive feature dataset using min-max normalization to obtain the normalized comprehensive dataset, the comprehensive feature dataset must cover multiple core parameters—including real-time acquired mold cavity temperature (80-120℃), stamping cylinder pressure (1.5-2.5MPa), and the first principal component (0.8-1.5), third principal component (0.7-1.2), and historical defect markers (0=no defects, 1=fatigue cracks, 2=dimensional deformation) fused in S106. Min-max normalization is performed using the formula: Mapping data from each dimension to the [0,1] interval eliminates the interference of dimensional differences on model training: for example, in the mold temperature dimension, the original value of 85℃ corresponds to... =80℃ =120℃, calculated as follows The stamping pressure of 2.1 MPa corresponds to =1.5MPa =2.5MPa, calculated as follows =0.6. The final normalized comprehensive dataset contains 6000 samples, and the data of each sample in all 5 dimensions are in the range [0,1], such as "timestamp 14:30:05 - temperature normalized value 0.125 - pressure normalized value 0.6 - first principal component normalized value 0.65 - third principal component normalized value 0.72 - defect label 0", which provides standardized input for model training.

[0096] Multiple sample subsets are randomly selected from the normalized comprehensive dataset through bootstrapping. Decision trees are constructed for each sample subset to form the initial random forest model. When bootstrapping, a sampling strategy with replacement is adopted—from 6000 normalized samples, 5000 samples are selected each time to form a subset (the sampling ratio is about 83%, and some samples are reserved for subsequent validation). A total of 10 samplings are performed to generate 10 sample subsets (the number of decision trees is set to 10 to balance model accuracy and computational efficiency). When constructing a single decision tree, the Gini coefficient is used as the node splitting criterion. Three features are randomly selected (to avoid a single feature dominating the split) to determine the splitting threshold. For example, when constructing a decision tree for a subset, three features are randomly selected: "normalized temperature value, first principal component, and normalized pressure value". By calculating the Gini coefficient, the splitting threshold for the root node is determined to be "first principal component > 0.6" (the Gini coefficient of the left and right subtrees decreases the most after splitting). Subsequent child nodes are split according to "normalized temperature value > 0.15" and "normalized pressure value > 0.55" respectively, ultimately generating a decision tree with a depth of 5. The 10 decision trees are combined in subset order to form an initial random forest model containing 10 base learners, with each tree independently outputting the defect category prediction result.

[0097] The training error of the initial random forest model on the validation set is calculated. If the training error is higher than a preset error threshold, the tree structure of the initial random forest model is optimized by adjusting the node splitting threshold of the decision tree and the pruning strategy. When iterating and updating the model, the validation set is taken from the unselected samples in the bootstrap sampling. The training error is calculated using the classification error rate (number of incorrectly predicted samples / total number of validation samples). The preset error threshold is set to 0.05 based on the accuracy requirement of the snap ring defect prediction (i.e., the error rate must be lower than 5%). During the initial model validation, 68 samples were incorrectly predicted out of 1000 samples, and the training error = 68 / 1020 ≈ 0.067 (higher than 0.05), requiring optimization to be initiated. During optimization, the node splitting threshold is first adjusted: the "first principal component > 0.6" of a certain decision tree root node is fine-tuned to ">0.58", reducing the error rate of the tree from 0.07 to 0.05. Then, a pre-pruning strategy is used to prune all branches with leaf nodes containing less than 5 samples (e.g., a branch containing only 3 samples has poor prediction stability) to reduce overfitting. After optimization, recalculate the verification error. If it is still higher than 0.05, repeat the adjustment until the error drops to 0.042 (below 0.05) after a certain iteration, then stop the tree structure optimization.

[0098] When the training error of the initial random forest model is lower than a preset error threshold, the iterative update stops. When the trained defect prediction model is obtained, the optimized structural parameters of the 10 decision trees need to be saved—including the node splitting features, splitting threshold, and defect category probability corresponding to each leaf node (e.g., if a leaf node contains 120 samples, of which 85 are labeled "fatigue crack", then the probability of the node outputting "fatigue crack" is 85 / 120≈0.708). The trained model needs to pass a consistency test: 200 new samples are randomly selected and input into the model. The 10 decision trees output the final category through "majority rule" voting. The prediction accuracy needs to be ≥95% (corresponding to the error threshold of 0.05). For example, if 192 out of 200 samples are predicted correctly, the accuracy is 96%, which meets the requirements, and the model can be used for subsequent real-time prediction.

[0099] After processing the real-time multimodal data through min-max normalization, it is input into the trained defect prediction model for forward calculation. When obtaining the model forward calculation result, the real-time multimodal data must be consistent with the dimensions of the comprehensive feature dataset. For example, select the real-time data collected at a certain moment, such as "mold temperature 86℃, stamping pressure 2.15MPa, first principal component 1.35, third principal component 1.0", and process it with the same normalization parameters: temperature normalization value = (86-80) / (120-80) = 0.15, pressure normalization value = (2.15-1.5) / (2.5-1.5) = 0.65, first principal component normalization value = (1.35-0.8) / (1.5-0.8) ≈ 0.786, third principal component normalization value = (1.0-0.7) / (1.2-0.7) = 0.6. The standardized data is input into the model, and the 10 decision trees output prediction results respectively: 6 predict "fatigue crack", 3 predict "no defects" and 1 predict "dimensional deformation". The final category is determined as "fatigue crack" according to the voting rules, and the category probabilities are output at the same time ("fatigue crack" probability = 6 / 10 = 0.6, "no defects" probability 0.3, "dimensional deformation" probability 0.1). The probability output and the category result together constitute the forward calculation result of the model.

[0100] In summary, step S106 eliminates data dimension interference through min-max standardization, providing a uniform scale input for model training; bootstrap sampling and multi-decision tree construction ensure the diversity and robustness of the initial model; training error judgment and tree structure optimization continuously improve model accuracy until the preset requirements are met; real-time data standardization and forward computation enable the engineering application of the model. The entire process closely follows the technical logic of "data standardization - model construction - optimization and verification - real-time prediction" in snap ring quality defect prediction.

[0101] In step S107, determining the type of potential quality defect based on the model's forward calculation results, triggering defect classification to obtain the final prediction result, and feeding back to adjust subsequent filtering parameters includes:

[0102] Extract the probability output from the forward calculation results of the model, compare the probability output with a preset output threshold, and determine the type of potential quality defect;

[0103] If the probability output exceeds a preset limit, defect type classification is triggered to determine the specific defect type. The specific defect type, probability output, and corresponding real-time multimodal data features are then integrated to obtain the final defect prediction result.

[0104] Based on the final defect prediction results, the compatibility between the current noise filtering parameters and the defect signal is analyzed. If there is significant noise interference in the defect signal, the noise threshold is adjusted.

[0105] Based on the adjusted noise threshold, the covariance matrix parameters of the Kalman filter are updated to provide optimized filtering parameters for noise filtering of subsequent multimodal data.

[0106] It should be noted that when extracting the probability output from the model's forward calculation results and comparing it with a preset output threshold to determine the type of potential quality defects, a preset output threshold must first be set based on the defect risk level of the snap ring production. Referring to the reliability requirements for predicting common defects such as fatigue cracks, dimensional deformation, and surface depressions in the process, the preset output threshold is set to 0.6. That is, when the probability output of a certain type of defect is ≥0.6, it is initially determined to be a potential defect type; if it is below 0.6, it is temporarily classified as "no clear defect" to avoid misjudgment due to low probability signals. When extracting the probability output, the focus should be on the model's output results for core defect categories, such as "fatigue crack probability 0.72, dimensional deformation probability 0.21, surface depression probability 0.07," and each should be compared with the threshold: if the fatigue crack probability 0.72 ≥ 0.6, the potential quality defect type is initially determined to be "fatigue crack risk"; if a set of results is "dimensional deformation probability 0.58, fatigue crack probability 0.35," since both are below 0.6, it is temporarily determined to be "no clear defect." For example, at a certain moment, the forward calculation results of the model are "fatigue crack probability 0.68, dimensional deformation probability 0.25, and surface depression probability 0.07". After comparing each probability with the preset output threshold of 0.6, the fatigue crack probability of 0.68 exceeds the threshold, and the potential quality defect type is initially judged to be "fatigue crack risk".

[0107] If the probability output exceeds a preset limit, defect type classification is triggered to determine the specific defect type. When integrating the specific defect type, probability output, and corresponding real-time multimodal data features to obtain the final defect prediction result, the preset limit must be higher than the output threshold, which is set to 0.7 (based on the technical indicator that "defect prediction accuracy must be ≥95%", when the probability exceeds 0.7, the actual defect matching rate can reach more than 96%, ensuring classification accuracy). After the classification is triggered, the specific defect type is first confirmed by matching the historical defect feature library. The historical defect feature library is constructed based on the defect records of the entire process of circlip stamping, heat treatment, and surface treatment (including defect generation process, related parameter fluctuations, and actual quality inspection results) and the "feature-defect" mapping rules. It stores feature combination thresholds according to defect type (fatigue crack, dimensional deformation, surface depression) and severity (mild, moderate, severe). The classification threshold for fatigue crack is clearly defined: mild corresponds to "first principal component 1.3-1.4, mold temperature fluctuation ±0.5-±0.8℃, stamping pressure 2.1-2.2MPa", and mild defects can be repaired by fine-tuning the tempering temperature. If the current data (first principal component 1.326, mold temperature fluctuation ±0.6℃, stamping pressure 2.12MPa) perfectly matches the threshold for "mild fatigue crack," with a matching degree of 92% (meeting the standard of ≥90% for validity), and combined with the process repair attributes of mild defects, the specific defect type is ultimately confirmed as "mild fatigue crack"—that is, the retaining ring experiences slight micro-stress concentration due to temperature and pressure fluctuations, without forming a visible crack, requiring fine-tuning of the heat treatment tempering temperature to eliminate stress. Then, by integrating this defect type, probability output, and real-time multimodal data features (such as mold temperature 85.3℃, stamping pressure 2.12MPa, first principal component 1.32), a structured result containing "defect category - confidence level - real-time parameters" is formed. For example, in the forward calculation results of a certain model, "fatigue crack probability 0.75" (exceeding the specified limit of 0.7), after triggering classification and matching the historical feature library, the specific defect type is confirmed as "fatigue crack (mild)". After integration, the final defect prediction result is "defect type: fatigue crack (mild), probability: 0.75, real-time features: mold temperature 85.3℃, stamping pressure 2.12MPa, first principal component 1.32".

[0108] Based on the final defect prediction result, the compatibility between the current noise filtering parameters and the defect signal is analyzed. If there is significant noise interference in the defect signal, when adjusting the noise threshold, the matching degree between the current noise filtering parameters (such as the process noise covariance Q and measurement noise covariance R of Kalman filtering) and the defect signal needs to be checked in reverse—"The allowable noise deviation of the snap ring defect signal is ±0.05℃ (temperature) and ±0.02MPa (pressure)." If the defect signal corresponding to the final prediction result (such as the temperature fluctuation signal associated with fatigue cracks) still has a deviation of 0.08℃ (exceeding the allowable range) after current filtering, it is determined that the current noise threshold (frequency domain noise amplitude threshold 0.2) is too high and cannot effectively suppress noise. During adjustment, dynamic correction is made according to the "deviation ratio": if the noise deviation of 0.08℃ exceeds the allowable value of 0.05℃ by more than 60%, the original noise threshold of 0.2 is lowered to 0.16 (a reduction of 20% to avoid over-filtering leading to loss of effective signals). For example, the final defect prediction result shows "fatigue crack (mild)", and the corresponding real-time temperature fluctuation signal still has a deviation of 0.09℃ after the current noise filtering (exceeding the allowable deviation ±0.05℃). The analysis determines that the current noise threshold of 0.2 is not suitable, so it is adjusted to 0.17 to enhance the filtering ability of temperature signal noise.

[0109] Based on the adjusted noise threshold, the covariance matrix parameters of the Kalman filter are updated to provide optimized filtering parameters for noise filtering of subsequent multimodal data. The core parameters of the Kalman filter (process noise covariance Q and measurement noise covariance R) need to be updated in conjunction with the noise threshold. The process noise covariance Q is positively correlated with the noise threshold. The original Q = 0.015, and the adjusted noise threshold of 0.17 is 15% lower than the original 0.2. Therefore, Q is updated to 0.015 × (1-15%) = 0.01275. The measurement noise covariance R is determined by the sensor accuracy. If the current noise originates from equipment vibration (unrelated to the sensor), R remains unchanged at 0.02, ensuring that parameter optimization focuses on noise source suppression. The updated parameters need to be stored in the filter parameter library and directly used for multimodal data processing in the next production batch (such as temperature and pressure signal filtering in subsequent stamping processes). For example, the adjusted noise threshold is 0.17, which is 15% lower than the original threshold of 0.2. Based on this, the process noise covariance Q of the Kalman filter is updated from 0.015 to 0.01275, while the measurement noise covariance R remains unchanged at 0.02. The updated parameters are stored in the parameter library for noise filtering of temperature and pressure signals in subsequent snap ring stamping processes.

[0110] In summary, step S107 achieves accurate identification of defect types through "probability threshold hierarchical judgment," which avoids misjudgment of low-probability signals and triggers fine classification through high thresholds. The integration of defect results with real-time data provides intuitive basis for operation and maintenance. The noise parameter adaptability analysis and covariance matrix update construct a "prediction-optimization" closed loop, ensuring that the noise filtering effect of subsequent data processing is more in line with the actual defect signal characteristics.

[0111] Reference Figure 2 The second embodiment of the present invention provides a machine learning-based intelligent prediction system for quality defects of snap rings, comprising:

[0112] The data acquisition module is used to acquire multimodal data from the snap ring production equipment;

[0113] The data processing module is used to perform preliminary aggregation and signal smoothing on the multimodal data to obtain a raw dataset containing material properties, and to perform noise filtering, iteratively adjusting the filtering parameters to obtain a clean dataset.

[0114] The feature optimization module is used to extract key features from the clean dataset, and after dimensionality reduction and weight optimization, a refined feature subset is obtained.

[0115] The edge processing module is used to select the data required for edge processing within the refined feature subset, and after compression encoding, signal deviation correction and real-time data integration, obtain a compressed feature package suitable for cloud upload;

[0116] The cloud fusion module is used to upload the compressed feature package to the cloud analysis platform for decoding, integrate historical data to obtain a comprehensive feature dataset, and match historical patterns to determine potential defect trends.

[0117] The model building and computation module is used to build and optimize the random forest model based on the comprehensive feature dataset to obtain the trained defect prediction model. It inputs real-time multimodal data for forward computation to obtain the model forward computation results.

[0118] The defect judgment and feedback module is used to judge the type of potential quality defects based on the forward calculation results of the model, trigger defect classification to obtain the final prediction result, and provide feedback to adjust subsequent filtering parameters.

[0119] It should be noted that the intelligent prediction device for quality defects of snap rings based on machine learning provided in this embodiment of the invention is used to execute all the process steps of the intelligent prediction method for quality defects of snap rings based on machine learning in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0120] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a machine learning-based intelligent prediction program for snap ring quality defects. When the processor executes the computer program, it implements the steps described in the various machine learning-based intelligent prediction method embodiments for snap ring quality defects, for example... Figure 1 The step S101 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data preprocessing module.

[0121] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0122] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0123] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the electronic device through various interfaces and lines.

[0124] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0125] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0126] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0127] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A machine learning-based intelligent prediction method for spring clip quality defects, characterized by, include: Obtain multimodal data from snap ring manufacturing equipment; The multimodal data is initially aggregated and smoothed to obtain a raw dataset containing material properties. Noise is then filtered, and the filtering parameters are iteratively adjusted to obtain a clean dataset. Key features of the clean dataset are extracted and then subjected to dimensionality reduction and weight optimization to obtain a refined feature subset. The data required for edge processing within the refined feature subset is selected, and after compression encoding, signal deviation correction, and real-time data integration, a compressed feature package suitable for cloud upload is obtained. The compressed feature package is uploaded to the cloud analysis platform for decoding, and historical data is integrated to obtain a comprehensive feature dataset. Historical patterns are matched to determine potential defect trends. Based on the comprehensive feature dataset, a random forest model is constructed and optimized to obtain a trained defect prediction model. Real-time multimodal data is input for forward computation to obtain the model's forward computation results. Based on the forward calculation results of the model, the potential quality defect type is determined, defect classification is triggered to obtain the final prediction result, and subsequent filtering parameters are adjusted accordingly. The process involves preliminary aggregation and signal smoothing of the multimodal data to obtain a raw dataset containing material properties, followed by noise filtering and iterative adjustment of the filtering parameters to obtain a clean dataset, including: The multimodal data is initially aggregated according to the collection timestamp to generate a raw dataset containing material properties; Kalman filtering is used to perform state vector prediction on the temperature fluctuation signal and pressure change signal in the multimodal data to obtain smoothed signal data. If the deviation between the smoothed signal data and the preset signal threshold exceeds the allowable range, then the smoothed signal data is analyzed in the frequency domain by Fourier transform to determine the initial noise level. The original dataset is decomposed in the frequency domain by Fourier transform to obtain a frequency domain signal dataset. The noise level of the frequency domain signal dataset is compared with a preset noise threshold. If the noise level of the frequency domain signal dataset exceeds the preset noise threshold, the Kalman filter parameters are updated and adjusted through the covariance matrix to obtain an optimized filter parameter set. Based on the optimized set of filtering parameters, the frequency domain signal dataset is subjected to secondary noise filtering using Kalman filtering to obtain a clean dataset with reduced noise. 2.The machine learning based intelligent prediction method of spring clip quality defects, according to claim 1, wherein, The acquisition of multimodal data from the snap ring manufacturing equipment includes: The multimodal data is collected from the snap ring production equipment, and the multimodal data includes real-time signals of temperature fluctuations and pressure changes. 3.The machine learning based intelligent prediction method of the spring clip quality defect, according to claim 1, characterized in that, The key features extracted from the clean dataset are then processed through dimensionality reduction and weight optimization to obtain a refined feature subset, including: By calculating the Pearson correlation coefficient, key features of the interaction relationship related to the quality defects of the snap ring in the clean dataset are extracted to form a multimodal key feature set; Principal component analysis is used to reduce the dimensionality of the multimodal key feature set. Principal components are calculated by eigenvalue decomposition to obtain the dimensionality-reduced feature vector set. The correlation between each feature in the feature vector set and the fatigue crack is compared with a preset correlation threshold. If the correlation between a certain feature and the fatigue crack is lower than the preset correlation threshold, the importance of the feature is adjusted to obtain an optimized feature weight set. Based on the optimized feature weight set, features whose weights meet the preset weight requirements are selected from the feature vector set through threshold filtering to obtain a refined feature subset.

4. The intelligent prediction method for quality defects of snap rings based on machine learning according to claim 1, characterized in that, The data selected for edge processing within the refined feature subset is compressed, encoded, corrected for signal deviation, and integrated in real time to obtain a compressed feature package suitable for cloud upload, including: By using threshold filtering, the edge processing data that matches the edge processing requirements is extracted from the refined feature subset, resulting in filtered edge data. In the local computation of the edge device, the filtered edge data is compressed and encoded using Huffman coding to generate an initial compressed feature packet. If the initial compressed feature packet has a potential signal deviation, the initial compressed feature packet is corrected for the signal deviation through state estimation to obtain a corrected compressed feature packet; The real-time collected multimodal data and the corrected compressed feature package are integrated according to data type and collection timestamp to generate a compressed feature package suitable for cloud uploading.

5. The intelligent prediction method for quality defects of snap rings based on machine learning according to claim 1, characterized in that, The step of uploading the compressed feature package to the cloud analysis platform for decoding, integrating historical data to obtain a comprehensive feature dataset, and matching historical patterns to determine potential defect trends includes: The compressed feature package is uploaded to a cloud analysis platform, and the compressed feature package is decoded by Huffman decoding to restore the feature vector set. Additional historical data on snap ring production is retrieved from the cloud database. This additional historical data is then fused with the restored feature vector set according to data dimensions and time series to obtain an expanded comprehensive feature dataset. The data in the comprehensive feature dataset is initially matched with the historical defect patterns in the historical data by traversing the tree structure, and the matching degree is calculated. If the matching degree exceeds a preset matching threshold, the corresponding historical defect pattern is determined as a potential defect trend; if the matching degree does not exceed the preset matching threshold, the comprehensive feature dataset is processed by Naive Bayes classification to obtain the potential defect trend.

6. The intelligent prediction method for quality defects of snap rings based on machine learning according to claim 1, characterized in that, The random forest model is constructed and optimized based on the comprehensive feature dataset to obtain a trained defect prediction model. Real-time multimodal data is then input for forward computation to obtain the model's forward computation results, including: The data of each dimension in the comprehensive feature dataset are normalized to obtain the normalized comprehensive dataset; Multiple sample subsets are randomly selected from the normalized comprehensive dataset by bootstrap sampling, and a decision tree is constructed for each sample subset to form an initial random forest model. Calculate the training error of the initial random forest model on the validation set. If the training error is higher than a preset error threshold, optimize the tree structure of the initial random forest model by adjusting the node splitting threshold of the decision tree and the pruning strategy, and iteratively update the model. When the training error of the initial random forest model is lower than a preset error threshold, the iterative update is stopped, and the defect prediction model that has been trained is obtained. After standardizing the real-time multimodal data, it is input into the trained defect prediction model for forward computation to obtain the model's forward computation results.

7. The intelligent prediction method for quality defects of snap rings based on machine learning according to claim 1, characterized in that, The process of determining the type of potential quality defects based on the forward calculation results of the model, triggering defect classification to obtain the final prediction result, and feeding back and adjusting the filtering parameters includes: Extract the probability output from the forward calculation results of the model, compare the probability output with a preset output threshold, and determine the type of potential quality defect; If the probability output exceeds a preset limit, defect type classification is triggered to determine the specific defect type. The specific defect type, probability output, and corresponding real-time multimodal data features are then integrated to obtain the final defect prediction result. Based on the final defect prediction results, the compatibility between the current noise filtering parameters and the defect signal is analyzed. If there is significant noise interference in the defect signal, the noise threshold is adjusted. Based on the adjusted noise threshold, the covariance matrix parameters of the Kalman filter are updated to provide optimized filtering parameters for noise filtering of subsequent multimodal data.

8. A machine learning-based intelligent prediction system for quality defects in snap rings, characterized in that, include: The data acquisition module is used to acquire multimodal data from the snap ring production equipment; The data processing module is used to perform preliminary aggregation and signal smoothing on the multimodal data to obtain a raw dataset containing material properties, and to perform noise filtering, iteratively adjusting the filtering parameters to obtain a clean dataset. The feature optimization module is used to extract key features from the clean dataset, and after dimensionality reduction and weight optimization, a refined feature subset is obtained. The edge processing module is used to select the data required for edge processing within the refined feature subset, and after compression encoding, signal deviation correction and real-time data integration, obtain a compressed feature package suitable for cloud upload; The cloud fusion module is used to upload the compressed feature package to the cloud analysis platform for decoding, integrate historical data to obtain a comprehensive feature dataset, and match historical patterns to determine potential defect trends. The model building and computation module is used to build and optimize the random forest model based on the comprehensive feature dataset to obtain the trained defect prediction model. It inputs real-time multimodal data for forward computation to obtain the model forward computation results. The defect judgment and feedback module is used to judge the type of potential quality defects based on the forward calculation results of the model, trigger defect classification to obtain the final prediction result, and provide feedback to adjust subsequent filtering parameters. The process involves preliminary aggregation and signal smoothing of the multimodal data to obtain a raw dataset containing material properties, followed by noise filtering and iterative adjustment of the filtering parameters to obtain a clean dataset, including: The multimodal data is initially aggregated according to the collection timestamp to generate a raw dataset containing material properties; Kalman filtering is used to perform state vector prediction on the temperature fluctuation signal and pressure change signal in the multimodal data to obtain smoothed signal data. If the deviation between the smoothed signal data and the preset signal threshold exceeds the allowable range, then the smoothed signal data is analyzed in the frequency domain by Fourier transform to determine the initial noise level. The original dataset is decomposed in the frequency domain by Fourier transform to obtain a frequency domain signal dataset. The noise level of the frequency domain signal dataset is compared with a preset noise threshold. If the noise level of the frequency domain signal dataset exceeds the preset noise threshold, the Kalman filter parameters are updated and adjusted through the covariance matrix to obtain an optimized filter parameter set. Based on the optimized set of filtering parameters, the frequency domain signal dataset is subjected to secondary noise filtering using Kalman filtering to obtain a clean dataset with reduced noise.

Images