A detection method for defects of a flexible printed circuit wire harness

Abstract

The present application provides a method for detecting defects in flexible printed circuit harnesses. By using a high-precision sensor array to collect data on the resistance changes of the harness in real time, combining infrared thermal imaging technology to obtain temperature distribution information, and using Fourier transform and deep learning models for data fusion analysis, precise positioning and classification of minute defects are achieved. The present invention also employs the Kalman filtering algorithm to dynamically predict the defect location and adaptively adjust the sensor configuration according to the detection results to improve the detection accuracy. This real-time monitoring method of multi-source data fusion can effectively identify hidden defects in the harness production process, such as microcracks and poor contacts, etc., providing precise data support for production quality control, helping to timely discover and solve potential problems, and improving the production efficiency and product reliability of flexible printed circuit harnesses.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a method for detecting defects in a flexible printed circuit harness. Background Art

[0002] As a key technology in the field of modern electronic manufacturing, flexible printed circuit harnesses are widely used in high-reliability scenarios such as smart devices, automotive electronics, and aerospace. The optimization of their production methods directly determines product quality and system stability. Flexible harnesses, with their lightness, high density, and bendability, have become a core component that promotes the miniaturization and integration of electronic devices. However, how to ensure the continuity and reliability of harnesses while achieving efficient production is a technical bottleneck that needs to be broken through.

[0003] Current production inspection methods, such as traditional end-to-end continuity testing, are easy to operate, but often have significant limitations when faced with high-density multi-layer flexible wiring harnesses. These methods can usually only determine whether the circuit is connected or disconnected, and cannot accurately locate the specific location of minor defects, and are not sensitive enough to detect hidden problems such as microcracks or poor contact. This extensive inspection method is difficult to meet the stringent requirements of quality control in complex wiring harness structures, resulting in potential defects that may evolve into major failures in subsequent use.

[0004] In the production process of flexible printed circuit harnesses, the core challenges faced by defect detection are mainly concentrated on two technical factors: first, the detection accuracy is not enough to capture physical damage or resistance changes at the micron level, and second, the detection range is difficult to cover the entire length of the harness and achieve real-time online monitoring. Due to the particularity of flexible materials and the complexity of multi-layer structures, tiny defects may be concealed in production, and the low resolution and offline operation mode of traditional detection technology make these problems difficult to be discovered and located in time. This not only increases the rework cost, but may also affect the final reliability of the product.

[0005] Therefore, how to quickly locate and identify hidden defects such as microcracks and breakpoints by improving detection accuracy and coverage in the production of flexible printed circuit harnesses has become a key issue in quality control. The solution to this problem requires a technical breakthrough in the limitations of traditional detection methods and the development of a systematic solution that can continuously monitor and accurately locate defects over the entire length. Summary of the invention

[0006] In order to solve the technical problems raised in the above background technology, the present invention provides a method for detecting defects of a flexible printed circuit harness, the method comprising:

[0007] S1. Obtain the real-time resistance data of the flexible printed circuit wire harness during the production process. Collect the resistance change values of each line segment through a high-precision sensor array, perform signal amplification processing for micron-level fluctuations, and obtain a preliminary resistance change distribution map.

[0008] S2. Extract the abnormal fluctuation area from the resistance change distribution map, analyze the signal frequency characteristics using Fourier transform, determine whether there is a periodic change caused by micro-defects, and determine the coordinates of potential physical damage locations.

[0009] S3. For the coordinates of the physical damage locations, deploy infrared thermal imagers within the entire length range of the wire harness to obtain the temperature distribution data on the surface of the wire harness during the production process. Through the spatial correspondence relationship between the thermal abnormal points and the resistance abnormal points, obtain a preliminary positioning set of hidden defects.

[0010] S4. Through the preliminary positioning set of hidden defects, call the pre-established deep learning model to perform fusion analysis on the infrared thermal imaging data and the resistance change data, judge the specific types of micro-cracks or poor contacts, and determine the defect classification results with high confidence.

[0011] S5. According to the defect classification results, generate a defect distribution vector for the entire length of the wire harness, and combine the time stamp sequence recorded in the real-time monitoring system to obtain the dynamic change trend of each defect during the production process.

[0012] S6. Extract the evolution characteristics of the defects from the dynamic change trend, use the Kalman filter algorithm to predict and correct the defect positions, and determine the precise defect coordinate set within the entire length of the wire harness for the rapid positioning requirement.

[0013] S7. Through the precise defect coordinate set, generate a real-time monitoring report, and combine the equipment operation parameters during the production process to judge whether the defects exceed the preset defect threshold range, and obtain the production link data that needs to be adjusted.

[0014] S8. According to the production link data, update the sensor configuration plan of the real-time monitoring system, increase the sampling frequency for areas with insufficient detection accuracy, obtain the optimized resistance change and temperature distribution data, and determine the new round of defect positioning results.

[0015] Optionally, step S1, obtain the real-time resistance data of the flexible printed circuit wire harness during the production process. Collect the resistance change values of each line segment through a high-precision sensor array, perform signal amplification processing for micron-level fluctuations, and obtain a preliminary resistance change distribution map, including:

[0016] Step S11. Collect the resistance data of each line segment through the sensor array, process the micron-level changes using high-precision acquisition technology, and obtain the real-time resistance data stream.

[0017] Step S12: Extract the resistance fluctuation characteristics from the real-time resistance data stream, use wavelet transform to amplify the signal for micron-level changes, and obtain the enhanced resistance fluctuation signal;

[0018] Step S13: According to the enhanced resistance fluctuation signal, use the difference method to calculate the resistance change value of each line segment to obtain a resistance change data set;

[0019] Step S14: If the fluctuation value in the resistance change data set exceeds the preset fluctuation value threshold, perform secondary calibration on the signal amplification result of the corresponding line segment using the least squares method to obtain the calibrated resistance change data;

[0020] Step S15: Use the K-means clustering algorithm to classify the calibrated resistance change data to determine the resistance change category of each line segment;

[0021] Step S16: Generate a preliminary resistance change distribution map through the resistance change category, and obtain the resistance change trend of each line segment in the distribution map;

[0022] Step S17: According to the resistance change trend, use the bilinear interpolation method to smooth the distribution map to obtain an optimized resistance change distribution map.

[0023] Optionally, in step S2: Extract the abnormal fluctuation area from the resistance change distribution map, use Fourier transform to analyze the signal frequency characteristics, judge whether there is a periodic change caused by a micro defect, and determine the coordinate of the potential physical damage position, including:

[0024] Step S21: Obtain the resistance change data through the distribution map and extract the abnormal fluctuation area;

[0025] Step S22: Use the fast Fourier transform (FFT) algorithm from the abnormal fluctuation area to obtain the signal frequency characteristics;

[0026] Step S23: Analyze the periodic change for the signal frequency characteristics to judge whether there is a regular fluctuation;

[0027] Step S24: If the periodic change exceeds the preset periodic change threshold, determine the existence of a micro defect;

[0028] Step S25: Obtain the position coordinate of the physical damage through the signal frequency corresponding to the micro defect;

[0029] Step S26: Use the bilinear interpolation method for coordinate mapping to determine the specific position of the physical damage in the distribution map;

[0030] Step S27: Judge the boundary of the damage range by comparing the position coordinate with the resistance change data.

[0031] Optionally, in step S27, by comparing the position coordinates with the resistance change data, the boundary of the damage range is determined, including:

[0032] Step S271, obtaining the resistance change data through the position coordinates and judging the boundary extraction result;

[0033] Step S272, extracting features from the resistance change data, performing dimensionality reduction using the principal component analysis algorithm, and obtaining the change analysis features;

[0034] Step S273, calculating the coordinate offset according to the change analysis features and determining the position mapping relationship;

[0035] Step S274, extracting the boundary point set based on the position mapping relationship and judging the boundary of the damage range;

[0036] Step S275, if the boundary point set exceeds the pre-set boundary point set threshold, then determine the range determination result; Step S276, according to the range determination result, use the bilinear interpolation method to interpolate the boundary points to obtain the damage range boundary coordinates;

[0037] Step S277, comparing the damage range boundary coordinates with the resistance change data to obtain the boundary judgment basis.

[0038] Optionally, in step S4, through the preliminary positioning set of hidden defects, a pre-established deep learning model is called to perform fusion analysis on the infrared thermal imaging data and the resistance change data, judge the specific types of microcracks or poor contacts, and determine the defect classification result with high confidence, including:

[0039] Step S41, through the preliminary positioning set of hidden defects, using an infrared thermal imager to obtain infrared thermal imaging data, using a resistance tester to obtain resistance change data, and determining the preliminary analysis basis;

[0040] Step S42, from the infrared thermal imaging data and the resistance change data, using a convolutional neural network model, pre-training the model on the training set, inputting the two types of data, and obtaining the feature extraction result;

[0041] Step S43, for the feature extraction result, using a support vector machine classifier to analyze the distribution characteristics of microcrack types and poor contacts, and judge the specific categories of defects;

[0042] Step S44, if the feature intensity of the microcrack type exceeds the preset feature intensity threshold, then adjust the weight parameters through the convolutional neural network model to determine the microcrack classification output;

[0043] Step S45, if the distribution characteristics of the poor contact are consistent with the resistance change data, then input the infrared thermal imaging data into the support vector machine classifier to obtain the poor contact classification output;

[0044] Step S46, based on the microcrack classification output and the poor contact classification output, use the weighted average method to fuse and analyze the process results, and judge the final type of defect classification;

[0045] Step S47, through the final type of defect classification, call the confidence evaluation method based on probability to determine the confidence output of the classification result.

[0046] Optionally, in step S46, based on the microcrack classification output and the poor contact classification output, use the weighted average method to fuse and analyze the process results, and judge the final type of defect classification, including:

[0047] Step S461, extract the feature vectors of the two types of data from the classification outputs of microcracks and poor contacts as the basis for preliminary fusion;

[0048] Step S462, for the extracted feature vectors, use the weighted average method to process the distribution characteristics of microcrack classification and poor contact to obtain the fused feature vectors;

[0049] Step S463, according to the fused feature vectors, adjust the parameters through a pre-established neural network model to judge the tendency of defect classification;

[0050] Step S464, according to the judgment result, obtain the preliminary type distribution of defect classification and determine the sorting result of type tendency;

[0051] Step S465, according to the sorting result, use a support vector machine classifier to process the output of the fusion process to obtain the adjusted type of defect classification;

[0052] Step S466, according to the adjusted type, fuse the analysis results again by the weighted average method to determine the final type of defect classification;

[0053] Step S467, according to the final type, extract the core features of the classification output from the result fusion to judge the integrity of business analysis.

[0054] Optionally, in step S5, according to the defect classification result, generate a defect distribution vector for the full length of the wire harness, and combine the time stamp sequence recorded in the real-time monitoring system to obtain the dynamic change trend of each defect during the production process, including:

[0055] Step S51, obtain the defect distribution data for the full length of the wire harness through the defect classification result and generate a distribution vector representation;

[0056] Step S52, extract the time stamp sequence from the real-time monitoring system, align the time stamp with the distribution vector in time to generate a time-defect distribution matrix;

[0057] Step S53: Based on the time-defect distribution matrix, calculate the change values of each defect at different time points to generate a time change sequence;

[0058] Step S54: According to the time change sequence, use the moving average method to extract the preliminary dynamic trend;

[0059] Step S55: For the preliminary dynamic trend, set a sliding window with a fixed size, calculate the change rate within the window to generate smoothed trend data;

[0060] Step S56: If there are outliers in the smoothed trend data, use the mean filtering method to replace the values of the outliers with the mean within the window to generate a stable trend result;

[0061] Step S57: According to the stable trend result, with time as the independent variable and defect distribution as the dependent variable, use the linear regression algorithm to calculate the long-term change slope of each defect to generate a trend direction value;

[0062] Step S58: According to the trend direction value, accumulate the slope values at each time point to generate an accumulated change slope sequence, and determine the dynamic change trend of each defect during the production process.

[0063] Optionally, in step S6, extract the evolution characteristics of the defects from the dynamic change trend, use the Kalman filter algorithm to predict and correct the defect positions, and for the fast positioning requirement, determine the set of precise defect coordinates within the entire length of the wire harness, including:

[0064] Step S61: Obtain the defect evolution data from the dynamic changes, and use the principal component analysis method to extract the key features to obtain an evolution feature set;

[0065] Step S62: Use the Kalman filter algorithm to filter the evolution feature set to obtain the preliminarily predicted defect positions;

[0066] Step S63: For the preliminarily predicted defect positions, use the least squares method for correction to determine the optimized defect position data;

[0067] Step S64: Extract the position change pattern from the optimized defect position data through linear regression analysis to obtain the position distribution along the entire length of the wire harness;

[0068] Step S65: If the position distribution exceeds the preset position distribution threshold, use the K-nearest neighbor algorithm to match the fast positioning requirement to obtain an adjusted set of defect positions;

[0069] Step S66: According to the position distribution along the entire length of the wire harness and the adjusted set of defect positions, use the interpolation method to determine the set of precise coordinates;

[0070] Step S67: Obtain the final defect location result through the mapping relationship between the coordinate set and the defect location.

[0071] Optionally, in step S8, according to the production link data, update the sensor configuration plan of the real-time monitoring system, increase the sampling frequency for areas with insufficient detection accuracy, obtain the optimized resistance change and temperature distribution data, and determine a new round of defect location results, including:

[0072] Step S81: Obtain the configuration data of the real-time monitoring system from the production link, generate sensor configuration update data, and extract preliminary optimization parameters;

[0073] Step S82: Update the sampling frequency of the monitoring system according to the sampling frequency adjustment value in the preliminary optimization parameters, obtain the adjusted resistance change data, and draw a resistance change trend chart;

[0074] Step S83: Combine the resistance change trend chart with the sampling frequency adjustment value to generate optimized temperature distribution data, and calculate the temperature distribution uniformity;

[0075] Step S84: If the temperature distribution uniformity is lower than the uniformity threshold, readjust the sensor configuration, obtain the updated monitoring data, and generate a preliminary defect location result;

[0076] Step S85: According to the preliminary defect location result, use the K-means algorithm to cluster the resistance change data and temperature distribution data, extract the defect distribution characteristics after clustering, and generate the accurate defect location;

[0077] Step S86: According to the accurate defect location, adjust the parameter configuration of the real-time monitoring system, obtain the optimized system operation data, and calculate the system stability;

[0078] Step S87: If the system stability meets the preset conditions, use the support vector machine algorithm to verify the defect location result, generate the verified defect distribution data, and determine the final optimization result.

[0079] Optionally, in step S86, according to the accurate defect location, adjust the parameter configuration of the real-time monitoring system, obtain the optimized system operation data, and calculate the system stability, including:

[0080] Step S861: Generate defect distribution data according to the accurate defect and location information, use the KMeans algorithm to cluster the defect distribution data, extract the feature distribution, and obtain a preliminary classification result;

[0081] Step S862: For the preliminary classification result, obtain dynamic update data from the real-time monitoring, calculate the data change trend, and determine the change characteristics;

[0082] Step S863: Update the parameter configuration according to the change characteristics, obtain the adjusted operation data, and judge the data consistency;

[0083] Step S864: If the data consistency is lower than the consistency threshold, regenerate the optimized operation data to obtain the stable parameters;

[0084] Step S865: Adjust the monitoring system through the stable parameters, obtain the optimized dynamic data, and calculate the system response time;

[0085] Step S866: Verify the operation data according to the system response time, generate the verified distribution data, and determine the final characteristic distribution;

[0086] Step S867: If the final characteristic distribution meets the preset conditions, use the support vector machine to classify the distribution data to obtain the classified operation status.

[0087] A detection method for defects of a flexible printed circuit wire harness provided by the present invention collects the wire harness resistance change data in real time through a high-precision sensor array, combines the infrared thermal imaging technology to obtain the temperature distribution information, and uses Fourier transform and deep learning models for data fusion analysis to achieve precise positioning and classification of micro defects. The present invention also uses the Kalman filter algorithm to dynamically predict the defect position and adaptively adjust the sensor configuration according to the detection results to improve the detection accuracy. This real-time monitoring method of multi-source data fusion can effectively identify hidden defects in the wire harness production process, such as micro cracks and poor contacts, etc., provides accurate data support for production quality control, helps to discover and solve potential problems in a timely manner, and improves the production efficiency and product reliability of the flexible printed circuit wire harness. Description of the Drawings

[0088] Figure 1 It is a flowchart of a detection method for defects of a flexible printed circuit wire harness of the present invention.

[0089] Figure 2 It is a flowchart of the second embodiment of a detection method for defects of a flexible printed circuit wire harness of the present invention. Detailed Embodiment

[0090] Next, the technical solutions of the present invention will be described clearly and completely in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work shall fall within the protection scope of the present invention.

[0091] Such as Figure 1 - Figure 2 , a detection method for defects of a flexible printed circuit wire harness provided by the present invention may specifically include:

[0092] S1. Obtain the real-time resistance data of the flexible printed circuit harness during the production process. Collect the resistance change values of each line segment through a high-precision sensor array, perform signal amplification processing for micrometer-level fluctuations, and obtain a preliminary resistance change distribution map.

[0093] Optionally, this step further includes:

[0094] Step S11. Collect the resistance data of each line segment through the sensor array, process the micrometer-level changes using high-precision acquisition technology, and obtain a real-time resistance data stream.

[0095] Step S12. Extract the resistance fluctuation characteristics from the real-time resistance data stream, use wavelet transform to perform signal amplification processing on the micrometer-level changes, and obtain an enhanced resistance fluctuation signal.

[0096] Step S13. According to the enhanced resistance fluctuation signal, use the difference method to calculate the resistance change value of each line segment, and obtain a resistance change data set.

[0097] Step S14. If the fluctuation value in the resistance change data set exceeds the preset fluctuation value threshold, then perform secondary calibration on the signal amplification result of the corresponding line segment using the least squares method to obtain calibrated resistance change data.

[0098] Step S15. Use the K-means clustering algorithm to classify the calibrated resistance change data and determine the resistance change category of each line segment.

[0099] Step S16. Generate a preliminary resistance change distribution map through the resistance change category, and obtain the resistance change trend of each line segment in the distribution map.

[0100] Step S17. According to the resistance change trend, use bilinear interpolation to smooth the distribution map and obtain an optimized resistance change distribution map.

[0101] Exemplarily, when collecting the resistance data of each line segment through the sensor array, one can imagine a high-precision acquisition technology applied to the microcircuit detection scenario.

[0102] Exemplarily, assume this is a manufacturing process of a flexible electronic skin. The sensor array is distributed on the tiny line segments on the skin surface, and the length of each line segment is only a few micrometers. During the acquisition process, the sensor records the resistance value at a frequency of 1000 times per second, capturing the micrometer-level changes caused by external pressure or temperature. This high-precision acquisition can ensure that the real-time resistance data stream contains sufficient details and provides a reliable basis for subsequent analysis. When extracting the resistance fluctuation characteristics from the real-time resistance data stream, wavelet transform is an effective tool.

[0103] It should be noted that by decomposing a signal into different frequency components, wavelet transform can amplify weak fluctuations at the micron scale.

[0104] For example, in a flexible electronic skin, when a certain line segment is slightly pressed, its resistance may increase slightly from 10 ohms to 10.2 ohms. Wavelet transform separates and enhances this subtle change, generating a clear fluctuation signal. The advantage of this signal amplification process is that even if the initial change is extremely small, it can be accurately captured and used for further analysis. When calculating the resistance change value based on the enhanced resistance fluctuation signal, the difference method can intuitively reflect the dynamic changes of each line segment.

[0105] Specifically, assume that the resistance value of a certain line segment changes from 10 ohms to 10.5 ohms within 1 second. The difference method calculates the change value as 0.5 ohms. By repeating this process for all line segments, a resistance change data set is obtained. This method is simple and efficient, can quickly quantify the change trend, and provides data support for subsequent judgment of the fluctuation value threshold. If the fluctuation value in the resistance change data set exceeds the preset fluctuation value threshold, for example, the set threshold is 0.3 ohms, then the signal of the over-standard line segment needs to be calibrated again.

[0106] In a possible implementation, the change value of a certain line segment is 0.5 ohms, exceeding the fluctuation value threshold. The least squares method is used to fit the amplified signal of this line segment to eliminate noise interference. Finally, the calibrated change value is adjusted to 0.45 ohms. This calibration can improve the accuracy of the data, avoid misjudgment caused by noise, and enhance the reliability of subsequent classification. When using the K-means clustering algorithm to classify the calibrated resistance change data, the line segments can be divided into three categories: stable, slightly fluctuating, and significantly fluctuating.

[0107] For example, the change values in the range of 0 - 0.1 ohms are classified as the stable category, 0.1 - 0.3 ohms as the slightly fluctuating category, and those exceeding 0.3 ohms as the significantly fluctuating category.

[0108] In one embodiment, among 10 line segments in a certain area, 8 belong to the stable category and 2 belong to the significantly fluctuating category. This classification helps to quickly identify abnormal line segments and lays a foundation for generating subsequent distribution maps. When generating a preliminary distribution map through the resistance change category, the change trend of each line segment can be intuitively displayed.

[0109] Preferably, assume that there are 100 line segments in a certain area of the electronic skin. The distribution map shows that the line segments near the edge are mostly in the significantly fluctuating category, and the central area is mostly in the stable category. This reflects the trend that the edge is more vulnerable to external forces and provides a basis for optimizing the design. When smoothing the distribution map using the bilinear interpolation method according to the resistance change trend.

[0110] It can be understood that this method makes the transition of the distribution map more natural through the weighted average of neighboring points.

[0111] For example, the change value of a certain line segment is 0.45 ohms, the adjacent line segment is 0.1 ohms, and the smoothed value after interpolation is about 0.3 ohms. The optimized distribution map is not only beautiful but also can more accurately reflect the overall change law, which helps to discover potential problems and improve the detection accuracy.

[0112] S2. Extract the abnormal fluctuation region from the resistance change distribution map, analyze the signal frequency characteristics using Fourier transform, judge whether there is a periodic change caused by a micro defect, and determine the coordinate of the potential physical damage position.

[0113] Optionally, this step further includes:

[0114] Step S21. Obtain the resistance change data through the distribution map and extract the abnormal fluctuation region.

[0115] Step S22. Use the fast Fourier transform (FFT) algorithm on the abnormal fluctuation region to obtain the signal frequency characteristics.

[0116] Step S23. Analyze the periodic change for the signal frequency characteristics and judge whether there is a regular fluctuation.

[0117] Step S24. If the periodic change exceeds the preset periodic change threshold, it is determined that there is a micro defect.

[0118] Step S25. Obtain the position coordinate of the physical damage through the signal frequency corresponding to the micro defect.

[0119] Step S26. Use the bilinear interpolation method for coordinate mapping to determine the specific position of the physical damage in the distribution map.

[0120] Step S27. Judge the boundary of the damage range by comparing the position coordinate with the resistance change data.

[0121] Exemplarily, after obtaining the resistance change data through the distribution map, extracting the abnormal fluctuation region is the key starting point for subsequent analysis.

[0122] It can be understood that the distribution map shows the resistance change of each line segment in the flexible printed circuit harness, and the abnormal fluctuation region often means potential problems.

[0123] For example, in the detection of a flexible electronic skin, the resistance change value of a certain area suddenly jumps from an average of 0.2 ohms to 0.8 ohms, which may indicate that there is an abnormality in this area. When extracting, a dynamic range can be set, for example, the area deviated more than 2 times from the mean value is marked as abnormal. This method can quickly lock the parts that need attention. When using the fast Fourier transform algorithm on the abnormal fluctuation region, the purpose is to convert the time-domain signal into a frequency-domain signal and extract the frequency characteristics.

[0124] Exemplarily, assume that the resistance value of a certain abnormal area shows irregular fluctuations within 1 second, and the fast Fourier transform can decompose the frequency components corresponding to these fluctuations.

[0125] Specifically, after the transformation, a significant 5 Hz frequency peak may be found, indicating that the fluctuations have a certain periodicity. The extraction of this frequency feature provides a quantitative basis for subsequent judgments. When analyzing the periodic changes of the signal frequency characteristics, it is necessary to determine whether the fluctuations are regular.

[0126] In a possible implementation, if the 5 Hz frequency peak continuously appears and the amplitude is stable, it can be considered that there are regular fluctuations. On the contrary, if the frequency distribution is scattered, for example, random peaks of 3 Hz, 7 Hz, and 10 Hz appear simultaneously, it may be just noise interference. This analysis can effectively distinguish between true periodic signals and disordered perturbations. If the periodic changes exceed the preset periodic change threshold, it can be inferred that there are micro defects.

[0127] Preferably, assume that the periodic change threshold is set to a frequency amplitude exceeding 0.5 units, and the 5 Hz peak in a certain area reaches 0.7 units, exceeding the periodic change threshold. At this time, it can be determined that there may be defects such as micro cracks or material fatigue in this area. This judgment method improves the sensitivity of defect detection. When obtaining the physical damage position coordinates through the signal frequency corresponding to the micro defects, the frequency feature plays a positioning role.

[0128] For example, the 5 Hz periodic fluctuations may correspond to the mechanical stress concentration area of a certain specific line segment.

[0129] In an embodiment, through a pre-established frequency and position mapping table, 5 Hz is associated with the 25th line segment in the distribution map. This method utilizes the correspondence between signal characteristics and physical positions. When using the bilinear interpolation method for coordinate mapping, the position of the damage in the distribution map can be determined more precisely.

[0130] Specifically, assume that the coordinates of the 25th line segment are x = 5, y = 3, and the frequency characteristics of the adjacent line segments are weak. After interpolation, the damage center coordinates may be x = 5.2, y = 3.1. This smooth mapping can refine the positioning accuracy and avoid position ambiguity. When judging the damage range boundary by comparing the position coordinates with the resistance change data, the original data can be combined for verification.

[0131] In an embodiment, the resistance change value of the line segment near the coordinates x = 5.2, y = 3.1 is 0.8 ohms, and it gradually drops below 0.3 ohms at the boundary. By comparison, the damage range can be delimited as an area centered on these coordinates with a radius of about 2 line segment units. This boundary judgment helps to comprehensively grasp the distribution of the damage.

[0132] It should be noted that each of the above steps is closely linked. From the extraction of the abnormal area to the determination of the boundary, a complete analysis chain is formed. The implementation of each technical theme closely revolves around the detection requirements of the flexible printed circuit harness, ensuring the rigor and practicality of the logic.

[0133] Optionally, in step S27, by comparing the position coordinates with the resistance change data to determine the boundary of the damage range, it further includes:

[0134] Step S271, obtaining the resistance change data through the position coordinates and judging the boundary extraction result.

[0135] Step S272, extracting features from the resistance change data and performing dimensionality reduction using the principal component analysis algorithm to obtain the change analysis features.

[0136] Step S273, calculating the coordinate offset according to the change analysis features and determining the position mapping relationship.

[0137] Step S274, extracting the boundary point set based on the position mapping relationship and judging the boundary of the damage range.

[0138] Step S275, if the boundary point set exceeds the pre-set boundary point set threshold, then determine the range determination result.

[0139] Step S276, according to the range determination result, using bilinear interpolation to interpolate the boundary points to obtain the damage range boundary coordinates.

[0140] Step S277, comparing the damage range boundary coordinates with the resistance change data to obtain the boundary judgment basis.

[0141] Exemplarily, when obtaining the resistance change data through the position coordinates and judging the boundary extraction result, it can be understood that the core of this process is to start from the known coordinate points and analyze in combination with the distribution of the resistance change.

[0142] In a possible implementation manner, assume that a certain position coordinate is x = 5, y = 3, and the corresponding resistance change value is 2.1 milliohms. By comparing with the resistance values of the surrounding coordinates, such as 1.8 milliohms at x = 4, y = 3 and 2.3 milliohms at x = 6, y = 3, it can be initially judged that this point may be near the boundary. This method relies on the continuity of the data and can quickly locate the abnormal area.

[0143] Specifically, when extracting features from the resistance change data and performing dimensionality reduction using the principal component analysis algorithm, the purpose is to simplify the high-dimensional data into a set of key features.

[0144] Exemplarily, assume that the original data contains multi-dimensional information such as resistance values, time, position, etc. Principal component analysis can transform these variables into a few principal components. For example, the first principal component reflects the overall change trend, and the second principal component highlights local fluctuations. This way of dimensionality reduction helps to reduce redundant information and facilitates subsequent analysis.

[0145] In one embodiment, when calculating the coordinate offset and determining the position mapping relationship according to the change analysis feature, the offset can be calculated through the distribution of eigenvalues.

[0146] For example, the value of the first principal component changes from 0.5 to 0.8 in a certain area. Combining with the coordinate grid, it is speculated that the offset is about 0.3 units. The establishment of this mapping relationship provides a basis for subsequent boundary extraction.

[0147] Preferably, when extracting the boundary point set based on the position mapping relationship and judging the boundary of the damage range, it can be screened by setting the boundary point set threshold.

[0148] For example, points with an offset greater than 0.2 are classified into the boundary point set. Suppose 10 boundary points are extracted in a certain area, distributed between x = 4 to 6 and y = 2 to 4. This indicates that the damage range may be concentrated within this grid. This method improves the accuracy of boundary recognition.

[0149] When determining the range result if the boundary point set exceeds the pre-set boundary point set threshold, it should be noted that the selection of the boundary point set threshold is crucial.

[0150] For example, if the preset upper limit of the boundary point set is 8, and 10 are actually extracted, it is confirmed that the range has exceeded the expectation. This determination method can effectively distinguish normal fluctuations from abnormal damages.

[0151] When obtaining the boundary coordinates by interpolating the boundary points using the bilinear interpolation method according to the range determination result, in one possible implementation, assume that interpolation is required between the boundary points x = 4, y = 2 and x = 6, y = 4. Combining with the resistance change values of 2.0 and 2.5 milliohms respectively, the value of the intermediate point x = 5, y = 3 can be estimated to be 2.25 milliohms, and the corresponding accurate coordinates are generated therefrom. This interpolation method can refine the boundary description.

[0152] For example, when obtaining the boundary judgment basis by comparing the boundary coordinates of the damage range with the resistance change data, the resistance value of the boundary coordinate x = 5, y = 3 is 2.25 milliohms, which is significantly different from the 1.5 milliohms of the surrounding non-boundary points, indicating that the boundary division is reasonable. This comparison verifies the reliability of the range determination and provides data support for subsequent analysis.

[0153] It can be understood that the implementation of each of the above steps focuses on the analysis of the relationship between resistance change and position, progressing step by step. From feature extraction to boundary interpolation, each step provides support for the final determination of the damage range. This method can improve the positioning accuracy and analysis efficiency in practical applications.

[0154] S3. For the physical damage position coordinates, deploy an infrared thermal imager within the full length range of the wire harness to obtain the temperature distribution data on the surface of the wire harness during the production process. Through the spatial correspondence relationship between the thermal anomaly points and the resistance anomaly points, obtain the preliminary positioning set of hidden defects.

[0155] Optionally, this step further includes:

[0156] Step S31. Deploy an infrared thermal imager along the full length of the wire harness to obtain the temperature distribution data on the surface of the wire harness during the production process, and generate a temperature distribution image.

[0157] Step S32. Extract the temperature values in the temperature distribution image, use a preset temperature threshold to judge the abnormal area, and determine the set of thermal anomaly points.

[0158] Step S33. Detect the resistance distribution within the full length range of the wire harness through a resistance measuring device, extract the resistance anomaly point data, and generate a set of resistance anomaly points.

[0159] Step S34. Perform a spatial comparison between the set of thermal anomaly points and the set of resistance anomaly points to obtain the set of overlapping areas.

[0160] Step S35. In the set of overlapping areas, use the K-means clustering algorithm to perform a preliminary positioning of the defects and generate a set of defect positions.

[0161] Step S36. Combine the temperature distribution image and the resistance distribution data to perform a spatial feature analysis on the areas in the set of defect positions, and generate a corrected set of positions.

[0162] Step S37. In the corrected set of positions, use the support vector machine algorithm to classify the defects and determine the final positioning result of the hidden defects.

[0163] Exemplarily, when deploying an infrared thermal imager along the full length of the wire harness, it can be understood that this method utilizes thermal imaging technology to capture the subtle changes in the surface temperature of the wire harness.

[0164] Exemplarily, on the production line of a flexible printed circuit harness, an infrared thermal imager is installed above the conveyor belt, covering the entire length range of the harness. Through real-time scanning, the acquired data reflects the temperature conditions of each part of the harness. For example, the surface temperature of a certain section of the harness is 35 degrees, while another section suddenly rises to 45 degrees. The generation of this temperature distribution image provides intuitive basic data for subsequent analysis. When extracting temperature values from the temperature distribution image, a reasonable temperature threshold needs to be set to determine the abnormal area.

[0165] For example, assume the normal temperature range is 30 to 40 degrees, and the temperature threshold is set such that areas exceeding 40 degrees are marked as abnormal.

[0166] In one embodiment, an image of a certain harness shows that the temperature in the area near the edge reaches 43 degrees, which is significantly high, and these points are classified into the set of thermal abnormal points. This method can quickly screen out potential problem areas and improve the detection efficiency. When detecting the harness with a resistance measurement device, the purpose is to obtain specific data on the resistance distribution.

[0167] Preferably, use a high-precision multimeter to measure the resistance of the harness section by section and record the resistance value of each section.

[0168] For example, the resistance value of a certain section jumps from the normal 0.3 ohms to 0.9 ohms, and such abnormal points are extracted to form a set of resistance abnormal points. The independence of this data provides a reliable basis for subsequent comparison. When making a spatial comparison between the set of thermal abnormal points and the set of resistance abnormal points, the overlap of the problem areas can be found.

[0169] Specifically, assume that the thermal abnormal points are concentrated in the 10th to 15th wire segments, and the resistance abnormal points also coincide in this range. Then these overlapping areas may hide defects. This comparison method utilizes the complementarity of multi-source data and enhances the accuracy of positioning. When using the K-means clustering algorithm to preliminarily locate the overlapping area, the algorithm will group the abnormal points according to the spatial distribution.

[0170] Exemplarily, assume there are 20 abnormal points in the overlapping area, the K value is set to 3, and after clustering, it is divided into three groups, which are concentrated near the 12th, 14th, and 16th wire segments respectively. This grouping provides a structured result for the preliminary positioning of defects and facilitates subsequent analysis. When conducting spatial feature analysis by combining the temperature distribution image and the resistance distribution data, the positioning result can be further corrected.

[0171] In a possible implementation, the 12th line segment after clustering shows a temperature of 42 degrees and a resistance of 0.8 ohms, while the temperature and resistance in the adjacent area gradually return to normal. By comparing these features, the corrected positioning may lock the defect center at the 12.5th line segment. This refinement process improves the accuracy of positioning. When using the support vector machine algorithm to classify defects, the purpose is to distinguish the types of defects.

[0172] For example, when inputting the corrected positioning set data and combining the temperature and resistance features, the algorithm may classify a certain defect as "material fatigue" and another as "poor contact".

[0173] In one embodiment, the features of the 12.5th line segment are marked as material fatigue, and this result provides a clear basis for judgment for the final hidden defect positioning. This classification method can effectively guide subsequent maintenance decisions.

[0174] S4. Through the preliminary positioning set of hidden defects, call the pre-established deep learning model to perform fusion analysis on the infrared thermal imaging data and the resistance change data, judge the specific types of microcracks or poor contacts, and determine the defect classification result with high confidence.

[0175] Optionally, this step further includes:

[0176] Step S41. Through the preliminary positioning set of hidden defects, use an infrared thermal imager to obtain infrared thermal imaging data, and use a resistance tester to obtain resistance change data to determine the basis for preliminary analysis.

[0177] Step S42. From the infrared thermal imaging data and the resistance change data, adopt a convolutional neural network model, pre-train the model on the training set, input the two types of data, and obtain the feature extraction result.

[0178] Step S43. For the feature extraction result, use a support vector machine classifier to analyze the distribution characteristics of microcrack types and poor contacts, and judge the specific categories of defects.

[0179] Step S44. If the feature intensity of the microcrack type exceeds the preset feature intensity threshold, then adjust the weight parameters through the convolutional neural network model to determine the microcrack classification output.

[0180] Step S45. If the distribution characteristics of the poor contact are consistent with the resistance change data, then input the infrared thermal imaging data into the support vector machine classifier to obtain the poor contact classification output.

[0181] Step S46. According to the microcrack classification output and the poor contact classification output, use the weighted average method to fuse and analyze the process results to judge the final type of defect classification.

[0182] Step S47: Based on the final type of defect classification, call the confidence evaluation method based on probability to determine the confidence output of the classification result.

[0183] Exemplarily, when obtaining infrared thermal imaging data and resistance change data through the preliminary positioning set of hidden defects, it can be understood that this step lays the foundation for subsequent analysis.

[0184] Exemplarily, on a wire harness production line, an infrared thermal imager is placed at a fixed position. When scanning a certain section of the wire harness, the change trend of the temperature from 38 degrees to 42 degrees is recorded. At the same time, a resistance tester measures section by section and finds that the resistance in a certain area rises from 0.4 ohms to 0.7 ohms. These data respectively reflect the abnormal fluctuations of temperature and resistance, providing the original basis for the next step of processing.

[0185] In a possible implementation manner, when extracting features from infrared thermal imaging data and resistance change data, a convolutional neural network model is used to process multi-dimensional information. The model has pre-learned a large number of wire harness defect samples on the training set, such as the distribution shape of the temperature abnormal area and the amplitude of the resistance mutation.

[0186] Specifically, when inputting a thermal imaging map showing that the temperature of a certain section reaches 44 degrees and combining it with the resistance data of 0.8 ohms, the model extracts the edge features and change trends of the abnormal area through the convolutional layer. This method can convert complex data into an analyzable feature set, facilitating subsequent classification. When using a support vector machine classifier for the feature extraction results, the purpose is to distinguish micro-cracks and poor contacts.

[0187] Preferably, the classifier sets the temperature gradient feature and resistance jump feature of micro-cracks according to the training data.

[0188] For example, if the temperature of a certain section of the wire harness rises rapidly from 36 degrees to 43 degrees and the resistance changes from 0.3 ohms to 0.9 ohms, the classifier determines it as the micro-crack type. In another embodiment, if the temperature only fluctuates slightly to 41 degrees while the resistance changes frequently, it may be a manifestation of poor contact. This classification method helps to quickly identify the defect category. If the intensity of the micro-crack feature exceeds the feature intensity threshold, such as the temperature difference exceeds 5 degrees or the resistance change is greater than 0.5 ohms, the convolutional neural network will adjust the weight parameters.

[0189] In one embodiment, when the temperature of a certain wire harness area reaches 45 degrees and the resistance is 1.0 ohm, the model clearly classifies it as a micro-crack through multiple iterations of optimization. This dynamic adjustment improves the adaptability of the classification.

[0190] It should be noted that if the distribution characteristics of poor contact are consistent with the resistance change, such as the resistance fluctuates repeatedly between 0.6 and 0.8 ohms, the thermal imaging data will be input into the support vector machine classifier again.

[0191] In a possible implementation, the temperature of a certain wire harness section is stable at 40 degrees, but the resistance data is abnormal. The classifier combines two types of information to confirm that it is a poor contact. This complementary verification improves the credibility of the result. When fusing the results using the weighted average method according to the classification outputs of microcracks and poor contacts, the contributions of the two types of data can be balanced.

[0192] For example, the classification weight of microcracks is set to 0.6, and that of poor contacts is 0.4. The final fused result tends to the more significant defect type.

[0193] Specifically, a certain area is determined to be a defect dominated by microcracks after comprehensive analysis. This method enhances the comprehensiveness of the judgment. When evaluating the confidence based on the final type of defect classification, a probability-based method is invoked.

[0194] Exemplarily, a wire harness defect is classified as a microcrack, and the calculated confidence reaches 0.9, indicating that the result is highly reliable. In another embodiment, the confidence of poor contact is 0.85. This quantitative evaluation provides a clear basis for subsequent decision-making and effectively guides the repair or optimization process.

[0195] Optionally, step S46, according to the classification outputs of microcracks and poor contacts, using the weighted average method to fuse the analysis process results and judge the final type of defect classification, further includes:

[0196] Step S461, extract the feature vectors of the two types of data from the classification outputs of microcracks and poor contacts as the basis for preliminary fusion.

[0197] Step S462, for the extracted feature vectors, use the weighted average method to process the distribution characteristics of microcrack classification and poor contact to obtain the fused feature vectors.

[0198] Step S463, according to the fused feature vectors, adjust the parameters through a pre-established neural network model to judge the tendency of defect classification.

[0199] Step S464, according to the judgment result, obtain the preliminary type distribution of defect classification and determine the sorting result of type tendency.

[0200] Step S465, according to the sorting result, use a support vector machine classifier to process the output of the fusion process to obtain the adjusted type of defect classification.

[0201] Step S466, according to the adjusted type, fuse the analysis results again using the weighted average method to determine the final type of defect classification.

[0202] Step S467, according to the final type, extract the core features of the classification output from the result fusion to judge the integrity of business analysis.

[0203] In a possible implementation, when extracting feature vectors from the outputs of micro-crack classification and poor contact classification, the principal component analysis (PCA) method can be adopted.

[0204] Specifically, for micro-crack classification, features such as crack length, width, and depth can be extracted; for poor contact, features such as contact area and pressure distribution can be extracted. These features form the basis for preliminary fusion.

[0205] Exemplarily, assume that the output of micro-crack classification contains 3 feature vectors [0.8, 0.6, 0.4], and the output of poor contact classification contains 2 feature vectors [0.7, 0.5]. When using the weighted average method, weights 0.6 and 0.4 can be assigned to micro-cracks and poor contact respectively, resulting in a fused feature vector [0.48, 0.36, 0.24, 0.28, 0.2].

[0206] It should be noted that the pre-established neural network model can be a multi-layer perceptron (MLP). By adjusting the number of neurons in the hidden layer and the activation function, the tendency of defect classification can be effectively judged.

[0207] For example, using the ReLU activation function and two hidden layers (each layer containing 10 neurons) can improve the model's fitting ability for non-linear features.

[0208] In one embodiment, the preliminary type distribution of defect classification may be presented as: micro-crack type A (40%), micro-crack type B (30%), poor contact type C (20%), poor contact type D (10%). Based on this distribution, the sorting result of type tendency can be determined.

[0209] Preferably, the support vector machine classifier can adopt the radial basis function (RBF) kernel function. By adjusting the kernel function parameters and the penalty factor, the classification effect can be optimized.

[0210] For example, setting the kernel function parameter γ = 0.1 and the penalty factor C = 1 can avoid overfitting while ensuring classification accuracy.

[0211] It can be understood that the final defect classification type is obtained through multiple fusions and adjustments. This method can comprehensively consider various features and classification results, improving the accuracy and reliability of classification. Extracting the core features of the classification output from the result fusion can help judge the integrity of business analysis and ensure that no important defect features or classification bases are missed.

[0212] S5. According to the defect classification result, generate a defect distribution vector for the entire length of the wire harness, and combine it with the time stamp sequence recorded in the real-time monitoring system to obtain the dynamic change trend of each defect during the production process.

[0213] Optionally, this step further includes:

[0214] Step S51: Obtain the defect distribution data of the full length of the wire harness through the defect classification result, and generate a distribution vector representation.

[0215] Step S52: Extract the time stamp sequence from the real-time monitoring system, align the time stamp with the distribution vector according to time, and generate a time-defect distribution matrix.

[0216] Step S53: Based on the time-defect distribution matrix, calculate the change value of each defect at different time points, and generate a time change sequence.

[0217] Step S54: According to the time change sequence, adopt the moving average method to extract the preliminary dynamic trend.

[0218] Step S55: For the preliminary dynamic trend, set a sliding window with a fixed size, calculate the change rate within the window, and generate smooth trend data.

[0219] Step S56: If there are abnormal points in the smooth trend data, then use the mean filtering method to replace the value of the abnormal point with the mean value within the window, and generate a stable trend result.

[0220] Step S57: According to the stable trend result, using time as the independent variable and defect distribution as the dependent variable, adopt the linear regression algorithm to calculate the long-term change slope of each defect, and generate a trend direction value.

[0221] Step S58: According to the trend direction value, accumulate the slope values of each time point, generate an accumulated change slope sequence, and determine the dynamic change trend of each defect in the production process.

[0222] Exemplarily, obtain the defect distribution data of the full length of the wire harness through the defect classification result, and generate a distribution vector representation.

[0223] It can be understood that this step integrates the scattered defect information into an overall view.

[0224] For example, on a wire harness production line, a certain section of the wire harness is 10 meters long. Through detection, 3 defects are found, located at 2 meters, 5 meters, and 8 meters respectively, and the type of each defect is known. The distribution vector can record these positions and their corresponding defect categories, forming an ordered data structure for subsequent analysis. Extract the time stamp sequence from the real-time monitoring system, align the time stamp with the distribution vector according to time, and generate a time-defect distribution matrix.

[0225] Specifically, the monitoring equipment on the production line records data every 5 minutes.

[0226] Exemplarily, a certain defect first appears at 9:00 and intensifies at 9:05. After the time stamp sequence is corresponded to the defect position, a two-dimensional matrix is formed, with the horizontal axis being time and the vertical axis being the defect distribution. This matrix intuitively reflects the change of the defect over time. Based on the time-defect distribution matrix, calculate the change value of each defect at different time points to generate a time change sequence.

[0227] In a possible implementation, the resistance value at a certain defect position is 0.5 ohms at 9:00 and rises to 0.8 ohms at 9:05, with a change value of 0.3 ohms. By comparing point by point in time, a sequence is obtained, clearly showing the evolution process of the defect. This sequence provides a basis for dynamic analysis. According to the time change sequence, the moving average method is used to extract the preliminary dynamic trend.

[0228] Preferably, take the average value of 3 time points to smooth the data.

[0229] For example, the resistance change sequence of a certain defect is 0.5, 0.8, 0.6 ohms, and the moving average result is 0.63 ohms. This method can reduce noise interference and highlight the trend direction, helping to more accurately grasp the defect change law. For the preliminary dynamic trend, set a sliding window of a fixed size, calculate the change rate within the window, and generate smoothed trend data.

[0230] In one embodiment, the sliding window is set to 5 minutes, and the resistance change rate of a certain defect within the window drops from 0.06 ohms / minute to 0.02 ohms / minute. This smoothing process can better reflect the stability and volatility of the defect, improving the readability of the trend. If there are abnormal points in the smoothed trend data, the mean filtering method is used to replace the value of the abnormal point with the mean value within the window to generate a stable trend result.

[0231] It should be noted that at a certain time point, the resistance suddenly increases to 1.2 ohms, while the data before and after are both around 0.7 ohms. Replace the abnormal value with the window mean value of 0.7 ohms. This processing can eliminate accidental errors and make the trend more valuable for reference. According to the stable trend result, with time as the independent variable and the defect distribution as the dependent variable, use the linear regression algorithm to calculate the long-term change slope of each defect to generate a trend direction value.

[0232] For example, the resistance value of a certain defect slowly rises from 0.4 ohms to 0.9 ohms over time, and the slope shows a positive value, indicating that the defect is deteriorating. This quantitative result provides a basis for predicting the development of the defect. According to the trend direction value, accumulate the slope values of each time point to generate a cumulative change slope sequence, and determine the dynamic change trend of each defect during the production process.

[0233] In a possible implementation, the cumulative slope of a certain defect gradually increases from 0 to 0.5, indicating that its impact is gradually expanding. This cumulative analysis can help identify key defects and optimize production decisions.

[0234] S6. Extract the evolution characteristics of defects from the dynamic change trend, use the Kalman filter algorithm to predict and correct the defect positions, and determine the set of accurate defect coordinates within the full length of the wire harness for fast positioning requirements.

[0235] Optionally, this step further includes:

[0236] Step S61. Obtain defect evolution data from the dynamic change, use the principal component analysis method to extract key features, and obtain the set of evolution characteristics.

[0237] Step S62. Use the Kalman filter algorithm to filter the set of evolution characteristics to obtain the preliminarily predicted defect positions.

[0238] Step S63. For the preliminarily predicted defect positions, use the least squares method for correction to determine the optimized defect position data.

[0239] Step S64. Extract the position change law from the optimized defect position data through linear regression analysis to obtain the position distribution of the full length of the wire harness.

[0240] Step S65. If the position distribution exceeds the preset position distribution threshold, use the K-nearest neighbor algorithm to match the fast positioning requirements and obtain the adjusted set of defect positions.

[0241] Step S66. According to the position distribution within the full length of the wire harness and the adjusted set of defect positions, use the interpolation method to determine the set of accurate coordinates.

[0242] Step S67. Obtain the final defect positioning result through the mapping relationship between the coordinate set and the defect positions.

[0243] Exemplarily, after obtaining the defect evolution data from the dynamic change, the principal component analysis method can be used to extract key features.

[0244] For example, on a wire harness production line, the defect data of a certain section of the wire harness includes multi-dimensional information such as resistance value, position offset, and surface wear degree. Through principal component analysis, these complex multi-variable data can be reduced in dimension, the main information can be retained, and a simplified set of evolution characteristics can be generated.

[0245] Specifically, assume that the defect evolution data of a certain wire harness contains 3 dimensions. After analysis, it is found that the contribution rate of the resistance change to the overall defect trend reaches 70%, the position offset contributes 20%, and the surface wear only accounts for 10%. This method can highlight the key factors and facilitate subsequent processing.

[0246] In a possible implementation, after the evolution feature set is generated, the Kalman filter algorithm is used for filtering processing to obtain the preliminarily predicted defect position.

[0247] Exemplarily, the resistance value of a certain defect fluctuates at multiple time points, such as 0.6 ohm, 0.7 ohm, 0.65 ohm. The Kalman filter predicts that the defect position at the next time point may be 5.2 meters by fusing historical data and the current measurement value. This method can effectively reduce noise interference and improve the stability of the prediction. For the preliminarily predicted defect position, the least squares method can be used for correction to optimize the defect position data.

[0248] It should be noted that assuming the predicted position is 5.2 meters, but the actual detection data shows that the defect is closer to 5.0 meters. The least squares method adjusts the predicted value to 5.05 meters by fitting multiple groups of data. This correction can improve the accuracy of the position and lay a foundation for subsequent analysis. Through linear regression analysis, the position change law can be extracted from the optimized defect position data to obtain the position distribution of the entire length of the wire harness.

[0249] In one embodiment, the entire length of a certain wire harness is 10 meters, and the defect position slowly moves from 2 meters to 2.5 meters. Linear regression shows that the position increases linearly with time. The extraction of this law helps to understand the movement characteristics of the defect. If the position distribution exceeds the preset position distribution threshold, for example, the defect deviation exceeds 0.3 meters, the K-nearest neighbor algorithm is used to match the rapid positioning requirement to obtain the adjusted defect position set.

[0250] Preferably, based on the similar defect patterns in the historical data, the K-nearest neighbor algorithm finds the nearest 3 neighbors and comprehensively judges that the defect position should be adjusted to 2.4 meters. This matching method can quickly respond to abnormal situations. According to the position distribution within the entire length of the wire harness and the adjusted defect position set, the interpolation method can be used to determine the precise coordinate set.

[0251] For example, there is no intermediate data between the two points of 2.4 meters and 5.05 meters for the defect position. The interpolation method calculates through linear interpolation that there may be potential defect coordinates at 3.5 meters. This method can fill the data gap and improve the integrity of the positioning. Through the mapping relationship between the coordinate set and the defect position, the final defect positioning result is obtained.

[0252] Specifically, the defect coordinate set of a certain wire harness is 2.4 meters, 3.5 meters, 5.05 meters, corresponding one by one to the resistance values of 0.7 ohm, 0.6 ohm, 0.8 ohm, forming a complete positioning view. This mapping can intuitively display the defect distribution and provide a reliable basis for production adjustment.

[0253] S7. Generate a real-time monitoring report based on the precise set of defect coordinates. Combine the equipment operation parameters during the production process to determine whether the defects exceed the preset defect threshold range, and obtain the production process data that needs to be adjusted.

[0254] Optionally, this step further includes:

[0255] Step S71. Obtain coordinate data from the set of defect coordinates, use the K-means clustering algorithm to cluster the coordinate data, determine the distribution characteristics of the defects, and obtain the defect aggregation area.

[0256] Step S72. Extract spatial features from the defect aggregation area, match them with the equipment operation parameters, judge the correlation between the defects and the equipment operation status, and obtain a preliminary correlation result.

[0257] Step S73. According to the preliminary correlation result, obtain the preset defect threshold range, and judge whether the defects exceed the defect threshold range through comparative analysis to obtain the defect data exceeding the standard.

[0258] Step S74. Extract the corresponding production process from the defect data exceeding the standard, use linear regression to analyze the correlation between the production process data and the defects, and obtain the set of affected processes.

[0259] Step S75. Obtain the change trend of the operation parameters from the set of affected processes, judge which parameters exceed the normal range, and obtain the set of parameters to be adjusted.

[0260] Step S76. According to the set of parameters to be adjusted, generate an adjustment plan in combination with the process data, determine the parameter optimization direction, and obtain the production process adjustment data.

[0261] Step S77. Extract the optimized operation parameters from the production process adjustment data, and judge whether the defect distribution is improved through simulation verification to obtain the final monitoring result.

[0262] Exemplarily, after obtaining the coordinate data from the set of defect coordinates, the K-means clustering algorithm can be used to group these data to identify the distribution characteristics of the defects.

[0263] For example, on a wire harness production line, assuming the set of defect coordinates includes multiple points such as 2.4 meters, 3.5 meters, and 5.0 meters, the K-means algorithm calculates the distances between these points and divides them into two clusters, one concentrated in the 2-3 meter interval and the other concentrated near 5 meters. This clustering method can intuitively reflect the defect aggregation area and facilitate subsequent analysis.

[0264] In a possible implementation, when extracting spatial features from the defect aggregation area, the distance and density between defect points can be concerned.

[0265] Specifically, the distance between defects in the 2-3 meter area is small and the density is high, while the defects near 5 meters are more scattered. These spatial features are matched with equipment operating parameters, such as the temperature or speed of the equipment during operation, to determine the correlation.

[0266] For example, assuming that when the temperature rises, the number of defects in the 2-3 meter area increases, the preliminary correlation results show that temperature may be an influencing factor.

[0267] It should be noted that, based on the preliminary correlation results, a preset defect threshold range can be set, such as a defect density of no more than 5 / m. Through comparative analysis, if the density in the 2-3 meter area reaches 7 / m, it will be classified as excessive defect data. This judgment method helps to screen out abnormal areas that need to be focused on. When extracting the corresponding production links from the excessive defect data, it can be traced back to the specific process.

[0268] For example, defects in the 2-3 meter area may occur in the insulation coating link. Linear regression is used to analyze the relationship between the temperature, pressure and other data of this link and the number of defects to obtain the set of influencing links.

[0269] Preferably, the analysis found that for every 10 degrees increase in temperature, the number of defects increased by 2, indicating that temperature is the key influencing factor.

[0270] In one embodiment, the change trend of the operating parameters is obtained from the set of influencing links, such as the temperature rising from 50 degrees to 70 degrees, which exceeds the normal range of 60 degrees. It is thus determined that the parameter set to be adjusted is temperature control. Such trend analysis can lock in the specific parameters that need to be optimized. When generating an adjustment plan based on the parameter set to be adjusted, the optimization direction can be proposed in combination with the link data.

[0271] For example, the temperature is reduced from 70 degrees to 55 degrees, and the pressure is adjusted to a stable value to form the production process adjustment data. This scheme design provides clear guidance for subsequent optimization. After extracting the optimized operating parameters from the production process adjustment data, simulation is used to verify whether the defect distribution has improved.

[0272] For example, after the temperature dropped to 55 degrees, the defect density in the 2-3 meter area dropped from 7 / meter to 3 / meter, indicating that the adjustment was effective. This simulation verification can provide a reliable basis for the final monitoring results and optimize production quality.

[0273] S8, based on the production data, updates the sensor configuration of the real-time monitoring system, increases the sampling frequency for areas with insufficient detection accuracy, obtains optimized resistance change and temperature distribution data, and determines a new round of defect location results.

[0274] Optionally, this step also includes:

[0275] Step S81: Obtain the configuration data of the real-time monitoring system from the production process, generate sensor configuration update data, and extract preliminary optimization parameters.

[0276] Step S82: Update the sampling frequency of the monitoring system according to the sampling frequency adjustment value in the preliminary optimization parameters, obtain the adjusted resistance change data, and draw a resistance change trend graph.

[0277] Step S83: Combine the resistance change trend graph and the sampling frequency adjustment value to generate optimized temperature distribution data, and calculate the temperature distribution uniformity.

[0278] Step S84: If the temperature distribution uniformity is lower than the preset uniformity threshold, readjust the sensor configuration, obtain the updated monitoring data, and generate a preliminary defect location result.

[0279] Step S85: According to the preliminary defect location result, use the K-means algorithm to cluster the resistance change data and the temperature distribution data, extract the defect distribution characteristics after clustering, and generate the precise defect location.

[0280] Step S86: According to the precise defect location, adjust the parameter configuration of the real-time monitoring system, obtain the optimized system operation data, and calculate the system stability.

[0281] Step S87: If the system stability meets the preset conditions, use the support vector machine algorithm to verify the defect location result, generate the verified defect distribution data, and determine the final optimization result.

[0282] Exemplarily, after obtaining the configuration data of the real-time monitoring system from the production process, sensor configuration update data can be generated.

[0283] For example, on a wire harness production line, the real-time monitoring system may include multiple resistance sensors and temperature sensors, and its configuration data covers the sampling frequency and sensitivity.

[0284] Exemplarily, assuming the initial sampling frequency is 10 times per second, by analyzing the real-time data in the production process, it is found that the resistance changes rapidly in some areas, and the preliminary optimization parameters suggest increasing the frequency to 15 times per second. This adjustment can capture key changes more timely.

[0285] In a possible implementation, after updating the monitoring system according to the sampling frequency adjustment value, the adjusted resistance change data can be obtained and a trend graph can be drawn.

[0286] Specifically, after adjusting the frequency to 15 times per second, the resistance value of a certain section of the wire harness fluctuates from 2 ohms to 2.5 ohms within 1 minute, and the trend graph shows that the fluctuation period is shortened. In this way, the law of resistance change can be observed more clearly.

[0287] It should be noted that when generating temperature distribution data by combining the resistance change trend graph and the sampling frequency adjustment value, the uniformity can be calculated through the collaborative work between sensors.

[0288] For example, the trend graph shows that the area with frequent resistance fluctuations corresponds to a temperature increase. The temperature distribution data indicates that the temperature of a certain wire harness section rises from 50 degrees to 65 degrees, while another section remains stable at 55 degrees, resulting in a decrease in uniformity. If the preset uniformity threshold is that the temperature difference does not exceed 8 degrees, the current difference of 10 degrees is lower than the requirement.

[0289] In one embodiment, if the temperature distribution uniformity does not meet the standard, the sensor configuration is readjusted. For example, the number of sensors in a certain area is increased from 2 to 3. The updated monitoring data shows that the temperature difference is reduced to 6 degrees. The preliminary defect location result generated thereby may point to the area with abnormal resistance.

[0290] Preferably, the K-means algorithm is used to cluster the resistance change data and the temperature distribution data to extract features.

[0291] For example, the point with a resistance value of 2.5 ohms and a temperature of 65 degrees is grouped into one cluster, and it is found that the defects are concentrated within a certain 2-meter section, and the precise location is confirmed.

[0292] It can be understood that after adjusting the monitoring system parameters according to the precise defect location, for example, further increasing the sampling frequency to 20 times per second in this area, the system operation data shows that the fluctuation detection is more stable. When calculating the system stability, if the resistance change amplitude is controlled within 0.2 ohms within 10 consecutive minutes, the preset condition is met. This kind of adjustment can improve the monitoring accuracy.

[0293] In one embodiment, if the system stability meets the standard, the support vector machine algorithm is used to verify the defect location result.

[0294] For example, by inputting the clustered defect distribution data, the algorithm confirms that the abnormal points within the 2-meter section are effectively separated from the normal points in other areas, generating the verified defect distribution data. The final optimization result shows that the adjusted configuration makes the defect detection more accurate, facilitating subsequent production optimization.

[0295] For example, from multiple aspects, increasing the number of sensors can improve the data resolution, while increasing the sampling frequency enhances the real-time performance. The combination of the two makes the correlation between the temperature distribution and the resistance change clearer. This multi-dimensional support method can effectively lock the defect location and optimize the system operation, providing reliable support for production quality.

[0296] Optionally, step S86, according to the precise defect location, adjusting the parameter configuration of the real-time monitoring system, obtaining the optimized system operation data, and calculating the system stability, further includes:

[0297] Step S861: Generate defect distribution data based on the precise defect and location information, cluster the defect distribution data using the KMeans algorithm, extract the feature distribution, and obtain the preliminary classification result.

[0298] Step S862: For the preliminary classification result, obtain the dynamically updated data from the real-time monitoring, calculate the data change trend, and determine the change characteristics.

[0299] Step S863: Update the parameter configuration according to the change characteristics, obtain the adjusted operation data, and judge the data consistency.

[0300] Step S864: If the data consistency is lower than the preset consistency threshold, regenerate the optimized operation data to obtain the stable parameters.

[0301] Step S865: Adjust the monitoring system through the stable parameters, obtain the optimized dynamic data, and calculate the system response time.

[0302] Step S866: Verify the operation data according to the system response time, generate the verified distribution data, and determine the final feature distribution.

[0303] Step S867: If the final feature distribution meets the preset conditions, classify the distribution data using the support vector machine to obtain the classified operation state.

[0304] Exemplarily, generating defect distribution data based on the precise defect and location information is the basic link for optimizing the real-time monitoring system.

[0305] For example, in semiconductor production, the defect distribution data may come from the tiny cracks or impurity points on the wafer surface, and the generation of these data depends on the signal changes captured by the sensors.

[0306] In a possible implementation, the wafer can be scanned by high-precision sensors to record the resistance and temperature values of each area, generating a defect distribution data table containing coordinate information. This data table can intuitively reflect the spatial distribution characteristics of the defects and provide a basis for subsequent analysis. When clustering the defect distribution data using the KMeans algorithm, the aim is to group similar defect features.

[0307] For example, in the wafer production scenario, the defect points with similar resistance change amplitudes can be grouped into one category, and features such as "concentrated distribution" or "dispersed distribution" can be extracted.

[0308] Exemplarily, if the resistance value fluctuation in a certain area is between 10 - 15 milliohms, while in another area it is between 5 - 8 milliohms, KMeans can separate these two types of defect points. This classification helps to identify potential causes of defects, such as local overheating of equipment or material inhomogeneity. When obtaining dynamically updated data from real-time monitoring for the preliminary classification results, it can be understood that the system needs to continuously track the changes in the production process.

[0309] Specifically, sensor data can be collected every 5 seconds to observe whether the resistance value shows an upward or downward trend over time.

[0310] In one embodiment, if the resistance value of a certain defect area gradually rises from 10 milliohms to 12 milliohms, it may indicate that the temperature in this area is accumulating and further attention is needed. The determination of this change characteristic provides data support for subsequent adjustments. After updating the parameter configuration according to the change characteristic, obtaining the adjusted operation data and judging the consistency are the keys to ensuring the reliability of the system.

[0311] For example, if the sampling frequency of a certain area is increased, from once per second to three times per second, the system will generate denser data points.

[0312] Preferably, the consistency can be judged by comparing the fluctuation ranges of the new and old data. For example, if the fluctuation of the new data is controlled within ±2 milliohms, while the old data is ±5 milliohms, it indicates that the adjustment is effective. This consistency check can significantly improve the monitoring accuracy. If the consistency is lower than the preset consistency threshold, when regenerating the optimized operation data, it can be achieved by increasing the number of sensors or adjusting the sampling timing.

[0313] In one possible implementation, if the data fluctuation in a certain area still exceeds the standard, 2 additional sensors can be deployed in this area to form a multi-point monitoring network. This method can effectively make up for the deficiencies of a single sensor and ensure the stability of the data. After adjusting the monitoring system through stable parameters, calculating the system response time is to evaluate the optimization effect.

[0314] For example, in the adjusted system, if the delay from the sensor receiving the signal to outputting the data is shortened from 200 milliseconds to 150 milliseconds, it indicates that the system responds more quickly to defects. This quick response can promptly capture abnormal situations in production. When verifying the operation data according to the system response time and generating the verified distribution data.

[0315] It should be noted that the accuracy of the distribution characteristics can be confirmed by comparing historical data and new data.

[0316] For example, if it is found after verification that the defect distribution in a certain area changes from "random" to "concentrated along the edge", it may indicate an abnormality in the edge part of the production equipment. This data after verification provides a reliable basis for the final decision. When using a support vector machine to classify the distribution data, if the final feature distribution meets the preset conditions, the operating state can be divided into two categories: "normal" and "abnormal".

[0317] Specifically, if the temperature distribution uniformity in a certain area reaches more than 95% and the resistance fluctuation is less than 10 milliohms, it can be classified as "normal".

[0318] In one embodiment, after learning through historical samples, the support vector machine can accurately identify abnormal areas. This classification method helps to quickly locate the root cause of problems and improve production efficiency.

[0319] A method for detecting defects in a flexible printed circuit harness provided by the present invention collects data on the change in the resistance of the harness in real time through a high-precision sensor array, obtains temperature distribution information by combining infrared thermal imaging technology, uses Fourier transform and a deep learning model for data fusion analysis to achieve precise positioning and classification of minor defects. The present invention also uses the Kalman filter algorithm to dynamically predict the defect position and adaptively adjusts the sensor configuration according to the detection results to improve the detection accuracy. This real-time monitoring method of multi-source data fusion can effectively identify hidden defects in the harness production process, such as microcracks and poor contacts, etc., provides accurate data support for production quality control, helps to timely discover and solve potential problems, and improves the production efficiency and product reliability of the flexible printed circuit harness.

[0320] The above content is only an example and explanation of the structure of the present invention. Those skilled in the art of the present technology make various modifications or supplements or use similar methods to replace the specific embodiments described, as long as they do not deviate from the structure of the invention or exceed the scope defined by this claims, they should all belong to the protection scope of the present invention.

Claims

1. A method for detecting defects in a flexible printed circuit harness, characterized in that, The method includes: S1. Obtain the real-time resistance data of the flexible printed circuit wire harness during the production process. Collect the resistance change values of each line segment through a high-precision sensor array, perform signal amplification processing on micron-level fluctuations, and obtain a preliminary resistance change distribution map. This step also includes: Step S11. Collect the resistance data of each line segment through the sensor array, process micron-level changes using high-precision acquisition technology, and obtain a real-time resistance data stream. Step S12. Extract the resistance fluctuation characteristics from the real-time resistance data stream, perform signal amplification processing on micron-level changes using wavelet transform, and obtain an enhanced resistance fluctuation signal. Step S13. According to the enhanced resistance fluctuation signal, calculate the resistance change value of each line segment using the difference method to obtain a resistance change data set. Step S14. If the fluctuation value in the resistance change data set exceeds the preset fluctuation value threshold, perform secondary calibration on the signal amplification result of the corresponding line segment using the least squares method to obtain calibrated resistance change data. Step S15. Use the K-means clustering algorithm to classify the calibrated resistance change data and determine the resistance change category of each line segment. Step S16. Generate a preliminary resistance change distribution map based on the resistance change category, and obtain the resistance change trend of each line segment in the distribution map. Step S17. According to the resistance change trend, perform smoothing processing on the distribution map using the bilinear interpolation method to obtain an optimized resistance change distribution map. S2. Extract the abnormal fluctuation area from the resistance change distribution map, analyze the signal frequency characteristics using Fourier transform, determine whether there is a periodic change caused by a minor defect, and determine the coordinates of potential physical damage locations. S3. For the coordinates of the physical damage location, deploy an infrared thermal imager within the full length range of the wire harness to obtain the temperature distribution data on the surface of the wire harness during the production process. Through the spatial correspondence relationship between the thermal anomaly point and the resistance anomaly point, obtain a preliminary positioning set of hidden defects. S4. Through the preliminary positioning set of hidden defects, call a pre-established deep learning model to perform fusion analysis on the infrared thermal imaging data and the resistance change data, determine the specific type of microcrack or poor contact, and obtain a defect classification result with high confidence. S5. According to the defect classification result, generate a defect distribution vector for the full length of the wire harness, and combine the time stamp sequence recorded in the real-time monitoring system to obtain the dynamic change trend of each defect during the production process. S6. Extract the evolution characteristics of the defects from the dynamic change trend, use the Kalman filter algorithm to predict and correct the defect positions, and determine the precise defect coordinate set within the full length of the wire harness for fast positioning requirements. S7. Through the precise defect coordinate set, generate a real-time monitoring report, combine the equipment operation parameters during the production process, determine whether the defect exceeds the preset defect threshold range, and obtain the production link data that needs to be adjusted. S8. According to the production link data, update the sensor configuration scheme of the real-time monitoring system, increase the sampling frequency for areas with insufficient detection accuracy, obtain optimized resistance change and temperature distribution data, and determine a new round of defect positioning results.

2. The method according to claim 1, wherein In step S2, an abnormal fluctuation region is extracted from the resistance change distribution map, and the signal frequency characteristics are analyzed by Fourier transform to determine whether there is a periodic change caused by a micro defect, and the coordinates of the potential physical damage position are determined, including: Step S21, obtaining resistance change data through the distribution map and extracting the abnormal fluctuation region; Step S22, using the fast Fourier transform (FFT) algorithm from the abnormal fluctuation region to obtain the signal frequency characteristics; Step S23, analyzing the periodic change for the signal frequency characteristics to determine whether there is a regular fluctuation; Step S24, if the periodic change exceeds the preset periodic change threshold, it is determined that a micro defect exists; Step S25, obtaining the position coordinates of the physical damage through the signal frequency corresponding to the micro defect; Step S26, using the bilinear interpolation method for coordinate mapping to determine the specific position of the physical damage in the distribution map; Step S27, judging the damage range boundary by comparing the position coordinates with the resistance change data.

3. The method according to claim 2, characterized in that, In step S27, judging the damage range boundary by comparing the position coordinates with the resistance change data, including: Step S271, obtaining the resistance change data through the position coordinates and judging the boundary extraction result; Step S272, extracting features from the resistance change data and using the principal component analysis algorithm for dimensionality reduction to obtain the change analysis features; Step S273, calculating the coordinate offset according to the change analysis features to determine the position mapping relationship; Step S274, extracting the boundary point set based on the position mapping relationship and judging the damage range boundary; Step S275, if the boundary point set exceeds the preset boundary point set threshold, the range determination result is determined; Step S276, according to the range determination result, using the bilinear interpolation method to interpolate the boundary points to obtain the damage range boundary coordinates; Step S277, comparing the damage range boundary coordinates with the resistance change data to obtain the boundary judgment basis.

4. The method according to claim 1, wherein In step S4, through the preliminary positioning set of hidden defects, a pre-established deep learning model is called to perform fusion analysis on the infrared thermal imaging data and the resistance change data to judge the specific type of microcrack or poor contact, and a high-confidence defect classification result is determined, including: Step S41, through the preliminary positioning set of hidden defects, using an infrared thermal imager to obtain infrared thermal imaging data and using a resistance tester to obtain resistance change data to determine the preliminary analysis basis; Step S42, from the infrared thermal imaging data and the resistance change data, using a convolutional neural network model to pre-train the model on the training set and inputting the two types of data to obtain the feature extraction result; Step S43, for the feature extraction result, using a support vector machine classifier to analyze the distribution characteristics of microcrack types and poor contact to judge the specific category of the defect; Step S44, if the feature intensity of the microcrack type exceeds the preset feature intensity threshold, adjusting the weight parameters through the convolutional neural network model to determine the microcrack classification output; Step S45, if the distribution characteristics of the poor contact are consistent with the resistance change data, inputting the infrared thermal imaging data into the support vector machine classifier to obtain the poor contact classification output; Step S46: Based on the microcrack classification output and the poor contact classification output, use the weighted average method to fuse and analyze the process results, and judge the final type of defect classification; Step S47: Through the final type of defect classification, call the probability-based confidence evaluation method to determine the confidence output of the classification result.

5. The method according to claim 4, characterized in that, The above-mentioned step S46, based on the microcrack classification output and the poor contact classification output, uses the weighted average method to fuse and analyze the process results, and judge the final type of defect classification, including: Step S461: Extract the feature vectors of the two types of data from the classification outputs of microcracks and poor contacts as the basis for preliminary fusion; Step S462: For the extracted feature vectors, use the weighted average method to process the distribution characteristics of microcrack classification and poor contact to obtain the fused feature vectors; Step S463: According to the fused feature vectors, adjust the parameters through a pre-established neural network model to judge the tendency of defect classification; Step S464: According to the judgment result, obtain the preliminary type distribution of defect classification and determine the sorting result of type tendency; Step S465: According to the sorting result, use the support vector machine classifier to process the output of the fusion process to obtain the adjusted type of defect classification; Step S466: According to the adjusted type, fuse the analysis results again by the weighted average method to determine the final type of defect classification; Step S467: According to the final type, extract the core features of the classification output from the result fusion to judge the integrity of business analysis.

6. The method according to claim 1, characterized in that, The above-mentioned step S5: According to the defect classification result, generate a defect distribution vector for the full length of the wire harness, and combine it with the time stamp sequence recorded in the real-time monitoring system to obtain the dynamic change trend of each defect during the production process, including: Step S51: Obtain the defect distribution data for the full length of the wire harness through the defect classification result and generate a distribution vector representation; Step S52: Extract the time stamp sequence from the real-time monitoring system, align the time stamp with the distribution vector in time, and generate a time-defect distribution matrix; Step S53: Based on the time-defect distribution matrix, calculate the change values of each defect at different time points to generate a time change sequence; Step S54: According to the time change sequence, use the moving average method to extract the preliminary dynamic trend; Step S55: For the preliminary dynamic trend, set a sliding window with a fixed size, calculate the change rate within the window, and generate smooth trend data; Step S56: If there are abnormal points in the smooth trend data, use the mean filtering method to replace the values of the abnormal points with the mean value within the window to generate a stable trend result; Step S57: According to the stable trend result, use the linear regression algorithm with time as the independent variable and defect distribution as the dependent variable to calculate the long-term change slope of each defect to generate a trend direction value; Step S58: According to the trend direction value, accumulate the slope values at each time point to generate an accumulated change slope sequence, and determine the dynamic change trend of each defect during the production process.

7. The method according to claim 1, characterized in that, In step S6, the evolution characteristics of defects are extracted from the dynamic change trend, and the Kalman filtering algorithm is used to predict and correct the defect positions. For the rapid positioning requirement, an accurate defect coordinate set within the full length of the wire harness is determined, including: Step S61, obtaining defect evolution data from the dynamic change, and using the principal component analysis method to extract key features to obtain an evolution feature set; Step S62, filtering the evolution feature set using the Kalman filtering algorithm to obtain the preliminarily predicted defect positions; Step S63, correcting the preliminarily predicted defect positions using the least squares method to determine the optimized defect position data; Step S64, extracting the position change law from the optimized defect position data through linear regression analysis to obtain the position distribution of the full length of the wire harness; Step S65, if the position distribution exceeds the preset position distribution threshold, using the K-nearest neighbor algorithm to match the rapid positioning requirement to obtain an adjusted defect position set; Step S66, determining the accurate coordinate set using the interpolation method according to the position distribution within the full length of the wire harness and the adjusted defect position set; Step S67, obtaining the final defect positioning result through the mapping relationship between the coordinate set and the defect position.

8. The method according to claim 1, characterized in that In step S8, according to the production link data, the sensor configuration scheme of the real-time monitoring system is updated, the sampling frequency is increased for the area with insufficient detection accuracy, the optimized resistance change and temperature distribution data are obtained, and a new round of defect positioning results are determined, including: Step S81, obtaining the configuration data of the real-time monitoring system from the production link, generating sensor configuration update data, and extracting preliminary optimization parameters; Step S82, updating the sampling frequency of the monitoring system according to the sampling frequency adjustment value in the preliminary optimization parameters, obtaining the adjusted resistance change data, and drawing a resistance change trend graph; Step S83, combining the resistance change trend graph and the sampling frequency adjustment value to generate the optimized temperature distribution data and calculating the temperature distribution uniformity; Step S84, if the temperature distribution uniformity is lower than the uniformity threshold, readjusting the sensor configuration to obtain updated monitoring data and generating a preliminary defect positioning result; Step S85, clustering the resistance change data and the temperature distribution data using the K-means algorithm according to the preliminary defect positioning result, extracting the defect distribution characteristics after clustering, and generating accurate defect positions; Step S86, adjusting the parameter configuration of the real-time monitoring system according to the accurate defect positions, obtaining the optimized system operation data, and calculating the system stability; Step S87, if the system stability meets the preset conditions, using the support vector machine algorithm to verify the defect positioning result, generating the verified defect distribution data, and determining the final optimization result.

9. The method according to claim 8, characterized in that, In step S86, according to the accurate defect positions, the parameter configuration of the real-time monitoring system is adjusted, the optimized system operation data is obtained, and the system stability is calculated, including: Step S861, generating defect distribution data according to the accurate defects and position information, clustering the defect distribution data using the KMeans algorithm, and extracting the feature distribution to obtain a preliminary classification result; Step S862: For the preliminary classification results, obtain dynamically updated data from real-time monitoring, calculate the data change trend, and determine the change characteristics; Step S863: According to the change characteristics, update the parameter configuration, obtain the adjusted operation data, and judge the data consistency; Step S864: If the data consistency is lower than the consistency threshold, regenerate the optimized operation data to obtain the stable parameters; Step S865: Through the stable parameters, adjust the monitoring system, obtain the optimized dynamic data, and calculate the system response time; Step S866: According to the system response time, verify the operation data, generate the verified distribution data, and determine the final feature distribution; Step S867: If the final feature distribution meets the preset conditions, use the support vector machine to classify the distribution data to obtain the classified operation status.

Images