Automobile cushion sponge quality detection method based on artificial intelligence

Abstract

The invention discloses an automobile cushion sponge quality detection method based on artificial intelligence, and particularly relates to the field of quality detection.The method comprises the steps that sponge full-surface data is obtained and preprocessed, then detection units are divided according to double-layer grids, surface damage, stains and density uniformity are detected based on an artificial intelligence model, and preliminary screening is executed; if not, stopping the detection; after the primary screening is qualified, collecting internal structure data of the sponge, extracting defect characteristics, and carrying out a cyclic indentation test to obtain physical performance parameters; single-mode quality functions of the surface, the internal structure and the physical performance are constructed respectively, after function data are preprocessed and subjected to environment correction, primary filtering is performed firstly, then a D-S evidence theory is applied to perform multi-mode fusion decision so as to output a final quality judgment result, and meanwhile, detection data feedback optimization is combined to form a quality control closed loop, so that the quality control accuracy is improved. And intelligent and accurate detection of the quality of the automobile cushion sponge is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of quality detection, and more particularly to a quality detection method for automobile cushion sponge based on artificial intelligence. BACKGROUND

[0002] As a core component of automobile interior, the surface state, internal structural integrity and physical and mechanical properties of automobile cushion sponge are directly related to the driving comfort, product durability and use safety, and are an important object of quality control of automobile parts. At present, the quality detection of automobile cushion sponge mainly focuses on the surface appearance and basic physical properties, and visual detection equipment is often used to collect surface images to assist in identifying obvious defects such as damage and stains. At the same time, mechanical testing instruments are used to obtain physical performance parameters such as elastic recovery rate and compression stiffness through indentation test to evaluate the use performance. With the application of artificial intelligence technology in the field of industrial detection, some detection processes have begun to introduce algorithm models to automatically identify surface defect features and reduce human intervention. In the industrial production scene, the detection technology as a whole is developing towards dataization and automation, and through the integration of data from multiple detection equipment, the quantitative evaluation of key quality indicators of sponge is initially realized to adapt to the basic requirements of the automobile industry for the stability of parts quality.

[0003] However, it still has some disadvantages in actual use, such as:

[0004] 1. The surface appearance defects or single physical performance detection of multi-focus automobile cushion sponge does not effectively detect the internal structure of the sponge, cannot cover the three core quality dimensions of surface, internal structure and physical performance, and cannot fully reflect the actual quality state of the sponge, thus having the problem of one-sided detection dimension.

[0005] 2. It relies on single detection equipment or manual visual judgment, lacks a multi-device collaborative collection and artificial intelligence model analysis mechanism, and manual subjective judgment is easily affected by experience and fatigue, resulting in errors. Single device cannot consider the accurate identification of geometric features, texture information and mechanical parameters, and the detection automation degree and precision are low.

[0006] 3. No detection data correction mechanism is established for temperature, humidity and other environmental factors. Temperature fluctuations can cause errors in the judgment of the uniformity of sponge density, and humidity changes can interfere with the internal defect recognition results. The detection data is greatly affected by the environment, and the stability and reliability of the data under different environments cannot be guaranteed.

[0007] 4. The quality judgment method of the prior art is relatively simple, and is mainly based on single index threshold value for direct judgment, lacks grading decision logic, does not introduce evidence fusion theory to process the uncertainty of multi-dimensional data, and is prone to misjudgment and omission, thus lacking in quality judgment precision. SUMMARY

[0008] In order to overcome the above-mentioned defects of the prior art, the present application provides an artificial intelligence-based automobile cushion sponge quality detection method, which solves the problems raised in the above background art through the following scheme.

[0009] To achieve the above object, the present application provides the following technical scheme: an artificial intelligence-based automobile cushion sponge quality detection method, comprising:

[0010] S1: surface defect detection and preliminary screening: a multi-view scanning system is built to obtain all surface data of the cushion sponge, which is divided into detection units according to double-layer grids after pretreatment; surface defect detection is performed through an artificial intelligence model, and preliminary screening is performed;

[0011] S2: internal quality data acquisition: internal structure data of the sponge is acquired by a sensor array, and internal defect features are extracted; sponge cycle indentation process data is acquired by a mechanical testing device, and physical performance parameters are obtained;

[0012] S3: construction of single-mode quality function: a surface quality function is constructed based on the surface defect detection result of S1, and internal structure quality function and physical performance quality function are respectively constructed based on the internal defect features and physical performance parameters acquired by S2;

[0013] S4: multi-modal data fusion and quality determination: the data of the three single-mode quality functions is pretreated and environment-corrected, and through two-level fusion decision, the obviously unqualified samples are first filtered by primary rules, and then the remaining samples are subjected to multi-modal fusion decision by using evidence fusion theory, and the final quality determination result is output; the model and function parameters are updated regularly, and feedback optimization is performed in combination with the detection data.

[0014] Preferably, the double-layer grid divided detection unit comprises:

[0015] Three calibration images of different angles of each plane of the target sponge cushion are acquired, and sponge point cloud data acquired by scanning six planes are acquired, which are spliced into a complete sponge three-dimensional image model after pretreatment by an algorithm; and the sponge surface is equally divided into nine large areas, numbered 1-9 from left to right and from top to bottom; each large area is further divided into nine small areas, which are compound numbered, and the partition result is bound with the three-dimensional image model and saved as a detection template with number.

[0016] Preferably, the surface defect detection comprises:

[0017] Detecting each small area of the detection template separately by deploying an artificial intelligence model, starting from the smallest number of small areas, traversing the entire detection template; through the damage detection model, output whether the current small area has damage; if there is no damage, deploy the stain detection model: if the small area has a stain, mark it as 1, otherwise mark it as 0, and real-time accumulate the number of marks as 1; at the same time, deploy the density detection model, output the density uniformity index of each minimum area.

[0018] Preferably, the preliminary screening comprises:

[0019] Damage veto: for any small area detected with damage, stop subsequent detection, and determine that the sponge is surface unqualified;

[0020] Stain determination: after completing all cell detection, count the number of small areas marked as 1, if more than 3, determine that the stain detection is unqualified, and terminate the process;

[0021] Density determination: if the density uniformity index difference of any adjacent small area in a large area is not higher than 3% and the density uniformity index difference of the diagonal small area is not higher than 5%, the large area is qualified, and the average value of the density uniformity index of the nine minimum cells in the large area is calculated;

[0022] If the average density uniformity index difference of any adjacent large area is not higher than 3% and the average density uniformity index difference of the diagonal large area is not higher than 5%, the density of the sponge cushion is determined to be qualified.

[0023] Preferably, the internal quality data collection comprises:

[0024] Structural data collection: scan the entire sponge cushion through a pre-set device to generate a complete internal structure three-dimensional image of the sponge, automatically segment the internal defect area of the three-dimensional image through defect parameter extraction, and calculate the cross-sectional area of each defect area, summarize the total defect area, and record the total area of the detection area;

[0025] Performance parameter collection: perform ten cycles of compression test on the sponge, record the initial thickness, and pre-set experimental parameters, take the 10th cycle data, record the thickness after compression, the thickness after rebound, and calculate the elastic recovery rate based on the thickness after compression and the thickness after rebound, take the maximum compression force of the 10th cycle, and calculate the compression stiffness.

[0026] Preferably, the surface quality function comprises:

[0027] According to the detection results of the preliminary screening stage, two key parameters are extracted: one is to determine the stain influence factor according to the number of small areas exceeding the standard covered by the stain, and the factor decreases linearly with the increase of the number of areas exceeding the standard, and the factor is 1 when there is no area exceeding the standard; the second is to calculate the density uniformity index of each large area based on the detection data, and then take the average of all large areas, and based on a large number of selected sponge samples covering different stain degrees and density distributions, the weights of the two parameters are determined through multiple linear regression, and finally the two parameters are weighted and summed according to the weights to form the surface quality function.

[0028] Preferably, the internal structure quality function comprises:

[0029] Based on the total defect area and the total area of the detection area obtained by collecting the internal quality data, the proportion of the total defect area to the total area of the detection area is calculated, according to the nonlinear characteristics of the influence of defects on quality, an exponential decay model is selected as the basic framework, different defect area proportion samples are selected, the model coefficients are determined through curve fitting, and finally the internal structure quality function is formed.

[0030] Preferably, the physical performance quality function comprises:

[0031] First, according to the industry standard, the standard values of elastic recovery rate and compression stiffness are set, for the elastic recovery rate, the ratio of the actual value to the standard value is calculated, if the ratio exceeds 1, then take 1; for the compression stiffness, the deviation degree of the actual value from the standard value is calculated, the influence of the deviation on the score is quantified through an exponential function, the greater the deviation, the lower the score, and the calculation results of the two indicators are multiplied to form the physical performance quality function.

[0032] Preferably, the pre-processing and environmental correction comprise:

[0033] Pre-processing: using the box plot method to process the extreme values in the surface quality function, the internal structure quality function and the physical performance quality function: calculating the quartiles of each function data, defining the abnormal value threshold range, the data exceeding the range is determined as an abnormal value, which is directly excluded and the exclusion reason is recorded; and using linear interpolation method to supplement the missing values;

[0034] Environmental correction: taking 20-25℃ as the standard temperature interval, when the detection environment temperature is between 20-25℃, no correction is needed, if higher than 25℃ or lower than 20℃, the density uniformity index is adjusted downward or upward according to the respective fixed proportion, and the correction coefficient is determined by fitting the relationship curve between temperature and density determination error through testing 100 standard sponge samples in the range of 15-30℃.

[0035] Taking 50% humidity as a standard benchmark, the deviation degree of the actual humidity from 50% is calculated first, and then a correction coefficient is determined according to the deviation degree, and the correction coefficient is multiplied by the original internal structure quality function value to obtain the corrected internal structure quality function value.

[0036] Preferably, the secondary fusion decision comprises:

[0037] First, rapid screening is performed through primary filtering: if the surface quality function value is lower than 0.6 or the internal structure quality function value is lower than 0.5, it is directly determined as unqualified; if the physical performance quality function value is lower than 0.7, secondary detection is started, and if it still does not meet the standard, it is determined as unqualified, and only when all the three meet the standard, it enters the advanced fusion.

[0038] In the advanced fusion stage, based on the D-S evidence theory, taking qualified, unqualified and uncertain as the identification framework, the evidence weight is assigned to the surface, internal structure and physical performance three single-mode quality functions respectively, then the evidence is corrected according to the respective unreliable coefficients, and then the corrected evidence is combined and the conflict coefficient is calculated, finally if the weight of qualified in the comprehensive evidence is more than 0.7 and the weight of unqualified is less than 0.2, it is determined as finally qualified, if the weight of qualified in the comprehensive evidence is not more than 0.4 and the weight of unqualified is not less than 0.5, it is determined as finally unqualified, if the above conditions are not met, it is determined as uncertain, and manual reinspection is required.

[0039] The technical effects and advantages of the present application are as follows:

[0040] 1. Full-dimension accurate coverage detection: The scheme breaks through the limitation of traditional single-dimension detection, collects full-surface data of the sponge through a multi-angle scanning system, accurately detects surface damage, stains and density uniformity, extracts internal bubble, layering and other defect features through an ultrasonic sensor array, obtains physical performance parameters such as elastic recovery rate and compression stiffness through experiments, and fully covers three core quality dimensions of surface, internal structure and physical performance, and fully reflects the actual quality state of the automobile cushion sponge.

[0041] 2. Artificial intelligence and multi-device cooperation to improve detection automation and precision: multiple types of detection devices are integrated and cooperatively calibrated, and based on an artificial intelligence model, the model is deployed to an edge computing unit to realize full-process automatic detection; multi-device cooperation avoids single-device deviation, and the artificial intelligence model accurately identifies small defects, greatly reduces the error of manual subjective intervention, and adapts to the high-efficiency detection needs of industrialized production.

[0042] 3. Introduce environmental parameter correction mechanism to ensure the reliability of detection data: Establish special correction logic for environmental interference: Take 20-25℃ as the standard temperature interval and 50% as the standard humidity reference, fit and correct the density uniformity index affected by temperature and the internal structure quality data affected by humidity by testing 100 standard sponge samples in the 15-30℃ interval, effectively avoid the detection deviation caused by temperature and humidity fluctuation, and ensure the stability and reliability of the detection data in different environments.

[0043] 4. Improve the accuracy of quality judgment through secondary fusion decision: First, quickly remove the samples with obvious unqualified surface quality, internal structure and physical performance through primary filtering to reduce invalid calculation; then, based on D-S evidence theory, assign evidence weight to the three-mode quality data and correct it, accurately judge the quality state of the edge sample through conflict coefficient processing and evidence combination, effectively reduce the misjudgment and omission, and ensure the accuracy of the final quality judgment result. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 It is the overall structure schematic diagram of the present application.

[0045] Fig. 2 It is the S1-S3 flowchart of the present application.

[0046] Fig. 3 It is the S4 flowchart of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely in the embodiments of the present application combined with the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Reference Figs. 1-3 The AI-based automobile cushion sponge quality detection method shown in the figure comprises:

[0049] S1: Surface defect detection and preliminary screening: Build a multi-view scanning system to obtain the full surface data of the sponge, and divide the detection unit according to the double-layer grid after pretreatment; detect the surface damage, stains and density uniformity through the artificial intelligence model, perform preliminary screening: if there is damage or stain exceeding the standard, it is judged as unqualified and terminated, and if the density uniformity meets the standard, it goes to S2;

[0050] S2: Internal structure quality data acquisition: Collect the internal structure data of the sponge by using sensor array and extract the internal defect features; collect the cycle indentation process data of the sponge by using mechanical testing equipment and obtain the physical performance parameters;

[0051] S3: Constructing single-modal quality function: constructing surface quality function based on surface defect detection results of S1, and constructing internal structure quality function and physical performance quality function based on internal defect features and physical performance parameters collected by S2;

[0052] S4: Multimodal data fusion and quality judgment: preprocessing and environmental correction of three single-modal quality function data, filtering obvious unqualified samples through primary rules first, then performing multimodal fusion decision on remaining samples by evidence fusion theory, and outputting final quality judgment results; periodically updating model and function parameters, and combining detection data for feedback optimization.

[0053] The specific steps are as follows:

[0054] S1: Surface defect detection and preliminary screening

[0055] S101: Surface data acquisition: a six-scan system composed of an integrated industrial CCD camera (its resolution is ≥1920×1080) and a line laser scanner (its accuracy is 0.05mm) is used, and a temperature and humidity sensor is matched; through Ethernet connection with an edge computing unit, the transmission bandwidth is not less than 100Mbps, the camera distortion is corrected by using a standard calibration plate, and the laser scanner is calibrated by using a standard step block, to ensure that the measurement error is ≤0.03mm; 3 calibration images of different angles (rotated by 0°, 30° and 60°, respectively) of the target sponge cushion are collected for each face, and the sponge point cloud data obtained by six-scan of the sponge cushion is executed Statistical Outlier Removal algorithm for denoising to remove environmental interference points; and the Iterative Closest Point algorithm is used to splice the six-scan point cloud to generate a complete three-dimensional image model of the sponge;

[0056] S102: Region division: the surface of the sponge is equally divided into 9 large regions, numbered 1-9 from left to right and top to bottom; each large region is further divided into 9 small regions, and the composite number is {(11, 12,..., 19), (21, 22,..., 29),..., (91, 92,..., 99)}; the division result is bound with the three-dimensional model and saved as a detection template with numbers.

[0057] S103: Surface defect detection: 10000 sponge defect sample images are collected, including damage, stains and uneven density, and divided into training set, validation set and test set according to the ratio of 7:2:1;

[0058] The improved YOLOv8, CNN and Transformer models are trained by using the PyTorch framework, and the test set accuracy is optimized to not less than 95%, 94% and 93%, respectively.

[0059] Convert the trained model into ONNX format, deploy it to the inference engine of the edge computing unit, set the batch size to 4 and the inference frame rate to ≥20fps;

[0060] Call the region division detection template to divide each surface image of the sponge cushion into 81 small region images, and perform three-model detection on each small region image:

[0061] Deploy the improved YOLOv8 model to detect damage: output whether there is damage in the current small region, if there is damage, stop subsequent detection of the target cushion sponge; if there is no damage, deploy the CNN model to detect stains: if there are stains in the small region, mark it as 1, otherwise mark it as 0, and accumulate the number of marks in real time; at the same time, deploy the Transformer model to detect density: output the density uniformity index of each minimum region at present;

[0062] S104: Preliminary screening:

[0063] Damage veto: for any small region detected with damage, stop subsequent detection and determine that the sponge is not qualified for the surface;

[0064] Stain determination: after completing all cell detection, count the number of small regions marked as 1, if more than 3, determine that the stain detection is unqualified, and terminate the process;

[0065] Density determination: if the density uniformity index difference of any adjacent small region in a large region is not higher than 3% and the density uniformity index difference of the diagonal small region is not higher than 5%, the large region is qualified, and the average value of the density uniformity index of the 9 minimum cells in the large region is calculated;

[0066] If the average density uniformity index difference of any adjacent large region is not higher than 3% and the average density uniformity index difference of the diagonal large region is not higher than 5%, the density of the sponge cushion is qualified;

[0067] S2: Internal structure data acquisition

[0068] S201: Internal structure data acquisition: fix the sponge that passes the preliminary screening on the scanning platform, ensure that the sponge surface is in close contact with the ultrasonic sensor array without gaps; and start the sensor control software, set the parameters: working frequency 8MHz, scanning step 0.5mm, sampling rate 100MHz;

[0069] Tomographic image generation: control the sensor array to move along the length direction with a step of 0.5mm, collect ultrasonic echo signals once for each movement; process the echo signals to convert them into gray-scale tomographic images with a resolution of 0.1mm x 0.1mm; splice the tomographic images in the order of scanning position to generate a complete internal structure three-dimensional image of the sponge;

[0070] Defect parameter extraction: load the three-dimensional image using ITK-SNAP software, and automatically segment the internal defect area by region growing method: set the gray threshold and seed point, and the seed point is used to automatically select the area with the lowest gray value; calculate the cross-sectional area of each defect area, denoted as: , and the total defect area is summarized: , and the total area of the detection area is recorded , i is the serial number, and j is the total number of defect areas;

[0071] S202: Physical performance parameter acquisition: through the deployed calibrated displacement sensor and electronic universal testing machine, ten cycle indentation tests are performed on the sponge cushion, the sponge is placed in the center of the lower clamp of the testing machine, the height of the upper clamp is adjusted to make the upper clamp contact with the surface of the sponge, and the pre-pressure is set to 5N, and the initial thickness is recorded, denoted as: ; set the test parameters in the testing machine control software: compression amount is 40% , compression speed is 50mm / min, cycle number is 10 times, and data sampling interval is 0.1s; execute 10 times of cycle indentation test: the upper clamp is pressed to the compression amount threshold at the set speed, and then rises to the pre-pressure state after 1s; record the force-displacement curve of each cycle in real time, take the 10th cycle data, and record the thickness after compression and the thickness after rebound, respectively denoted as: 、 ; and based on the thickness after compression and the thickness after rebound, the elastic recovery rate is calculated, denoted as: R, ;

[0072] Compression stiffness K: take the maximum compression force of the 10th cycle , and the compression stiffness K is calculated, ;

[0073] S3: Constructing a single modal quality function: converting the original data of each detection link into a unified quantitative index, which is used to objectively describe the quality level of the sponge in the three dimensions of surface state, internal structure and physical performance, and the specific analysis is as follows:

[0074] S301: Constructing a surface quality function: comprehensively evaluating the stain distribution and density uniformity of the sponge surface, and quantifying the overall quality of the surface;

[0075] Stain influence factor: based on the marked results of stain determination in S1, the mathematical function is: , the value decreases linearly with the increase of the number of stained cells, reflecting the negative impact of stains on surface quality; when there is no stain, the maximum value 1 is taken, representing the optimal stain dimension;

[0076] Density uniformity index: based on the density uniformity index of each minimum area determined in step S1, the average value of the index of 9 small areas in each large area is calculated according to the grouping of large areas, and the average value of 9 large areas is taken as the density uniformity index of the whole, denoted as: ; The closer the value is to 1, the more uniform the density distribution of the sponge surface represents;

[0077] Surface quality function: , wherein , are the stain influence factor coefficient and the density uniformity index coefficient respectively, which are determined by the multi-source linear regression analysis of 1000 historical samples, and the values are 0.4 and 0.6 respectively;

[0078] S302: Construct internal structure quality function: based on the area ratio of internal defects, the integrity of the internal structure of the sponge is quantified, and the exponential function is used to accurately fit the nonlinear characteristics of the influence of internal defects on the mechanical properties of the sponge. The specific mathematical function is: , wherein is the total defect area ratio coefficient, and the correlation analysis is carried out on 50 sample data covering different defect area ratios with known internal defects. When the value is 0.8, the fitting degree of the function value and the actual mechanical performance attenuation trend is the highest;

[0079] S303: Construct physical performance quality function: comprehensively evaluate the elastic recovery ability and structural stiffness of the sponge, and quantify whether its mechanical performance meets the use requirements of automobile seat cushion. The specific analysis is as follows:

[0080] Based on the internal quality data collected in S2, the elastic recovery rate ratio and the compression stiffness deviation term are calculated. Elastic recovery rate and compression stiffness are the core indicators of sponge physical performance, and both of them need to meet the standard. The coupling is used for comprehensive evaluation of physical performance quality, and the specific mathematical function is as follows:

[0081]

[0082] , wherein is the industry standard elastic recovery rate, is the preset standard compression stiffness range;

[0083] S4: Multi-modal data fusion and quality judgment:

[0084] The quality functions of surface, internal structure and physical performance are standardized and fused with evidence fusion theory, which is the core link of finally outputting the comprehensive quality judgment result of sponge. Through the logic from data preprocessing to hierarchical decision, the transformation from single dimension quality evaluation to global quality judgment is realized, and the specific analysis is as follows:

[0085] S401: Function value preprocessing: eliminate abnormal interference in raw data, correct the influence of environmental factors on detection results, ensure the comparability and accuracy of the three quality functions, and provide reliable input for subsequent fusion;

[0086] Use box plot method to process extreme values in surface quality function, internal structure quality function and physical performance quality function: calculate the quartiles of each function data: Q1 (lower quartile, 25th percentile), Q3 (upper quartile, 75th percentile); Define the threshold range of outliers: , wherein (IQR) is the interquartile range.

[0087] Processing logic: data beyond this range is determined as an outlier, directly excluded and recorded for exclusion reason; Linear interpolation method is used to supplement the missing values;

[0088] Environmental parameter correction:

[0089] Temperature correction for density uniformity index: when the temperature deviates from the standard range of 0-25℃, the sponge will expand and contract with heat, which may lead to misjudgment of texture characteristics, especially when the temperature is greater than 25℃, the influence is more significant, and the correction formula is: Through the test of 100 standard sponge samples in the range of 15-30℃, the relationship curve between temperature and density determination error is fitted to determine the correction coefficient;

[0090] Humidity correction for internal structure quality function: when the humidity is greater than 50%, the sponge will absorb moisture, which will change the speed of ultrasonic wave propagation, resulting in overestimation of internal defect area; The correction formula is: ; Where T is the environmental temperature of the target sponge cushion measured by the temperature and humidity sensor, and H is the measured humidity.

[0091] S402: Secondary fusion decision: adopt hierarchical logic from primary filtering to advanced fusion, first quickly exclude obvious unqualified samples, then fine fusion judgment for edge samples, and balance efficiency and accuracy;

[0092] Primary filtering: based on the threshold of single modal quality function for preliminary screening:

[0093] Surface quality filtering: if the corrected , it is directly judged as unqualified, corresponding to serious surface stains or poor density uniformity, without further detection;

[0094] Internal structure filtering: if the corrected , it is directly judged as unqualified, corresponding to internal defect area ratio not less than 8%, and mechanical properties cannot meet the requirements;

[0095] Physical performance filtering: if , start secondary physical performance detection, if the secondary detection is not qualified , determine as unqualified; if , enter advanced fusion;

[0096] S403: D-S evidence theory advanced fusion: for the primary filtered qualified samples, fuse the three quality functions by D-S evidence theory, solve the uncertainty problem of multi-source information, and output the comprehensive quality judgment:

[0097] D-S evidence theory advanced fusion is the core link of multi-dimensional information comprehensive decision for the primary filtered qualified samples. Through the logic of evidence modeling, conflict processing and combination judgment, the quality information of surface, internal structure and physical performance of three modalities is converted into the final quality conclusion. This process fully considers the uncertainty and reliability difference of each modality, solves the one-sidedness problem of single dimension judgment, and its specific process is as follows:

[0098] A1: Evidence modeling: identify framework and single modality evidence distribution

[0099] Construct the identification framework: , wherein, qualified means that the sponge cushion meets the automobile cushion use standard in surface quality, internal structure and physical performance three dimensions; unqualified means that at least one dimension does not meet the standard and cannot be used for production; uncertain means that the existing data is not enough to determine qualified or unqualified, and further verification is needed.

[0100] Single modality evidence distribution rule: convert the three single modality quality function values obtained in S3 into the confidence degree of the three elements in the identification framework, and the distribution rule is: for any modality function value: ; wherein F represents S, I and P; its distribution logic is: the qualified confidence degree is directly equal to the function value; the remaining confidence degree , 90% is allocated to unqualified to strengthen the sensitivity to defects, and 10% is allocated to uncertain to reserve the fault tolerance space of detection error.

[0101] A2: Evidence discount correction: weight adjustment based on modality reliability, the inherent reliability of different detection modalities is different, which needs to be corrected by unreliable coefficient to ensure that the weight of high reliability modality is more significant;

[0102] Unreliable coefficient setting: surface detection : ; internal detection : ; physical performance detection : ;

[0103] Its correction formula is: for the evidence of the tth modality , the corrected evidence is For: ; where A is the elements in the evidence;

[0104] Revision logic: keep the original evidence of the proportion; distribute the confidence of the proportion equally to the three elements.

[0105] A3: Evidence combination: conflict handling and multi-source information fusion Combine the three revised evidences through the Dempster rule to obtain the comprehensive confidence;

[0106] Calculate the conflict coefficient: quantify the degree of contradiction between evidences: ; where B, C, and D are the propositions of the three evidences; : indicates complete agreement between evidences; indicates complete conflict;

[0107] Combination rule: low conflict : use the Dempster orthogonal sum rule to superimpose the consistent confidence, and distribute the conflict part in proportion: , , is the comprehensive confidence;

[0108] High conflict : use the Yager improved rule to include the conflict part into the uncertainty, and avoid misjudgment caused by conflict: , , .

[0109] A4: Final decision: decision output based on comprehensive confidence According to the combined , , , the following rules are used for decision:

[0110] Determination of qualified: need to meet: and ;

[0111] Determination of unqualified: if , or , directly determine unqualified;

[0112] Determination of uncertainty: except for the above two cases, it is determined to be uncertain, and manual re-inspection needs to be started.

[0113] And for every 1000 batches of defect samples, the detection model is fine-tuned by transfer learning, and the gradient descent method is used to recalibrate the , , ​ , and theoretical evidence .

[0114] The present application builds a multi-device scanning system, adopts double-layer grid division detection unit, detects the damage, stain and density uniformity of the sponge surface by minimum unit cell first, collects the internal structure data and physical performance parameters after screening qualified samples; then constructs the single-mode quality function of surface, interior and physical performance based on the above data, after data preprocessing and environment correction, carries out multi-modal fusion decision through D-S evidence theory, finally determines whether the quality is qualified, and forms a quality control closed loop through regular model updating and process feedback, realizes intelligent and accurate detection and optimization of sponge quality.

[0115] Secondly: the present application discloses only the structure involved in the present application, other structures can refer to the general design, and the same embodiment and different embodiments of the present application can be combined with each other under the condition of no conflict;

[0116] Finally: the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for detecting the quality of automotive seat cushion foam based on artificial intelligence, characterized in that, include: S1: Surface defect detection and preliminary screening: A multi-view scanning system was built to obtain all surface data of the cushion sponge. After preprocessing, the detection units were divided according to a double-layer grid. Surface defects are detected using an artificial intelligence model, and preliminary screening is performed. S2: Internal quality data acquisition: Sensor arrays are used to acquire data on the internal structure of the sponge and extract internal defect features; data on the cyclic indentation process of the sponge are acquired through mechanical testing equipment to obtain physical performance parameters; S3: Constructing a single-modal quality function: Based on the surface defect detection results of S1, a surface quality function is constructed. Based on the internal defect features and physical performance parameters collected in S2, an internal structure quality function and a physical performance quality function are constructed respectively. S4: Multimodal Data Fusion and Quality Judgment: Preprocess and environmental correction are performed on the three single-modal quality function data. Then, through a two-level fusion decision, obviously unqualified samples are first filtered out by primary rules. The evidence fusion theory is then used to perform multimodal fusion decision on the remaining samples, and the final quality judgment result is output. Regularly update the model and function parameters, and optimize based on the feedback from the detection data.

2. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The dual-layer mesh division and detection unit includes: Three calibrated images of each plane of the target sponge cushion from different angles were collected, along with sponge point cloud data obtained from scanning six planes. After preprocessing, the sponge was stitched together using an algorithm to form a complete three-dimensional image model. The sponge surface was then divided into nine large regions, numbered 1-9 from left to right and top to bottom. Each large region was further divided into nine smaller regions, which were then composite-numbered. The partitioning results were then bound to the three-dimensional image model and saved as a numbered detection template.

3. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The surface defect detection includes: The deployed artificial intelligence model detects each small area of ​​the detection template individually, starting from the smallest number of the small area and traversing the entire detection template; the damage detection model outputs whether the current small area is damaged; if there is no damage, the stain detection model is deployed: if there is a stain in the small area, it is marked as 1, otherwise it is marked as 0, and the number of marked as 1 is accumulated in real time; at the same time, the density detection model is deployed to output the density uniformity index of each current small area.

4. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The preliminary screening includes: Damage rejection: If any small area is found to be damaged, subsequent testing will be stopped, and the sponge will be deemed to have a surface defect. Stain detection: After all cells have been detected, count the number of small areas marked with 1. If the number exceeds 3, the stain detection is considered unqualified and the process is terminated. Density determination: If the density uniformity index difference between any adjacent small areas within a large area is not higher than 3% and the density uniformity index difference between diagonal small areas is not higher than 5%, then the large area is qualified, and the average density uniformity index of the 9 smallest cells within the large area is calculated. If the difference in the average density uniformity index of any adjacent large area is not higher than 3% and the difference in the average density uniformity index of the diagonal large area is not higher than 5%, then the density of the sponge cushion is deemed qualified.

5. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The internal quality data collection includes: Structural data acquisition: The entire sponge cushion is scanned using a preset device to generate a complete three-dimensional image of the sponge's internal structure. By extracting defect parameters, the internal defect areas are automatically segmented from the three-dimensional image, and the cross-sectional area of ​​each defect area is calculated. The total defect area is then summarized and the total area of ​​the detected area is recorded. Performance parameter acquisition: The sponge was subjected to ten cycles of indentation test. The initial thickness was recorded and the experimental parameters were preset. The data of the 10th cycle was taken, and the thickness after compression and the thickness after rebound were recorded. The elastic recovery rate was calculated based on the thickness after compression and the thickness after rebound. The maximum compression force of the 10th cycle was taken and the compression stiffness was calculated.

6. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The surface quality function includes: Based on the test results of the initial screening stage, two key parameters were extracted: First, the stain influence factor was determined according to the number of small areas exceeding the standard covered by stains. When there were areas exceeding the standard, the factor decreased linearly with the increase of the number of areas, and the factor was 1 when there were no areas exceeding the standard. Second, based on the test data, the density uniformity index of each large area was calculated, and then the average value of all large areas was taken. Based on a large number of selected sponge samples covering different stain levels and density distributions, the weights of the two parameters were determined by multiple linear regression. Finally, the two parameters were weighted and summed to form the surface quality function.

7. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The internal structural quality function includes: Based on the total defect area and the total area of ​​the detection area obtained from the internal quality data collection, the ratio of the total defect area to the total area of ​​the detection area is calculated. According to the nonlinear characteristics of the impact of defects on quality, the exponential decay model is selected as the basic framework. Samples with different defect area ratios are selected, and the model coefficients are determined by curve fitting. Finally, the internal structural quality function is formed.

8. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The physical performance quality function includes: First, set standard values ​​for elastic recovery rate and compressive stiffness according to industry standards. For elastic recovery rate, calculate the ratio of actual value to standard value. If the ratio exceeds 1, take 1. For compressive stiffness, calculate the degree of deviation between actual value and standard value. Quantify the impact of deviation on score through exponential function. The greater the deviation, the lower the score. Multiply the calculation results of the two indicators to form the physical performance quality function.

9. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The pretreatment and environmental correction include: Preprocessing: Box plot method is used to handle extreme values ​​in surface quality function, internal structure quality function, and physical performance quality function: quartiles of each function data are calculated, outlier threshold range is defined, data exceeding the range are identified as outliers, directly removed and the reason for removal is recorded; and linear interpolation is used to supplement missing values. Environmental correction: 20-25℃ is used as the standard temperature range. When the ambient temperature is between 20-25℃, no correction is required. If it is higher than 25℃ or lower than 20℃, the density uniformity index is adjusted downward or upward according to their respective fixed proportions. The correction coefficient is determined by testing 100 standard sponge samples in the range of 15-30℃, fitting the relationship curve between temperature and density judgment error, and determining the correction coefficient value. Using 50% humidity as the standard benchmark, first calculate the deviation of the actual humidity from 50%, then determine the correction factor based on the deviation, and multiply the correction factor by the original internal structure quality function value to obtain the corrected internal structure quality function value.

10. The method for detecting the quality of automotive seat cushion foam based on artificial intelligence according to claim 1, characterized in that: The secondary fusion decision includes: First, a preliminary screening is performed: if the surface quality function value is below 0.6 or the internal structure quality function value is below 0.5, it is directly judged as unqualified; if the physical performance quality function value is below 0.7, a secondary test is initiated, and if it still fails to meet the standard, it is judged as unqualified. Only when all three items meet the standard can it proceed to advanced fusion. The advanced fusion stage is based on the DS evidence theory, using qualified, unqualified, and uncertain as the identification framework. Evidence weights are assigned to the three single-modal quality functions of surface, internal, and physical performance, and then the evidence is corrected according to its respective unreliability coefficient. Subsequently, the corrected evidence is combined and the conflict coefficient is calculated. Finally, if the weight of qualified in the integrated evidence is greater than 0.7 and the weight of unqualified is less than 0.2, it is judged as qualified in the final stage. If the weight of qualified in the integrated evidence is not more than 0.4 and the weight of unqualified is not less than 0.5, it is judged as unqualified in the final stage. If none of the above conditions are met, it is judged as uncertain and manual re-inspection is required.

Images