A robot sorting matching mold temperature field linkage automobile injection quality inspection system

Abstract

The application discloses a kind of robot sorting matching mould temperature field linkage automobile injection quality inspection system, it is related to automobile injection quality inspection technical field, the module includes data acquisition module, resource division module and quality inspection identification module;Its technical key points are: by identifying target quality inspection interval, and judge whether to satisfy multidimensional space condition, if satisfying multidimensional space condition, establish correlation mapping, generate abnormal correlation score, and divide multiple time-frequency resource pool, by at least one sorting time sequence in first, second time-frequency resource pool is sent, to drive robot sorting;After determining the time-frequency resource in first time-frequency resource pool, select remaining time-frequency resource to execute quality inspection evaluation strategy: according to the preset temperature field linkage quality inspection model, determine the quality inspection defect type of each target quality inspection area in each target automobile injection product;The application improves the accuracy of injection product identification under the cooperation of preemptive resource scheduling.

Description

Technical Field

[0001] This invention relates to the field of automotive injection molding quality inspection technology, specifically to a robot sorting system with mold temperature field linkage for automotive injection molding quality inspection. Background Technology

[0002] With the rapid development of the automotive industry and precision manufacturing technology for auto parts, high-precision online quality inspection and high-speed robotic adaptive sorting of automotive injection molded parts have become core technology requirements on industrial automated production lines. In the mold opening and demolding stage of injection molding machines, the use of machine learning in conjunction with industrial robots for defect detection and automated sorting and removal of injection molded parts has been widely applied.

[0003] However, existing sorting and quality inspection technologies still have the following limitations: On the one hand, existing image acquisition devices use fixed shooting angles and focal lengths, lacking the ability to adaptively adjust to complex spatial conditions; fixed image acquisition modes are prone to edge defocusing, perspective distortion, or loss of detail, and the system cannot automatically switch the acquisition zoom mode according to the spatial condition verification results; at the same time, in this case, existing technologies rely on single-dimensional surface optical images, lacking linkage with the physical characteristics of the internal thermal field of the cavity. Although visual inspection algorithms have been introduced, these algorithms only perform defect analysis on the two-dimensional surface images after demolding, and cannot establish a spatial correlation mapping between surface image features and multi-point temperature field data of the cavity at the moment of mold opening. It is difficult to effectively identify internal hidden defects or thermal damage caused by local heat accumulation, resulting in insufficient robustness of subsequent quality inspection defect judgment. On the other hand, existing processing logic must wait for machine learning to complete complex quality inspection reasoning for all regions of the entire image and output the final defect label before generating control instructions to drive the robot to perform unloading or rejection actions. Within the extremely short mold opening and demolding cycle of the injection molding machine, this can easily cause the robot's sorting response to lag. For example, in the inspection of large-sized or structurally complex injection molded parts such as car dashboards, door panels, or bumpers, the image usually contains dozens of clips, mounting holes, and exterior trim areas. The reasoning time for all regions to be calculated is often hundreds of milliseconds or even several seconds. It is impossible to achieve decoupling and preemptive scheduling of algorithm computing resources and robot mathematical control time-frequency resources. It is extremely easy for improper gripping timing of the robotic arm to cause damage to the injection molded parts or production line stoppage, reducing the yield rate and response speed of the production line. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a robot sorting system with mold temperature field linkage for automotive injection molding quality inspection. The system uses a data acquisition module to synchronously collect multi-point temperature field data and injection images of the mold cavity during mold opening. A resource allocation module performs decreasing priority checks based on multi-dimensional spatial conditions of geometric, material, and process constraints. If the conditions are not met, the system adaptively switches to a zoom mode; if they are met, a temperature field correlation mapping is established and a time-frequency resource pool is allocated, enabling preemptive scheduling of the sorting sequence. A quality inspection identification module uses a temperature field linkage quality inspection model to splice and fuse fluctuation characteristics with the temporal characteristics of abnormal temperature field environments. An integrated feature encoder and classification encoder accurately output the final quality inspection defect type, thus solving the problems mentioned in the background technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This application provides a robot sorting and mold temperature field linkage quality inspection system for automotive injection molding, the system comprising: The data acquisition module acquires multi-point temperature field data of the mold cavity and the injection molding image of the car to be inspected in the mold-open state; The resource partitioning module identifies the target quality inspection area based on the automotive injection molding image to be inspected, and determines whether the target quality inspection area meets the multi-dimensional spatial conditions. If the multi-dimensional spatial conditions are met, it establishes an association mapping based on the temperature field data, generates an anomaly association score, and divides the area into multiple time-frequency resource pools. It sends at least one sorting sequence to the first and second time-frequency resource pools to drive the robot to sort. If the multi-dimensional spatial conditions are not met, it continuously switches the acquisition zoom mode. After the time-frequency resources in the first time-frequency resource pool are determined, the quality inspection identification module selects the remaining time-frequency resources to execute the quality inspection evaluation strategy: according to the preset temperature field linkage quality inspection model, it detects each target automotive injection molded part and determines the quality inspection defect type of each target quality inspection area in each target automotive injection molded part.

[0006] Furthermore, multi-point temperature field data of the mold cavity and images of the automotive injection molding to be inspected are acquired in the mold-open state, including: In response to the mold opening command, the PLC controller synchronously sends hardware trigger pulse signals to the infrared thermal imager and industrial camera. Infrared thermal imagers are used to measure the temperature distribution at multiple points in the mold cavity in real time, generating temperature field data. Using an industrial camera, the injection-molded automotive parts are illuminated by light sources from multiple angles to capture a first image of the injection-molded automotive parts from multiple angles. All target injection-molded automotive parts in the first image are identified. Each target injection-molded automotive part includes at least one target quality inspection area. Based on the multi-point temperature distribution of the target quality inspection area, the acquisition zoom mode is matched to drive the industrial camera to capture a second image of the injection-molded automotive parts, which is then used as the final image to be inspected. The acquisition zoom mode includes a first zoom mode and a second zoom mode.

[0007] Furthermore, establish association mappings, including: Determine the target gate of the target automotive injection molded part, map the target gate as the root topology node, map each target quality inspection area as a slave topology node, and configure the runner geometry distance and wall thickness parameters on each topology edge; By performing multi-layer wavelet decomposition on the automotive injection molding image to be inspected, the time-frequency energy distribution is extracted. Combined with the synchronously acquired multi-point temperature field data of the mold cavity, the first ratio of the temperature distribution from each topological node to the geometric distance of the flow channel, and the first product of the wall thickness parameter and the energy distribution are calculated. The binary group consisting of the first ratio and the first product in the preset time period is marked as the correlation analysis group, and the standard score of the correlation analysis group in the correlation analysis space of the current measurement period is marked as the abnormal correlation score.

[0008] Furthermore, multiple time-frequency resource pools are divided, including: Under the condition of satisfying the spatial acquisition constraints, the abnormal correlation score and the minimum score threshold are compared: if the abnormal correlation score is less than or equal to the minimum score threshold, the current temperature field environment is determined to be normal, the first time-frequency resource pool is allocated for the current sorting task, and temperature field sensing is continuously performed. If the abnormal correlation score is greater than the minimum score threshold, the current temperature field environment is determined to be abnormal, and the current sorting task is classified into the second time-frequency resource pool. At the same time, the real-time robot status of the sorting task is identified, and the variable stiffness clamping force and the upper limit of the transport acceleration of the robot sorting mold are obtained. By monitoring the time-frequency resource status of the second time-frequency resource pool, control resources are allocated to the sorting tasks in the second time-frequency resource pool first, and the robot end effector is driven to grasp the target automotive injection molded parts with variable stiffness clamping force. The center of mass offset trajectory is retrieved and the sorting is carried out according to a smooth curve that is less than the upper limit of the transport acceleration.

[0009] Furthermore, determine whether the target quality inspection area meets the multi-dimensional spatial conditions, including: Multidimensional space conditions include geometric constraints, material constraints, and process constraints, and the corresponding judgment priorities decrease in that order. Geometric constraints include: positional tolerance range and dimensional range; Material constraints include: color difference range; Process constraints include: maximum gate mark and maximum weld line position offset; If the judgment result does not satisfy at least one constraint condition, then mark it as not satisfying the multidimensional space condition.

[0010] Furthermore, the pre-defined temperature field linkage quality inspection model has a built-in feature encoder and a classification encoder, including: The initial quality inspection defect type label is output, and the feature threshold within the boundary of the normalized quality inspection feature signal is used as the weight coefficient to determine the defect status score. The fluctuation feature and timestamp are extracted from the defect status score by the feature encoder, and the temperature field data under the same timestamp are retrieved. The time-series features of abnormal temperature field environment are statistically analyzed. The fluctuation feature and time-series feature are spliced ​​to form a fused feature, which is then input into the classification encoder to determine the final quality inspection defect type.

[0011] Furthermore, the step of obtaining the feature threshold within the boundary of the first quality inspection feature signal includes: Based on the vertical direction of the target quality inspection area, the energy values ​​of each pixel in the same scanning direction in the automotive injection molding image corresponding to each target injection molded part are accumulated to obtain the cumulative energy value of each pixel; where the scanning direction is the vertical direction of the extension direction of the target quality inspection area; Based on the cumulative energy values ​​of each pixel in the same scanning direction parallel to the target quality inspection area, the characteristic signal curves of each scanning direction are plotted; where the parallel direction is the parallel direction of the target quality inspection area. Extract the maximum energy value of the feature signal curve, determine the feature threshold of the corresponding feature signal curve based on the maximum energy value, determine the target pixel in the feature signal curve of the corresponding scan line based on the feature threshold, and connect the target pixel along the extension direction of the target quality inspection interval to form the quality inspection feature signal boundary.

[0012] Furthermore, based on the matched acquisition zoom mode, the zoom is adjusted to the first focal length using the first zoom mode during the first time period, and to the second focal length using the second zoom mode during the second time period.

[0013] (III) Beneficial Effects This invention provides a robot sorting system with mold temperature field linkage for automotive injection molding quality inspection, which has the following advantages: 1. This invention simultaneously acquires the temperature field of multiple points in the cavity under mold opening conditions and the injection molding image of the car to be inspected. It uses multi-dimensional spatial conditions of geometric constraints, material constraints and process constraints to perform priority verification. When the verification fails, it triggers the dynamic switching of the acquisition zoom mode. To a certain extent, it can terminate subsequent invalid calculations in time to save computing power. At the same time, when there are local anomalies, it automatically drives the focal length mode to perform secondary accurate image acquisition, which significantly improves the image acquisition quality of the target quality inspection area and the utilization efficiency of edge computing resources. 2. This invention maps the target quality inspection area with multiple temperature fields, quantifies and generates anomaly correlation scores, and divides them into multiple time-frequency resource pools. By prioritizing the sending of sorting sequences in the first and second time-frequency resource pools to drive robot sorting, the robot sorting response delay is shortened to a certain extent, effectively avoiding clamping damage or injection molding production line stoppages caused by delays in robotic arm control sequences. Under the coordination of seizing remaining time-frequency resources, a quality inspection evaluation strategy is executed. A temperature field linkage quality inspection model with built-in feature encoders and classification encoders is introduced. The fluctuation characteristics of defect status scores and the time sequence characteristics of abnormal temperature field environments are fused in a high dimension, which can accurately output quality inspection defect types including no defects, first defect problems, and second defect problems, significantly improving the robustness and accuracy of the model in identifying complex injection molding defects. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating two-dimensional monitoring and task interruption recovery timing control according to an exemplary embodiment; Figure 2 This is a schematic diagram of a module according to an exemplary embodiment. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] Research and Development Concept: The data acquisition module synchronously collects multi-point temperature field data and injection images of the mold cavity under mold opening conditions; the resource allocation module performs decreasing priority verification based on multi-dimensional spatial conditions of geometric, material, and process constraints. If the conditions are not met, the zoom mode is adaptively switched; if the conditions are met, a temperature field correlation mapping is established and a time-frequency resource pool is allocated to achieve preemptive scheduling of sorting time sequence; through the quality inspection identification module, the temperature field linkage quality inspection model is used to splice and fuse the fluctuation characteristics and the time sequence characteristics of abnormal temperature field environment, and the built-in feature encoder and classification encoder accurately output the final quality inspection defect type; the algorithm computing power and robot physical control are decoupled in time and frequency, which significantly improves the online inspection accuracy and automated sorting efficiency of precision injection molded parts.

[0017] Example: This embodiment of the invention provides a robot sorting system with mold temperature field linkage for automotive injection molding quality inspection; Figure 1 This is a schematic diagram illustrating two-dimensional monitoring and task interruption recovery timing control according to an exemplary embodiment. Figure 2 This is a schematic diagram of a module according to an exemplary embodiment; please refer to... Figures 1 to 2The system includes a data acquisition module, a resource allocation module, and a quality inspection and identification module, and the data acquisition module, resource allocation module, and quality inspection and identification module are connected to each other. The following is an explanation of each module involved: Data acquisition module: Acquires multi-point temperature field data of the mold cavity and images of the automotive injection molding to be inspected in the mold-open state; In response to the mold opening command, the PLC controller establishes a data handshake with the injection molding machine controller via industrial Ethernet. The PLC controller synchronously sends hardware trigger pulse signals to the industrial camera and infrared thermal imager. The PLC sends high-level TTL hardware pulse signals to the external trigger pin of the industrial camera and the synchronization interface of the infrared thermal imager through the high-speed hardware output port. By aligning with the millisecond-level timestamp through direct hardware wiring, it ensures that the mold opening action of the injection molding machine is synchronized with the above devices at the millisecond level, eliminating modal asynchrony caused by mechanical vibration and data transmission delay. The infrared thermal imager is deployed directly above or to the side of the injection molding machine mold. It has a built-in blackbody calibration unit and is used to measure the temperature distribution of multiple points on the inner wall of the mold cavity and the surface of the hot injection molded part in real time without contact when the mold is opened. It outputs a thermal imaging matrix containing the physical temperature value of each pixel and generates temperature field data containing a two-dimensional spatial grid temperature distribution. An industrial camera is mounted at the end of a robotic arm or directly in front of the mold. Under the same pulse trigger, it sequentially illuminates the automotive injection molded part with light sources at 0°, 45°, and 75° to capture multi-angle images of the automotive injection molded part, obtaining a first automotive injection molded image from multiple angles. The first automotive injection molded image and temperature field data are transmitted to an edge computing node via a GigE interface. The edge node performs white balance correction, histogram contrast enhancement, and Gaussian noise reduction on the first automotive injection molded image. Based on the processed first automotive injection molded image, the target automotive injection molded part and its target quality inspection area are identified. Specific steps include: identifying all target automotive injection molded parts within the current mold cavity, such as bumpers and snap-on panels, through template matching or geometric contour detection in the first automotive injection molded image; automatically defining at least one target quality inspection area on each target automotive injection molded part, with the areas between targets typically representing areas of abrupt changes in injection molded part wall thickness, gate areas, or locations of easily shrinking reinforcing ribs. Based on the multi-point temperature distribution of the target quality inspection area, a corresponding acquisition zoom mode is matched, including a first zoom mode and a second zoom mode. According to the matched acquisition zoom mode, the zoom is adjusted to the first focal length in the first time period and to the second focal length in the second time period. The temperature field data is projected onto the target quality inspection area according to the spatial pixel coordinate system. The real-time maximum temperature value, average temperature value, and temperature gradient change rate of each target quality inspection area are extracted, and a three-dimensional monitoring vector is constructed: maximum temperature value, average temperature value, and temperature gradient change rate. A preset standard three-dimensional monitoring vector is included, including the standard maximum temperature value, standard average temperature value, and standard temperature gradient change rate. The three-dimensional monitoring vector is compared with the preset vector one by one: the difference between the maximum temperature value and the standard maximum temperature value, the difference between the average temperature value and the standard average temperature value, and the difference between the temperature gradient change rate and the standard temperature gradient change rate are calculated. If the difference between the three is greater than or equal to 0, it is determined that the current temperature distribution is uniform, and the first zoom mode is automatically matched. The PLC keeps the industrial camera lens in a low magnification and large depth of field state to maintain the existing field of view. If there is a difference result less than 0, the current temperature is determined to be locally abnormal. The second zoom mode is automatically matched, and a pulse drive command is sent to the motor driver of the industrial camera to drive the stepper motor to adjust the focal length and focus ring of the camera lens, and fine-tune and magnify the field of view to focus on the target quality inspection area with the local high temperature abnormality. It should be noted that the steps for obtaining the standard three-dimensional monitoring vector are as follows: Based on historical statistics, while the injection molding machine is in a steady-state production state, multiple sets of temperature field data of calibrated automotive injection molded parts that have passed offline quality inspection are continuously collected at the moment of demolding. The preliminary mean and standard deviation of the highest temperature value, average temperature value, and temperature gradient change rate in the multiple sets of calibration data are calculated. The 3σ criterion is used to remove temperature field data with a deviation of more than three times the standard deviation. The arithmetic mean of the statistically obtained highest temperature value, average temperature value, and temperature gradient change rate are calculated. The obtained results are marked as the standard highest temperature value, standard average temperature value, and standard temperature gradient change rate, respectively, to form the preset standard three-dimensional monitoring vector. In addition, by continuously running K-mold injection molding products, such as K being 100, and setting a sliding window, the temperature field data corresponding to the latest qualified injection molding products are continuously included in the sliding window and the oldest data is removed over time. The preset standard three-dimensional monitoring vector is updated in real time to adaptively eliminate parameter drift caused by changes in environmental temperature and humidity.

[0018] By simultaneously acquiring the temperature field of multiple points in the cavity under mold opening conditions and the injection molding image of the car to be inspected, and using multi-dimensional spatial conditions of geometric constraints, material constraints and process constraints to perform priority verification, the dynamic switching of the acquisition zoom mode is triggered when the verification fails. To a certain extent, this timely termination of subsequent invalid calculations saves computing power. At the same time, when there are local anomalies, the zoom mode is automatically driven for secondary accurate image acquisition, which significantly improves the image acquisition quality of the target quality inspection area and the utilization efficiency of edge computing resources.

[0019] Resource partitioning module: Based on the image of the automotive injection molding to be inspected, the module identifies the target quality inspection area and determines whether the target quality inspection area meets the multi-dimensional spatial conditions. If the multi-dimensional spatial conditions are met, the module combines temperature field data to establish an association mapping, generates an anomaly association score, and divides the area into multiple time-frequency resource pools. At least one sorting sequence is sent to the first and second time-frequency resource pools to drive the robot to sort. If the multi-dimensional spatial conditions are not met, the module continuously switches the acquisition zoom mode. Determining whether the target quality inspection area meets multi-dimensional spatial conditions includes geometric constraints, material constraints, and process constraints, with the priority decreasing in the following order: geometric constraints > material constraints > process constraints. The following is an explanation of some of the terms involved: Geometric constraints include: positional tolerance range and dimensional range; specifically, positional tolerance range: the maximum allowable offset of key feature points, feature lines, or feature surfaces on the target quality inspection area relative to the reference coordinate system, such as the positional tolerance of the center of a mounting hole or the profile tolerance of an assembly surface; when constructing geometric constraint verification, positional tolerance determines the alignment accuracy of the structural assembly; smaller positional tolerance requirements can ensure extremely high assembly accuracy, but will increase the requirements for the dimensional stability of injection molding; larger positional tolerances may lead to subsequent assembly jamming or loosening; Size range: The allowable variation range of independently measurable linear or angular dimensions on the target quality inspection area, usually defined by an upper and lower limit size, such as the thickness of reinforcing ribs, the diameter of through holes, and the cantilever length of clips; measured values ​​within this range will be retained by the system and judged as qualified; measured values ​​less than the lower limit or greater than the upper limit will be judged as out of tolerance, for example: measured values ​​less than the lower limit size are judged as shrinkage and material shortage, and measured values ​​greater than the upper limit size are judged as bulging and expansion; Material constraints include: color difference range and density uniformity lower limit; specifically, color difference range: refers to the maximum allowable Euclidean distance in Lab color space between the surface color of the target quality inspection area and the standard color chart under a standard D65 light source and a specific viewing angle; Process constraints include: maximum gate mark and maximum weld line position offset. Specifically, maximum gate mark refers to the maximum allowable height of any protrusion or depression remaining on the product surface after gate removal. For example, after the submarine gate of a car exterior rearview mirror housing is removed, a protrusion with a height of about 0.35mm remains on the B side of the injection molded part. This side will later be used to install a lens adjustment motor. The motor mounting surface requires the gate residual height to be less than 0.2mm, otherwise it will cause the motor to be misaligned after installation. Maximum weld line position offset refers to the maximum allowable offset of the actual position of the weld line formed by the convergence of two or more melt fronts relative to the reference position predicted by mold flow analysis. By prioritizing geometric constraint verification, subsequent verifications are terminated directly if the geometric constraints are not met to save computing power; material constraint verification is triggered only after the geometric constraints are passed; process constraint verification is finally performed only after both geometric and material constraints are passed; if the judgment result does not meet at least one constraint condition, it is marked as not meeting the multidimensional space condition. Establishing an association mapping includes: determining the target gate of the target automotive injection molded part, mapping the target gate as a root topology node, mapping each target quality inspection area as a slave topology node, and configuring runner geometric distance and wall thickness parameters on each topology edge; wherein, the runner geometric distance refers to the curve length traversed by the melt from the center position of the injection point of the target gate, along the geometric center layer of the mold cavity, following the actual melt flow path to the geometric center position of the target quality inspection area; the wall thickness parameter is used to characterize the segmented distribution of the nominal thickness of the product in the direction perpendicular to the cavity surface along the melt flow path corresponding to the above runner geometric distance; The time-frequency energy distribution is extracted by multi-level wavelet decomposition of the image of the car injection molding to be inspected. The specific steps include: performing a three-level wavelet transform on the image of the car injection molding to be inspected. First layer: The image of the car injection molding to be detected is converted into an M×N grayscale image. Using the selected db4 wavelet basis, one-dimensional low-pass filtering Lo_D and high-pass filtering Hi_D are performed along the row and column directions of the grayscale image, respectively. Then, downsampling is performed at intervals to obtain four sub-band images, namely: the first layer low-frequency approximate sub-band LL1, the first layer horizontal high-frequency sub-band LH1, the first layer vertical high-frequency sub-band HL1, and the first layer diagonal high-frequency sub-band HH1. Among them, the size of LL1 is approximately (M / 2)×(N / 2), which contains the smooth main information of the image; the size of LH1 is approximately (M / 2)×(N / 2), which highlights the edges and texture changes in the horizontal direction; the size of HL1 is approximately (M / 2)×(N / 2), which highlights the edges and texture changes in the vertical direction; and the size of HH1 is approximately (M / 2)×(N / 2), which highlights the details and noise in the diagonal direction. Second layer: The same filtering and downsampling operation is performed again on the low-frequency approximate sub-band LL1 obtained from the first layer decomposition to obtain four sub-band images of the second layer: the second layer low-frequency approximate sub-band LL2, the second layer horizontal high-frequency sub-band LH2, the second layer vertical high-frequency sub-band HL2, and the second layer diagonal high-frequency sub-band HH2, each with a size of approximately (M / 4)×(N / 4). Third layer: Continue to perform the same operation on LL2 to obtain four sub-band images of the third layer: the third layer low-frequency approximate sub-band LL3, the third layer horizontal high-frequency sub-band LH3, the third layer vertical high-frequency sub-band HL3, and the third layer diagonal high-frequency sub-band HH3, each with a size of approximately (M / 8)×(N / 8). After three layers of wavelet decomposition, the original image is converted into a set of 3×3+1=10 sub-band images, including one bottom-level first low-frequency approximate sub-band LL3, and nine detail sub-bands of different scales and directions, such as: LH1, HL1, HH1, LH2, HL2, HH2, LH3, HL3, HH3. Each sub-band is essentially a wavelet coefficient matrix, where each element is called a wavelet coefficient, and its absolute value reflects the degree of grayscale change of the image at that location at that scale and in that direction. The energy distribution map of each of the nine sub-bands is calculated, and the local energy map of each sub-band is normalized by max-min, mapping the energy values ​​to the interval [0, 1]. By summing the normalized energy values ​​of each sub-band, the energy distribution of the target quality inspection area is constructed. It should be noted that in the subsequent calculation process, the numerical values ​​involved need to be dimensionless to make the calculation results have physical meaning. The specific steps are not described in detail here. Combining the multi-point temperature field data of the mold cavity acquired simultaneously, the first ratio of the temperature distribution from each topological node to the geometric distance of the flow channel and the first product of the wall thickness parameter and the energy distribution are calculated. The binary pair formed by the first ratio and the first product in the preset time period is marked as the correlation analysis group, and the standard score of the correlation analysis group in the correlation analysis space of the current calculation period is marked as the abnormal correlation score. In the correlation analysis space, the horizontal axis represents the abnormality of the temperature process, and the vertical axis represents the abnormality of the surface texture result. This space is pre-divided into several regions, and each region corresponds to a different correlation level. Definition of Standard Score: The abnormal association score is defined as the standardized value of the Mahalanobis distance of the association analysis group relative to the center point of the normal region in the association analysis space. Determination of the normal region: The normal region is determined based on the statistical distribution of a large number of known good product samples through association analysis. Specifically, during the system debugging phase, several products that have been manually confirmed to be qualified are collected, and the first ratio and first product of each of them from the topological nodes are calculated to form a scatter cloud of good products in the association analysis space. A two-dimensional Gaussian fit is performed on the scatter cloud to obtain the center point and covariance matrix of the good product distribution. The normal region is usually defined as an elliptical region with the center point as the center and a radius of 3 times the standard deviation. Multiple time-frequency resource pools are divided, including: under the condition of satisfying spatial acquisition constraints, comparing the abnormal correlation score and the minimum score threshold: if the abnormal correlation score is less than or equal to the minimum score threshold, it indicates that the correlation analysis group falls within the normal area, and the temperature field correlation pattern of the slave topology node is consistent with that of good products. In this case, the current temperature field environment is determined to be normal, and the first time-frequency resource pool is allocated for the current sorting task, and temperature field sensing continues; if the abnormal correlation score is greater than the minimum score threshold, it indicates that the correlation analysis group deviates significantly from the normal area, and there is an abnormal correlation between the temperature field and texture of the slave topology node, i.e., the two do not match. In this case, the current temperature field environment is determined to be abnormal, and the current sorting task is allocated to the second time-frequency resource pool; where the minimum score threshold is usually set to 3. The steps for obtaining the minimum score threshold include: based on historical statistics, collecting multi-point temperature field data of historical injection molded parts under normal temperature field environment and automotive injection images as a benchmark sample set; extracting the abnormal correlation scores of each topological node in each benchmark sample, calculating the mean and standard deviation of the data, and setting the minimum score threshold as the sum of the mean and a multiple of the standard deviation; it should be noted that the value of the multiple is just an example, and the specific setting should be based on the actual situation, which will not be elaborated here. In addition, based on the specific location of the correlation analysis group in the correlation analysis space, a preliminary diagnosis of the quality inspection defect type can be achieved, generating initial quality inspection defect type labels: When the location of the correlation analysis group meets one of the following conditions, it is determined that the topology node has an abnormal cooling process, and the abnormality has formed a detectable defect on the product surface texture: If the coordinates of all points in the current correlation analysis group are greater than the coordinates of the center point of the good product distribution, this indicates that the actual cooling rate of the mold in this area is significantly faster than theoretically expected, and the surface texture energy of the product is significantly higher than the normal baseline. This may be because the melt solidifies prematurely due to the mold being too cold, and the flow front cannot fully fuse, forming cold material marks or microcracks on the surface. These defects are manifested in the image as an abnormal increase in high-frequency texture energy. If only the x-coordinates of all points in the current correlation analysis group are less than the x-coordinates of the center point of the good product distribution, this indicates that the mold in this area has an overheating trend, but has not yet reached the level that causes surface shrinkage marks. In the above cases, the robot sorts the current injection molded product to the control buffer area with time-of-life monitoring function. The system automatically sets a second inspection of the product after a preset time-of-life cycle to capture delayed warping, cracking, or discoloration defects. If the vertical coordinate of the landing point of the correlation analysis group deviates significantly from the center point coordinate of the good product distribution, this landing point indicates that there is an abnormality in the surface texture of the product. However, this abnormality cannot be explained by the mold temperature field data. The robot will sort the current product to the manual re-inspection area and generate a text prompt, suggesting that the mold cavity status or raw material batch information be checked. Simultaneously, upon receiving a sorting decision instruction for the current injection molded part, such as isolation and re-inspection, the system first reads the joint encoder, servo driver feedback, and end effector force sensor data of the robot currently performing the sorting task via the industrial real-time bus to identify the robot's real-time state. The real-time robot state includes, but is not limited to: the robot's current joint angle, the end effector's pose in the base coordinate system, the current speed of each joint, and whether the end effector is currently in an unloaded or holding state. This information is used to obtain the robot's instantaneous kinematic and dynamic parameters, providing a reference value for the subsequent calculation of variable stiffness clamping force and handling acceleration upper limit, ensuring that the generation of subsequent control instructions is based on the robot's actual state. By identifying the real-time robot state of the sorting task, the system obtains the variable stiffness clamping force and handling acceleration upper limit of the robot's sorting mold. After setting a context checkpoint for the sorting tasks in the first time-frequency resource pool and saving their current data frame state, their execution is interrupted to release time-frequency resources. By monitoring the time-frequency resource status of the second time-frequency resource pool, control resources are allocated to the sorting tasks in the second time-frequency resource pool first, driving the robot's end effector to grasp the target automotive injection molded parts with variable stiffness clamping force. When the tasks in the second time-frequency resource pool are completed and the time-frequency resources are released, the execution of the interrupted tasks in the first time-frequency resource pool is automatically resumed from the context checkpoint using the released time-frequency resources. For example: Set times t0, t1, t3, and t4. From t0 to t1: The first time-frequency resource pool sorting task executes normally, occupying control resources. At time t1: The second time-frequency resource pool sorting task becomes urgent. The first time-frequency resource pool sorting task saves its context checkpoint and is interrupted, while the second time-frequency resource pool sorting task jumps into the queue. From t1 to t3: Resources are concentrated to ensure the second time-frequency resource pool sorting task. At time t3: The second time-frequency resource pool completes its task and releases time-frequency resources. The first time-frequency resource pool sorting task automatically resumes execution from the checkpoint. The allocation of control resources for sorting tasks in the second time-frequency resource pool includes: monitoring the usage of current control computing resources and communication bandwidth resources in both time and frequency domains; in the time domain, monitoring the queue length and average waiting time of the current sorting task; in the frequency domain, monitoring whether there are periodic peaks in the communication load rate of the robot control bus; if there are periodic peaks in the communication load rate, it means that the bus bandwidth resources are tight, which may cause robot movement jitter due to data packet loss or delay; the time-frequency resource status reflects the real-time availability of system control resources. By dynamically adjusting the priority weight of the task scheduler, sorting tasks in the second time-frequency resource pool are placed at the front of the current execution queue to ensure that the sorting action of injection molded products is completed in the shortest time after the temperature field data is collected, avoiding missing the best time tracking start time due to waiting for the regular sorting queue; The force control is transmitted to the end effector via the robot controller's force control interface. It should be noted that the end effector can be a pneumatic gripper, an electric servo gripper, or a vacuum suction cup, etc. For example, for an electric servo gripper, the controller adjusts the output torque of the gripper motor through a current loop or torque loop closed-loop adjustment, ensuring that the actual gripping force applied to the product surface accurately tracks the target value. For example, for a pneumatic gripper, the system adjusts the air supply pressure through an electro-proportional valve, thereby indirectly controlling the gripping force. During force control, force sensors installed at the gripper tips or the base of the suction cup provide real-time feedback on the actual gripping force, forming a force control closed loop. This ensures that the gripping force is neither too small, causing the product to slip, nor too large, causing damage to the product surface. The robot retrieves the center of mass offset trajectory and sorts the parts according to a smooth curve that is less than the upper limit of the handling acceleration. The robot sorts the target injection molded parts to the target location, such as the qualified part bin, the waste part bin, the manual re-inspection area, or the time-tracking buffer area, according to the center of mass offset trajectory. It then releases the variable stiffness clamping force to complete a complete sorting action. The robot records the execution time, actual clamping force curve, and handling acceleration curve of this sorting action and stores them in the sorting log of the product for subsequent quality traceability and process analysis. The following is an explanation of some of the terms involved: Center of mass offset trajectory: refers to the motion path of the end effector tool's center point in Cartesian space during the robot's handling process; Smooth curve handling: This refers to the robot's motion trajectory being interpolated using an S-shaped velocity curve. The acceleration of the S-shaped velocity curve is continuous in the initial acceleration phase and the final deceleration phase, which can effectively avoid inertial impact caused by sudden acceleration changes and further reduce the risk of products slipping or being damaged due to shaking during handling. Less than the upper limit of handling acceleration: During the entire handling trajectory execution, the robot controller monitors the actual acceleration of the end effector in real time, and ensures that the actual acceleration is always less than the upper limit of handling acceleration through feedforward control and speed limiting; once the actual acceleration is detected to be close to the upper limit, the controller automatically scales the trajectory speed proportionally to ensure safe handling and sorting.

[0020] By mapping the target quality inspection area with multiple temperature fields, anomaly correlation scores are generated quantitatively, and multiple time-frequency resource pools are divided accordingly. By prioritizing the sending of sorting sequences in the first and second time-frequency resource pools to drive robot sorting, the robot sorting response latency is shortened to a certain extent.

[0021] Quality Inspection Identification Module: After the time-frequency resources in the first time-frequency resource pool are determined, the remaining time-frequency resources are selected to execute the quality inspection evaluation strategy: each target automotive injection molded part is detected according to the preset temperature field linkage quality inspection model, and the quality inspection defect type of each target quality inspection area in each target automotive injection molded part is determined; The pre-defined temperature field linkage quality inspection model incorporates a feature encoder and a predictive encoder. It includes: generating a preliminary quality inspection defect type label for each target quality inspection area of ​​the target automotive injection molded part based on the partitioned diagnostic results of the aforementioned correlation analysis space; wherein, the quality inspection defect type label is a classification label, and its possible values ​​include, but are not limited to: no defects, cold material marks, weld line offset, texture loss, etc.; after outputting the initial quality inspection defect type label, the feature threshold within the normalized quality inspection feature signal boundary is used as a weighting coefficient to determine the defect status score; it should be noted that a higher feature threshold means that the defect signal needs to be stronger to be judged as a defect, i.e., the lower the detection sensitivity; conversely, a lower feature threshold means higher detection sensitivity; by using the feature threshold as a weighting coefficient, the defect judgment scale for different areas and different products can be adaptively adjusted. The step of obtaining the feature threshold within the boundary of the first quality inspection feature signal includes: according to the vertical direction of the target quality inspection interval, accumulating the energy values ​​of each pixel in the same scanning direction in the automotive injection molding image corresponding to each target injection molded product, and obtaining the cumulative energy value of each pixel; wherein, the scanning direction is the vertical direction of the extension direction of the target quality inspection area; it should be noted that the energy value at this point is the energy value of the corresponding pixel in the energy distribution map of the nine sub-bands calculated above, and the physical meaning of the cumulative energy value is: the overall intensity of the defect texture signal on the scan line, and the more obvious the defect of the scan line, the greater its cumulative energy value; Based on the cumulative energy value of each pixel in the same scanning direction parallel to the target quality inspection area, the characteristic signal curves of each scanning direction are plotted with the position index in the parallel direction as the horizontal axis and the cumulative energy value corresponding to the scan line as the vertical axis; where the parallel direction is the parallel direction of the target quality inspection area. Extract the maximum energy value of the feature signal curve, and determine the feature threshold of the corresponding feature signal curve based on the maximum energy value. Specifically, using the maximum energy value as the benchmark, sort the data points on the feature curve from largest to smallest, and take the u-th percentile value as the feature threshold; where u is between 90 and 95; for example, the value of u is determined according to the surface texture level of the target quality inspection area: for leather interior parts, u=90; for high-gloss exterior parts, u=95; the basis is that, according to actual measurement and statistics, under defect-free conditions, the normal fluctuation range of the cumulative feature value of each scan line in the target quality inspection area is concentrated in the range of about 90% of the mean. Scan lines greater than this range are candidate target quality inspection areas with texture abnormalities. Based on the feature threshold, the target pixel in the corresponding scan line is determined in the feature signal curve of the corresponding scan line. The target pixel is connected along the extension direction of the target quality inspection interval. The target pixel determined on each scan line is connected sequentially along the extension direction of the target quality inspection interval to form one or two boundary lines, thus forming the quality inspection feature signal boundary. The feature encoder includes an input layer, a first hidden layer, a second hidden layer, and an output layer; Input Layer: The input layer receives a sequence of defect status scores for T consecutive time periods, with a sequence length of T. Each time step contains one defect status score value. The T score values ​​are arranged into a one-dimensional vector in chronological order. The timestamp corresponding to each score value is normalized and then used as an auxiliary feature to be concatenated into the input vector, forming an input vector of length 2T. The number of neurons in the input layer is set to 2T, where T is typically 32, i.e., a 64-dimensional input vector, to capture sufficiently long historical trend information. First hidden layer: The first hidden layer maps the input layer scoring sequence to a high-dimensional space. Through fully connected operations and the ReLU activation function, it captures the fluctuation characteristics of the defect status score along the time axis, including the combined characteristics of the continuous increase of the score mean and the gradual increase of the score standard deviation, the regularity of the score oscillation, and the characteristic of the score changing abruptly at a certain moment. The first hidden layer has 64 neurons. Second hidden layer: The second hidden layer reduces the dimensionality of the high-dimensional fluctuation features extracted by the first hidden layer. Through fully connected operations, it selects the fluctuation features that are most discriminative for predicting the type of quality inspection defects, removes noise and redundant information, and avoids overfitting of the subsequent classification model. The second hidden layer has 16 neurons. Output layer: The output layer maps the fluctuation features filtered by the second hidden layer into a low-dimensional first feature vector. This first feature vector is a compact representation of the defect state fluctuation features of the current consecutive T cycles. The output layer is set with 8 neurons, that is, the dimension of the first feature vector is 8 dimensions. Temperature field data at the same timestamp were retrieved, and the temporal characteristics of abnormal temperature field environments were statistically analyzed. Specifically, temperature field data of multiple points in the mold cavity with the same timestamp for the above T consecutive time periods were retrieved, and the temporal characteristics of abnormal temperature field environments in each period were statistically analyzed. The statistical content includes: the proportion of abnormal measurement points, the average abnormal amplitude, and the number of abnormal duration periods, totaling 4-dimensional features. The 4-dimensional features of the T periods were arranged in chronological order to form a temporal feature vector of length 4T. The fluctuation characteristics and time-series characteristics are concatenated to form a 16-dimensional fused feature, which is then input into the classification encoder to determine the final quality inspection defect type. Classification encoder: includes an input layer, a first hidden layer, a second hidden layer, and an output layer; Input layer: The input layer receives the above 12-dimensional fused feature vector and has 12 neurons; First hidden layer: The first hidden layer maps the fused features to a high-dimensional space. Through fully connected operations and the ReLU activation function, it captures the complex nonlinear mapping relationship between the features of each dimension in the fused features and the quality inspection defect type. This includes the lag correlation characteristic where the temperature field anomaly time sequence features lead the defect fluctuation features by several cycles. For example, after the mold temperature field becomes abnormal, it usually takes 3 to 5 time cycles for the injection molded product defect status score to show a significant response. The first hidden layer has 32 neurons. Second hidden layer: The second hidden layer reduces the dimensionality of the high-dimensional features extracted by the first hidden layer, selects the feature combination that is most decisive for the identification of quality inspection defect types, and suppresses the interference of redundant features on the classification results; the second hidden layer has 16 neurons. The output layer maps the output of the second hidden layer to 3 neurons through a fully connected operation. 3 equals the preset total number of quality inspection defect types, including at least one of the following types: no defect, first defect problem, and second defect problem. Non-defect issues: used to characterize the corresponding target quality inspection area that has been identified, and the target quality inspection area meets the multi-dimensional spatial conditions and has no thermal damage; The first defect is that the corresponding target quality inspection area was not identified. Second defect problem: used to indicate that the corresponding target quality inspection area has been identified, but the target quality inspection area has at least one abnormality, including but not limited to cold material marks and shrinkage marks; During training, historical data was divided into training, validation, and test sets in a 7:2:1 ratio; each training sample was labeled with one of the three defect types mentioned above; the Adam optimizer was used for parameter optimization, with a learning rate of 0.001 and a batch size of 32; the maximum number of training epochs was 200, and an early stopping mechanism was introduced: training was terminated when the validation set loss value did not decrease for 10 consecutive epochs; the classification cross-entropy loss function was used to accurately measure the difference between the final quality inspection defect type distribution and the true defect type labels.

[0022] By implementing a quality inspection and evaluation strategy in collaboration with the remaining time and frequency resources, and introducing a temperature field linkage quality inspection model with built-in feature encoder and classification encoder, the fluctuation characteristics of defect status scores and the temporal characteristics of abnormal temperature field environment are spliced ​​and fused in high dimension. This enables the accurate output of quality inspection defect types, including no defects, first defect problems, and second defect problems, which significantly improves the robustness and accuracy of the model in identifying complex injection molding defects.

[0023] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0025] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0026] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A robot sorting and mold temperature field linkage automotive injection molding quality inspection system, characterized in that, The system includes: The data acquisition module acquires multi-point temperature field data of the mold cavity and the injection molding image of the car to be inspected in the mold-open state; The resource partitioning module identifies the target quality inspection area based on the automotive injection molding image to be inspected, and determines whether the target quality inspection area meets the multi-dimensional spatial conditions. If the multi-dimensional spatial conditions are met, it establishes an association mapping based on the temperature field data, generates an anomaly association score, and divides the area into multiple time-frequency resource pools. It sends at least one sorting sequence to the first and second time-frequency resource pools to drive the robot to sort. If the multi-dimensional spatial conditions are not met, it continuously switches the acquisition zoom mode. After the time-frequency resources in the first time-frequency resource pool are determined, the quality inspection identification module selects the remaining time-frequency resources to execute the quality inspection evaluation strategy: according to the preset temperature field linkage quality inspection model, it detects each target automotive injection molded part and determines the quality inspection defect type of each target quality inspection area in each target automotive injection molded part.

2. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 1, characterized in that, Acquire multi-point temperature field data of the mold cavity in the mold-open state and the injection molding image of the automobile to be inspected, including: In response to the mold opening command, the PLC controller synchronously sends hardware trigger pulse signals to the infrared thermal imager and industrial camera. Infrared thermal imagers are used to measure the temperature distribution at multiple points in the mold cavity in real time, generating temperature field data. Using an industrial camera, the injection-molded automotive parts are illuminated by light sources from multiple angles to capture a first image of the injection-molded automotive parts from multiple angles. All target injection-molded automotive parts in the first image are identified. Each target injection-molded automotive part includes at least one target quality inspection area. Based on the multi-point temperature distribution of the target quality inspection area, the acquisition zoom mode is matched to drive the industrial camera to capture a second image of the injection-molded automotive parts, which is then used as the final image to be inspected. The acquisition zoom mode includes a first zoom mode and a second zoom mode.

3. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 1, characterized in that, Establishing association mappings includes: Determine the target gate of the target automotive injection molded part, map the target gate as the root topology node, map each target quality inspection area as a slave topology node, and configure the runner geometry distance and wall thickness parameters on each topology edge; By performing multi-layer wavelet decomposition on the automotive injection molding image to be inspected, the time-frequency energy distribution is extracted. Combined with the synchronously acquired multi-point temperature field data of the mold cavity, the first ratio of the temperature distribution from each topological node to the geometric distance of the flow channel, and the first product of the wall thickness parameter and the energy distribution are calculated. The binary group consisting of the first ratio and the first product in the preset time period is marked as the correlation analysis group, and the standard score of the correlation analysis group in the correlation analysis space of the current measurement period is marked as the abnormal correlation score.

4. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 3, characterized in that, Divide the time-frequency resource pools into multiple pools, including: Under the condition of satisfying the spatial acquisition constraints, the abnormal correlation score and the minimum score threshold are compared: if the abnormal correlation score is less than or equal to the minimum score threshold, the current temperature field environment is determined to be normal, the first time-frequency resource pool is allocated for the current sorting task, and temperature field sensing is continuously performed. If the abnormal correlation score is greater than the minimum score threshold, the current temperature field environment is determined to be abnormal, and the current sorting task is classified into the second time-frequency resource pool. At the same time, the real-time robot status of the sorting task is identified, and the variable stiffness clamping force and the upper limit of the transport acceleration of the robot sorting mold are obtained. By monitoring the time-frequency resource status of the second time-frequency resource pool, control resources are allocated to the sorting tasks in the second time-frequency resource pool first, and the robot end effector is driven to grasp the target automotive injection molded parts with variable stiffness clamping force. The center of mass offset trajectory is retrieved and the sorting is carried out according to a smooth curve that is less than the upper limit of the transport acceleration.

5. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 1, characterized in that, Determine whether the target quality inspection area meets the multidimensional spatial conditions, including: Multidimensional space conditions include geometric constraints, material constraints, and process constraints, and the corresponding judgment priorities decrease in that order. Geometric constraints include: positional tolerance range and dimensional range; Material constraints include: color difference range; Process constraints include: maximum gate mark and maximum weld line position offset; If the judgment result does not satisfy at least one constraint condition, then mark it as not satisfying the multidimensional space condition.

6. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 2, characterized in that, The pre-defined temperature field linkage quality inspection model has a built-in feature encoder and a classification encoder, including: The initial quality inspection defect type label is output, and the feature threshold within the boundary of the normalized quality inspection feature signal is used as the weight coefficient to determine the defect status score. The fluctuation feature and timestamp are extracted from the defect status score by the feature encoder, and the temperature field data under the same timestamp are retrieved. The time-series features of abnormal temperature field environment are statistically analyzed. The fluctuation feature and time-series feature are spliced ​​to form a fused feature, which is then input into the classification encoder to determine the final quality inspection defect type.

7. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 6, characterized in that, The steps for obtaining the feature threshold within the boundary of the first quality inspection feature signal include: Based on the vertical direction of the target quality inspection area, the energy values ​​of each pixel in the same scanning direction in the automotive injection molding image corresponding to each target injection molded part are accumulated to obtain the cumulative energy value of each pixel; where the scanning direction is the vertical direction of the extension direction of the target quality inspection area; Based on the cumulative energy values ​​of each pixel in the same scanning direction parallel to the target quality inspection area, the characteristic signal curves of each scanning direction are plotted; where the parallel direction is the parallel direction of the target quality inspection area. Extract the maximum energy value of the feature signal curve, determine the feature threshold of the corresponding feature signal curve based on the maximum energy value, determine the target pixel in the feature signal curve of the corresponding scan line based on the feature threshold, and connect the target pixel along the extension direction of the target quality inspection interval to form the quality inspection feature signal boundary.

8. The robot sorting and mold temperature field linkage automotive injection molding quality inspection system according to claim 2, characterized in that, Based on the matched acquisition zoom mode, the zoom is adjusted to the first focal length using the first zoom mode during the first time period, and then adjusted to the second focal length using the second zoom mode during the second time period.

Images