A method and system for adjusting parameters of an insert-molded injection-molded product

Abstract

The application discloses a kind of burying forming injection product parameter adjustment method and system, it is related to industrial production field, comprising: obtaining current forming process data and extracting first multidimensional process feature;Based on the first multidimensional process feature, the high correlation degree historical forming process data matched, the historical forming process data include forming defect annotation result;According to the second multidimensional process feature of historical forming process data, forming defect annotation result and current first multidimensional process feature, construct associated feature information;Determine current forming defect annotation result based on associated feature information;According to the determination result, generate process parameter adjustment instruction, real-time adjustment injection process parameter.The method is matched by multidimensional process feature association and historical data, realizes the intelligentization and precision of parameter adjustment, can effectively reduce the forming defect incidence, improves product quality stability.

Description

Technical Field

[0001] This invention relates to the field of industrial production, and more specifically, to a method and system for adjusting parameters of embedded injection molded products. Background Technology

[0002] In embedded injection molding production, product quality is significantly affected by injection molding process parameters. Traditional parameter adjustment methods often rely on manual experience or simple parameter feedback, making it difficult to fully consider the complex process characteristics and relationships during molding. This results in low parameter adjustment accuracy, slow response, and a tendency for molding defects such as weld lines and shrinkage marks, affecting product quality stability and production efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for adjusting parameters of embedded injection molded products.

[0004] In a first aspect, embodiments of the present invention provide a method for adjusting parameters of an embedded injection molded product, comprising:

[0005] Acquire the current molding process data, and extract the first multi-dimensional process feature from the current molding process data;

[0006] Based on the first multidimensional process feature matching, multiple historical molding process data with high correlation to the current molding process data are matched, and each historical molding process data has a corresponding molding defect annotation result;

[0007] Based on the second multi-dimensional process features corresponding to each historical molding process data and the molding defect labeling results corresponding to each historical molding process data, as well as the first multi-dimensional process features of the current molding process data, we construct associated feature information.

[0008] Based on the associated feature information, the forming defect labeling results of the current forming process data are determined;

[0009] Based on the identified molding defect labeling results, corresponding injection molding equipment process parameter adjustment instructions are generated, and the process parameters of the current or subsequent embedded molding injection process are adjusted in real time according to the adjustment instructions.

[0010] In one possible implementation, the first multidimensional process feature includes a first process timing feature, and the second multidimensional process feature includes a second process timing feature.

[0011] The step of constructing associated feature information based on the second multi-dimensional process features corresponding to each of the historical molding process data, the molding defect annotation results corresponding to each of the historical molding process data, and the first multi-dimensional process features of the current molding process data includes:

[0012] Parameter integration mapping is performed on multiple first process timing features to obtain target process stage features;

[0013] The target process stage features are projected onto a preset working condition semantic domain by a preset feature projection unit to obtain the first stage mapping vector group.

[0014] For each historical molding process data, the second process timing feature corresponding to the second process timing feature is respectively modeled by working condition mapping to obtain the second stage mapping vector group corresponding to each historical molding process data.

[0015] The molding process data process task identifier of the current molding process data is obtained, as well as the process parameters obtained by converting the sensor data of the current molding process data;

[0016] The process status of the current molding process data is generated based on the operating conditions of the current molding process data.

[0017] The first process description data is generated based on the molding process data process task identifier, the process parameters, and the process status.

[0018] Based on the first working condition identifier used to distinguish different information under the same molding process data, and the second stage mapping vector group, the second process description data and the molding defect annotation result corresponding to the same historical molding process data, a working condition instance is constructed to obtain multiple working condition instances corresponding to multiple historical molding process data.

[0019] The multiple working condition instances are spliced ​​together to obtain comprehensive feature information, wherein the multiple working condition instances in the comprehensive feature information are demarcated by a second working condition identifier used to distinguish process description data of different molding processes.

[0020] Construct a target operating condition sample based on the first operating condition identifier, the first stage mapping vector group, and the first process description data;

[0021] The associated feature information is generated based on the comprehensive feature information and the target working condition sample.

[0022] In one possible implementation, the first multidimensional process feature includes target process stage features and first working condition semantic features; matching multiple historical molding process data with high correlation to the current molding process data based on the first multidimensional process feature includes:

[0023] Obtain a pre-established process stage feature library and a working condition semantic feature library. The process stage feature library includes the feature association relationship between each candidate molding process data and the process stage features corresponding to the candidate molding process data. The working condition semantic feature library includes the feature association relationship between each candidate molding process data and the working condition semantic features corresponding to the candidate molding process data.

[0024] Based on the target process stage characteristics, the process stage feature library and the working condition semantic feature library are matched respectively, and based on the first working condition semantic feature, the process stage feature library and the working condition semantic feature library are matched respectively, to obtain multiple target candidate molding process data that are highly correlated with the current molding process data;

[0025] Based on the matching degree of molding process data between the multiple target candidate molding process data and the current molding process data, the historical molding process data is extracted from the multiple target candidate molding process data.

[0026] In one possible implementation, the plurality of target candidate molding process data includes a plurality of target candidate molding process data obtained by matching each matching mode; the step of extracting the historical molding process data from the plurality of target candidate molding process data based on the molding process data matching degree between the plurality of target candidate molding process data and the current molding process data includes:

[0027] For each matching mode, the matching degree between the current molding process data and each target candidate molding process data is calculated.

[0028] For each target candidate molding process data, calculate the comprehensive matching degree between the target candidate molding process data and the current molding process data in each matching mode;

[0029] The historical molding process data is extracted from the multiple target candidate molding process data based on the comprehensive matching degree.

[0030] In one possible implementation, the step of extracting a first multidimensional process feature from the current molding process data includes:

[0031] Multiple molding process data sampling points are extracted from the current molding process data, and the multiple molding process data sampling points are segmented into multiple data units;

[0032] Target molding process data sampling points are extracted from the multiple data units, and process stage feature extraction is performed on the target molding process data sampling points to obtain the first process timing feature;

[0033] Obtain the first process description data corresponding to the current molding process data, and extract the first working condition semantic features from the first process description data.

[0034] The first multidimensional process feature is obtained based on the first process timing feature and the first operating condition semantic feature.

[0035] In one possible implementation, determining the molding defect annotation result of the current molding process data based on the associated feature information includes:

[0036] The associated feature information is reconstructed into working condition vectors to obtain an associated working condition vector group, wherein the working condition vectors are the input features adapted to the defect annotation parsing model;

[0037] The associated working condition vector group is input into the defect annotation parsing model. The defect annotation parsing model is obtained by locking the core parameter set of the preset inference model and adjusting the adjustable parameter unit of the inference model according to the training associated working condition vector group. The adjustable parameter unit is related to the lightweight weight branch embedded in the inference model.

[0038] Obtain the target molding defect annotation results of the current molding process data output by the defect annotation parsing model.

[0039] In one possible implementation, the training steps of the defect annotation parsing model include:

[0040] Acquire the first training process stage features, first training process description data, and first training forming defect annotation results corresponding to the first training forming process data, as well as the second training process stage features, second training process description data, and second training forming defect annotation results corresponding to the second training forming process data.

[0041] According to the preset feature projection unit, the first training process stage features and the second training process stage features are respectively subjected to domain transformation processing to obtain the first training condition mapping vector group and the second training condition mapping vector group. Training-related feature information is constructed based on the first training condition mapping vector group, the first training process description data and the first training forming defect annotation result, as well as the second training condition mapping vector group and the second training process description data.

[0042] The training-related feature information is reconstructed into working condition vectors to obtain the training-related working condition vector group.

[0043] The lightweight weighted branch is embedded in the inference model to embed the adjustable parameter unit in the inference model through the lightweight weighted branch;

[0044] The core parameter set of the inference model is locked, and the adjustable parameter unit of the inference model is adjusted according to the training associated working condition vector group and the second training formed defect annotation result to obtain the defect annotation parsing model.

[0045] In one possible implementation, before performing domain transformation processing on the process stage features according to a preset feature projection unit, the method further includes:

[0046] Obtain a snapshot of the training process state and the corresponding training process description data;

[0047] The training process state snapshot is used to extract features to obtain training process state features, and the training process state features are input to the original feature projection unit so that the original feature projection unit projects the training process state features to the input preset working condition semantic domain of the inference model.

[0048] Obtain the training condition semantic features corresponding to the feature projection request, and input the training condition semantic features and the target training visual condition semantic features output by the feature projection unit into the inference model, wherein the model parameters of the inference model are locked.

[0049] Obtain the training prediction process description data output by the inference model, and train the original feature projection unit based on the training process description data and the training prediction process description data to obtain the feature projection unit.

[0050] In one possible implementation, adjusting the adjustable parameter units of the inference model based on the training-related working condition vector group to obtain the defect annotation parsing model includes:

[0051] The training associated working condition vector group is input into the inference model embedded with the lightweight weight branch, and the predicted training defect labeling result output by the inference model is obtained.

[0052] Based on the difference between the predicted training defect labeling results and the second trained defect labeling results, the adjustable parameter units of the inference model are adjusted to obtain the defect labeling parsing model;

[0053] The method further includes:

[0054] Based on the difference between the predicted training defect annotation results and the second trained defect annotation results, the module parameters of the feature projection unit are adjusted.

[0055] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0056] Compared to existing technologies, the beneficial effects provided by this invention include: The method and system for adjusting parameters of embedded injection molded products disclosed in this invention acquires current molding process data and extracts a first multi-dimensional process feature; based on the first multi-dimensional process feature, it matches highly correlated historical molding process data, which includes molding defect annotation results; it constructs associated feature information based on the second multi-dimensional process feature of the historical molding process data, the molding defect annotation results, and the current first multi-dimensional process feature; it determines the current molding defect annotation result based on the associated feature information; and it generates process parameter adjustment instructions based on the determination result, adjusting the injection molding process parameters in real time. This method, through multi-dimensional process feature association and historical data matching, achieves intelligent and precise parameter adjustment, effectively reducing the incidence of molding defects and improving product quality stability. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating the steps of a method for adjusting parameters of an embedded injection molded product according to an embodiment of the present invention.

[0059] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0061] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0062] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating a method for adjusting parameters of an embedded injection molded product according to an embodiment of the present disclosure. The method for adjusting parameters of an embedded injection molded product will be described in detail below.

[0063] Step S201: Obtain the current molding process data, and extract the first multi-dimensional process feature from the current molding process data;

[0064] Step S202: Match multiple historical molding process data that are highly correlated with the current molding process data according to the first multi-dimensional process feature; each of the historical molding process data has a corresponding molding defect annotation result.

[0065] Step S203: Construct associated feature information based on the second multi-dimensional process features corresponding to each historical molding process data, the molding defect labeling results corresponding to each historical molding process data, and the first multi-dimensional process features of the current molding process data.

[0066] Step S204: Determine the molding defect labeling result of the current molding process data based on the associated feature information;

[0067] Step S205: Based on the identified molding defect labeling results, generate corresponding injection molding equipment process parameter adjustment instructions, and adjust the process parameters of the current or subsequent embedded molding injection process in real time according to the adjustment instructions.

[0068] In this embodiment of the invention, the following example uses the injection molding process of an automotive electronic sensor housing as an example. The specific parameters, batch identifiers, model structures, etc. involved are all illustrative and do not represent real experimental data. The purpose is to clearly demonstrate the method and process.

[0069] Obtain the current molding process data and extract the first multi-dimensional process feature:

[0070] The server connects to the injection molding production equipment via industrial communication protocols to collect real-time data on the current molding process, including dynamic time-series data and static process description data. Dynamic time-series data consists of key process parameters within a molding cycle (e.g., 30-40 seconds), such as melt temperature (typically 240-260℃), injection pressure (0-150MPa), screw position (0-200mm), and injection speed (0-100mm / s), with a sampling frequency of approximately 100Hz, generating thousands of sampling points per cycle. Static process description data includes product model (e.g., "sensor housing A"), material type (e.g., "PA66 + glass fiber reinforced material"), mold information (e.g., "multi-cavity, point gate structure"), and equipment operating status (e.g., "normal production, no alarms").

[0071] The server segments and extracts features from the dynamic time-series data: the sampling points are divided into data units according to the molding stage (injection, holding pressure, cooling), and key sampling points of each stage (such as the pressure peak point in the injection stage, the switching point in the holding pressure stage, and the temperature stabilization point in the cooling stage) are extracted by thresholding or gradient analysis. Then, the first process time-series features are obtained by using a time-series feature extraction model (such as a CNN-LSTM hybrid model), which includes time-series parameters such as injection peak pressure (approximately 140-160 MPa), pressure rise rate (approximately 30-40 MPa / s), holding pressure decay rate (approximately 0.4-0.6 MPa / s), and cooling temperature gradient (approximately 3-5℃ / s). The dimensions are usually 100-150.

[0072] Regarding time series characteristics:

[0073] For static process description data, the server first performs structured processing: material type is mapped to material property parameters (e.g., melting point approximately 250-260℃, melt flow index (MFI) approximately 20-30 g / 10 min), mold information is associated with mold structure parameters (e.g., number of cavities 2-4, gate diameter approximately 1-1.5 mm), and equipment status is converted into equipment health indicators (e.g., motor load rate 60-70%, temperature control accuracy ±1℃). Then, semantic coding models (e.g., BERT or industry pre-trained models) are used to extract semantic features for the first operating condition, including material flowability level (e.g., medium, 0.7-0.9), mold complexity score (e.g., medium-low, 0.5-0.7), and equipment stability index (e.g., good, 0.8-0.95), etc., with dimensions typically ranging from 200 to 300. Finally, the server concatenates the first process time sequence features with the first operating condition semantic features to form a first multidimensional process feature vector with dimensions of approximately 300-450.

[0074] Match historical molding process data that is highly correlated with the current molding process data:

[0075] Based on the first multi-dimensional process features, the server matches highly correlated historical molding process data from the historical process database. The historical database stores historical production batch data of similar products (such as sensor housings, connectors, etc.). Each data entry includes process timing features, working condition semantic features, and corresponding molding defect annotation results (such as "no defects", "weld marks", "shrinkage marks", "flash", etc.). The data volume is usually from hundreds to thousands of batches.

[0076] The server calls two pre-built feature libraries: a process stage feature library (stores the association between process time sequence features and batch identifiers of historical batches) and a working condition semantic feature library (stores the association between working condition semantic features and batch identifiers of historical batches). Using a multi-pattern matching method, the server first filters candidate batches with high similarity (e.g., cosine similarity ≥ 0.75) in the process stage feature library based on the first process time sequence feature, and then filters candidate batches with high similarity (e.g., cosine similarity ≥ 0.8) in the working condition semantic feature library based on the first working condition semantic feature, thus obtaining two candidate sets.

[0077] After deduplicating the two candidate sets, the server calculates the overall matching degree for each candidate batch: taking into account both temporal feature similarity (weight can be set to 0.5-0.7, as temporal parameters have a more direct impact on defects) and semantic feature similarity (weight can be set to 0.3-0.5), the overall matching degree is obtained by weighted summation. The batches are sorted from high to low based on their overall matching degree, and the top 3-5 batches are selected as historical molding process data with a high correlation to the current molding process data. For example, a certain historical batch may be selected and associated with its molding defect annotation results (such as "weld mark, cavity 2") because its material type and mold structure are consistent with the current batch (high semantic similarity) and its temporal parameters such as injection pressure and holding time are similar (high temporal similarity).

[0078] Constructing associated feature information:

[0079] The server constructs associated feature information based on the first multidimensional process features of the current molding process data, the second multidimensional process features of the historical molding process data (i.e., the process sequence features and working condition semantic features of historical batches), and the historical molding defect annotation results.

[0080] First, domain transformation is performed on the process stage features: through a preset feature projection unit (such as a domain adaptation model trained based on GAN), the current first process time series features and the historical second process time series features are projected onto a unified preset working condition semantic domain (which includes dimensions related to defect risk assessment), to obtain the current first stage mapping vector group and the historical second stage mapping vector group. The vector dimension is usually 500-600 dimensions, including sub-vectors related to defects such as pressure stability, temperature uniformity, and cooling efficiency (for example, the pressure stability sub-vector value corresponding to the "weld mark" defect in a certain historical batch is relatively high, about 0.7-0.9).

[0081] Secondly, process description data and working condition instances are generated: the current first process description data integrates product model, real-time process parameters (e.g., current melt temperature approximately 250℃, injection speed approximately 80-90mm / s), production batch (e.g., batches 100-150), and other information; the historical second process description data consists of the process parameters for the corresponding batch (e.g., the melt temperature of a certain historical batch is slightly lower than the current temperature, approximately 245-250℃). The server constructs a working condition instance for each historical batch, including a first working condition identifier (e.g., "weld line defect_injection stage") that distinguishes different information within the same batch, a second-stage mapping vector group, second process description data, and molding defect annotation results; multiple historical working condition instances are sorted and concatenated according to their comprehensive matching degree, and delimited by a second working condition identifier (e.g., "batch separator_historical batch 1") to form comprehensive feature information.

[0082] Finally, construct the target working condition sample and combine associated features: construct the target working condition sample based on the current first working condition identifier (such as "current production_injection stage"), the first stage mapping vector group and the first process description data, and then combine the target working condition sample with the comprehensive feature information through the feature alignment algorithm (unified dimension) to generate associated feature information (the dimension is usually 1500-2500).

[0083] Determine the molding defect annotation results of the current molding process data:

[0084] Based on associated feature information, the server determines the target molding defect annotation result of the current molding process through a defect annotation parsing model. This model is a lightweight adaptation model: based on a pre-trained general inference model (such as ResNet, Transformer, etc.), its core parameters (such as the parameters of the first 90% of the network layers) are locked, and only adjustable parameter units related to lightweight weight branches (embedded at the end of the model, with a parameter scale of 1%-5% of the core model) are adjusted to adapt to injection molding defect scenarios and reduce training costs.

[0085] The server first preprocesses the associated feature information: standardization (e.g., mean 0, variance 1) and dimensionality reduction (e.g., PCA) to obtain a set of associated working condition vectors adapted to the defect annotation parsing model. Then, the associated working condition vectors are input into the model, which uses lightweight weighted branches to compare the correlation between current and historical features (e.g., the current pressure stability sub-vector value is close to the sub-vector value of historical "weld mark" batches), and combines this with historical defect annotation results to infer the current defect risk. For example, if the model identifies that the "weld mark risk sub-vector" value in the current first-stage mapping vector set is in the medium range (approximately 0.6-0.8), and the real-time injection speed is slightly lower than that of historical defect-free batches, it comprehensively determines that the current molding process has a target molding defect annotation result of "potential weld mark (cavity 2, medium risk level)".

[0086] Generate process parameter adjustment instructions and adjust the molding process:

[0087] Based on the identified molding defect annotations, the server generates and executes process parameter adjustment instructions for the injection molding equipment, adjusting the process parameters of the current or subsequent molding processes in real time.

[0088] The server queries a pre-defined "defect-parameter adjustment rule library," which stores typical adjustment strategies corresponding to different defect types and risk levels. For example, for "weld marks (medium risk)," the rule library might suggest the following adjustment strategies: appropriately increase the melt temperature (e.g., 5-10°C to improve material flowability), increase the injection speed (e.g., 5%-10% to shorten filling time), and extend the holding pressure time (e.g., 1-2 seconds to replenish the melt). Combining the current equipment status and material characteristics (e.g., material heat resistance, equipment heating capacity), the server selects specific adjustment parameters (e.g., increasing melt temperature by 6-8°C, increasing injection speed by 7-8%), and generates an adjustment instruction, including the parameter name, target value, and execution timing (e.g., "effective in the next molding cycle").

[0089] The server sends adjustment commands to the injection molding machine controller via an industrial bus (such as Profinet or Modbus). The controller then performs parameter adjustments during a specified molding cycle (e.g., the next molding cycle): the heating system increases power to raise the melt temperature, the servo motor adjusts its speed to increase the injection speed, and the timer extends the holding pressure stage. After adjustment, the server continuously monitors the molding process data and product quality for subsequent molding cycles (e.g., through visual inspection or manual sampling) to verify whether defects have been improved (e.g., weld lines changing from "potentially visible" to "invisible," conforming to industry standards).

[0090] The above process integrates data collection, feature extraction, historical matching, defect judgment and parameter adjustment through the server, realizing intelligent optimization of the embedded molding injection process. The specific values ​​and model structures involved are for illustrative purposes only, and can be adjusted according to product type and equipment characteristics in actual applications.

[0091] In this embodiment of the invention, the first multidimensional process feature includes a first process timing feature, and the second multidimensional process feature includes a second process timing feature. The construction of associated feature information based on the second multidimensional process feature corresponding to each historical molding process data, the molding defect labeling result corresponding to each historical molding process data, and the first multidimensional process feature of the current molding process data can be implemented through the following example.

[0092] A first-stage mapping vector group is obtained by performing working condition mapping modeling on the first process timing features, and a second-stage mapping vector group is obtained by performing working condition mapping modeling on the second process timing features corresponding to each historical molding process data.

[0093] Obtain the first process description data corresponding to the current molding process data and the second process description data corresponding to each of the historical molding process data;

[0094] The associated feature information is constructed based on the second stage mapping vector group, the second process description data, and the molding defect annotation results corresponding to each of the historical molding process data, as well as the first stage mapping vector group and the first process description data.

[0095] In this embodiment of the invention, taking the embedded injection molding production of automotive electronic sensor housings as an example, the server executes the following process for constructing associated feature information:

[0096] The process timing characteristics are modeled using operating condition mapping to obtain a set of stage mapping vectors:

[0097] The server, based on the first process timing features of the current molding process data (including timing parameters such as peak pressure of 145-155 MPa in the injection stage, pressure decay rate of 0.4-0.6 MPa / s in the holding stage, and temperature gradient of 3-5℃ / s in the cooling stage), calls a preset feature projection unit (a domain transformation model trained based on injection molding defect scenarios) to perform condition mapping modeling. This model projects the timing features from the original process parameter domain to a unified preset condition semantic domain (a 512-dimensional space including defect risk assessment dimensions), outputting a first-stage mapping vector set. This vector set includes: a pressure stability sub-vector (describing the impact of injection pressure fluctuations on weld lines, current value 0.65-0.75), a temperature uniformity sub-vector (describing the impact of melt temperature distribution on shrinkage marks, current value 0.70-0.80), and a cooling efficiency sub-vector (describing the impact of cooling rate on warpage, current value 0.80-0.90), etc., with an overall dimension of 512.

[0098] Simultaneously, the server performs the same mapping modeling on the second process timing characteristics (injection peak pressure of 148MPa and holding pressure decay rate of 0.55MPa / s for historical batch A; injection peak pressure of 152MPa and holding pressure decay rate of 0.45MPa / s for historical batch B, etc.) of the three highly correlated historical molding process data (such as historical batches A, B, and C), respectively, to obtain second-stage mapping vector groups. For example, historical batch A has a "weld mark (cavity 2)" defect, and its pressure stability sub-vector value in the second-stage mapping vector group reaches 0.82 (higher than the current 0.72); historical batch B is marked with "shrinkage mark (cavity 1)," and its temperature uniformity sub-vector value is 0.65 (lower than the current 0.75).

[0099] Obtain current and historical process description data:

[0100] The server generates the first process description data of the current molding process, which integrates information such as product model "SensorCase-001", real-time process parameters (melt temperature 250-255℃, injection speed 80-90mm / s, current production mold number 120-130), and equipment status "normal operation (no alarm)", and stores it in the form of a structured dictionary.

[0101] Simultaneously, the server retrieves the second process description data of three highly correlated historical batches from the historical database: the second process description data of historical batch A includes melt temperature of 248-252℃, injection speed of 75-85mm / s, production mold number 90-100, and the molding defect labeling result "weld mark (cavity 2, high risk level)"; the second process description data of historical batch B includes melt temperature of 253-257℃, injection speed of 85-95mm / s, production mold number 110-120, and the labeling result "shrinkage mark (cavity 1, medium risk level)"; the second process description data of historical batch C includes melt temperature of 250-254℃, injection speed of 80-90mm / s, and the labeling result "no defects".

[0102] Integrating multi-source information to construct associated feature information:

[0103] The server first constructs a "process instance" for each historical batch. Taking historical batch A as an example, its process instance includes: a second-stage mapping vector group (512-dimensional, including a pressure stability sub-vector of 0.82), second-stage process description data (melt temperature 248-252℃, injection speed 75-85mm / s), and the molding defect annotation result "weld mark (cavity 2, high risk level)". A first process identifier, "Historical Batch A_Weld Mark Defect", is added to distinguish different process stages within the same batch. Similarly, process instances for historical batches B and C are constructed, with identifiers "Historical Batch B_Shrinkage Mark Defect" and "Historical Batch C_No Defect", respectively.

[0104] Subsequently, the server concatenates the three historical operating condition instances in order of comprehensive matching degree (historical batches A > B > C), and uses the second operating condition identifier "batch separator" to separate the different batch instances, forming a 1536-dimensional "historical comprehensive feature block" (3 instances × 512 dimensions).

[0105] Finally, the server combines the current first-stage mapping vector group (512 dimensions) with the first process description data (structured dictionary, including melt temperature 250-255℃, injection speed 80-90mm / s, production cycle 120-130) to generate the "current working condition sample" and adds the first working condition identifier "current production_target sample".

[0106] The server uses a feature alignment algorithm (converting the structured dictionary into a 128-dimensional feature vector) to concatenate the "current operating condition sample" (512+128=640 dimensions) with the "historical comprehensive feature block" (1536 dimensions), unifying the dimensions to 2176 dimensions. This forms associated feature information that includes the current and historical process time-series feature mapping vectors, process description data, and historical defect annotation results, which is used for subsequent defect determination.

[0107] Through the above process, the server achieves multi-dimensional correlation between the current molding process and historical defect data, laying the foundation for accurately determining the current defect risk.

[0108] In this embodiment of the invention, the number of channels of the first process timing feature is multiple; the first stage mapping vector group is obtained by performing working condition mapping modeling on the first process timing feature, which can be implemented through the following example.

[0109] Parameter integration mapping is performed on multiple first process timing features to obtain target process stage features;

[0110] The target process stage features are projected onto a preset working condition semantic domain by a preset feature projection unit to obtain the first stage mapping vector group.

[0111] In an embodiment of the present invention, for example, during the current molding process, the first process timing features collected by the server include three channels: Channel 1 is the timing features of the injection stage (including 100Hz sampling data of injection pressure, injection speed, and melt temperature, with a duration of 5s and a total of 500 sampling points), Channel 2 is the timing features of the holding pressure stage (including 100Hz sampling data of holding pressure, holding time, and temperature fluctuation, with a duration of 7s and a total of 700 sampling points), and Channel 3 is the timing features of the cooling stage (including 100Hz sampling data of mold temperature, cooling time, and screw position, with a duration of 18s and a total of 1800 sampling points), with a total of three channels (multiple channels).

[0112] The server performs parameter integration mapping on the first process timing features of the three channels: for channel 1, it extracts the peak injection pressure of 150 MPa, the average velocity of 85 mm / s, and the temperature variance of 2.5℃²; for channel 2, it extracts the holding pressure decay rate of 0.5 MPa / s and the holding time percentage of 25% (7s / 30s); for channel 3, it extracts the mold temperature gradient of 4℃ / s and the cooling efficiency of 0.85 (normalized to 0-1). These seven key parameters are then weighted and fused (injection stage weight 0.4, holding pressure weight 0.3, cooling pressure weight 0.3) to obtain 64-dimensional target process stage features, encompassing integrated process parameters for the entire injection-holding-cooling process.

[0113] Subsequently, the server invokes a preset feature projection unit (a GAN domain transformation model trained based on injection molding defect scenarios) to project the 64-dimensional target process stage features onto a preset working condition semantic domain (512 dimensions, including a defect risk assessment dimension). This projection process maps process parameters to defect-related semantic features through a domain adaptation algorithm, outputting a first-stage mapping vector group (512 dimensions), including a pressure stability sub-vector (0.72, describing weld line risk), a temperature uniformity sub-vector (0.78, describing shrinkage mark risk), and other defect-related sub-vectors, completing the working condition mapping modeling.

[0114] In this embodiment of the invention, the construction of the associated feature information based on the second stage mapping vector group, the second process description data, and the molding defect annotation result corresponding to each of the historical molding process data, as well as the first stage mapping vector group and the first process description data, can be implemented through the following example.

[0115] Based on the first working condition identifier used to distinguish different information under the same molding process data, and the second stage mapping vector group, the second process description data and the molding defect annotation result corresponding to the same historical molding process data, a working condition instance is constructed to obtain multiple working condition instances corresponding to multiple historical molding process data.

[0116] The multiple working condition instances are spliced ​​together to obtain comprehensive feature information, wherein the multiple working condition instances in the comprehensive feature information are demarcated by a second working condition identifier used to distinguish process description data of different molding processes.

[0117] Construct a target operating condition sample based on the first operating condition identifier, the first stage mapping vector group, and the first process description data;

[0118] The associated feature information is generated based on the comprehensive feature information and the target working condition sample.

[0119] In this embodiment of the invention, for example, the server constructs a working condition instance for three highly correlated historical molding process data (historical batches A, B, and C). A first working condition identifier is used to distinguish different information within the same historical batch. For example, the first working condition identifier for historical batch A (labeled "weld mark (cavity 2)") is set to "Historical A_Injection Stage_Weld Mark Working Condition," for historical batch B (labeled "shrinkage mark (cavity 1)") it is "Historical B_Pressure Holding Stage_Shrinkage Mark Working Condition," and for historical batch C (labeled "No Defects") it is "Historical C_Cooling Stage_Normal Working Condition." Each working condition instance includes: a second-stage mapping vector group (512 dimensions, such as the pressure stability sub-vector of historical A being 0.82), second-process description data (melt temperature of historical A: 248℃, injection speed: 75mm / s), and molding defect labeling results (such as "weld mark (cavity 2, high risk)").

[0120] The server sorts and concatenates the three working condition instances according to their overall matching degree, and uses the second working condition identifier ("batch separator_history A" "batch separator_history B") as the boundary to form comprehensive feature information (3×512=1536 dimensions).

[0121] The first working condition identifier of the current molding process is set as "Current Production_Full Stage_Target Working Condition". The server combines the first stage mapping vector group (512-dimensional, pressure stability sub-vector 0.72) and the first process description data (current melt temperature 252℃, injection speed 85mm / s, production mold number 120) to construct the target working condition sample (512-dimensional vector + structured description data).

[0122] Finally, the server combines the comprehensive feature information (1536 dimensions) with the target working condition sample (512 dimensions + structured data) using a feature alignment algorithm, unifying the dimensions to 2048 dimensions, and generating associated feature information, including current and historical process mapping features, descriptive data, and defect annotation relationships.

[0123] In this embodiment of the invention, the step of obtaining the first process description data corresponding to the current molding process data and the second process description data corresponding to each of the historical molding process data can be implemented through the following examples.

[0124] The molding process data process task identifier of the current molding process data is obtained, as well as the process parameters obtained by converting the sensor data of the current molding process data;

[0125] The process status of the current molding process data is generated based on the operating conditions of the current molding process data.

[0126] The first process description data is generated based on the molding process data, the process task identifier, the process parameters, and the process status.

[0127] In an embodiment of the present invention, for example, the server first obtains the molding process data process task identifier of the current molding process data, which is “SensorCase_20240701_10”, corresponding to the product model “Automotive Electronic Sensor Housing A”, and is used to uniquely identify the current production task.

[0128] Subsequently, the server collects real-time data from the injection molding machine's sensor system and converts it into process parameters: the peak injection pressure is 152MPa and the holding pressure is 102MPa, obtained through pressure sensor data conversion; the melt temperature is 253℃ and the mold temperature is 85℃, obtained through temperature sensor data conversion; the injection speed is 88mm / s and the screw stroke is 195mm, obtained through displacement sensor data conversion; and the current molding cycle is 30.5s, obtained through timer data conversion.

[0129] Next, the server generates the process status based on the current molding process data: the current production batch is the 150th batch, the equipment controller reports no alarm code (alarm code "0000"), the material barrel melt state monitoring value is 0.92 (standardized 0-1, 1 is the best), and the overall process status is determined to be "normal production (batch 150, no equipment alarm, material melt state is good)".

[0130] Finally, the server integrates the molding process data, including the process task identifier "SensorCase_20240701_10", process parameters (injection peak pressure 152MPa, melt temperature 253℃, etc.), and process status "normal production (150th mold...)", to generate the first process description data, which is stored in the form of a structured dictionary, containing the task identifier, real-time parameter list, and production status description.

[0131] In this embodiment of the invention, the first multidimensional process feature includes target process stage features and first working condition semantic features; the implementation can be carried out by matching multiple historical molding process data that are highly correlated with the current molding process data according to the first multidimensional process feature.

[0132] Obtain a pre-established process stage feature library and a working condition semantic feature library. The process stage feature library includes the feature association relationship between each candidate molding process data and the process stage features corresponding to the candidate molding process data. The working condition semantic feature library includes the feature association relationship between each candidate molding process data and the working condition semantic features corresponding to the candidate molding process data.

[0133] Based on the target process stage characteristics, the process stage feature library and the working condition semantic feature library are matched respectively, and based on the first working condition semantic feature, the process stage feature library and the working condition semantic feature library are matched respectively, to obtain multiple target candidate molding process data that are highly correlated with the current molding process data;

[0134] Based on the matching degree of molding process data between the multiple target candidate molding process data and the current molding process data, the historical molding process data is extracted from the multiple target candidate molding process data.

[0135] In an embodiment of the invention, for example, the first multi-dimensional process feature of the current molding process includes target process stage features (64 dimensions, integrating key parameters of the injection, holding, and cooling stages, such as peak injection pressure of 150 MPa, holding pressure decay rate of 0.5 MPa / s, and mold temperature gradient of 4 °C / s) and first working condition semantic features (256 dimensions, including semantic encoding of material flowability grade of 0.85, mold complexity score of 0.6, and equipment stability index of 0.9). The server matches highly correlated historical data according to the following steps:

[0136] Obtain a pre-established library of process features:

[0137] The server calls two feature libraries:

[0138] Process Stage Feature Library: Stores the correlation between 800+ alternative molding process data (historical batches) of similar products (sensor housings) from the past year and their corresponding process stage features. Each record includes: alternative data identifiers (such as "LOT20240301" and "LOT20240415"), and 64-dimensional process stage features (such as the process stage features of historical batch A including injection peak pressure of 148MPa and holding pressure decay rate of 0.55MPa / s).

[0139] Working condition semantic feature library: Stores the association between the same candidate data and the corresponding working condition semantic features. Each record contains: candidate data identifier, 256-dimensional working condition semantic features (e.g., the semantic features of historical batch A include material flowability 0.85 and mold complexity 0.6, which are consistent with the current one).

[0140] Multi-feature cross-matching of target candidate data:

[0141] The server matches two feature libraries respectively using the target process stage features and the semantic features of the first working condition to filter target candidate molding process data:

[0142] Target process stage feature matching:

[0143] Matching process stage feature library: Calculate the cosine similarity between the current 64-dimensional target process stage features and the process stage features of each candidate data in the library, filter candidate data with similarity ≥ 0.75, and obtain historical batches A (similarity 0.85), B (0.82), C (0.78), and D (0.75).

[0144] Matching the semantic feature library of working conditions: Although the semantic feature library of working conditions stores semantic features, similarity is calculated by feature transformation (mapping the process stage features to semantic feature dimensions) to filter out historical batches A (0.76) and B (0.74, below the threshold), and only A is retained.

[0145] First working condition semantic feature matching:

[0146] Matching the semantic feature library of working conditions: Calculate the cosine similarity between the current 256-dimensional semantic features of the first working condition and the semantic features of each candidate data in the library, filter the candidate data with similarity ≥ 0.8, and obtain the historical batches A (0.90), B (0.88), E (0.85), and F (0.80).

[0147] Matching process stage feature library: similarity is calculated by mapping semantic features to process stage feature dimensions, and historical batches A (0.82) and B (0.80) are selected.

[0148] After merging and deduplication, the target candidate molding process data are A, B, C, D, E, and F (6 batches).

[0149] Extracting historical molding process data based on matching degree:

[0150] The server calculates the overall matching degree between each target candidate data and the current forming process data:

[0151] Single feature matching degree weighting: target process stage feature matching degree weight 0.6 (process parameters have a more direct impact on defects), first working condition semantic feature matching degree weight 0.4 (semantic information provides scenario constraints).

[0152] Overall matching degree calculation:

[0153] Historical Batch A: Process stage matching degree 0.85 (from process stage feature library) + semantic matching degree 0.90 (from working condition semantic feature library) → Overall matching degree = 0.85×0.6 + 0.90×0.4 = 0.87;

[0154] Historical Batch B: Process stage matching degree 0.82 + semantic matching degree 0.88 → overall matching degree = 0.82 × 0.6 + 0.88 × 0.4 = 0.844;

[0155] Historical batch E: process stage matching degree 0.72 (failed process stage library screening but semantic library matching) + semantic matching degree 0.85 → overall matching degree = 0.72×0.6 + 0.85×0.4 = 0.772.

[0156] Sort by overall matching degree from high to low, the server extracts the top 3 (A, B, E) as historical molding process data with high correlation to the current molding process data, and their molding defect labeling results are "weld marks (cavity 2)", "shrinkage marks (cavity 1)" and "no defects" respectively.

[0157] In this embodiment of the invention, the plurality of target candidate molding process data includes a plurality of target candidate molding process data obtained by matching each matching mode; the step of extracting the historical molding process data from the plurality of target candidate molding process data based on the molding process data matching degree between the plurality of target candidate molding process data and the current molding process data can be implemented through the following example.

[0158] For each matching mode, the matching degree between the current molding process data and each target candidate molding process data is calculated.

[0159] For each target candidate molding process data, calculate the comprehensive matching degree between the target candidate molding process data and the current molding process data in each matching mode;

[0160] The historical molding process data is extracted from the multiple target candidate molding process data based on the comprehensive matching degree.

[0161] In an embodiment of the invention, for example, the first multi-dimensional process feature of the current molding process includes target process stage features (64 dimensions, integrating parameters for all stages such as peak injection pressure of 150MPa, holding pressure decay rate of 0.5MPa / s, and mold temperature gradient of 4℃ / s) and first working condition semantic features (256 dimensions, including semantic encoding of material flowability grade of 0.85, mold complexity score of 0.6, and equipment stability index of 0.9). The server matches highly correlated historical data according to the following steps:

[0162] Determine the matching pattern and calculate the single-pattern matching degree:

[0163] The server calculates the matching degree between the current molding process data and the target candidate molding process data (historical batches A, B, C, and D) for four matching modes:

[0164] Mode 1 (Target Process Stage Feature Matching to Process Stage Feature Library): The process stage feature library stores the process stage features (64 dimensions) of historical batches. The server calculates the cosine similarity between the current 64-dimensional target process stage features and each candidate data in the library. For example: historical A has a matching degree of 0.85 (injection pressure 148MPa, holding pressure decay rate 0.55MPa / s, close to the current one), B has 0.82, C has 0.78, and D has 0.75 (all ≥ 0.75 threshold).

[0165] Mode 2 (Target Process Stage Feature Matching Operating Condition Semantic Feature Library): The features of the current target process stage are mapped to semantic feature dimensions through domain transformation, and the similarity is calculated with the semantic features of historical batches in the operating condition semantic feature library. For example: Historical A has a matching degree of 0.76 (the process parameters are close to the semantic features of historical A after semantic mapping), while B has a matching degree of 0.74 (<0.75 threshold, so it is discarded).

[0166] Mode 3 (First Working Condition Semantic Feature Matching Working Condition Semantic Feature Library): The working condition semantic feature library stores 256-dimensional semantic features from historical batches. The server calculates the cosine similarity between the current semantic feature and each candidate data in the library. For example: Historical A has a matching degree of 0.90 (material and mold parameters are completely consistent), B has 0.88, C has 0.79 (<0.8 threshold, discarded), and D has 0.82.

[0167] Mode 4 (First Working Condition Semantic Feature Matching Process Stage Feature Library): Maps the current first working condition semantic features to process stage feature dimensions, and calculates the similarity with the process stage features of historical batches in the process stage feature library. For example: historical A has a matching degree of 0.82, while B has a matching degree of 0.80 (≥0.8 threshold).

[0168] Calculate the overall matching degree of the target candidate data:

[0169] The server assigns weights to each pattern (pattern 1 with a weight of 0.3, pattern 3 with a weight of 0.4, and patterns 2 and 4 each with a weight of 0.15, as direct matching to the corresponding database is more accurate), and calculates the overall matching degree of each target candidate data using a weighted average.

[0170] History A: Pattern 1 (0.85×0.3) + Pattern 2 (0.76×0.15) + Pattern 3 (0.90×0.4) + Pattern 4 (0.82×0.15) = 0.255 + 0.114 + 0.36 + 0.123 = 0.852;

[0171] History B: Pattern 1 (0.82×0.3) + Pattern 3 (0.88×0.4) + Pattern 4 (0.80×0.15) = 0.246 + 0.352 + 0.12 = 0.718 (Pattern 2 did not meet the standard and is not included);

[0172] Historical D: Pattern 1 (0.75×0.3) + Pattern 3 (0.82×0.4) = 0.225 + 0.328 = 0.553 (Patterns 2 and 4 did not meet the standard).

[0173] Extracting historical data based on overall matching score:

[0174] The server sorts data by overall matching degree: Historical A (0.852) > Historical B (0.718) > Historical D (0.553). Historical batches A and B with an overall matching degree ≥ 0.7 are selected as historical molding process data with high correlation to the current molding process data. Among them, Historical A is marked "weld mark (cavity 2)" and Historical B is marked "shrinkage mark (cavity 1)" to provide a basis for subsequent defect judgment.

[0175] In this embodiment of the invention, the step of extracting multi-dimensional process features from the current molding process data to obtain the first multi-dimensional process feature can be implemented through the following example.

[0176] Multiple molding process data sampling points are extracted from the current molding process data, and the multiple molding process data sampling points are segmented into multiple data units;

[0177] Target molding process data sampling points are extracted from the multiple data units, and process stage feature extraction is performed on the target molding process data sampling points to obtain the first process timing feature;

[0178] Obtain the first process description data corresponding to the current molding process data, and extract the first working condition semantic features from the first process description data.

[0179] The first multidimensional process feature is obtained based on the first process timing feature and the first operating condition semantic feature.

[0180] In this embodiment of the invention, for example, during the current molding process, the dynamic timing data collected by the server consists of real-time process parameters within a 30-second molding cycle, with a sampling frequency of 100Hz, generating a total of 3000 molding process data sampling points (multiple sampling points). The server divides the sampling points into three data units according to the molding stage: injection stage (0-5 seconds, 500 sampling points, corresponding to the melt filling cavity process), holding pressure stage (5-12 seconds, 700 sampling points, corresponding to the melt shrinkage process), and cooling stage (12-30 seconds, 1800 sampling points, corresponding to the product solidification process), with the number of data units consistent with the molding stage.

[0181] Extracting the timing features of the first process:

[0182] The server extracts target molding process data sampling points from three data units: For the injection stage data unit, the peak injection pressure point (150MPa at 2.3 seconds), average speed point (85mm / s), and temperature fluctuation variance (2.5℃²) are extracted using the gradient threshold method (pressure change rate > 30MPa / s); for the holding pressure stage data unit, the holding pressure decay rate (0.5MPa / s) and holding time percentage (25%, 7 seconds / 30 seconds) are extracted through decay curve fitting; for the cooling stage data unit, the mold temperature gradient (4℃ / s) and cooling efficiency (0.85, normalized to 0-1) are extracted through temperature gradient calculation. These seven key parameters are then used to extract temporal features using a CNN-LSTM model, resulting in 128-dimensional first-process temporal features, including injection stage pressure fluctuation features (e.g., peak pressure 150MPa, average speed 85mm / s), holding pressure stage stability features (e.g., decay rate 0.5MPa / s), and cooling stage efficiency features (e.g., temperature gradient 4℃ / s).

[0183] Extract semantic features for the first working condition:

[0184] The server retrieves the first process description data corresponding to the current molding process data, including: the molding process data process task identifier "SensorCase_20240615_08" (uniquely identifying the current production task); material parameters (PA66+GF30, melting point 255℃, MFI 25g / 10min); mold parameters (2 cavities, gate diameter 1.2mm, 4 cooling water channels); and equipment status (spindle motor load rate 65%, temperature control accuracy ±1℃). After structuring the first process description data, semantic features of the working condition are extracted using a BERT pre-trained model: semantic information such as material flowability level (0.85), mold complexity score (0.6), and equipment stability index (0.9) are encoded into a 256-dimensional first working condition semantic feature vector, realizing the vectorization of unstructured process description.

[0185] Combined first multi-dimensional process features:

[0186] The server combines the 128-dimensional first-process time-series features (time-series dynamic parameters) with the 256-dimensional first-condition semantic features (static semantic information) using a feature concatenation algorithm to form a 384-dimensional first-multidimensional process feature vector. This vector simultaneously includes dynamic process parameter features of the molding process (such as injection pressure and holding pressure decay rate) and static scene semantic features (such as material flowability and mold complexity), providing comprehensive feature support for subsequent matching with historical data.

[0187] In this embodiment of the invention, the step of determining the molding defect labeling result of the current molding process data based on the associated feature information can be implemented through the following example.

[0188] The associated feature information is reconstructed into working condition vectors to obtain an associated working condition vector group, wherein the working condition vectors are the input features adapted to the defect annotation parsing model;

[0189] The associated working condition vector group is input into the defect annotation parsing model. The defect annotation parsing model is obtained by locking the core parameter set of the preset inference model and adjusting the adjustable parameter unit of the inference model according to the training associated working condition vector group. The adjustable parameter unit is related to the lightweight weight branch embedded in the inference model.

[0190] Obtain the target molding defect annotation results of the current molding process data output by the defect annotation parsing model.

[0191] In an embodiment of the present invention, for example, during the current molding process, the server has constructed associated feature information (2048-dimensional vector), including the first stage mapping vector group of the current molding process (512-dimensional, pressure stability sub-vector 0.72, temperature uniformity sub-vector 0.75), the first process description data (melt temperature 252℃, injection speed 85mm / s, production mold number 120), and the comprehensive feature information (1536-dimensional, including the annotation results and mapping features of historical A "weld mark (cavity 2, high risk)" and historical B "shrinkage mark (cavity 1, medium risk)").

[0192] The associated working condition vector group is obtained by reconstructing the working condition vector:

[0193] The server reconstructs the working condition vector from the 2048-dimensional associated feature information: First, it performs standardization (adjusting the feature values ​​to a mean of 0 and a variance of 1 to eliminate the influence of dimensions), then uses PCA (Principal Component Analysis) to retain 95% of the feature contribution rate, compressing the dimension from 2048 to 512, thus obtaining the associated working condition vector set adapted to the defect annotation parsing model. This vector set contains the correlation between current and historical process features, such as: pressure stability sub-vector (0.72, reflecting current injection pressure fluctuations), temperature uniformity sub-vector (0.75, reflecting melt temperature distribution), and historical weld line defect correlation sub-vector (0.88, reflecting similarity to historical weld line A feature), etc., totaling 512 dimensions, which are directly used as model input.

[0194] Input defect annotation parsing model:

[0195] The server invokes a defect annotation and parsing model, which is built on the ResNet-50 inference model. The core parameter set (the first 49 convolutional layers, responsible for basic feature extraction) has been pre-trained and locked. Only the lightweight weight branch (adjustable parameter unit) after the last fully connected layer is retained for adjustment. The lightweight weight branch contains 5120 parameters (512 input dimensions × 10 defect categories, such as "weld marks", "shrinkage marks", "flashes", etc.), which are directly related to defect parsing tasks such as weld mark risk assessment and shrinkage mark location determination. During training, the weight of this branch is adjusted by historical associated working condition vector groups (including annotation results) to adapt to injection molding defect scenarios.

[0196] The server inputs a 512-dimensional associated working condition vector group into the model: The model extracts general features (such as pressure fluctuation curves and temperature distribution gradients) through a pre-locked core parameter layer, and then focuses on defect-related features through lightweight weighted branches (such as the cosine similarity of the current pressure stability sub-vector 0.72 and the historical A weld line sub-vector 0.82 reaching 0.91, which is higher than the similarity of other defects such as shrinkage marks and flash (0.65-0.75), to achieve accurate matching between defect categories and risk levels.

[0197] Obtain the target forming defect annotation results:

[0198] The defect annotation and parsing model outputs the defect probability distribution of the current molding process through the softmax activation function: "weld mark" probability 0.72 (cavity 2), "shrinkage mark" probability 0.21 (cavity 1), and "no defect" probability 0.07. The server selects the defect category with the highest probability, combines it with the risk level assessment (current pressure stability subvector 0.72 < historical A 0.82, the risk level is determined to be "medium"), and outputs the target molding defect annotation result: "weld mark (potential, cavity 2, medium risk level)", thus completing the defect determination of the current molding process.

[0199] In this embodiment of the invention, the training steps of the defect annotation parsing model can be implemented through the following example.

[0200] Acquire the first training process stage features, first training process description data, and first training forming defect annotation results corresponding to the first training forming process data, as well as the second training process stage features, second training process description data, and second training forming defect annotation results corresponding to the second training forming process data.

[0201] According to the preset feature projection unit, the first training process stage features and the second training process stage features are respectively subjected to domain transformation processing to obtain the first training condition mapping vector group and the second training condition mapping vector group. Training-related feature information is constructed based on the first training condition mapping vector group, the first training process description data and the first training forming defect annotation result, as well as the second training condition mapping vector group and the second training process description data.

[0202] The training-related feature information is reconstructed into working condition vectors to obtain the training-related working condition vector group.

[0203] The lightweight weighted branch is embedded in the inference model to embed the adjustable parameter unit in the inference model through the lightweight weighted branch;

[0204] The core parameter set of the inference model is locked, and the adjustable parameter unit of the inference model is adjusted according to the training associated working condition vector group and the second training formed defect annotation result to obtain the defect annotation parsing model.

[0205] In this embodiment of the invention, taking the training of a defect annotation and analysis model for the injection molding production of automotive electronic sensor housings as an example, the server executes the following steps. The training data, model parameters, etc. involved are all illustrative examples, intended to clearly demonstrate the training process.

[0206] Obtain training and shaping process data and corresponding feature and annotation results:

[0207] The server retrieves two types of training and shaping process data from the historical database:

[0208] First training and shaping process data: Select historical batches with clearly marked defects (e.g., "LOT20240310", abbreviated as "training batch X"), including:

[0209] The first training process stage features 64 dimensions, integrating all process parameters such as peak injection pressure of 148MPa, holding pressure decay rate of 0.55MPa / s, and mold temperature gradient of 4.2℃ / s.

[0210] First training process description data: material type "PA66+GF30", melt temperature 248℃, injection speed 75mm / s, production mold number 90;

[0211] The first training molding defect labeling result: "Weld line (severe, cavity 2, high risk level)" (defect type, location and risk confirmed by manual re-inspection).

[0212] Second training and shaping process data: Select historical batches with no defects or other defects (e.g., "LOT20240420", abbreviated as "training batch Y"), including:

[0213] Characteristics of the second training process stage: 64 dimensions, peak injection pressure 152MPa, holding pressure decay rate 0.45MPa / s, mold temperature gradient 3.8℃ / s;

[0214] Second training process description data: material type "PA66+GF30", melt temperature 255℃, injection speed 90mm / s, production mold number 150;

[0215] The second training result for defect labeling: "No defects" (visual inspection + manual sampling confirmation).

[0216] A total of 1000+ sets of training data (including 500+ defective samples and 500+ normal samples) were selected and divided into training set and validation set in a 7:3 ratio.

[0217] Constructing training-related feature information:

[0218] The server invokes a preset feature projection unit (a domain transformation model trained based on GAN, which has been pre-trained and adapted to the injection molding process domain) to perform domain transformation processing on the features from the first and second training process stages:

[0219] Project the 64-dimensional first training process stage features of training batch X onto the preset working condition semantic domain (512 dimensions, including defect risk assessment dimension) to obtain the first training working condition mapping vector group, in which the pressure stability sub-vector value reaches 0.82 (corresponding to high risk of weld lines).

[0220] The 64-dimensional second training process stage features of training batch Y are projected in the same way to obtain the second training condition mapping vector group, with a pressure stability sub-vector value of 0.35 (corresponding to low risk).

[0221] The server constructs training-related feature information based on the first training condition mapping vector group (512-dimensional), the first training process description data (melt temperature 248℃, injection speed 75mm / s), the first training molding defect annotation result ("weld line (severe, cavity 2)"), the second training condition mapping vector group (512-dimensional), and the second training process description data (melt temperature 255℃, injection speed 90mm / s): it constructs "weld line defect_training condition instance" (including the first training condition mapping vector group, process description data, and annotation result) for training batch X, and "no defect_training condition instance" for training batch Y, and splices them according to defect type (using "training batch separator" as the boundary) to form training-related feature information (each group of training data corresponds to a 512-dimensional mapping vector + structured description data).

[0222] Reconstruct the training associated working condition vector set:

[0223] The server reconstructs the working condition vector from the training-related feature information to adapt it to the model input format.

[0224] Standardization: The 512-dimensional training condition mapping vector group is standardized with "mean 0, variance 1" to eliminate the dimensional differences of different process parameters (such as pressure unit MPa, temperature unit ℃).

[0225] Dimensionality reduction: The PCA algorithm is used to retain 95% of the feature contribution rate, reducing the 512-dimensional vector to 256 dimensions (model input adaptation dimension), while retaining key features of defect risk (such as pressure stability sub-vector and temperature uniformity sub-vector).

[0226] Combining structured data: Key parameters (such as melt temperature and injection speed) in the training process description data are encoded into 64-dimensional feature vectors, which are then concatenated with the 256-dimensional mapping vectors after dimensionality reduction to form a 320-dimensional training-related working condition vector group (model input dimension).

[0227] Embedding lightweight weighted branches into the inference model:

[0228] The server uses ResNet-50 as the base inference model, which has been pre-trained on general image classification tasks and has strong feature extraction capabilities. The server locks the core parameter set of ResNet-50 (the first 49 convolutional layers, responsible for basic feature extraction, with approximately 23 million parameters) (freezing the weights and not participating in training), embedding only lightweight weight branches at the end of the model as adjustable parameter units.

[0229] Lightweight weighted branch structure: It contains 1 fully connected layer (inputting a 320-dimensional training associated working condition vector group, outputting a 10-dimensional defect category probability, corresponding to 10 common injection molding defects such as "weld marks", "shrinkage marks" and "flashes"), with a parameter scale of 320×10=3200 (0.014% of the core model parameters, lightweight design to reduce training cost).

[0230] Branch Function: Focus on injection molding defect analysis task, and learn the correlation between features such as "pressure stability sub-vector → weld line" and "temperature uniformity sub-vector → shrinkage mark" and defects by adjusting the weights.

[0231] Lock the core parameters and adjust the adjustable parameter units:

[0232] The server locks the core parameter set of the inference model (the first 49 layers of ResNet-50) (weight gradients are set to 0 and not updated), and adjusts the adjustable parameter units of the lightweight weight branches only by training the associated working condition vector set and the second training defect annotation results (labels such as "no defects").

[0233] Forward propagation: Input the 320-dimensional training associated working condition vector group into the model, extract basic features (such as pressure fluctuation curves and temperature distribution gradients) in the core parameter layer, and calculate the probability distribution of defect categories (such as "weld marks" probability 0.85 and "shrinkage marks" probability 0.10) in the lightweight weight branch.

[0234] Loss calculation: The cross-entropy loss function is used to compare the model output probability with the defect labeling results of the second training (such as the label [0,0,1,...,0] corresponding to "no defect") and calculate the loss value (initial loss value 1.2, target convergence to below 0.05);

[0235] Backpropagation: The Adam optimizer (learning rate 0.001, decay rate 0.9) is used to update the gradients of only the 3200 adjustable parameters of the lightweight weight branch for 5000 iterations (training rounds) until the loss function converges (final loss value 0.048).

[0236] Obtain the defect annotation analysis model:

[0237] After training, the server saves the optimal parameters (3200 weight values) of the lightweight weighted branch and combines them with the locked ResNet-50 core parameter set to obtain a defect annotation and parsing model. This model can directly input new associated working condition vector sets (such as the 320-dimensional vector of the current molding process) and output the probability distribution of defect categories and risk levels, achieving accurate annotation and parsing of injection molding defects.

[0238] Through the above steps, the server completes the training of the defect annotation and analysis model. The model achieves a defect recognition accuracy of 92% on the validation set (the recognition rate of major defects such as weld lines and shrinkage marks is >95%), meeting the real-time defect judgment requirements of injection molding production.

[0239] In this embodiment of the invention, before performing domain transformation processing on the process stage features according to the preset feature projection unit, the following implementation method is also provided.

[0240] Obtain a snapshot of the training process state and the corresponding training process description data;

[0241] The training process state snapshot is used to extract features to obtain training process state features, and the training process state features are input to the original feature projection unit so that the original feature projection unit projects the training process state features to the input preset working condition semantic domain of the inference model.

[0242] Obtain the training condition semantic features corresponding to the feature projection request, and input the training condition semantic features and the target training visual condition semantic features output by the feature projection unit into the inference model, wherein the model parameters of the inference model are locked.

[0243] Obtain the training prediction process description data output by the inference model, and train the original feature projection unit based on the training process description data and the training prediction process description data to obtain the feature projection unit.

[0244] In this embodiment of the invention, for example, before performing domain transformation processing on the process stage features, the server needs to first train the original feature projection unit so that it can accurately project the process stage features to the preset working condition semantic domain. The following uses the scenario of "weld-line defect recognition" in the injection molding production of automotive electronic sensor housings as an example to explain the training process in detail:

[0245] Obtain training process state snapshots and training process description data:

[0246] The server retrieves over 1000 sets of injection molding process training data from the historical database. Each set includes a snapshot of the training process status and corresponding training process description data.

[0247] Training process state snapshots: Select high-frequency time series data snapshots of key stages within the injection molding cycle, such as the pressure-velocity-temperature curve snapshots (including peak injection pressure of 145-155MPa and velocity of 80-90mm / s) during the injection stage (0-5s, 100Hz sampling, 500 points), the pressure decay curve snapshots (decay rate of 0.4-0.6MPa / s) during the holding pressure stage (5-12s, 700 points), and the temperature gradient curve snapshots (gradient of 3-5℃ / s) during the cooling stage (12-30s, 1800 points). Each snapshot is a 2000-dimensional time series vector (500+700+1800-dimensional splicing).

[0248] Training process description data: Structured process information corresponding to the snapshot, including material parameters (PA66+GF30, melting point 255℃, MFI 25g / 10min), mold parameters (2 cavities, gate diameter 1.2mm), equipment parameters (FANUCα-S50iA, heating coil power 3kW), and defect annotation association data (such as the pressure fluctuation characteristics of the snapshot corresponding to "weld line (cavity 2)"), stored in a structured dictionary.

[0249] Extract the state features of the training process and project them onto the preset working condition semantic domain:

[0250] The server extracts features from the training process state snapshots: a CNN-LSTM hybrid model is used to extract features from the 2000-dimensional time-series snapshot data, retaining 7 key parameters such as injection peak pressure (148MPa), pressure rise rate (35MPa / s), holding pressure decay rate (0.5MPa / s), and mold temperature gradient (4℃ / s), which are combined into 64-dimensional training process state features (consistent with the feature dimensions of subsequent process stages to ensure domain transformation adaptation).

[0251] The server calls the original feature projection unit (an untrained GAN domain transformation model, containing a generator G and a discriminator D, with 64-dimensional state features as input and 512-dimensional preset working condition semantic domain vector as output). The 64-dimensional training process state features are input into the generator G and projected onto the preset working condition semantic domain (512-dimensional, including the defect risk assessment dimension) through a randomly initialized weight matrix to obtain the target training visual working condition semantic features (initial projection result, with pressure stability sub-vector values ​​of 0.5-0.6, and low correlation with actual defect features).

[0252] The inference model with locked input parameters obtains the predicted process description data:

[0253] The server extracts training condition semantic features from the training process description data: the structured description data (materials, molds, equipment parameters) is encoded into a 256-dimensional semantic vector through the BERT model, which includes semantic information such as material flowability level 0.85, mold complexity 0.6, and equipment stability 0.9.

[0254] The server concatenates the 256-dimensional training process semantic features with the 512-dimensional target training visual process semantic features output by the original feature projection unit (unifying them to 512 dimensions through feature alignment), inputting a parameter-locked inference model (ResNet-50, with the first 49 layers of the core parameter set pre-trained and locked, used only for evaluating projection effects). Based on the input semantic and projection features, the model outputs training prediction process description data (e.g., predicting melt temperature of 240℃ and injection speed of 70mm / s, which deviates from the actual training process description data of 250℃ and 85mm / s).

[0255] Train the original feature projection units to obtain feature projection units:

[0256] The server uses the difference between the training prediction process description data and the actual training process description data as the loss signal to train the original feature projection unit:

[0257] Loss calculation: The MSE (mean squared error) loss function is used to compare the predicted process parameters (such as melt temperature 240℃) with the actual training process parameters (250℃) to calculate the loss value (initial loss value 15.6℃², target convergence to below 1.0℃²).

[0258] Backpropagation optimization: The weight matrix of the original feature projection unit generator G is adjusted using the Adam optimizer (learning rate 0.0005) to make the target training visual working condition semantic features closer to the distribution of the training working condition semantic features (e.g., improving the correlation between the pressure stability sub-vector and the weld line defect). After 5000 iterations, the loss value is reduced to 0.8℃², at which point the generator G can accurately project the 64-dimensional training process state features to the preset working condition semantic domain (e.g., the pressure stability sub-vector value of the weld line sample is increased from 0.55 to 0.82, and the matching degree with the training working condition semantic features reaches 0.92).

[0259] After training, the original feature projection unit is optimized into a feature projection unit, which has the ability to stably project process stage features (such as subsequent target process stage features) to the preset working condition semantic domain, providing reliable domain transformation support for the construction of subsequent associated feature information.

[0260] In this embodiment of the invention, the step of adjusting the adjustable parameter unit of the inference model according to the training associated working condition vector group to obtain the defect annotation parsing model can be implemented through the following example.

[0261] The training associated working condition vector group is input into the inference model embedded with the lightweight weight branch, and the predicted training defect labeling result output by the inference model is obtained.

[0262] Based on the difference between the predicted training defect labeling results and the second trained defect labeling results, the adjustable parameter units of the inference model are adjusted to obtain the defect labeling parsing model;

[0263] The method further includes:

[0264] Based on the difference between the predicted training defect annotation results and the second trained defect annotation results, the module parameters of the feature projection unit are adjusted.

[0265] In an embodiment of the present invention, for example, in the current training scenario, the server has constructed a training associated working condition vector group (512-dimensional, containing features of training batch X "weld marks (severe, cavity 2, high risk)" and training batch Y "no defects": training batch X pressure stability sub-vector 0.78, temperature uniformity sub-vector 0.65; training batch Y pressure stability sub-vector 0.35, temperature uniformity sub-vector 0.88), and embedded a lightweight weighted branch (5120 parameters, 512 inputs × 10 defect categories) into the inference model (ResNet-50, the first 49 layers of the core parameter set are locked).

[0266] Input the model to obtain the prediction training defect annotation results:

[0267] The server inputs a 512-dimensional training-related working condition vector set into the inference model embedded with a lightweight weighted branch. The model extracts basic features (such as pressure fluctuation curves and temperature gradients) through the core parameter layer, and the lightweight weighted branch calculates the defect probability distribution based on the initial weights. For example, for training batch X (actually labeled "weld mark (severe, cavity 2, high risk)"), the model outputs a predicted training defect labeling result of "weld mark (cavity 2, medium risk)" with a probability of 0.62; for training batch Y (actually labeled "no defect"), the prediction is "no defect" with a probability of 0.75. At this point, there is a difference between the prediction and the actual labeling (risk level deviation, insufficient probability).

[0268] Adjusting the adjustable parameter unit yields the defect annotation analytical model:

[0269] The server calculates the difference between the predicted training defect annotation results and the second training final defect annotation results (actual annotations): using the cross-entropy loss function, the loss value for training batch X is 1.2 (probability 0.62 vs. target 1.0), the loss value for training batch Y is 0.3 (probability 0.75 vs. target 1.0), and the total loss is 0.75. The Adam optimizer (learning rate 0.001) is used to backpropagate and update only the adjustable parameter units (5120 parameters) of the lightweight weight branch: focusing on adjusting the risk level weight of "weld mark - cavity 2" (increasing from 0.5 to 0.8) and the confidence weight of the "no defect" category (increasing from 0.6 to 0.9). After 3000 iterations, the prediction probability of training batch X increases to 0.92 (risk level "high"), the prediction probability of training batch Y reaches 0.95, and the total loss decreases to 0.05, obtaining the defect annotation parsing model.

[0270] Synchronously adjust the module parameters of the feature projection unit:

[0271] The server analysis revealed the root cause of the prediction discrepancy: the predicted risk level of training batch X was too low, stemming from the pressure stability sub-vector value of 0.78 in the target training visual condition semantic features output by the feature projection unit (lower than the actual high-risk feature value of 0.85 for weld lines). Therefore, the module parameters (weight matrix of generator G) of the feature projection unit were fine-tuned based on the loss value: the projection weights were adjusted using gradient descent, optimizing the pressure stability sub-vector of training batch X from 0.78 to 0.83 (closer to the actual high-risk feature), while maintaining the stability of other features such as the temperature uniformity sub-vector. After adjustment, the correlation between the training associated condition vector group and the actual defect features improved, and the model validation set accuracy increased from 85% to 92%, ensuring the coordinated optimization of feature projection and defect analysis.

[0272] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 performs the aforementioned method for adjusting parameters of embedded injection molded products. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0273] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for adjusting parameters of an embedded injection molded product, characterized in that, include: Acquire the current molding process data, and extract the first multi-dimensional process feature from the current molding process data; Based on the first multidimensional process feature matching, multiple historical molding process data with high correlation to the current molding process data are matched, and each historical molding process data has a corresponding molding defect annotation result; Based on the second multi-dimensional process features corresponding to each of the historical molding process data and the molding defect labeling results corresponding to each of the historical molding process data, as well as the first multi-dimensional process features of the current molding process data, we construct associated feature information. Based on the associated feature information, the forming defect labeling results of the current forming process data are determined; Based on the identified molding defect labeling results, corresponding injection molding equipment process parameter adjustment instructions are generated, and the process parameters of the current or subsequent embedded molding injection process are adjusted in real time according to the adjustment instructions. The first multidimensional process feature includes a first process timing feature, and the second multidimensional process feature includes a second process timing feature; The step of constructing associated feature information based on the second multi-dimensional process features corresponding to each of the historical molding process data, the molding defect annotation results corresponding to each of the historical molding process data, and the first multi-dimensional process features of the current molding process data includes: Parameter integration mapping is performed on multiple first process timing features to obtain target process stage features; The target process stage features are projected onto a preset working condition semantic domain by a preset feature projection unit to obtain the first stage mapping vector group. For each historical molding process data, the second process timing feature corresponding to the second process timing feature is respectively modeled by working condition mapping to obtain the second stage mapping vector group corresponding to each historical molding process data. The molding process data process task identifier of the current molding process data is obtained, as well as the process parameters obtained by converting the sensor data of the current molding process data; The process status of the current molding process data is generated based on the operating conditions of the current molding process data. The first process description data is generated based on the molding process data process task identifier, the process parameters, and the process status. Based on the first working condition identifier used to distinguish different information under the same molding process data, and the second stage mapping vector group, the second process description data and the molding defect annotation result corresponding to the same historical molding process data, a working condition instance is constructed to obtain multiple working condition instances corresponding to multiple historical molding process data. The multiple working condition instances are spliced ​​together to obtain comprehensive feature information, wherein the multiple working condition instances in the comprehensive feature information are demarcated by a second working condition identifier used to distinguish process description data of different molding processes. Construct a target operating condition sample based on the first operating condition identifier, the first stage mapping vector group, and the first process description data; The associated feature information is generated based on the comprehensive feature information and the target working condition sample.

2. The method according to claim 1, characterized in that, The first multidimensional process feature includes target process stage features and first operating condition semantic features; Based on the first multi-dimensional process feature, multiple historical molding process data with high correlation to the current molding process data are matched, including: Obtain a pre-established process stage feature library and a working condition semantic feature library. The process stage feature library includes the feature association relationship between each candidate molding process data and the process stage features corresponding to the candidate molding process data. The working condition semantic feature library includes the feature association relationship between each candidate molding process data and the working condition semantic features corresponding to the candidate molding process data. Based on the target process stage characteristics, the process stage feature library and the working condition semantic feature library are matched respectively, and based on the first working condition semantic feature, the process stage feature library and the working condition semantic feature library are matched respectively, to obtain multiple target candidate molding process data that are highly correlated with the current molding process data; Based on the matching degree of molding process data between the multiple target candidate molding process data and the current molding process data, the historical molding process data is extracted from the multiple target candidate molding process data.

3. The method according to claim 2, characterized in that, The multiple target candidate molding process data includes multiple target candidate molding process data obtained by matching each matching mode; the step of extracting the historical molding process data from the multiple target candidate molding process data based on the molding process data matching degree between the multiple target candidate molding process data and the current molding process data includes: For each matching mode, the matching degree between the current molding process data and each target candidate molding process data is calculated. For each target candidate molding process data, calculate the comprehensive matching degree between the target candidate molding process data and the current molding process data in each matching mode; The historical molding process data is extracted from the multiple target candidate molding process data based on the comprehensive matching degree.

4. The method according to claim 1, characterized in that, The step of extracting multi-dimensional process features from the current molding process data to obtain the first multi-dimensional process feature includes: Multiple molding process data sampling points are extracted from the current molding process data, and the multiple molding process data sampling points are segmented into multiple data units; Target molding process data sampling points are extracted from the multiple data units, and process stage feature extraction is performed on the target molding process data sampling points to obtain the first process timing feature; Obtain the first process description data corresponding to the current molding process data, and extract the first working condition semantic features from the first process description data. The first multidimensional process feature is obtained based on the first process timing feature and the first operating condition semantic feature.

5. The method according to claim 1, characterized in that, The step of determining the molding defect annotation result of the current molding process data based on the associated feature information includes: The associated feature information is reconstructed into working condition vectors to obtain an associated working condition vector group, wherein the working condition vectors are the input features adapted to the defect annotation parsing model; The associated working condition vector group is input into the defect annotation parsing model. The defect annotation parsing model is obtained by locking the core parameter set of the preset inference model and adjusting the adjustable parameter unit of the inference model according to the training associated working condition vector group. The adjustable parameter unit is related to the lightweight weight branch embedded in the inference model. Obtain the target molding defect annotation results of the current molding process data output by the defect annotation parsing model.

6. The method according to claim 5, characterized in that, The training steps of the defect annotation and parsing model include: Acquire the first training process stage features, first training process description data, and first training forming defect annotation results corresponding to the first training forming process data, as well as the second training process stage features, second training process description data, and second training forming defect annotation results corresponding to the second training forming process data. According to the preset feature projection unit, the first training process stage features and the second training process stage features are respectively subjected to domain transformation processing to obtain the first training condition mapping vector group and the second training condition mapping vector group. Training-related feature information is constructed based on the first training condition mapping vector group, the first training process description data and the first training forming defect annotation result, as well as the second training condition mapping vector group and the second training process description data. The training-related feature information is reconstructed into working condition vectors to obtain the training-related working condition vector group. The lightweight weighted branch is embedded in the inference model to embed the adjustable parameter unit in the inference model through the lightweight weighted branch; The core parameter set of the inference model is locked, and the adjustable parameter unit of the inference model is adjusted according to the training associated working condition vector group and the second training formed defect annotation result to obtain the defect annotation parsing model.

7. The method according to claim 6, characterized in that, Before performing domain transformation processing on the process stage features according to the preset feature projection unit, the method further includes: Obtain a snapshot of the training process state and the corresponding training process description data; The training process state snapshot is used to extract features to obtain training process state features, and the training process state features are input to the original feature projection unit so that the original feature projection unit projects the training process state features to the input preset working condition semantic domain of the inference model. Obtain the training condition semantic features corresponding to the feature projection request, and input the training condition semantic features and the target training visual condition semantic features output by the feature projection unit into the inference model, wherein the model parameters of the inference model are locked. Obtain the training prediction process description data output by the inference model, and train the original feature projection unit based on the training process description data and the training prediction process description data to obtain the feature projection unit.

8. The method according to claim 6, characterized in that, The step of adjusting the adjustable parameter units of the inference model based on the training associated working condition vector group to obtain the defect annotation parsing model includes: The training associated working condition vector group is input into the inference model embedded with the lightweight weight branch, and the predicted training defect labeling result output by the inference model is obtained. Based on the difference between the predicted training defect labeling results and the second trained defect labeling results, the adjustable parameter units of the inference model are adjusted to obtain the defect labeling parsing model; The method further includes: Based on the difference between the predicted training defect annotation results and the second trained defect annotation results, the module parameters of the feature projection unit are adjusted.

9. A server system, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-8.

Images