Thermosetting material detection method and device and electronic equipment

Abstract

The invention provides a thermosetting material detection method and device and electronic equipment, and the method comprises the steps: obtaining the original index parameters of a sample thermosetting material, introducing a double calibration module to carry out the preprocessing of the original index parameters, carrying out the preprocessing of the original data from two dimensions, can recognize and inhibit a high-uncertainty sample, and improves the detection precision. And the feature robustness is improved. And the model is guided to learn an abnormal boundary, the abnormal recognition capability of the model is enhanced, and the stability and generalization capability of subsequent model training are remarkably improved. An improved framework based on space-time fusion modeling and an anomaly scoring mechanism is constructed, the prediction capability and the anomaly recognition capability are considered, and particularly, excellent performance is shown in a high-missing-rate environment. In addition, quality control feedback adjustment can be carried out on the detection result, and real-time abnormal response is achieved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and specifically to a method, apparatus, and electronic device for testing thermosetting materials. Background Technology

[0002] Thermosetting materials are a class of polymeric materials that undergo irreversible cross-linking reactions upon heating or the addition of a curing agent, forming a three-dimensional network structure. This structure endows them with excellent mechanical strength, dimensional stability, heat resistance, and chemical corrosion resistance, making them widely used in electronic packaging, composite materials, automotive parts, aerospace structural components, adhesives (such as urea-formaldehyde resin, phenolic resin, and epoxy resin), and engineered wood products. Urea-formaldehyde resin, due to its good bonding properties, thermosetting properties, and cost-effectiveness, is widely used in engineered wood products, building adhesives, and decorative materials. However, its production process is characterized by large quality fluctuations and complex process parameters, making it highly susceptible to factors such as batch differences in raw materials, fluctuations in environmental conditions, and equipment status, which can lead to unstable product performance or even substandard quality.

[0003] Currently, the industry generally uses traditional sampling inspection and fixed rule control methods to control the quality of raw materials and products. This method not only has a low inspection frequency and delayed feedback, but also cannot dynamically respond to real-time anomalies that occur during the production process. In addition, when faced with practical problems such as the lack of high-frequency data, multivariate interference, and nonlinear spatiotemporal changes, the robustness and predictive ability of traditional methods are clearly insufficient, making it difficult to meet the requirements of modern intelligent manufacturing for high precision, real-time performance, and adaptability.

[0004] The development of artificial intelligence, especially deep learning technology, has provided a new solution for intelligent quality detection and control in the urea-formaldehyde resin production process. By performing real-time modeling and analysis of multi-source sensor data, dynamic sensing, anomaly identification, and feedback adjustment of the thermosetting material's properties can be achieved. However, existing methods lack the ability to model anomalies and uncertainties, lack effective anomaly simulation mechanisms, and are insufficient in modeling when faced with missing data and strong spatiotemporal coupling relationships. Therefore, a systematic and model-based improvement method is urgently needed. Summary of the Invention

[0005] This application provides a method, apparatus, and electronic equipment for testing thermosetting materials, in order to solve the technical problems of difficulty in quality testing, anomaly identification, and process control of thermosetting materials.

[0006] This application provides a method for testing thermosetting materials. The method includes: acquiring the original index parameters of a sample thermosetting material; preprocessing the original index parameters using a dual calibration module to obtain preprocessed index parameters; obtaining a thermosetting material testing model by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on the preprocessed index parameters; testing the thermosetting material to be tested according to the thermosetting material testing model to obtain testing results, and performing quality control feedback adjustment on the testing results to adjust the quality of the testing results.

[0007] In one embodiment of this application, the original index parameters are preprocessed using a dual calibration module to obtain preprocessed index parameters, including: identifying uncertain samples in the original index parameters; correcting the uncertain samples using a calibration loss function to obtain a corrected index dataset; and introducing perturbations into the corrected index dataset to obtain preprocessed index parameters.

[0008] In one embodiment of this application, identifying uncertain samples in the original index parameters includes: extracting time-series data to be processed from the original index parameters; performing time-series feature extraction on the time-series data to be processed through a time-series modeling network to obtain a time-series extraction result; mapping the time-series extraction result to a low-dimensional embedding space through a first projection head to obtain a first mapping result; and mapping the time-series extraction result to a low-dimensional embedding space through a second projection head to obtain a second mapping result; calculating the squared Euclidean distance between the first mapping result and the center of the hypersphere to obtain a first distance; and calculating the squared Euclidean distance between the second mapping result and the center of the hypersphere to obtain a second distance; the first distance and the second distance are used to determine the deviation of the time-series data to be processed from the center of the hypersphere; if the first distance follows a Gaussian distribution, the distance probability feature of the time-series data to be processed is obtained by combining the second distance to determine the uncertain samples.

[0009] In one embodiment of this application, perturbation is introduced into the corrected indicator dataset to obtain preprocessed indicator parameters, including: perturbing the corrected indicator dataset according to a pre-set anomaly type generation rule to obtain a perturbed dataset; setting a first label value for the corrected indicator dataset and a second label value for the perturbed dataset; and identifying the anomaly boundary of the corrected indicator dataset based on the first label value and the second label value using a supervised loss function to obtain the preprocessed indicator parameters.

[0010] In one embodiment of this application, a thermosetting material detection model is obtained by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on the preprocessed index parameters. The model includes: interpolating the preprocessed index parameters to obtain continuously differentiable paths; constructing a variable space correlation graph and encoding the continuously differentiable paths into spatial hidden states according to spatial structure relationships through graph convolution; generating temporal hidden states based on the time derivatives and multivariate control functions of the spatial hidden states using neural differential equations; backpropagating based on a mask loss function and simultaneously processing the parameters of the graph convolution and the neural differential equations to enable the spatial hidden states and temporal hidden states to co-fit and detect; decoding the temporal hidden states to obtain a predicted value sequence; and identifying anomalies in the predicted value sequence using a distributed anomaly scoring mechanism to obtain the thermosetting material detection model.

[0011] In one embodiment of this application, an anomaly identification is performed on the predicted value sequence using a distributed anomaly scoring mechanism to obtain a thermosetting material detection model. This includes: under the assumption that the predicted value sequence conforms to a Gaussian distribution, statistical analysis of the historical distribution of the predicted value sequence is performed using a rolling window, and univariate anomalies are quantified using negative log-likelihood to obtain variable anomaly scores; structural anomaly scores are calculated using the spatiotemporal hidden states, and multivariate anomaly scores and structural anomaly scores are linearly weighted and aggregated to output a global anomaly index, thereby obtaining the thermosetting material detection model.

[0012] In one embodiment of this application, after obtaining the detection result, the method further includes: extracting the dynamic features of the current thermosetting material from the detection result; scoring the probability distribution of the predicted values ​​of each variable in the dynamic features of the current thermosetting material to obtain a probability score; and scoring the overall spatial structure in the dynamic features of the current thermosetting material to obtain a structural score; comparing the probability score with the probability score of the historical standard sample; and comparing the structural score result with the structural score of the historical standard sample to perform quality detection on the current thermosetting material.

[0013] In one embodiment of this application, the quality control feedback adjustment of the detection result includes: if the current quality detection result of the thermosetting material is abnormal, then according to a preset automatic adjustment strategy, the adjustment command is fed back to the original thermosetting material ratio and process parameter control terminal; receiving newly generated thermosetting material index parameters, and iteratively updating the thermosetting material detection model through the newly generated thermosetting material index parameters; and performing quality detection on the thermosetting material to be tested through the updated thermosetting material detection model, so as to perform quality control feedback adjustment of the detection result.

[0014] The thermosetting material testing device provided in this application includes: an acquisition module for acquiring the original index parameters of a sample thermosetting material; a data preprocessing module for preprocessing the original index parameters through a dual calibration module to obtain preprocessed index parameters; a training module for obtaining a thermosetting material testing model by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on the preprocessed index parameters; and an identification module for detecting the thermosetting material to be tested according to the thermosetting material testing model, obtaining detection results, and performing quality control feedback adjustment on the detection results to adjust the quality of the detection results.

[0015] This application provides an electronic device, including a processor, a memory, and a communication bus; the communication bus is used to connect the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the above-described thermosetting material detection method.

[0016] The beneficial effects of this application are as follows: After obtaining the original index parameters of the thermosetting materials of the samples, a dual calibration module is introduced to preprocess the original index parameters. Preprocessing the original data from two dimensions can identify and suppress samples with high uncertainty, improving feature robustness. Furthermore, it guides the model to learn anomaly boundaries, enhancing the model's anomaly recognition ability and significantly improving the stability and generalization ability of subsequent model training. An improved framework based on spatiotemporal fusion modeling and anomaly scoring mechanism is constructed, balancing prediction and anomaly recognition capabilities, and exhibiting excellent performance, especially in environments with high missing rates. Moreover, this application can perform quality control feedback adjustment on the detection results, achieving real-time anomaly response.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0019] Figure 1 This is a flowchart illustrating a method for testing thermosetting materials in an exemplary embodiment of this application;

[0020] Figure 2 This is a block diagram of a thermosetting material testing device illustrated in an exemplary embodiment of this application;

[0021] Figure 3This is a schematic diagram illustrating the structure of an electronic device as shown in an exemplary embodiment of this application. Detailed Implementation

[0022] The embodiments of this application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be understood that the preferred embodiments are only for illustrating this application and are not intended to limit the scope of protection of this application.

[0023] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0025] The embodiments of this application respectively propose a method for testing thermosetting materials, a device for testing thermosetting materials, and an electronic device. These embodiments will be described in detail below.

[0026] Please see Figure 1 , Figure 1 A flowchart illustrating a method for testing thermosetting materials according to an embodiment of this application is shown. Figure 1 As shown, the method includes at least steps S110 to S130, which are described in detail below:

[0027] Step S110: Obtain the original index parameters of the sample thermosetting material.

[0028] In the preparation of urea-formaldehyde resin thermosetting materials, the physicochemical properties of the original thermosetting material directly affect the final quality of the product. To ensure the accuracy of subsequent modeling and quality control, comprehensive and high-frequency data collection of the original thermosetting material is necessary. Due to its excellent bonding properties, thermosetting properties, and cost-effectiveness, urea-formaldehyde resin is widely used in fields such as engineered wood products, building adhesives, and decorative thermosetting materials. As an example, the sample thermosetting material can be urea-formaldehyde resin.

[0029] It should be noted that the thermosetting materials listed above, such as urea-formaldehyde resin, are merely one example presented in this application, and there is no limitation on the thermosetting materials used in the samples.

[0030] In one embodiment of this application, the original index parameters of the sample thermosetting material (urea-formaldehyde resin) are automatically acquired using a variety of highly sensitive industrial-grade sensor devices. These industrial-grade sensor devices include: 1. Near-infrared spectrometer: used for non-destructive testing of chemical composition; 2. Gas or liquid chromatograph: used to analyze molecular components, free formaldehyde, etc.; 3. Rotational or vibrational viscometer: used to acquire rheological performance parameters in real time.

[0031] In one embodiment of this application, the original index parameters of the sample thermosetting material include, but are not limited to, the key index parameters of each batch of original thermosetting material. For example, the key index parameters include: 1. Solid content: measures the proportion of non-volatile solid substances in the resin; 2. Viscosity: reflects the flow characteristics and has a direct impact on processing and molding; 3. Formaldehyde emission: closely related to environmental protection indicators and user safety; 4. Molecular weight distribution: reveals the degree of polymerization reaction and the performance of thermosetting materials.

[0032] Step S120: The original index parameters are preprocessed using the dual calibration module to obtain the preprocessed index parameters.

[0033] To further enhance the quality and effectiveness of the data preprocessing stage, this application employs a dual calibration module to perform anomaly modeling and discrimination optimization on the original thermosetting material characteristic data before model training. The dual calibration module consists of two parts: uncertainty modeling calibration and native anomaly calibration, which respectively enhance the preprocessing of the original features from the perspectives of "data distribution consistency" and "anomaly feature simulation".

[0034] In one embodiment of this application, the original index parameters are preprocessed by a dual calibration module to obtain preprocessed index parameters, including: identifying uncertain samples in the original index parameters; correcting the uncertain samples by a calibration loss function to obtain a corrected index dataset; and introducing perturbation into the corrected index dataset to obtain the preprocessed index parameters.

[0035] As an example, uncertainty modeling calibration identifies uncertain samples in the original index parameters; the uncertain samples are corrected using a calibration loss function to obtain a corrected index dataset; and perturbations are introduced into the corrected index dataset through native anomaly calibration to obtain preprocessed index parameters. A dual calibration module is proposed to improve the quality consistency and anomaly resistance of the data preprocessing stage. By introducing a combined mechanism of uncertainty modeling calibration and native anomaly calibration, the original data is preprocessed from two dimensions: uncertainty modeling calibration uses dual projection heads to model the distance distribution from samples to the center of the hypersphere, identifying and suppressing high-uncertainty samples and improving feature robustness; native anomaly calibration generates six types of native anomaly samples through perturbation design, guiding the model to learn anomaly boundaries, enhancing the model's anomaly recognition ability, and significantly improving the stability and generalization ability of subsequent model training.

[0036] In one embodiment of this application, identifying uncertain samples in the original index parameters includes: extracting time-series data to be processed from the original index parameters; extracting time-series features from the time-series data to be processed using a time-series modeling network to obtain a time-series extraction result; mapping the time-series extraction result to a low-dimensional embedding space using a first projection head to obtain a first mapping result; and mapping the time-series extraction result to a low-dimensional embedding space using a second projection head to obtain a second mapping result; calculating the squared Euclidean distance between the first mapping result and the center of the hypersphere to obtain a first distance; and calculating the squared Euclidean distance between the second mapping result and the center of the hypersphere to obtain a second distance; the first distance and the second distance are used to determine the deviation of the time-series data to be processed from the center of the hypersphere; if the first distance follows a Gaussian distribution, the distance probability feature of the time-series data to be processed is obtained by combining the second distance to determine the uncertain samples.

[0037] As a concrete example, uncertainty modeling is used to identify and suppress "high uncertainty samples" in the original data caused by factors such as sensor errors and environmental interference. Anomaly removal and enhancement processing is achieved through feature distance modeling.

[0038] Single-class distance modeling under the Gaussian distribution assumption calculates the distance from the sample to the center of the hypersphere. Defined as:

[0039] ;

[0040] Where s is an original time series sample; For the sample The representation after temporal feature extraction is output by a temporal modeling network (such as the output of a Transformer encoder). is the first projection head, used to map temporal features to a low-dimensional embedding space; c is the center of the hypersphere, used to measure the central tendency of samples in the embedding space; The first distance is the squared Euclidean distance between sample s and center c, which is embedded under the first projection head and reflects its "deviation". The distance follows a Gaussian distribution ,in:

[0041] ;

[0042] in, The second distance is the square of the Euclidean distance under the second projection head, used as a distance assessment for the auxiliary viewpoint. The sum of the first and second distances is used as the mean of a Gaussian distribution to describe the expected distance of normal samples. The square of the difference between the first and second distances is used as the variance of the Gaussian distribution to measure the consistency and uncertainty of distance calculations. The sample distance (first distance) The Gaussian distribution it follows is modeled to determine its probabilistic characteristics in the embedding space.

[0043] Dual projector heads and Time series features Mapping to a lower-dimensional space allows for the separate calculation of the distances twice:

[0044] ;

[0045] For the second projection head, another set of low-dimensional embedding representations is generated; where the hypersphere center... The initialization method is as follows:

[0046] ;

[0047] in, The initial projection head is used to generate the feature map at the center of the hypersphere. Model the network for timing in the initial state; This is the number of samples used to calculate the center of the hypersphere, i.e., the size of the training sample set; To include all samples The average value of the features after initial mapping is used as the initialization of the hypersphere center.

[0048] Calibration loss function By combining the Gaussian distribution assumption and dual-projection distance, high-uncertainty samples are suppressed and normal samples are enhanced.

[0049] ;

[0050] The contribution of high-uncertainty samples (such as anomalous contamination points) to the loss is reduced by using an exponential term; Directly penalize inconsistencies to improve predictive confidence for normal samples.

[0051] In one embodiment of this application, perturbation is introduced into the corrected indicator dataset to obtain preprocessed indicator parameters, including: perturbing the corrected indicator dataset according to a pre-set anomaly type generation rule to obtain a perturbed dataset; setting a first label value for the corrected indicator dataset and setting a second label value for the perturbed dataset; and identifying the anomaly boundary of the corrected indicator dataset based on the first label value and the second label value using a supervised loss function to obtain the preprocessed indicator parameters.

[0052] As a concrete example, a perturbation is introduced during the data preprocessing stage through a native anomaly calibration module to generate simulated anomaly samples, thereby improving the subsequent model's ability to identify real anomalies. According to pre-defined anomaly type generation rules, the corrected indicator dataset is perturbed to obtain a perturbed dataset. Perturbation operations are designed based on the time series anomaly types to generate a simulated anomaly sample set. For example, the perturbation rules can be: 1. Point anomaly: Modify the random dimension of the last observation to an extreme value. 2. Context anomaly: Modify the last observation based on the historical mean offset. 3. Collective anomaly: Modify a continuous segment to an extreme value.

[0053] The corrected index dataset can be normal samples, and the perturbation dataset can be abnormal samples. The first label value is... The second label value is 1. Combined with the classification header... Normal samples (corrected indicator dataset) and abnormal samples (perturbation dataset) are mapped to labels -1 (first label value) and 1 (second label value), respectively, and the following supervised loss is designed:

[0054] ;

[0055] Normal sample label is The abnormal sample label is ; The first indicator function represents the normal sample type; when the sample... Belongs to the normal sample set When the time condition is met, the value is 1; otherwise, it is 1. ; The second indicator function represents the type of anomalous sample; when the sample Belongs to the simulated anomaly sample set When the time condition is met, the value is 1; otherwise, it is 0. The supervised loss function of the native anomaly calibration module is used to optimize the classifier's ability to distinguish between normal samples and simulated anomaly samples. Characterization of the joint sample set (normal sample set) and simulated anomaly sample set Samples on ) Taking the mathematical expectation means calculating the average loss over the entire training set; For the sample The temporal feature representation is usually extracted by a temporal coding network; The classification head network maps temporal features to scalar classification outputs to distinguish between normal and abnormal samples. The supervision label for the normal samples (corrected indicator dataset) has a value of -1; This is the supervisory label for the anomalous samples (perturbation dataset), with a value of +1; The mean squared error between the predicted value and label -1 for normal samples; The mean squared error between the predicted value of the outlier sample and the label + 1; To simulate anomaly sample sets, an anomaly sample set generated by perturbation operations (such as point anomalies, context anomalies, and collective anomalies); This is the normal sample set, derived from the undisturbed portion of the real data.

[0056] As can be seen, the aforementioned supervision mechanism significantly enhances the discrimination performance by guiding the model to identify "abnormal boundaries" in the data through explicit labels. In summary, the dual calibration module performs structured calibration and anomaly intervention on the original features during the data preprocessing stage, which not only improves data quality and consistency but also provides a solid data foundation for the robust training of the thermosetting material detection model.

[0057] To effectively identify anomalous fluctuations in the quality of urea-formaldehyde resin thermosetting materials, especially in complex contexts with missing data and strong spatiotemporal dependencies, an improved multivariate time-series anomaly detection framework is proposed. This method integrates dynamic graph neural network modeling and a distributed anomaly scoring mechanism, demonstrating good performance in handling high missing rates and non-stationary dynamic characteristics.

[0058] Step S130: Based on the preprocessed index parameters, dynamic graph neural network modeling and distributed anomaly scoring mechanism are integrated to obtain the thermosetting material detection model. An improved multivariate time-series anomaly detection framework is proposed, constructing an improved framework based on spatiotemporal fusion modeling and anomaly scoring mechanism. This framework effectively addresses issues such as missing values ​​and nonlinear dynamic changes in multi-source sensor data; the non-stationarity of anomaly distributions and the difficulty in identifying multiple types of anomalies. It balances predictive and anomaly identification capabilities, exhibiting excellent performance, especially in environments with high missing values.

[0059] The thermosetting material testing model can achieve accurate modeling of the spatiotemporal behavior of complex quality data through the coordinated use of three parts: continuous path construction, graph convolutional spatial modeling, and differential equation temporal modeling.

[0060] In one embodiment of this application, a thermosetting material detection model is obtained by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on preprocessed index parameters. This includes: interpolating the preprocessed index parameters to obtain continuously differentiable paths; constructing a variable space correlation graph and encoding the continuously differentiable paths into spatial hidden states according to spatial structure relationships through graph convolution; generating temporal hidden states based on the time derivatives and multivariate control functions of the spatial hidden states using neural differential equations; backpropagating based on a mask loss function and simultaneously processing the parameters of graph convolution and neural differential equations to enable co-fitting detection between the spatial and temporal hidden states; decoding the temporal hidden states to obtain a predicted value sequence; and identifying anomalies in the predicted value sequence using a distributed anomaly scoring mechanism to obtain the thermosetting material detection model. By fusing information from the dual calibration module output and using data provided by the native anomaly calibration module to train the anomaly recognition network, the dual projection distance score is also introduced into the rolling distribution likelihood value to co-construct the final anomaly discrimination function, thereby significantly improving the detection accuracy and robustness of the model in environments with high missing rates and multiple anomalies. A dynamic graph neural control differential equation modeling method is proposed to capture the spatial dependence and temporal dynamic characteristics of data. This method constructs dynamic graph neural control differential equations by: using graph convolution to model the spatial dependence structure between the original thermosetting material properties; and using neural control differential equations to capture the temporal evolution trajectory of variables. This enables continuous modeling and trend perception of high-dimensional, dynamic, and asynchronous quality data. This method enhances the model's ability to express complex production process behaviors.

[0061] As an example, continuous path construction is performed first. To handle missing values ​​in the observation sequence, natural cubic spline interpolation is first used to construct continuously differentiable paths from the original sequence. This continuous path serves as the input basis for subsequent differential modeling.

[0062] Next, spatial process modeling is performed. Graph convolution is used to model the spatial correlations between variables, capturing the structural dependencies between the original thermosetting material properties. The evolution of the spatial process is as follows:

[0063] ;

[0064] For a moment , This is the initial time point; In order to be in The spatial hidden state at a given time represents the structural representation between variables; Initial time point The space is hidden; The continuous time path is a continuous input sequence obtained through cubic spline interpolation, used to handle missing values; The time derivative of a continuous path is used to construct neural control differential equations; The hidden state is defined for the graph convolution driving function. The direction of updates; This is the set of learnable parameters in a graph convolutional network. From arrive The integral process is used to construct the differential form of the hidden state evolution.

[0065] Among them, graph convolution operator Defined as:

[0066] ;

[0067] in, This is the first fully connected neural network layer, used to perform non-linear transformations on the graph convolution output; Given an adjacency matrix with self-loops, The matrix is ​​an identity matrix to ensure that nodes can incorporate their own information; This is the original adjacency matrix, which describes the spatial connectivity between variables.

[0068] If the graph structure is unknown, the adjacency matrix... Node embedding matrix generate:

[0069] ;

[0070] Where E is the node embedding matrix, and each row represents the embedding representation of a node in the low-dimensional space; is the degree matrix of the graph, used to normalize the graph convolution; This is the symmetric normalized adjacency matrix in graph convolution; For hidden state An intermediate representation after a fully connected layer and ReLU activation; This is a learnable weight matrix in graph convolution, used to control the aggregate weights of adjacent nodes.

[0071] Then, time process modeling is performed to output spatial process. As input, the evolutionary behavior of variables over time is modeled using neural control differential equations (NCDE):

[0072] ;

[0073] For a moment , This is the initial time point; For time The hidden state of time process at a given moment represents the dynamic representation of the variable at that moment; Initial time point Temporally hidden state; This is the control function, defined by the neural network, used to specify the direction of state updates; This is the set of parameters for the time-process control function, i.e., the trainable weights in the control network.

[0074] Among them, the control function It is composed of multiple fully connected subnetworks:

[0075] ;

[0076] For the i-th variable, a fully connected neural network layer is used to process the temporal features of that variable; Let represent the hidden state of the i-th variable at time t; Introduce nonlinear modeling capabilities for nonlinear activation functions (such as ReLU); The concatenation operator is used to join the control function outputs of various variables to form a unified vector; The concatenated control function vector serves as the driving function for the neural differential equation.

[0077] Finally, joint optimization is performed using a masked loss function to synchronously optimize the parameters of the spatial and temporal modules through backpropagation, achieving spatiotemporal consistency binding. The spatial and temporal processes jointly constitute an augmented differential system. To prevent missing values ​​from interfering with training, a masked loss function is defined, optimizing only the valid observations.

[0078] ;

[0079] The average mask L1 loss at time t; For a moment ; For a variable i to be a valid observation at that time step, i.e., a masked variable, if variable i is a valid observation at time t, then Otherwise ; q is the variable index (from 1 to N), and N is the total number of variables; Let i be the actual observed value of variable i at time t; Let i be the model prediction value of variable i at time t; The L1 (absolute value) error between the predicted and actual values; The number of variables involved in the calculation (which can be considered as a normalization factor).

[0080] This paper proposes a dynamic graph neural control differential equation modeling method to capture the spatial dependence and temporal dynamic characteristics of data. By constructing dynamic graph neural control differential equations, it uses graph convolution to model the spatial dependence structure between the original thermosetting material properties and uses neural control differential equations to capture the temporal evolution trajectory of variables, thus achieving continuous modeling and trend perception of high-dimensional, dynamic, and asynchronous quality data. This method enhances the model's ability to express complex production process behaviors. The mechanism significantly reduces the impact of noise and improves the model's generalization ability to incomplete data.

[0081] To accurately quantify the degree of anomaly at each moment, a scoring method based on a rolling Gaussian distribution is proposed. This mechanism does not rely on the current observations but only utilizes the historical distribution of predicted values, making it particularly suitable for scenarios with unstable sensors or high missing rates.

[0082] In one embodiment of this application, an anomaly identification of the predicted value sequence is performed through a distributed anomaly scoring mechanism to obtain a thermosetting material detection model. This includes: under the assumption that the predicted value sequence conforms to a Gaussian distribution, the historical distribution of the predicted value sequence is statistically analyzed through a rolling window, and univariate anomalies are quantified through negative log-likelihood to obtain variable anomaly scores; structural anomaly scores are calculated through spatiotemporal hidden states, and multivariate anomaly scores and structural anomaly scores are linearly weighted and aggregated to output a global anomaly index to obtain the thermosetting material detection model.

[0083] This scoring mechanism is a downstream component of dynamic graph neural network modeling. Based on the predicted value sequence output by the spatiotemporal hidden state, it quantifies the degree of anomaly at each moment and is suitable for scenarios with unstable sensors and high missing rates.

[0084] As an example, the scoring mechanism includes variable-level anomaly likelihood scoring and global anomaly scoring. First, there's variable-level anomaly likelihood scoring, which includes variable-level anomaly likelihood scoring and negative log-Gaussian likelihood scoring; the variable-level anomaly likelihood scoring is implemented as follows: For each variable, the predicted value... Calculate the historical mean and standard deviation within the sliding window W:

[0085] ;

[0086] Let be the predicted value of variable i at time t; W is the length of the sliding window, used to statistically analyze historical predicted values; j is the time offset within the window. Let variable i be at time i The predicted value; Let be the historical mean of variable i at time t within a rolling window; Let be the historical standard deviation of variable i at time t within a rolling window; This is the standard deviation operator, which calculates the standard deviation of the predicted value sequence within the sliding window.

[0087] The negative log-Gaussian likelihood scoring is implemented as follows: Assuming the predicted values ​​follow a Gaussian distribution, the negative log-likelihood is used to measure the degree of anomaly of the current predicted values:

[0088] ;

[0089] The abnormal score of variable i at time t is given by the larger the value, the greater the deviation from the normal distribution. To measure forecast uncertainty (the larger the standard deviation, the higher the uncertainty); The constant term represents the normal distribution; the squared term measures the degree of deviation between the current predicted value and the historical mean. Let be the historical mean of variable i at time t within a rolling window; Let be the rolling window historical standard deviation of variable i at time t.

[0090] Secondly, there is the global anomaly scoring, which involves aggregating the anomaly scores of all variables through linear weighting to form a global anomaly index for the current time step:

[0091] ;

[0092] ;

[0093] For the current moment The global anomaly score; N is the total number of variables; i is the i-th variable; For variables At any moment Negative log-likelihood anomaly score calculated based on rolling Gaussian distribution; For variables At any moment Structural deviation score (i.e., structural anomaly score); is the fusion coefficient, which controls the weight ratio of the structure score in the total anomaly score; s is the original sample (e.g., the time-series feature input of the original thermosetting material at a certain moment). Feature representations extracted for temporal modeling networks; The first projection head network maps features to the anomaly detection space; is the second projection head network, used for parallel modeling of auxiliary anomaly distances; c is the center of the hypersphere, representing the reference point of the "normal distribution" in the projection space; is the Euclidean norm, used to measure the distance between a sample and the center point; Score(s) is the structural anomaly score of sample s, indicating the degree to which it deviates from the normal distribution; To calibrate the loss function; This is the anomaly calibration loss function.

[0094] It is evident that by proposing an anomaly detection mechanism that integrates "dual-projection distance score" and "rolling distribution likelihood value," the calibration loss function and anomaly calibration loss function are introduced into anomaly detection and synergistically fused with the rolling Gaussian distribution negative log-likelihood score. This leverages structural deviation information to compensate for the statistical model's omission of specific anomalies, resulting in a more robust and discriminative global anomaly score. The fusion mechanism significantly improves the accuracy and interpretability of identifying various complex anomalies (such as mutations, drifts, and collective anomalies).

[0095] Step S140: The thermosetting material to be tested is tested according to the thermosetting material testing model to obtain the test results, and the quality control feedback adjustment is performed on the test results to adjust the quality of the test results.

[0096] In one embodiment of this application, after obtaining the detection result, the method further includes: extracting the dynamic features of the current thermosetting material from the detection result; scoring the probability distribution of the predicted values ​​of each variable in the dynamic features of the current thermosetting material to obtain a probability score; and scoring the overall spatial structure in the dynamic features of the current thermosetting material to obtain a structural score; comparing the probability score with the probability score of historical standard samples; and comparing the structural score result with the structural score of historical standard samples to perform quality detection on the current thermosetting material.

[0097] As an example, after model training, the model is deployed in a real-world production environment and integrated with a sensor system to build a real-time quality monitoring system, enabling intelligent perception and anomaly warning throughout the entire production process. For instance, the trained thermosetting material detection model can be embedded into an edge computing device or production control system. The model receives and processes data input from various sensors in real time. It analyzes the dynamic characteristics of the current batch of thermosetting materials in real time, comparing them with the distribution of historical normal samples; it uses a dual mechanism of structural scoring and probability scoring to assess whether the current state is abnormal. Once an abnormal score is detected exceeding a set threshold, an alarm mechanism is immediately triggered; a prompt message is sent to the control terminal, while simultaneously marking suspected abnormal indicators and trend changes; it supports human-machine collaborative intervention, assisting operators in preliminary diagnosis and handling.

[0098] In one embodiment of this application, quality control feedback adjustment of the detection results includes: if the current quality detection result of the thermosetting material is abnormal, then according to a preset automatic adjustment strategy, feeding back the adjustment command to the original thermosetting material ratio and process parameter control terminal; receiving newly generated thermosetting material index parameters, and iteratively updating the thermosetting material detection model using the newly generated thermosetting material index parameters; and performing quality detection on the thermosetting material to be tested using the updated thermosetting material detection model, so as to perform quality control feedback adjustment of the detection results.

[0099] As a concrete example, based on the real-time output of the test results, an automatic control strategy feeds back to the original thermosetting material ratio and process parameter control terminals, achieving closed-loop quality control. The parameter adjustment range includes: 1. Original thermosetting material ratio: such as adjusting the molar ratio of urea to formaldehyde; 2. Reaction conditions: such as temperature profile, pH value, catalyst dosage, etc.; 3. Mechanical conditions: such as stirring speed, reaction time, molding pressure, etc. The control method employs an automated execution unit (such as a PLC controller) to adjust parameters based on model prediction output.

[0100] In one embodiment of this application, the thermosetting material testing model continuously receives new data during operation; incremental learning or transfer learning techniques are used to optimize model parameters to adapt to equipment aging or batch changes in thermosetting material raw materials; ultimately realizing an adaptive, highly robust, and highly generalizable intelligent quality control closed-loop system.

[0101] exist Figure 1 The technical solution presented describes a method that, after obtaining the original index parameters of the thermosetting material samples, preprocesses these parameters using a dual calibration module. This two-dimensional preprocessing of the original data helps identify and suppress samples with high uncertainty, improving feature robustness. Furthermore, it guides the model to learn anomaly boundaries, enhancing its anomaly detection capability and significantly improving the stability and generalization ability of subsequent model training. An improved framework based on spatiotemporal fusion modeling and anomaly scoring mechanisms is constructed, balancing predictive and anomaly detection capabilities, and demonstrating excellent performance, especially in environments with high missing rates. Moreover, this application enables quality control feedback adjustment of the detection results, achieving real-time anomaly response.

[0102] Figure 2 A block diagram of a thermosetting material testing apparatus according to an exemplary embodiment of this application is shown. (Reference) Figure 2 As shown, a thermosetting material testing device 200 according to an embodiment of this application includes: an acquisition module 210, a data preprocessing module 220, a model training module 230, and a thermosetting material testing module 240. The acquisition module is used to acquire the original index parameters of the sample thermosetting material; the data preprocessing module is used to preprocess the original index parameters through a dual calibration module to obtain preprocessed index parameters; the model training module is used to obtain a thermosetting material testing model by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on the preprocessed index parameters; and the thermosetting material testing module is used to test the thermosetting material to be tested according to the thermosetting material testing model, obtain the test results, and perform quality control feedback adjustment on the test results to adjust the quality of the test results.

[0103] In one embodiment of this application, the data preprocessing module 220 is used to identify uncertain samples in the original index parameters; to correct the uncertain samples by calibrating the loss function to obtain the corrected index dataset; and to introduce perturbation into the corrected index dataset to obtain the preprocessed index parameters.

[0104] In one embodiment of this application, the data preprocessing module 220 is used to extract the time series data to be processed from the original index parameters; to extract time series features from the time series data to be processed through a time series modeling network to obtain a time series extraction result; to map the time series extraction result to a low-dimensional embedding space through a first projection head to obtain a first mapping result; and to map the time series extraction result to a low-dimensional embedding space through a second projection head to obtain a second mapping result; to calculate the square of the Euclidean distance between the first mapping result and the center of the hypersphere to obtain a first distance; and to calculate the square of the Euclidean distance between the second mapping result and the center of the hypersphere to obtain a second distance; the first distance and the second distance are used to determine the deviation of the time series data to be processed from the center of the hypersphere; if the first distance follows a Gaussian distribution, the distance probability feature of the time series to be processed is obtained by combining the second distance to determine the uncertain samples.

[0105] In one embodiment of this application, the data preprocessing module 220 is further configured to perturb the corrected indicator dataset according to a pre-set anomaly type generation rule to obtain a perturbed dataset; set a first label value for the corrected indicator dataset and a second label value for the perturbed dataset; and identify the anomaly boundary of the corrected indicator dataset based on the first label value and the second label value using a supervised loss function to obtain the preprocessed indicator parameters.

[0106] In one embodiment of this application, the model training module 230 is used to interpolate the preprocessed index parameters to obtain continuously differentiable paths; construct a variable space correlation graph, and encode the continuously differentiable paths into spatial hidden states according to spatial structure relationships through graph convolution; generate temporal hidden states based on the time derivatives and multivariate control functions of the spatial hidden states through neural differential equations; backpropagate based on the mask loss function, and simultaneously process the parameters of graph convolution and neural differential equations, so that the spatial hidden states and temporal hidden states are co-fitted for detection; decode the temporal hidden states to obtain a predicted value sequence; and perform anomaly identification on the predicted value sequence through a distributed anomaly scoring mechanism to obtain a thermosetting material detection model.

[0107] In one embodiment of this application, the model training module 230, under the assumption that the predicted value sequence conforms to a Gaussian distribution, statistically analyzes the historical distribution of the predicted value sequence through a rolling window and quantifies univariate anomalies through negative log-likelihood to obtain variable anomaly scores; calculates structural anomaly scores through spatiotemporal hidden states, linearly weights and aggregates multivariate anomaly scores and structural anomaly scores, and outputs global anomaly indicators to obtain a thermosetting material detection model.

[0108] In one embodiment of this application, the thermosetting material testing device 200 further includes a quality testing module 250. The quality testing module 250 is used to extract the dynamic features of the current thermosetting material from the testing results; score the probability distribution of the predicted values ​​of each variable in the dynamic features of the current thermosetting material to obtain a probability score; and score the overall spatial structure in the dynamic features of the current thermosetting material to obtain a structure score; compare the probability score with the probability score of historical standard samples; and compare the structure score result with the structure score of historical standard samples to perform quality testing on the current thermosetting material.

[0109] In one embodiment of this application, the thermosetting material testing device 200 further includes a quality feedback adjustment module 260. The quality feedback adjustment module 260 is used to, if the current quality test result of the thermosetting material is abnormal, feed back the adjustment command to the original thermosetting material ratio and process parameter control terminal according to the preset automatic adjustment strategy; receive newly generated thermosetting material index parameters, and iteratively update the thermosetting material testing model through the newly generated thermosetting material index parameters; and perform quality testing on the thermosetting material to be tested through the updated thermosetting material testing model, so as to perform quality control feedback adjustment on the test results.

[0110] It should be noted that the apparatus and method provided in the above embodiments belong to the same concept, and the specific ways in which each module and unit performs operations have been described in detail in the method embodiments, and will not be repeated here. In practical applications, the apparatus provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the apparatus can be divided into different functional modules to complete all or part of the functions described above, and this is not a limitation.

[0111] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the structure of an electronic device according to an exemplary embodiment of this application. It should be noted that... Figure 3 The electronic device 300 shown is merely an example and should not impose any limitations on the functionality or usage area of ​​the embodiments of this application.

[0112] like Figure 3As shown, the electronic device 300 includes a processor 301, a memory 302, and a communication bus 303; the communication bus 303 is used to connect the processor 301 and the memory 302; the processor 301 is used to execute a computer program stored in the memory 302 to implement one or more methods as described in the above embodiments.

[0113] The electronic device provided in this application includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic device performs the various steps of the above method.

[0114] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0115] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0116] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for testing thermosetting materials, characterized in that, The method includes: Obtain the original index parameters of the sample thermosetting material; The original index parameters are preprocessed using a dual calibration module to obtain the preprocessed index parameters. Based on the preprocessed index parameters, a detection model for thermosetting materials is obtained by integrating dynamic graph neural modeling and a distributed anomaly scoring mechanism. The thermosetting material to be tested is tested according to the aforementioned thermosetting material testing model to obtain the test results, and the quality control feedback adjustment is performed on the test results to adjust the quality of the test results.

2. The method for testing thermosetting materials according to claim 1, characterized in that, The original index parameters are preprocessed using a dual calibration module to obtain preprocessed index parameters, including: Identify uncertain samples in the original index parameters; The uncertain samples are corrected by calibrating the loss function to obtain the corrected index dataset; The corrected index dataset is perturbed to obtain preprocessed index parameters.

3. The method for testing thermosetting materials according to claim 2, characterized in that, Identifying uncertain samples in the original index parameters includes: Extract the time series data to be processed from the original index parameters; Temporal features are extracted from the temporal data to be processed using a temporal modeling network to obtain temporal extraction results. The timing extraction result is mapped to a low-dimensional embedding space using a first projection head to obtain a first mapping result; and the timing extraction result is mapped to a low-dimensional embedding space using a second projection head to obtain a second mapping result. The square of the Euclidean distance between the first mapping result and the center of the hypersphere is calculated to obtain the first distance; and the square of the Euclidean distance between the second mapping result and the center of the hypersphere is calculated to obtain the second distance; the first distance and the second distance are used to determine the deviation of the time series data to be processed from the center of the hypersphere; If the first distance follows a Gaussian distribution, the distance probability characteristics of the time series to be processed are obtained by combining the second distance, and uncertain samples are identified.

4. The method for testing thermosetting materials according to claim 2, characterized in that, The corrected index dataset is perturbed to obtain preprocessed index parameters, including: According to the pre-defined anomaly type generation rules, the corrected index dataset is perturbed to obtain the perturbed dataset; A first label value is set for the corrected index dataset, and a second label value is set for the perturbation dataset; By using a supervised loss function, the abnormal boundaries of the corrected index dataset are identified based on the first label value and the second label value, and the preprocessed index parameters are obtained.

5. The method for testing thermosetting materials according to claim 1, characterized in that, Based on the preprocessed index parameters, a thermosetting material detection model is obtained by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism, including: Interpolate the preprocessed index parameters to obtain a continuously differentiable path; Construct a variable space association graph, and encode the continuous differentiable paths into spatial hidden states according to the spatial structure relationship through graph convolution; The temporal hidden state is generated using neural differential equations based on the temporal derivative and multivariable control function of the spatial hidden state. Backpropagation is based on the mask loss function, and the parameters of the graph convolution and the neural differential equation are processed simultaneously, so that the spatial hidden state and the temporal hidden state are co-fitted for detection. Decode the time-hidden state to obtain a sequence of predicted values; Anomalies in the predicted value sequence are identified using a distributed anomaly scoring mechanism to obtain a thermosetting material detection model.

6. The method for testing thermosetting materials according to claim 5, characterized in that, Anomaly identification is performed on the predicted value sequence using a distributed anomaly scoring mechanism to obtain a thermosetting material detection model, including: Assuming the predicted value sequence conforms to the Gaussian distribution assumption, the historical distribution of the predicted value sequence is statistically analyzed using a rolling window, and univariate anomalies are quantified using negative log-likelihood to obtain variable anomaly scores; The structural anomaly score is calculated by the spatiotemporal hidden state, and the multivariate anomaly score and the structural anomaly score are linearly weighted and aggregated to output a global anomaly index, thereby obtaining a thermosetting material detection model.

7. The method for testing thermosetting materials according to claim 6, characterized in that, After obtaining the test results, the following is also included: Extract the dynamic characteristics of the current thermosetting material from the detection results; The probability distribution of the predicted values ​​of each variable in the dynamic characteristics of the current thermosetting material is scored to obtain a probability score; and, The overall spatial structure of the current thermosetting material in its dynamic characteristics is evaluated to obtain a structural score. The probability score is compared with the probability score of the historical standard sample; and the structure score is compared with the structure score of the historical standard sample to perform quality testing on the current thermosetting material.

8. The method for testing thermosetting materials according to claim 7, characterized in that, The quality control feedback adjustment of the detection results includes: If the quality test result of the current thermosetting material is abnormal, the control command will be fed back to the original thermosetting material ratio and process parameter control terminal according to the preset automatic control strategy. Receive newly generated thermosetting material index parameters, and iteratively update the thermosetting material detection model using the newly generated thermosetting material index parameters; The updated thermosetting material testing model is used to perform quality testing on the thermosetting material to be tested, so as to adjust the quality control feedback based on the test results.

9. A testing device for thermosetting materials, characterized in that, The thermosetting material testing device includes: The acquisition module is used to acquire the original index parameters of the sample thermosetting material; The data preprocessing module is used to preprocess the original index parameters through the dual calibration module to obtain the preprocessed index parameters; The model training module is used to obtain a thermosetting material detection model by fusing dynamic graph neural modeling and a distributed anomaly scoring mechanism based on the preprocessed index parameters. The thermosetting material testing module is used to test the thermosetting material to be tested according to the thermosetting material testing model, obtain the test results, and perform quality control feedback adjustment on the test results to adjust the quality of the test results.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the thermosetting material testing method as described in any one of claims 1 to 8.

Images