A Method and System for Optimizing Oscillating Pressure Sintering Process Based on Negative Sample Learning
By constructing a negative sample database and a time-series classification model, combined with cost-sensitive learning and oversampling, abnormal risk prediction and automatic intervention in the oscillating pressure sintering process were realized. This solved the problems of unused knowledge of failure cases and lack of advance prediction, and improved production efficiency and product yield.
Patent Information
- Application Number
- CN202610979229.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-02
- Publication Date
- 2026-11-17
AI Technical Summary
The valuable knowledge from failure cases in existing oscillating pressure sintering processes has not been effectively utilized, leading to repeated trial and error and high R&D costs. There is a lack of ability to predict abnormal operating conditions in advance, and the failure modes are complex, making it difficult for conventional machine learning methods to effectively learn failure modes.
A negative sample database is constructed to record structured information and multi-source sensor data of failed cases. An anomaly risk prediction is performed using a time-series classification model, and automatic intervention is executed when a risk is predicted. Cost-sensitive learning and oversampling are used to handle sample imbalance. The database is deployed on edge computing units or cloud platforms for real-time early warning and intervention.
It enables early detection and proactive avoidance of abnormal risks during the oscillating pressure sintering process, reduces the cost of repeated trial and error, improves product yield, and has the ability to accumulate independent knowledge and continuously optimize.
Smart Images

Figure CN123120126A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of materials preparation and artificial intelligence process optimization technology, specifically relating to a method and system for predicting and proactively intervening in abnormal risks in real time during oscillating pressure sintering by constructing a database and training a time-series classification model using sintering failure cases (negative samples). This invention is applicable to the optimization of oscillating pressure sintering processes for various materials, including ceramics, metals, cermet composites, and cemented carbides. Background Technology
[0002] Oscillating pressure sintering is an advanced material densification technology that superimposes periodic dynamic pressure oscillations on traditional hot pressing sintering. By adjusting parameters such as pressure amplitude, oscillation frequency, pressure average, and waveform, this technology can significantly promote particle rearrangement, eliminate residual porosity, and inhibit abnormal grain growth, thereby obtaining high-density, high-performance sintered products. However, the oscillating pressure sintering process still faces the following technical challenges related to failure cases in actual production: (1) Valuable knowledge contained in failed cases is not effectively utilized, and repeated trial and error leads to high R&D costs. In the process of oscillating pressure sintering, the occurrence of failed cases (such as cracking, insufficient density, uneven structure, etc.) is often accompanied by expensive material costs, equipment occupation time and manpower investment. In traditional process development, failed cases are simply recorded as "scrap," and the knowledge of the correlation between process parameters and failure modes contained therein is not extracted in a structured way and reused systematically. The same mistakes may occur repeatedly between different operators and different batches, resulting in extended R&D cycles and high costs. Especially in oscillating pressure sintering, there is a complex nonlinear relationship between pressure oscillation parameters (amplitude, frequency, waveform) and defects such as cracking, which is difficult to summarize systematically by human experience alone.
[0003] (2) Existing control systems lack the ability to predict abnormal operating conditions in advance and can only respond passively after defects occur. Most mainstream sintering control systems (including existing AI-assisted control systems) adopt an open-loop or semi-closed-loop mode of "setting process → execution → detection results". Even in systems with feedback learning capabilities (such as Chinese patent document CN117109324A), the model update is based on the overall result after sintering is completed, and it is impossible to judge in real time whether the current state is developing towards failure during the sintering process. When defects such as cracking actually occur, they are often irreversible, causing the sample to be scrapped. Therefore, there is an urgent need for a technical means that can predict abnormal risks several minutes or even earlier during the sintering process and actively adjust process parameters to avoid defects.
[0004] (3) The unique failure modes and pressure oscillation parameter characteristics of oscillating pressure sintering make it difficult to directly apply general anomaly detection methods. Compared with conventional hot pressing sintering, oscillating pressure sintering introduces dynamic pressure parameters (amplitude, frequency, waveform), making its failure modes more complex. For example, cracking may originate from fatigue effects caused by excessive pressure amplitude or inappropriate frequency, or from the coupling effect of heating rate and pressure oscillation. These failure modes often exhibit specific precursory characteristics in the time-series changes of sensor data (pressure, temperature, displacement, acoustic emission) (such as pressure waveform distortion, abnormal fluctuations in displacement rate, and sudden increases in acoustic emission signal). However, there is currently no method specifically designed for oscillating pressure sintering, using time-series sensor data from historical failure cases to train a model and achieve online risk prediction.
[0005] (4) The number of successful and failed cases is severely imbalanced, making it difficult for conventional machine learning methods to effectively learn failure patterns. In actual production, the number of successful sintering batches far exceeds the number of failed batches (the ratio of positive to negative samples can often reach 10:1 or even higher). Conventional classification models trained on such imbalanced datasets tend to favor the majority class (successful cases), resulting in extremely low recall rates for failed cases, which is difficult to meet the reliability requirements of industrial applications. Therefore, it is necessary to design specialized sample imbalance handling strategies (such as cost-sensitive learning, oversampling, etc.) and targeted model architectures.
[0006] In summary, no existing technology discloses a method or system that can systematically use negative samples from oscillating pressure sintering to predict timing anomalies and provide real-time early warning and proactive intervention during the sintering process. Summary of the Invention
[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for optimizing the oscillating pressure sintering process based on negative sample learning.
[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: An optimization method for oscillating pressure sintering process based on negative sample learning includes the following steps: S1: Construct a negative sample database, which records structured information of sintering failure cases. Each negative sample includes at least: process parameter values at the time of failure, failure type, and multi-source sensor data sequence within a preset time window before the failure. The failure type includes one or more of cracking, insufficient density, and uneven structure. The multi-source sensor data sequence includes pressure value sequence, temperature value sequence, and displacement value sequence. S2: Train a time-series classification model using the negative sample database and the positive sample database. The time-series classification model takes multi-source sensor data sequences as input and outputs anomaly risk category or anomaly risk probability. Cost-sensitive learning or oversampling methods are used to handle the imbalance between positive and negative samples during training. S3: During the oscillating pressure sintering process, the current multi-source sensor data sequence is collected in real time and input into the trained time series classification model to predict whether there is an abnormal risk in the current sintering state and the specific type of abnormal risk. S4: If an abnormal risk is predicted and the probability of the abnormal risk exceeds a preset threshold, an intervention operation will be automatically executed. The intervention operation includes at least one of the following: reducing the heating rate, reducing the pressure amplitude, reducing the pressure frequency, extending the heat preservation time, and issuing an alarm signal.
[0009] As a further improvement to the above technical solution: In step S1, the multi-source sensing data sequence further includes one or more of the following: acoustic emission signal sequence, infrared thermal imaging temperature distribution sequence, or furnace atmosphere concentration sequence.
[0010] The negative sample database also records textual descriptions of the reasons for failure; the reasons for failure include one or more of the following: excessively rapid heating rate, pressure amplitude exceeding the material strength limit, uneven pressure distribution due to mold wear, and abnormal sintering activity of powder batches.
[0011] The temporal classification model is a long short-term memory network model, a temporal convolutional network model, or a Transformer model based on an attention mechanism.
[0012] The positive sample database records the process parameters and multi-source sensor data sequences of successful sintering cases; the time-series classification model is trained using joint training of positive and negative samples, where the loss weight of negative samples is higher than that of positive samples.
[0013] The loss function expression for the cost-sensitive learning is: in, For batch size, For the first The true label of each sample For category The corresponding loss weights are set as follows: the weight for positive sample classes is set to 1, and the weight for negative sample classes is set to a value greater than 1. The total number of categories, For the model to the first The output log probability of each category.
[0014] The oversampling method uses the SMOTE algorithm to synthesize new negative samples. The formula for generating the new samples is: in, For a selected negative sample feature vector, The feature vector of the next nearest negative sample to the given negative sample. A random number between 0 and 1.
[0015] When the time-series classification model is a long short-term memory network model, its unit state update follows the following formula: in, For a moment The input sensor data vector, In hidden state, For the unit state, The activation value for the forget gate. This is the activation value of the input gate. This is the activation value of the output gate. and For trainable weights and biases, Here is the Sigmoid function, and ⊙ represents element-wise multiplication.
[0016] The preset threshold Adjustments will be made dynamically based on the current sintering stage. in, Based on the threshold, , The adjustment amount is positive, and ≥ , This is the threshold for the end of the initial sintering stage. This is the threshold for the start time of the final sintering stage; Alternatively, the preset threshold is based on the most recent Accuracy of the prediction Automatic adjustment: in, For learning rate, For the target accuracy, For the most recent The actual accuracy of the prediction.
[0017] In step S4, reducing the heating rate specifically involves multiplying the current heating rate by a coefficient of 0.5 to 0.8; reducing the pressure amplitude specifically involves reducing the current pressure amplitude by 0.2t to 0.5t, or reducing it to 50% to 80% of the current pressure amplitude; reducing the pressure frequency specifically involves reducing the current pressure frequency by 0.1Hz to 0.3Hz; and extending the heat preservation time specifically involves increasing the heat preservation time by 10% to 30% based on the original heat preservation time.
[0018] The method is used for oscillatory pressure sintering of silicon carbide ceramics, silicon nitride ceramics, titanium alloys, nickel-based superalloys, or cemented carbides.
[0019] This invention also provides an oscillating pressure sintering process optimization system based on negative sample learning, comprising: The negative sample construction module is used to record structured information of sintering failure cases and build a negative sample database. The time-series classification model, deployed on edge computing units or cloud platforms, is trained using the methods described above and is used to predict abnormal risks in the current sintering state in real time. The early warning and intervention module is used to automatically execute the intervention operation described in step S4 when an abnormal risk is predicted and the probability exceeds a preset threshold. The model update module is used to add newly occurring failure or success cases to the corresponding database and periodically retrain or fine-tune the time series classification model.
[0020] As a further improvement to the above technical solution: The early warning and intervention module includes: a human-machine interface that displays early warning information and suggested operations, and an automatic control command sent to the actuator of the sintering equipment.
[0021] The model update module is configured to automatically trigger incremental training whenever M negative samples or N positive samples are added, where M≤10 and N≤50.
[0022] Compared with the prior art, the advantages of the present invention are as follows: The present invention provides an optimization method for oscillating pressure sintering process based on negative sample learning. In step S1, a negative sample database is constructed, recording multi-source sensor data sequences (pressure, temperature, and displacement values) and failure types before failures occur. This allows the implicit abnormal evolution processes in historical failure cases to be structurally preserved, transforming discarded "negative knowledge" into computable data assets. This directly solves the problems of "unutilized knowledge inherent in failure cases and high costs of repeated trial and error." In step S2, a time-series classification model is trained using negative and positive samples, and cost-sensitive learning or oversampling is employed to address sample imbalance. This solves the problem of "severe imbalance in the number of positive and negative samples, making it difficult for conventional models to learn failure modes," enabling the model to learn from scarce data. The system effectively learns the temporal characteristics of anomaly precursors from failed samples. By using S3 to collect current sensor data sequences in real time during sintering and inputting them into the model to predict anomaly risks and specific types, it achieves early detection of failure modes such as cracking, insufficient density, and uneven microstructure, solving the problem of "lacking the ability to predict abnormal working conditions in advance and only being able to respond passively." By using S4 to automatically execute intervention operations (reducing the heating rate, reducing the pressure amplitude, reducing the pressure frequency, extending the holding time, or issuing an alarm) when the predicted risk probability exceeds the threshold, the prediction results are transformed into proactive control actions. Process parameters are adjusted before defects form to avoid risks, ultimately achieving the technical effects of reducing the cost of repeated trial and error, improving product yield, and avoiding the recurrence of the same failures.
[0023] The oscillating pressure sintering process optimization system based on negative sample learning of this invention records structured information of sintering failure cases and establishes a negative sample database through a negative sample construction module, enabling the system to autonomously accumulate knowledge from failures and solving the problem of "failure case knowledge not being utilized"; through a time-series classification model deployed on edge computing units or cloud platforms, the system predicts the abnormal risks of the current sintering state in real time, and can use the time-series patterns of historical failure cases to continuously assess the risks of the ongoing sintering process, solving the problem of "lack of advance prediction capability"; through an early warning and intervention module, when the predicted risk probability exceeds a threshold... The automatic execution of the intervention operation described in step S4 directly couples the risk prediction results into the control link of the actuator, forming a closed loop from data acquisition and model inference to equipment regulation, solving the problem of "only being able to passively respond after a defect occurs"; through the model update module, newly occurring failure or success cases are added to the database and the model is retrained or fine-tuned periodically. The system has the self-optimization capability to continuously evolve with the accumulation of production data, solving the problems of "conventional models cannot effectively learn failure modes" and "knowledge cannot be accumulated", and finally achieving the technical effect of proactive early warning and automatic avoidance of abnormal working conditions during oscillating pressure sintering. Attached Figure Description
[0024] Figure 1This is an overall flowchart of the oscillating pressure sintering process optimization method based on negative sample learning of the present invention.
[0025] Figure 2 This is a flowchart illustrating the construction process of the negative sample database for this invention.
[0026] Figure 3 This is a flowchart of the training process for the LSTM-based temporal classification model of this invention.
[0027] Figure 4 This is a block diagram of the architecture of the oscillating pressure sintering process optimization system based on negative sample learning of the present invention. Detailed Implementation
[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the implementation of the present invention is not limited thereto.
[0029] Example 1: like Figure 1 As shown, this embodiment provides an optimization method for oscillating pressure sintering process based on negative sample learning, which is used in the oscillating pressure sintering process of SiC ceramic materials to predict cracking risk in advance and automatically intervene.
[0030] 1. Construction of the negative sample database, the construction process is as follows: Figure 2 As shown Data from over 600 sintering experiments conducted in a 30T oscillating pressure sintering furnace were collected, including 540 successful cases and 60 failed cases. The failures were characterized by: 32 cases of cracking, 18 cases of insufficient density, and 10 cases of uneven microstructure.
[0031] For each failed case, construct negative samples according to the following specifications: Process parameter values at the time of failure: record the temperature (°C), average pressure (t), pressure amplitude (t), oscillation frequency (Hz), and heating rate (°C / min) at the time of failure.
[0032] Failure types: cracking, insufficient density, uneven structure.
[0033] The multi-source sensor data sequence within a preset time window before failure: sensor data from 5 minutes prior to the failure occurred, with a sampling frequency of 1Hz, meaning each negative sample contains 300 time-step data points. Forced recording fields include: pressure value sequence (t), temperature value sequence (°C), and displacement value sequence (mm). Some negative samples also additionally record acoustic emission signal sequence (mV), infrared thermal imaging temperature distribution sequence (°C), and furnace atmosphere concentration sequence (ppm).
[0034] Textual description of the reasons for failure (to aid knowledge accumulation): for example, "heating rate 30℃ / min, exceeding the recommended value of 20℃ / min", "pressure amplitude 2.5t, material limit 1.8t", "uneven pressure distribution caused by mold wear", "abnormal sintering activity of powder batch", etc.
[0035] The positive sample database similarly records the sensor data sequence and the final density value achieved for successful sintering cases, while the failure type field is set to "none".
[0036] 2. Time-series classification model architecture, training process as follows: Figure 3 As shown This embodiment uses a Long Short-Term Memory (LSTM) network as the temporal classification model.
[0037] The model structure is as follows: Input layer: time step T=300, feature dimension D=3 (pressure, temperature, displacement). In other embodiments using extended sensor data, D can be extended to 6.
[0038] LSTM layer: Two-layer LSTM, each layer has 128 cells, and the cell state update follows the formula: in, For a moment The input sensor data vector, In hidden state, For the unit state, The activation value for the forget gate. This is the activation value of the input gate. This is the activation value of the output gate. and For trainable weights and biases, Here, is the Sigmoid function, and ⊙ represents element-wise multiplication. Through this LSTM layer, the model can automatically capture temporal precursor features such as pressure waveform distortion and abnormal displacement rates before cracking.
[0039] Dropout layer: Add Dropout between LSTM layers and between LSTM layers and fully connected layers, with a dropout rate of 0.3, to prevent overfitting.
[0040] Fully connected layer: 128 → 64 → K, where K=3 (corresponding to three types of output: normal, cracking risk, and insufficient density risk; tissue asymmetry can be classified separately or merged into the insufficient density category).
[0041] Output layer: Softmax activation, outputting the probability of each class.
[0042] 3. Handling of Imbalanced Positive and Negative Samples In this embodiment, the ratio of positive to negative samples is approximately 9:1 (540:60). A combined strategy of cost-sensitive learning and SMOTE oversampling is employed.
[0043] (1) Cost-sensitive loss function The loss function uses weighted cross-entropy, and its expression is: in, For batch size, For the first The true label of each sample For category The corresponding loss weights. Set the weights for positive samples (normal samples). Negative sample weights (cracking risk, insufficient density risk) This makes the model pay a higher price for misjudging negative samples, effectively improving its ability to identify rare failure cases.
[0044] (2) SMOTE oversampling For minority classes (cracked, insufficient density), synthesis is performed in the feature space, and the formula for generating new samples is: in, and These are the feature vectors of two negative samples within the same risk category. For a selected negative sample feature vector, This is the feature vector of the nearest neighboring negative sample. In this embodiment, the 300×3 time series data of each negative sample are extracted into eight statistical features: mean pressure, pressure variance, maximum-minimum pressure difference, temperature slope, and final displacement value. These features are then synthesized using SMOTE, increasing the number of negative samples to about half that of the positive samples (i.e., about 270 samples), while retaining the original negative samples, thereby alleviating overfitting.
[0045] (3) Joint training of positive and negative samples During training, both a positive sample database (successful cases) and a negative sample database enhanced by SMOTE are used simultaneously, with the negative samples having a higher loss weight.
[0046] 4. Model Training The Adam optimizer was used with a learning rate of 0.001, a batch size of 64, 200 training epochs, and early stopping (the training stopped if the validation set loss did not decrease for 20 consecutive epochs). The training / validation / test sets were split into 70% / 15% / 15% sets, maintaining the proportion of each class.
[0047] After training, performance was evaluated on the test set: the negative sample recall (true positive rate) reached 88%, and the positive sample false positive rate (false positive rate) was 4.2%. Compared with the baseline model that did not use cost-sensitive learning and SMOTE (recall rate of only 62%), the recall rate of this embodiment was improved by 26 percentage points, demonstrating the effectiveness of the imbalance handling strategy.
[0048] 5. Real-time prediction and anomaly risk assessment The trained LSTM model was deployed on the edge computing industrial computer (CPU: Intel i7, RAM: 16GB) of the sintering furnace. During the sintering process, the system continuously received data from pressure, temperature, and displacement sensors (1Hz). A window of data was accumulated every 300 time steps (5 minutes), with a sliding window step size of 1 second, meaning a prediction was made every second. The model output three types of probabilities: , , .
[0049] Abnormal risk assessment: If > or If the threshold is greater than θ, an abnormal risk is identified, and the risk type is the corresponding defect type. θ is the base threshold, initially set to 0.7.
[0050] 6. Dynamic threshold adjustment To improve the prediction sensitivity at different sintering stages, this embodiment employs two threshold adjustment strategies for users to choose from and switch between.
[0051] Strategy 1 (Piecewise Function): in =0.7, =0.2, =0.1. That is, the initial threshold is reduced to 0.5 (to increase sensitivity to abnormal temperature rise), the final threshold is reduced to 0.6 (to prevent overheating), and it is maintained at 0.7 in the middle stage. Temperature-time correspondence: the initial stage is room temperature to 900℃, the middle stage is 900~1600℃, and the final stage is above 1600℃ until the end of heat preservation.
[0052] Strategy Two (Online Adjustment Based on Accuracy): For every 10 batches of sintering completed, the statistical prediction accuracy rate is... (Based on actual results) If If the value is less than 0.85, the threshold will be updated proportionally. in, The learning rate is 0.05 in this example; For the target accuracy, this example uses a value of 0.90; For the most recent The actual accuracy of the prediction.
[0053] For example, if the accuracy rate of the most recent 10 batches is 0.80, then: The threshold is slightly increased to reduce false alarms; if the accuracy remains below the target for an extended period, the threshold is gradually adaptive to allow the system to adapt to equipment drift or material changes.
[0054] In actual operation, users can choose to enable Strategy 1, Strategy 2, or a combination of both (using Strategy 1 initially and then fine-tuning with Strategy 2 later). In this embodiment, enabling both simultaneously yields the best results.
[0055] 7. Specific implementation of intervention procedures When the risk probability exceeds a threshold, the system executes intervention actions based on the risk type and the current stage. The parameter ranges for the intervention actions are shown in the table below: All intervention commands are sent to the sintering furnace's PLC via the OPC UA protocol, with higher priority than conventional AI real-time closed-loop control commands, forming a "safety-first" nested control architecture. For example, when the model predicts a cracking risk probability of 0.8, the PLC immediately reduces the heating rate from 20℃ / min to 12℃ / min and the pressure amplitude from 1.5t to 1.2t, while simultaneously triggering an alarm. In practice, cracking was successfully avoided.
[0056] 8. Continuous Model Updates After each sintering cycle, regardless of success or failure, the system automatically adds complete process data (sensor sequence, process parameters, and final result) to the corresponding database. The system is configured to automatically trigger incremental training whenever M negative samples (M=5) or N positive samples (N=20) are added. Incremental training fine-tunes the existing model, reducing the learning rate to 0.0001 and training for 10 epochs. The updated model is evaluated on the validation set; if its performance is not lower than the original model, the online model is automatically replaced; otherwise, the original model is retained, and engineers are notified to check the data quality.
[0057] This continuous update mechanism enabled the system to improve the negative sample recall rate from 88% to 92% and reduce the false positive rate from 4.2% to 3.1% after 6 months of use, demonstrating the self-optimization effect of "becoming more accurate with use".
[0058] This embodiment was continuously operated in SiC ceramic production for 3 months, processing a total of 500 batches. The system issued 35 warnings, of which 30 warnings were prevented by operator or automatic intervention (28 preventing cracking, 2 preventing insufficient density), 3 warnings were not addressed and resulted in cracking, and 2 were false alarms (triggered without actual risk). The overall warning accuracy rate was 89%, and the defect prevention rate was 86%. Compared with batches produced during the same period without this method, the product yield increased from 92.5% to 97.8%, the scrap rate due to cracking decreased by 75%, and the average number of trial and error attempts for developing new process formulations decreased from 15 to 5, significantly reducing R&D costs.
[0059] Example 2: This embodiment, as an alternative to LSTM, uses Temporal Convolutional Network (TCN) as a temporal classification model, which is suitable for scenarios with extremely high requirements for real-time inference (such as requiring prediction every 0.5 seconds) and a large amount of training data.
[0060] Dilated causal convolution is used, with dilation factors of 1, 2, 4, and 8 for each layer, kernel size of 3, and number of channels of 64.
[0061] Four residual blocks are stacked, each containing two layers of dilated convolutions, followed by weight normalization and ReLU activation.
[0062] Global average pooling is followed by a fully connected layer (64→K) and Softmax output.
[0063] Training and imbalance handling: Same as in Example 1, and will not be repeated here. TCN training speed is 40% faster than LSTM, with a single inference time of only 5ms.
[0064] Threshold adjustment and intervention: Same as in Example 1, and will not be repeated here.
[0065] On the same dataset, TCN has a recall rate of 86%, slightly lower than LSTM's 88%, but it has a faster inference speed, meeting the requirement of a 0.5-second prediction cycle.
[0066] Example 3: This embodiment provides an optimization system for oscillating pressure sintering process based on negative sample learning. The system composition is as follows: Figure 4 As shown: Negative Sample Construction Module: Running on the data server, this module provides a web interface for operators to annotate failure cases. When a sintering result is determined to be a failure, the operator selects the failure type (cracking / insufficient density / uneven structure) in the interface and enters the failure reason text (optional). The system automatically extracts the sensor data sequence from the historical database for the 5 minutes prior to the failure, forming a complete negative sample record, which is then stored in the MySQL database. Batch import of historical failure data is supported.
[0067] Temporal classification model: The LSTM model trained in Example 1 was converted to TensorFlow Lite format and deployed on an edge computing industrial control computer (Intel i7, 16GB RAM, no GPU). The model's single inference time is <20ms, meeting the requirement of 1Hz real-time prediction. At the same time, a deeper LSTM model (4 layers, 256 units) is maintained on the cloud server for global retraining once a week.
[0068] Early warning and intervention module: Human-machine interface: The current risk probability curve is displayed in real time on the sintering monitoring screen. , When the threshold is exceeded, a red warning box will pop up, displaying suggested actions (such as "It is recommended to reduce the pressure amplitude by 0.3t").
[0069] Automatic intervention controller: Communicates with the sintering furnace PLC via Modbus TCP protocol. When the risk probability exceeds the threshold and persists for 3 seconds (to avoid false alarms due to transient noise), it automatically sends an intervention command (such as a pressure amplitude reduction command). The operator can set the "automatic intervention" mode or "alarm-only" mode on the interface.
[0070] Model update module: Deployed on a cloud server, configured to automatically trigger incremental training whenever M=5 negative samples or N=20 positive samples are added (ensuring M≤10, N≤50). Incremental training fine-tunes the existing model for 5 epochs, reducing the learning rate to 0.0001. Global retraining is performed every Monday morning (using all accumulated data, retraining for 200 epochs). After training, the new model is packaged and pushed to all edge devices, achieving synchronized model updates across the entire plant. The previous version of the model is also retained, supporting one-click rollback.
[0071] System workflow (in conjunction with Example 1): The operator inputs material information and target density through the target setting module.
[0072] After sintering begins, the time-series classification model predicts risks in real time.
[0073] If an alert is issued, the alert and intervention module will automatically implement intervention or prompt the operator.
[0074] After sintering, the feedback update module adds the data to the database based on the actual results (success / failure) and triggers retraining when the threshold is met.
[0075] Practical Application Results: This system has been tested for three months in the oscillating pressure sintering production of SiC ceramics and Ti-6Al-4V titanium alloys, processing over 800 batches. The system successfully issued early warnings of cracking precursors 67 times, of which 62 instances were prevented by automatic intervention, while 5 instances resulted in cracking due to the operator setting the mode to "alarm only" and thus failing to intervene automatically. Insufficient density warnings occurred 28 times, all of which were avoided by automatically increasing the pressure amplitude or extending the holding time. The overall product yield improved from 92% to 97.5%, and the average number of trial and error attempts for developing new materials decreased from 12 to 4. The system model update module triggered 23 incremental training iterations and 6 global retraining iterations, reducing the model's false negative rate from the initial 12% to 5% and the false positive rate from 8% to 3%, fully demonstrating the advantages of continuous self-optimization.
[0076] Example 4: This embodiment demonstrates how the above method can be applied to different materials, such as silicon nitride ceramics, titanium alloys, nickel-based superalloys, and cemented carbides.
[0077] Migration steps: Collect a small amount of sintering data of the target material (at least 20 successful cases and 5 failed cases are recommended) and construct an initial negative sample database according to the specifications of Example 1.
[0078] The LSTM model from Example 1 was used, but the parameters of the first two LSTM layers were frozen, and only the last layer and the fully connected layer were fine-tuned (transfer learning). The learning rate was set to 0.0005, and training was performed for 20 epochs.
[0079] For failure modes specific to the target material (such as β-phase grain coarsening in titanium alloys and Co pooling in cemented carbide), add corresponding categories to the failure types (failure types can be expanded and are not limited to cracking / insufficient density / inhomogeneous microstructure).
[0080] The parameter range for intervention operations needs to be adjusted according to the physical properties of the target material: for example, titanium alloys are more sensitive to the heating rate, and the reduction coefficient can be set to 0.5~0.7; the pressure amplitude adjustment range of cemented carbide can be reduced to 0.1t~0.3t.
[0081] For Ti-6Al-4V titanium alloy, after using only 25 initial negative samples (including 15 cracked samples and 10 samples with insufficient density) for transfer learning, the recall rate reached 81% in the first online run. After one month of operation (accumulating 50 negative samples), the recall rate increased to 89%, achieving results close to those of SiC ceramics. This demonstrates the material versatility of the present invention.
[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for optimizing oscillating pressure sintering process based on negative sample learning, characterized in that, Includes the following steps: S1: Construct a negative sample database, which records structured information of sintering failure cases. Each negative sample includes at least: process parameter values at the time of failure, failure type, and multi-source sensor data sequence within a preset time window before the failure. The failure type includes one or more of cracking, insufficient density, and uneven structure. The multi-source sensor data sequence includes pressure value sequence, temperature value sequence, and displacement value sequence. S2: Train a time-series classification model using the negative sample database and the positive sample database. The time-series classification model takes multi-source sensor data sequences as input and outputs anomaly risk category or anomaly risk probability. Cost-sensitive learning or oversampling methods are used to handle the imbalance between positive and negative samples during training. S3: During the oscillating pressure sintering process, the current multi-source sensor data sequence is collected in real time and input into the trained time series classification model to predict whether there is an abnormal risk in the current sintering state and the specific type of abnormal risk. S4: If an abnormal risk is predicted and the probability of the abnormal risk exceeds a preset threshold, an intervention operation will be automatically executed. The intervention operation includes at least one of the following: reducing the heating rate, reducing the pressure amplitude, reducing the pressure frequency, extending the heat preservation time, and issuing an alarm signal.
2. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, In step S1, the multi-source sensing data sequence further includes one or more of the following: acoustic emission signal sequence, infrared thermal imaging temperature distribution sequence, or furnace atmosphere concentration sequence.
3. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, The negative sample database also records textual descriptions of the reasons for failure; the reasons for failure include one or more of the following: excessively rapid heating rate, pressure amplitude exceeding the material strength limit, uneven pressure distribution due to mold wear, and abnormal sintering activity of powder batches.
4. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, The temporal classification model is a long short-term memory network model, a temporal convolutional network model, or a Transformer model based on an attention mechanism.
5. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, The positive sample database records the process parameters and multi-source sensor data sequences of successful sintering cases; the time-series classification model is trained using joint training of positive and negative samples, where the loss weight of negative samples is higher than that of positive samples.
6. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1 or 5, characterized in that, The loss function expression for the cost-sensitive learning is: in, For batch size, For the first The true label of each sample For category The corresponding loss weights are set as follows: the weight for positive sample classes is set to 1, and the weight for negative sample classes is set to a value greater than 1. The total number of categories, For the model to the first The output log probability of each category.
7. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1 or 5, characterized in that, The oversampling method uses the SMOTE algorithm to synthesize new negative samples. The formula for generating the new samples is: in, For a selected negative sample feature vector, The feature vector of the next nearest negative sample to the given negative sample. A random number between 0 and 1.
8. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 4, characterized in that, When the time-series classification model is a long short-term memory network model, its unit state update follows the following formula: in, For a moment The input sensor data vector, In hidden state, For the unit state, The activation value for the forget gate. This is the activation value of the input gate. This is the activation value of the output gate. and For trainable weights and biases, Here is the Sigmoid function, and ⊙ represents element-wise multiplication.
9. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1 or 6, characterized in that, The preset threshold Adjustments will be made dynamically based on the current sintering stage. in, Based on the threshold, , The adjustment amount is positive, and ≥ , This is the threshold for the end of the initial sintering stage. This is the threshold for the start time of the final sintering stage; Alternatively, the preset threshold is based on the most recent Accuracy of the prediction Automatic adjustment: in, For learning rate, For the target accuracy, For the most recent The actual accuracy of the prediction.
10. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, In step S4, reducing the heating rate specifically involves multiplying the current heating rate by a coefficient of 0.5 to 0.8; reducing the pressure amplitude specifically involves reducing the current pressure amplitude by 0.2t to 0.5t, or reducing it to 50% to 80% of the current pressure amplitude; reducing the pressure frequency specifically involves reducing the current pressure frequency by 0.1Hz to 0.3Hz; and extending the heat preservation time specifically involves increasing the heat preservation time by 10% to 30% based on the original heat preservation time.
11. The method for optimizing oscillating pressure sintering process based on negative sample learning according to claim 1, characterized in that, The method is used for oscillatory pressure sintering of silicon carbide ceramics, silicon nitride ceramics, titanium alloys, nickel-based superalloys, or cemented carbides.
12. A system for optimizing an oscillating pressure sintering process based on negative sample learning, characterized in that, include: The negative sample construction module is used to record structured information of sintering failure cases and build a negative sample database. A time-series classification model, deployed on an edge computing unit or cloud platform, is trained using the method described in any one of claims 1 to 11, and is used to predict abnormal risks in the current sintering state in real time. The early warning and intervention module is used to automatically execute the intervention operation described in step S4 of claim 1 when an abnormal risk is predicted to exist and the probability exceeds a preset threshold. The model update module is used to add newly occurring failure or success cases to the corresponding database and periodically retrain or fine-tune the time series classification model.
13. The oscillating pressure sintering process optimization system based on negative sample learning according to claim 12, characterized in that, The early warning and intervention module includes: a human-machine interface that displays early warning information and suggested operations, and an automatic control command sent to the actuator of the sintering equipment.
14. The oscillating pressure sintering process optimization system based on negative sample learning according to claim 12, characterized in that, The model update module is configured to automatically trigger incremental training whenever M negative samples or N positive samples are added, where M≤10 and N≤50.
Citation Information
Patent Citations
Kiln control method based on visual identification and kiln capable of automatically adjusting temperature
CN117109324A
