Method and system for monitoring and optimizing powder metallurgy sintering process of wear resistant materials

Abstract

The present application relates to the technical field of process control, in particular to a wear-resistant material powder metallurgy sintering process monitoring and optimization method and system. The method comprises: acquiring multi-source time series data of temperature and pressure, combining phase diagram of materials to label phase state coding, and generating multi-modal time series feature sequence; inputting the sequence into a time series prediction network, weighting feature channels based on a parameter sensitivity vector, and processing the sequence using a stage gating causal attention mechanism to obtain a context vector containing process stage relationships; predicting performance indicators such as hardness and density based on the vector, and calculating a performance fluctuation gradient matrix; solving the optimal process parameter adjustment amount according to the matrix, and updating the parameter sensitivity vector for the next cycle. That is, the scheme of the present application forms a closed-loop control system by fusing phase state information and identifying stage association, thereby improving product quality consistency and yield, and improving the accuracy of model prediction and control.

Description

Technical Field

[0001] This invention relates to the field of process control technology. More specifically, this invention relates to a method and system for monitoring and optimizing the sintering process of wear-resistant powder metallurgy. Background Technology

[0002] The sintering process in powder metallurgy is a core step that determines the performance of the final product. This process involves heat-treating a pressed powder blank under pre-set high temperature, high pressure, and a specific atmosphere to promote physicochemical reactions between powder particles, thereby obtaining a dense body with target density, hardness, and microstructure. During sintering, changes in process parameters such as temperature, pressure, and atmosphere directly affect key performance indicators of the final product, such as density and hardness. However, traditional sintering process control methods in existing technologies struggle to effectively address the impact of uncertainties such as raw material batch fluctuations and equipment status changes. This results in poor product performance consistency, low yield, and the inability to achieve online optimization.

[0003] To address the aforementioned issues, existing technologies propose utilizing time series analysis models, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), to learn the mapping relationship between sintering process data and product performance. These technical solutions directly process raw time-series data collected by sensors and perform unified modeling of the entire sintering process, attempting to establish a correlation between process parameters and product quality, thereby improving the level of intelligent process control.

[0004] However, the aforementioned existing technologies still have significant shortcomings in practical applications. First, the modeling approach that directly processes raw time-series data lacks a deep integration of the sintering process mechanism. It fails to consider that the material undergoes different phase states during sintering, such as solid, liquid, and solid-liquid coexistence phases, and this phase transition information, which plays a decisive role in the final microstructure and properties, is not reflected in the model. Second, the approach of modeling the entire process uniformly ignores the stage-specific characteristics of the sintering process itself, namely the heating, holding, and cooling stages. Since the dominant physical mechanisms and key influencing parameters differ at different stages, this uniform approach cannot effectively identify cross-stage dependencies and long-range effects, thus limiting the accuracy of model prediction and control and making it difficult to meet the requirements of high-quality and high-efficiency production. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for monitoring and optimizing the sintering process of wear-resistant powder metallurgy, in order to solve the problems of poor accuracy of model prediction and poor control precision in the prior art; to this end, this invention provides solutions in the following two aspects.

[0006] In a first aspect, the method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials provided by the present invention includes the following steps:

[0007] Multi-source time-series data, including temperature and pressure, are acquired during the sintering process. Based on the temperature and pressure data and a preset material phase diagram, a material phase code is assigned to the data at each time step. The multi-source time-series data and the phase code are then fused to generate a multimodal time-series feature sequence. This multimodal time-series feature sequence is input into a time-series prediction network. Based on the parameter sensitivity vector obtained from the previous control cycle, the feature channels corresponding to controllable process parameters in the multimodal time-series feature sequence are weighted. A stage-gated causal attention mechanism is used to process the weighted multimodal time-series feature sequence. The heating, holding, and cooling stages are divided based on the temperature change rate, and the stage is further divided into stages. Attention scores are calculated for data points within segments and between stages to obtain context vectors that include process stage relationships. Based on these context vectors, predicted values ​​of performance indicators, including at least hardness and density, are calculated in parallel, and the partial derivatives of each predicted performance indicator with respect to each controllable process parameter are calculated to obtain a performance fluctuation gradient matrix. Based on the performance fluctuation gradient matrix, the predicted performance indicators, and the preset performance target range, the optimal process parameter adjustment amount is obtained. A comprehensive sensitivity is calculated for each controllable process parameter based on the performance fluctuation gradient matrix, and the comprehensive sensitivities of each controllable process parameter are combined to update the parameter sensitivity vector for the next control cycle.

[0008] Preferably, the step of generating the multimodal time-series feature sequence includes: using the collected real-time temperature and pressure values ​​as two-dimensional coordinates, querying the corresponding phase region on a preset material phase diagram, and assigning a unique thermal code to the phase region as the phase code; concatenating the unique thermal code with the temperature and pressure values ​​at the same time step and the set values ​​of each controllable process parameter to obtain an increased-dimensional feature vector, which serves as the data point of the multimodal time-series feature sequence.

[0009] Preferably, the time-series prediction network adopts a Transformer encoder structure.

[0010] Preferably, the step-gated causal attention mechanism for processing the weighted feature sequence includes: calculating the temperature change rate between consecutive time steps, dividing the sintering process into a heating stage, a holding stage, and a cooling stage; within each stage, the data point only calculates the attention weight for the previous data point in the current stage; at the stage boundary, the gating unit calculates the attention weight for the first data point of the current stage to the last data point of the previous stage.

[0011] Preferably, the step of dividing the temperature stage is as follows: when the temperature change rate is greater than a preset first positive threshold, it is determined to be a heating stage; when the temperature change rate is less than a preset second negative threshold, it is determined to be a cooling stage; otherwise, it is determined to be a heat preservation stage.

[0012] Preferably, the step of obtaining the optimal process parameter adjustment amount includes: constructing an optimization objective function, the optimization objective function including a penalty term for performance index prediction values ​​exceeding a preset performance target range, and a regularization penalty term for the magnitude of the process parameter adjustment amount; and obtaining the optimal process parameter adjustment amount by minimizing the optimization objective function.

[0013] Preferably, the optimization objective function Defined as:

[0014] ,in, This is the optimized vector of predicted performance metrics. and These are the lower and upper bound vectors of the preset performance target range, respectively. Let be the vector of adjustment amounts for the optimal process parameters to be solved. The regularization coefficient is used to control the magnitude of the parameter adjustment. It represents the square of the L2 norm of a vector.

[0015] Preferably, the step of updating the parameter sensitivity vector includes: for each controllable process parameter in the performance fluctuation gradient matrix, extracting its corresponding column vector; calculating the L2 norm of the column vector and using it as the comprehensive sensitivity of the controllable process parameter; and combining the comprehensive sensitivities of all controllable process parameters into a new vector as the parameter sensitivity vector for the next control cycle.

[0016] Preferably, the step of calculating the predicted value of the performance index includes: inputting the context vector in parallel to multiple independent prediction head networks composed of fully connected layers, with each prediction head outputting a predicted value of a performance index.

[0017] In the second aspect, the monitoring and optimization system for the powder metallurgy sintering process of wear-resistant materials includes:

[0018] The processor; the memory storing computer instructions for monitoring and optimizing the sintering process of wear-resistant powder metallurgy, wherein when the computer instructions are executed by the processor, the system performs the aforementioned system method for monitoring and optimizing the sintering process of wear-resistant powder metallurgy.

[0019] The beneficial effects of this invention are as follows: By fusing material physical phase information during the sintering process with multi-source time-series data and utilizing a stage-gating mechanism to identify the intrinsic connections and long-range effects of different stages such as heating, holding, and cooling, the accuracy of final performance prediction is significantly improved. By predicting multiple performance indicators such as hardness and density in parallel and calculating the performance fluctuation gradient matrix, a quantitative basis for the coordinated control of multi-objective performance is provided. Based on the optimal process parameter adjustment amount obtained from the gradient solution, the sintering process can be directly optimized online to correct potential deviations. The updated parameter sensitivity vector further enhances the targeting of subsequent control, forming a closed-loop control system from prediction to control to learning, thereby improving product quality consistency and yield. Attached Figure Description

[0020] Figure 1 This schematically illustrates the step-by-step flowchart of the method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials in this embodiment;

[0021] Figure 2 The schematic diagram illustrates the structural block diagram of the wear-resistant material powder metallurgy sintering process monitoring and optimization system in this embodiment. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 As shown, the method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials in this embodiment includes the following steps:

[0024] Step S1: Obtain multi-source time-series data including temperature and pressure during the sintering process; based on the temperature and pressure data and combined with a preset material phase diagram, label the data of each time step with a material phase code, and fuse the multi-source time-series data with the phase code to generate a multi-modal time-series feature sequence.

[0025] Specifically, thermocouples, pressure sensors, and oxygen probes are arranged inside the sintering furnace to collect real-time, timestamped raw data streams, including atmospheric parameters such as furnace temperature, chamber pressure, and oxygen partial pressure. A temperature-pressure two-dimensional phase diagram is pre-established for specific wear-resistant materials (e.g., WC-Co cemented carbide), which delineates the material's state regions under different temperature and pressure conditions, such as the solid phase region, liquid phase region, and solid-liquid coexistence region.

[0026] The steps for generating a multimodal time-series feature sequence include: For each sampling time point, using the real-time temperature and pressure values ​​at that moment as two-dimensional coordinates, locating the region where it is located in a preset material phase diagram. The system assigns a unique thermal code as its physical phase code for each phase region; for example, the solid phase region corresponds to the code [1,0,0], the solid-liquid coexistence region corresponds to [0,1,0], and the liquid phase region corresponds to [0,0,1]. The multi-source time-series data vector at that time point is concatenated with the corresponding physical phase code vector to obtain a fused feature vector. The fused feature vectors of all time points are arranged in chronological order to finally form a multimodal time-series feature sequence for analysis in subsequent steps.

[0027] For example, materials exhibit different phases under different temperatures and pressures, such as solid, liquid, and solid-liquid coexistence phases. These phases directly determine the microstructure and properties of the material. Assume the material exists in three states: solid, solid-liquid coexistence phase, and liquid. At time step t, the sensor acquires a temperature of 1700 degrees Celsius and a pressure of 1 standard atmosphere. By querying a preset phase diagram, this coordinate point falls within the solid-liquid coexistence region. If the unique thermal codes for the three phases are 100 for solid, 010 for solid-liquid coexistence phase, and 001 for liquid, then the physical phase code for the current time step is 010. If the sensor also acquires a protective atmosphere argon concentration of 99.9%, a laser power setting of 300 watts, and a scanning speed of 1000 mm / s, then the fused multimodal feature vector is a vector composed of 1700, 1, 99.9, 300, 1000, 0, 1, and 0.

[0028] Step S2: Input the multimodal time-series feature sequence into the time-series prediction network. Based on the parameter sensitivity vector obtained in the previous control cycle, weight the feature channels in the multimodal time-series feature sequence corresponding to the controllable process parameters. Use a stage-gated causal attention mechanism to process the weighted multimodal time-series feature sequence. Divide the heating, holding and cooling stages based on the temperature change rate, and calculate the attention score for the data points within and between stages to obtain a context vector that includes the process stage relationships.

[0029] Specifically, the time-series prediction network preferably employs a Transformer encoder structure. Before inputting the multimodal time-series feature sequence into the network, the sensitivity values ​​of each controllable process parameter, such as heating power and pressurization rate, are read from the parameter sensitivity vector calculated and stored at the end of the previous control cycle. The feature channels corresponding to the controllable parameters in the multimodal time-series feature sequence are multiplied by their corresponding sensitivity values ​​to enhance the representation weight of key parameters in the model.

[0030] Calculate the first-order difference of temperature with respect to time for the entire time series, i.e., the rate of temperature change. Set a positive threshold and a negative threshold, and classify the series points with a rate of temperature change greater than the positive threshold as the heating stage, less than the negative threshold as the cooling stage, and those in between as the holding stage.

[0031] In the causal attention module, a learnable stage-gated matrix is ​​introduced for the interaction between the query vector Q and the key vector K. The element values ​​of this matrix depend on the combination of stages to which Q and K belong. For example, the influence of the heating stage on the heat preservation stage, or the influence between data points within the heat preservation stage. By multiplying this gating matrix element-wise with the standard attention score matrix, the intensity of information transmission between different stages is adjusted, and a weighted sum is obtained to obtain a context vector that incorporates the phase-specific patterns of the process. Specifically, the stage-gated causal attention mechanism divides the entire time series into several independent stages. Within each stage, data points only focus on historical data occurring before them within that stage. At the boundary between two stages, a gating unit ensures that the first data point of the new stage receives the last data point of the previous stage.

[0032] More specifically, stage-gated causal attention mechanisms are further divided into intra-stage causal attention and inter-stage causal attention. In intra-stage causal attention, for any time step t, the scope of attention computation is limited to its own stage S, and it can only focus on time steps earlier than t within that stage. ,in and Inter-stage causal attention: For the first data point in stage S, in addition to calculating its attention with itself according to the rules, an attention score with the last data point of the previous stage is also calculated through a gating mechanism.

[0033] In an optional embodiment, a stage-gated causal attention mechanism is used to process the weighted feature sequence, including: calculating the temperature change rate between consecutive time steps, dividing the sintering process into a heating stage, a holding stage, and a cooling stage; within each stage, the attention weight of the data point is calculated only for the previous data point in the current stage; at the stage boundary, the attention weight of the first data point in the current stage is calculated for the last data point in the previous stage through a gating unit.

[0034] The specific steps for dividing the temperature stages are as follows: when the temperature change rate is greater than a preset first positive threshold, it is determined to be a heating stage; when the temperature change rate is less than a preset second negative threshold, it is determined to be a cooling stage; otherwise, it is determined to be a heat preservation stage.

[0035] Specifically, a typical heat treatment process can be divided into stages such as heating, holding, and cooling. The system state patterns within each stage are similar, while the relationships between different stages have specific key nodes. Traditional attention mechanisms may cause the model to focus on temporally close but physically unrelated data from the distant past, such as incorrectly focusing on details of the heating stage during the cooling stage, causing information interference. Taking a heat treatment process as an example, assuming a heating threshold of 0.5°C per second and a cooling threshold of -0.5°C per second, when the system temperature rises from 500°C to 510°C in 10 seconds, the temperature change rate is 1°C per second, exceeding the heating threshold, thus indicating the start of the heating stage. During this stage, when predicting the state at the 5th second, the focus is on the data from the first 4 seconds, ignoring data from earlier cooling or holding stages. When the heating stage ends and the holding stage begins, for example, when the temperature stabilizes at 800°C, the first data point of the holding stage will pass through a gating unit, allowing special attention to the last data point of the heating stage, i.e., the system state at 800°C.

[0036] Step S3: Based on the context vector, calculate the predicted values ​​of performance indicators, including at least hardness and density, in parallel, and calculate the partial derivative of each performance indicator prediction value with respect to each controllable process parameter to obtain the performance fluctuation gradient matrix.

[0037] Specifically, the final context vector generated in step S2 is simultaneously input into two or more independent prediction head networks. These prediction heads consist of fully connected layers, and each prediction head is responsible for outputting a specific performance index. For example, the first prediction head network outputs the predicted hardness value of the material after sintering, and the second prediction head network outputs the predicted density value.

[0038] Since the entire calculation process from input to output consists of differentiable neural network modules, the built-in automatic differentiation function of the deep learning framework is utilized. This function can calculate the partial derivative of each performance index prediction value (such as hardness and density) with respect to each controllable process parameter in the input sequence (such as the target temperature curve and the target pressure curve set at each time point).

[0039] All calculated partial derivative values ​​are combined into a matrix, where rows represent different performance indicators (such as hardness and density) and columns represent different controllable process parameters. This final generated matrix is ​​the performance fluctuation gradient matrix.

[0040] Step S4: Based on the performance fluctuation gradient matrix, the predicted value of the performance index, and the preset performance target range, the optimal process parameter adjustment amount is obtained; a comprehensive sensitivity is calculated for each controllable process parameter based on the performance fluctuation gradient matrix, and the comprehensive sensitivity of each controllable process parameter is combined and updated into a parameter sensitivity vector for the next control cycle.

[0041] Specifically, the objective function aims to minimize the weighted squared error between the predicted values ​​of each performance index and the central value of the target interval, while adding a regularization term to penalize excessive parameter adjustments. Using the performance fluctuation gradient matrix, the change in the predicted performance values ​​can be linearly approximated as a function of the parameter adjustments. By solving a constrained quadratic programming problem, a set of optimal process parameter adjustments is obtained, such as increasing the heating rate of the next stage by 5 degrees Celsius per minute and increasing the holding pressure by 2 MPa. These adjustments are then sent to the PLC controller of the sintering furnace for execution.

[0042] Extract the column vector corresponding to each controllable process parameter in the performance fluctuation gradient matrix, and calculate the L2 norm or weighted norm of this column vector. Use this as an indicator of the parameter's overall performance impact under the current process state, i.e., the overall sensitivity. Combine the overall sensitivities of all controllable process parameters into a new vector, and update it to the parameter sensitivity vector used at the start of the next control cycle.

[0043] In an optional embodiment, the step of obtaining the optimal process parameter adjustment amount includes: constructing an optimization objective function, the optimization objective function including a penalty term for performance index prediction values ​​exceeding a preset performance target range, and a regularization penalty term for the magnitude of the process parameter adjustment amount; and obtaining the optimal process parameter adjustment amount by minimizing the optimization objective function.

[0044] Optimize objective function Defined as:

[0045] ,

[0046] in, This is the optimized vector of predicted performance metrics. and These are the lower and upper bound vectors of the preset performance target range, respectively. Let be the vector of adjustment amounts for the optimal process parameters to be solved. The regularization coefficient is used to control the magnitude of the parameter adjustment. It represents the square of the L2 norm of a vector.

[0047] In actual production, the goal is not to adjust performance indicators to a precise value, but rather to bring them into an acceptable range. At the same time, adjustments to process parameters should not be too large, otherwise it may lead to equipment oscillation or production instability. Suppose we need to control two performance indicators: tensile strength and elongation. The target is tensile strength between 1100 and 1200 MPa, and elongation between 15% and 20%. The current predicted performance is tensile strength of 1080 MPa and elongation of 21%, both outside the target range. The first term of the optimization function incurs a penalty because 1100 minus 1080 is greater than zero. The second term incurs another penalty because 21% minus 20% is greater than zero. The third term ensures that the calculated adjustments to process parameters, such as increasing laser power and decreasing powder feeding rate, are not excessively large. By minimizing the objective function, an optimal adjustment scheme is calculated, such as increasing laser power by 10 watts and decreasing powder feeding rate by 0.2 grams per minute, so that the adjusted predicted performance is closest to or even enters the target range, while ensuring a smooth adjustment.

[0048] In an optional embodiment, the step of updating the parameter sensitivity vector includes: for each controllable process parameter in the performance fluctuation gradient matrix, extracting its corresponding column vector; calculating the L2 norm of the column vector and using it as the comprehensive sensitivity of the controllable process parameter; and combining the comprehensive sensitivities of all controllable process parameters into a new vector as the parameter sensitivity vector for the next control cycle.

[0049] The influence of each process parameter is not static; it can change with variations in the working environment or material state. By calculating the overall sensitivity, the control weight of each parameter can be understood in real time. For example, suppose there are two controllable process parameters, heating temperature and holding time, which together affect the product's hardness and toughness. At a certain moment, the first column of the calculated performance fluctuation gradient matrix, corresponding to the partial derivative vector of heating temperature, is 10, -0.1, indicating that for every 1°C increase in temperature, hardness increases by 10 units and toughness decreases by 0.1 units. The second column, corresponding to the partial derivative vector of holding time, is 2, 0.5. Calculations show that the overall sensitivity of heating temperature is approximately 10, and the overall sensitivity of holding time is approximately 2.06. This indicates that, under the current conditions, the impact of heating temperature on the overall product performance is far greater than that of holding time. The new parameter sensitivity vector will be updated to [10, 2.06] and used to guide the next round of control decisions, so that when adjustments are needed, the system will prioritize adjusting the more sensitive heating temperature, thereby achieving more efficient control.

[0050] This invention also provides a monitoring and optimization system for the powder metallurgy sintering process of wear-resistant materials. For example... Figure 2As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the above-described method for monitoring and optimizing the sintering process of wear-resistant powder metallurgy materials according to the present invention.

[0051] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0052] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0053] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0054] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for monitoring and optimizing the sintering process of wear-resistant materials using powder metallurgy, characterized in that, Includes the following steps: Acquire multi-source time-series data including temperature and pressure during the sintering process; based on the temperature and pressure data and combined with the preset material phase diagram, label the data of each time step with a material phase code, and fuse the multi-source time-series data with the phase code to generate a multi-modal time-series feature sequence; The multimodal time-series feature sequence is input into the time-series prediction network. Based on the parameter sensitivity vector obtained from the previous control cycle, the feature channels corresponding to the controllable process parameters in the multimodal time-series feature sequence are weighted. The weighted multimodal time-series feature sequence is processed by a stage-gated causal attention mechanism. The heating, holding and cooling stages are divided based on the temperature change rate. Attention scores are calculated for data points within and between stages to obtain a context vector that includes the relationship between process stages. Based on context vectors, the predicted values ​​of performance indicators, including at least hardness and density, are calculated in parallel. The partial derivatives of the predicted value of each performance indicator with respect to each controllable process parameter are calculated to obtain the performance fluctuation gradient matrix. Based on the performance fluctuation gradient matrix, the predicted performance index values, and the preset performance target range, the optimal process parameter adjustment amount is obtained by solving the following steps: constructing an optimization objective function, which includes a penalty term for the predicted performance index values ​​exceeding the preset performance target range and a regularization penalty term for the magnitude of the process parameter adjustment amount; and minimizing the optimization objective function to obtain the optimal process parameter adjustment amount. Optimize objective function for: ； in, This is the optimized vector of predicted performance metrics. and These are the lower and upper bound vectors of the preset performance target range, respectively. Let be the vector of adjustment amounts for the optimal process parameters to be solved. The regularization coefficient is used to control the magnitude of the parameter adjustment. It is the square of the L2 norm of the vector; Calculate a comprehensive sensitivity for each controllable process parameter based on the performance fluctuation gradient matrix; combine the comprehensive sensitivities of each controllable process parameter to update the parameter sensitivity vector for the next control cycle, including: for each controllable process parameter in the performance fluctuation gradient matrix, extract its corresponding column vector; calculate the L2 norm of the column vector as the comprehensive sensitivity of the controllable process parameter; combine the comprehensive sensitivities of all controllable process parameters into a new vector as the parameter sensitivity vector for the next control cycle.

2. The method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials according to claim 1, characterized in that, The steps for generating multimodal temporal feature sequences include: The collected real-time temperature and pressure values ​​are used as two-dimensional coordinates. The corresponding phase region is queried on the preset material phase diagram, and a unique thermal code is assigned to the phase region as the phase code. The unique thermal encoding is concatenated with the temperature and pressure values ​​at the same time step, as well as the set values ​​of each controllable process parameter, to obtain an increased-dimensional feature vector, which serves as the data point of the multimodal time-series feature sequence.

3. The method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials according to claim 1, characterized in that, The time-series prediction network adopts a Transformer encoder structure.

4. The method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials according to claim 1, characterized in that, The process of using a stage-gated causal attention mechanism to process the weighted feature sequence includes: Calculate the rate of temperature change between consecutive time steps and divide the sintering process into a heating stage, a holding stage, and a cooling stage. Within each stage, attention weights are calculated only for previous data points within the current stage; At the stage boundary, attention weights are calculated for the first data point of the current stage and the last data point of the previous stage using gating units.

5. The method for monitoring and optimizing the powder metallurgy sintering process of wear-resistant materials according to claim 4, characterized in that, The specific steps for dividing the temperature stages are as follows: When the temperature change rate is greater than a preset first positive threshold, it is determined to be a heating stage; when the temperature change rate is less than a preset second negative threshold, it is determined to be a cooling stage; otherwise, it is determined to be a heat preservation stage.

6. The method for monitoring and optimizing the sintering process of wear-resistant material powder metallurgy according to claim 1, characterized in that, The steps for calculating the predicted values ​​of performance metrics include: The context vector is input in parallel to multiple independent prediction head networks composed of fully connected layers, and each prediction head outputs a predicted value of a performance metric.

7. A monitoring and optimization system for the sintering process of wear-resistant powder metallurgy, characterized in that, include: processor; A memory storing computer instructions for monitoring and optimizing the sintering process of wear-resistant powder metallurgy, wherein when the computer instructions are executed by the processor, the system performs the method for monitoring and optimizing the sintering process of wear-resistant powder metallurgy according to any one of claims 1-6.

Images