Intelligent real-time judgment method for terminal point of ultrahigh-pressure mercury lamp zirconium wire high-temperature activation process

Abstract

The invention provides a method based on multi-modal data acquisition and intelligent feature fusion, and the method comprises the steps: collecting and aligning temperature and wide-spectrum spectral signals and process time sequence data in real time, carrying out the denoising and normalization preprocessing, and achieving the recognition of a process stage through a pre-training time sequence classification network. And calling a special feature extraction model according to stages to calculate key criterion features. A dynamic comprehensive judgment score is generated through multi-model fusion weighting, and a judgment threshold is adaptively adjusted in combination with historical data, so that continuous intelligent judgment of an activation end point and process termination closed-loop control are realized. According to the method, the accuracy and the real-time performance of end point judgment are improved, the adaptability to working condition changes is enhanced, stage confidence distribution is output online based on a pre-training time sequence classification network, and a special feature extraction module is dynamically scheduled according to the stage confidence distribution, so that key physical indexes of each stage are specifically captured, and the accuracy and the instantaneity of end point judgment are improved. And the correlation and the sensitivity of the criterion are remarkably improved, so that the accuracy and the robustness of activation end point recognition are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent determination technology for high-precision optical device manufacturing processes, and in particular to an intelligent real-time determination method for the endpoint of a high-temperature activation process for zirconium filaments in ultra-high pressure mercury lamps. Background Technology

[0002] Currently, the endpoint determination in the high-temperature activation process of zirconium filaments for ultra-high pressure mercury lamps mainly relies on fixed judgment logic and static threshold settings. In practical applications, threshold comparisons based on single sensor data (such as temperature curves or spectral characteristics) are commonly used, or machine learning models are employed to classify process signals to determine the activation endpoint. This type of technical approach can achieve a certain level of accuracy in relatively simple and stable processes such as heating, holding, and cooling. However, in scenarios with complex multi-stage processes or significant changes in operating conditions, the system's judgment accuracy and response consistency face significant bottlenecks.

[0003] In recent years, with the increasing demands of optical device manufacturing, the industry has gradually introduced methods such as high-frequency data acquisition, time series analysis, and multimodal feature fusion. These methods collect multi-source data such as temperature signals and spectral reflectance through sensor networks and employ neural networks and ensemble learning to improve process discrimination capabilities. Some technical solutions attempt to use deep learning models (such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or combinations thereof) to perform end-to-end modeling of the entire process, achieving endpoint identification and anomaly alarms. However, these methods generally have the following characteristics: First, the judgment logic is fixed, meaning model training and judgment threshold setting are usually completed under static process conditions, making dynamic adjustment according to the actual process stage is difficult; second, there is insufficient decoupling of judgment criteria between stages, with different process stages often using the same criterion system, making it difficult to fully utilize the differences in features within and outside the stage; third, they heavily rely on highly complex algorithms or large amounts of labeled data, resulting in weak model interpretability and engineering controllability, making it difficult to adapt to the stable operation requirements of complex actual process scenarios.

[0004] In the practical application of zirconium filament activation technology for ultra-high pressure mercury lamps, existing technologies are mainly applicable to endpoint determination in single-stage processes. For example, setting a fixed threshold based solely on the rate of temperature change or spectral peak signals can only achieve limited accuracy in processes with small fluctuations in process characteristics and few anomalies. For complex processes with alternating heating, holding, and cooling stages and significant dynamic changes in process characteristics, fixed determination logic often suffers from threshold drift, false triggering, or missed determinations. Furthermore, its robustness is further reduced by changes in environmental parameters and equipment aging. While some deep learning or fusion perception methods can perform multi-feature discrimination under rich data conditions, their model versatility and engineering generalization are limited, making it difficult to meet real-time and adaptive requirements without relying on large-scale reinforcement learning or knowledge graph modeling.

[0005] In summary, existing technologies for endpoint determination in multi-stage processes have the following technical shortcomings: First, they cannot achieve dynamic perception and adaptive determination of process stages, and the determination algorithms and threshold settings lack specificity during stage transitions; second, fixed thresholds result in rigid determination standards, which are prone to misjudgment or determination lag due to fluctuations between process batches, which is not conducive to the consistency and high-precision control of complex processes; third, the determination mechanism lacks interpretability and closed-loop feedback, making it difficult to provide data basis for process optimization and model iteration. Summary of the Invention

[0006] This application provides an intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in ultra-high pressure mercury lamps, aiming to solve the problems existing in the prior art mentioned in the background.

[0007] This application provides a method for intelligent real-time determination of the endpoint of a high-temperature activation process for a zirconium filament in an ultra-high pressure mercury lamp, specifically including: S1: Real-time acquisition of temperature time series data and broadband spectral data stream during the high-temperature activation process of zirconium filament in ultra-high pressure mercury lamp, and recording of time stamp information of process stages to obtain the raw dataset; S2: Denoise and normalize the original dataset to generate a standardized temperature curve sequence and spectral reflectance matrix; S3: The process stage identification engine, built using a pre-trained time series classification network, performs online inference on the standardized temperature curve sequence and outputs the category confidence distribution of the process stage at the current moment. S4: Determine the current dominant process stage based on the category confidence distribution, and calculate the key criterion features by calling the corresponding dedicated feature extraction sub-model group based on the stage category; S5: Input the special features of each stage into the multi-model output fusion module, and combine the confidence weight of the current dominant process stage to weight and integrate the features of each criterion to generate a dynamic comprehensive judgment score. S6: Based on the current dominant process stage, load the corresponding dynamic judgment threshold range from the preset rule base, and perform drift compensation on the upper and lower limits of the threshold range to generate an adaptive judgment threshold. S7: Determine whether the dynamic comprehensive judgment score continuously exceeds the adaptive judgment threshold and maintains it within a preset time window. If the condition is met, generate an activation endpoint confirmation signal and trigger a process termination command. S8: While generating the activation endpoint confirmation signal, record the main contribution characteristics and their weight change trends of each stage in this activation process.

[0008] The beneficial technical effects of this invention are as follows: 1) This invention introduces a lightweight process stage identification engine and a differentiated feature extraction sub-model group to achieve accurate perception and criterion decoupling of the dynamic characteristics of multiple stages in the high-temperature activation process of zirconium wire. Based on a pre-trained time series classification network, this invention outputs the stage confidence distribution online and dynamically schedules a dedicated feature extraction module accordingly, so that key physical indicators of each stage can be captured in a targeted manner, significantly improving the relevance and sensitivity of the criteria, thereby greatly improving the accuracy and robustness of activation endpoint identification. 2) A comprehensive judgment framework integrating dynamic threshold adjustment mechanism and weighted multi-model output was constructed. Combined with historical batch data drift compensation strategy, adaptive decision-making capability was achieved in non-steady-state process environment. This invention automatically loads the corresponding threshold range according to the currently identified process stage and introduces a trend correction algorithm based on historical statistics to adjust the upper and lower limit boundaries in real time to ensure that the judgment criteria always fit the current working conditions. At the same time, the fusion module records the main contribution features of each stage and their weight evolution path through interpretable feedback path, which improves the quality control level of the manufacturing process. 3) This invention adopts a modular decoupling design driven by stage perception, which avoids the high computational overhead and deployment difficulty brought about by complex reinforcement learning or large-scale neural network joint modeling, and has the advantages of low latency, high efficiency and easy integration. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of calculating the category confidence distribution of the current process stage in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the calculation of key criterion features in an embodiment of the present invention. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0012] like Figure 1 As shown, this embodiment provides an intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in ultra-high pressure mercury lamps, specifically including: S1: Real-time acquisition of temperature time series data and broadband spectral data stream during the high-temperature activation process of zirconium filament in ultra-high pressure mercury lamp, and simultaneous recording of time stamp information of process stages to form a raw dataset containing multimodal sensing signals and operating context. S2: The original dataset is preprocessed by denoising and normalization to eliminate environmental interference and sensor drift effects, and to generate a standardized temperature curve sequence and spectral reflectance matrix, which serve as the basic input for subsequent identification and feature extraction. S3: Construct a process stage identification engine based on a pre-trained time series classification network, input the standardized temperature curve sequence into the engine, perform online inference, and output the category confidence distribution of the current process stage, wherein the process stage includes heating stage, heat preservation stage and cooling stage; S4: Determine the current dominant process stage based on the confidence distribution of the categories, and call the corresponding dedicated feature extraction sub-model group based on the stage category to calculate key criteria features for different stages: extract the temperature rise slope and spectral reflectance inflection point for the heating stage, extract the spectral energy concentration and thermal radiation stability index for the heat preservation stage, and extract the cooling rate and spectral attenuation consistency coefficient for the cooling stage. S5: Input the special features of each stage into the multi-model output fusion module, combine the confidence weight of the current dominant process stage to weight and integrate each criterion feature, and generate a dynamic comprehensive judgment score. This score is a quantitative indicator that reflects the degree to which the activation process is close to the end point. S6: Based on the current dominant process stage, load the corresponding dynamic judgment threshold range from the preset rule base, and use historical batch data to compensate for the drift of the upper and lower limits of the threshold range to generate an adaptive judgment threshold that is suitable for the current working conditions. S7: Determine whether the dynamic comprehensive judgment score continuously exceeds the adaptive judgment threshold and is maintained within a preset time window. If the condition is met, generate an activation endpoint confirmation signal and trigger a process termination command.

[0013] S8: While generating the activation endpoint confirmation signal, record the main contribution characteristics and their weight change trends of each stage in this activation process to form an interpretable feedback log to support subsequent process optimization and model iteration.

[0014] In this embodiment, step S1 involves real-time acquisition of temperature time-series data and broadband spectral data streams during the high-temperature activation process of the zirconium filament in the ultra-high pressure mercury lamp, and simultaneous recording of time stamp information for each process stage, to form a raw dataset containing multimodal sensing signals and operating context. Specifically, this includes the following steps: S1.1: Acquire the temperature simulation signal output from the thermocouple sensor arranged inside the high-temperature activation furnace, perform analog-to-digital conversion processing based on the high-precision ADC module to generate a discrete temperature time-series sampling sequence, which serves as the basic physical quantity input reflecting the heating process of the zirconium wire; S1.2: Acquire the broadband (200–800 nm) reflectance spectrum simulation signal collected by the fiber optic spectrometer, perform wavelength calibration and photocurrent digitization processing using the embedded spectral acquisition unit, generate a spectral reflectance vector sequence with wavelength index alignment relationship, and form an optical response input characterizing the surface state evolution of the zirconium wire; S1.3: Based on the process control timing signal output by the PLC control system, extract the preset time window boundary information of the current operating stage, and generate a process stage time mark sequence synchronized with the physical sampling clock. This sequence provides macroscopic process position markings for the heating section, heat preservation section and cooling section. S1.4: Perform timestamp alignment processing on the temperature time-series sampling sequence, spectral reflectance vector sequence and process stage time stamp sequence, and use interpolation compensation algorithm to fill the data asynchrony gap caused by sampling frequency differences to generate a multimodal sensing data frame set under a unified time reference. S1.5: The multimodal sensing data frame set is packaged and stored according to the batch number and furnace ID to generate an original dataset with working condition context labels. This dataset serves as a unified input source for subsequent denoising normalization, stage identification and feature extraction.

[0015] In this embodiment, step S2 involves denoising and normalizing the original dataset to eliminate environmental interference and sensor drift, generating a standardized temperature curve sequence and spectral reflectance matrix as the basic input for subsequent identification and feature extraction. Specifically, this includes the following steps: S2.1: Acquire the raw time-series data stream from the temperature sensor and broadband spectrometer, and perform low-frequency noise suppression processing on the temperature time-series data based on the sliding window mean filtering algorithm to generate a smoothed temperature signal sequence as a preliminary noise reduction output; S2.2: The original broadband spectral data stream is decomposed and reconstructed using a wavelet threshold denoising algorithm. The high-frequency detail coefficients of the spectrum are extracted using the Daubechies wavelet basis function and soft threshold quantization is performed to remove random noise components and obtain the denoised spectral reflectance time series matrix. S2.3: The smoothed temperature signal sequence is input to the adaptive anomaly detection module. This module identifies anomalous sampling points that deviate significantly from the normal trend based on the Local Outlier Factor (LOF) algorithm and repairs them using linear interpolation to generate a continuous and abrupt clean temperature curve sequence. S2.4: Based on the minimum-maximum normalization method, scale unification processing is performed on the clean temperature curve sequence and the denoised spectral reflectance time series matrix respectively, mapping them to the [0,1] interval to eliminate dimensional differences and generate a standardized temperature curve sequence and a standardized spectral reflectance matrix. S2.5: Align and fuse the standardized temperature curve sequence with the standardized spectral reflectance matrix according to the timestamp to generate a structured multimodal standardized dataset, which serves as the unified input data source for the next stage process stage identification engine and feature extraction sub-model group.

[0016] In this embodiment, in step S3, a process stage identification engine is constructed based on a pre-trained time series classification network. The standardized temperature curve sequence is input into the engine, online inference is performed, and the category confidence distribution of the current process stage is output. The process stage includes a heating stage, a holding stage, and a cooling stage, such as... Figure 2 As shown, the specific steps include the following: S3.1: Based on the standardized temperature curve sequence archived in historical batches and the corresponding manually labeled process stage, a training dataset is constructed using a time series classification task. Each sample contains a fixed-length sliding window slice and the corresponding stage category label to support supervised learning modeling of the process stage recognition engine. Furthermore, by using label mapping rules, each sliding window segment is bound to the stage label that appears most frequently within its time coverage range, thus realizing the correspondence between segments and stage categories and generating label vectors.

[0017] Furthermore, a time-series data structuring encapsulation algorithm is used to index and store all window segments by batch number and furnace ID to generate a training sample set, ensuring the traceability and batch correlation of the model input.

[0018] Furthermore, the feature consistency test method is applied to verify whether the temperature curve segments in the training sample set conform to a unified standardized scale. Inconsistent samples are re-normalized to eliminate scale bias and improve the stability of model training.

[0019] S3.2: A lightweight Temporal Fusion Transformer (TFT) network architecture is adopted as the basic model. A temporal feature extraction structure including static variable embedding, known future input encoding and multi-head attention mechanism is designed. The temperature sequence in the training dataset is trained end-to-end, the cross-entropy loss function is optimized, and a pre-trained process stage recognition engine with the ability to recognize stage evolution patterns is generated. Furthermore, a low-dimensional vectorized mapping of static process contexts such as batch ID and furnace ID is achieved through a static variable embedding module, and then time-step splicing and fusion are performed with dynamic temperature sequence features to obtain an input tensor with global operating condition awareness.

[0020] Furthermore, the known future input encoding module enables the encoding and expression of future process control variables, and integrates them with the current temperature sequence features in the time dimension to improve the model's ability to predict the stage evolution trend.

[0021] Furthermore, a multi-head attention mechanism is used to achieve weighted modeling of the global dependency of the temperature sequence, and the correlation weight matrix between different time steps is calculated to capture the influence weight of cross-stage features on the current stage determination.

[0022] Furthermore, a gated residual network (GRN) is used to achieve nonlinear transformation of cross-module features and enhance interaction between channels, generating a high-dimensional stage discrimination feature vector after feature fusion and attention weighting.

[0023] Using cross-entropy loss function Optimize model parameters, cross-entropy loss function The calculation formula is: ; in, The total number of samples, For the first The predicted probability values ​​of the true stage category in each sample are used to iteratively update the network weight matrix using the gradient descent method, which significantly improves the accuracy of stage identification.

[0024] S3.3: The real-time standardized temperature curve sequence generated in step S2 is fragmented according to the preset sliding window length, and then restored to the original dimension range through the inverse transformation of the normalization parameters and re-standardized into the model input format to ensure the consistency between the input signal and the training distribution, so as to obtain a temperature sequence block that can be used for online inference. S3.4: Input the temperature sequence block into the process stage recognition engine that has completed offline training, perform forward propagation calculation through the encoder-decoder structure, and output the probability distribution vector of the three types of process stages corresponding to each time step, namely the category confidence distribution, which represents the uncertainty estimate of the stage of the currently activated process. The stage context enhancement of the temperature input signal is achieved by using the static variable embedding and time position encoding module in the encoder structure, forming an embedding representation matrix that combines physical quantity characteristics and temporal position information.

[0025] Furthermore, a multi-head attention mechanism is used to perform global feature weight calculation on the embedding representation matrix within the encoder, capturing the correlation of temperature change patterns across time steps, and obtaining a high-dimensional temporal feature tensor weighted by attention.

[0026] Furthermore, the high-dimensional feature tensor is recursively parsed using the gated recurrent unit (GRU) sequence modeling module in the decoder to generate a sequence of stage-related feature vectors for the corresponding time step, providing input for probability mapping calculation.

[0027] Furthermore, the relevant feature vector sequence of each stage is input into the Softmax normalized mapping function, and the probability distribution of each time step in the three process stages of heating, holding and cooling is calculated according to the following formula.

[0028] S3.5: Perform moving average filtering on the category confidence distribution of outputs from multiple consecutive inference cycles to suppress misjudgments caused by stage jumps due to sensor noise or brief disturbances, and generate a smooth and stable dominant stage confidence trajectory as a reliable operating condition criterion input for the next stage feature extraction sub-model call.

[0029] In this embodiment, in step S4, the current dominant process stage is determined based on the category confidence distribution, and the corresponding dedicated feature extraction sub-model group is called based on the stage category to calculate key criterion features for different stages: for the heating stage, the temperature rise slope and spectral reflectance inflection point are extracted; for the heat preservation stage, the spectral energy concentration and thermal radiation stability index are extracted; and for the cooling stage, the cooling rate and spectral attenuation consistency coefficient are extracted. Figure 3 As shown, the specific steps include the following: S4.1: Based on the confidence distribution of process stage categories output in S3, the maximum confidence selection strategy is adopted to determine the current dominant process stage, that is, the category with the highest confidence is used as the judgment result of the current stage, and the stage identifier is output to provide the working condition context for feature extraction for subsequent stage adaptation. Furthermore, the dimension containing the highest confidence value is identified through a vector comparison method, and the index value of that dimension is used as the stage category number.

[0030] Furthermore, a stage mapping table retrieval method is used to map category numbers to dominant process stage identifiers and generate explicit stage symbol data.

[0031] Furthermore, by using the working condition context binding process, the current stage identifier is directly associated with the subsequent special feature extraction sub-model selection process, thus establishing a synchronization relationship between data flow and control flow.

[0032] S4.2: Based on the dominant process stage identifier determined in S4.1, load the corresponding stage's feature extraction sub-model from the pre-configured dedicated feature extraction sub-model library: if it is the heating stage, call the heating stage feature extraction sub-model; if it is the heat preservation stage, call the heat preservation stage feature extraction sub-model; if it is the cooling stage, call the cooling stage feature extraction sub-model, to ensure that the feature extraction process matches the current process dynamic characteristics.

[0033] Furthermore, the feature sub-model is accurately invoked in memory through the sub-model index loading algorithm, and an executable object of the feature extraction algorithm module that matches the current dynamic working condition is obtained.

[0034] Furthermore, the I / O interface binding method enables seamless transmission and binding between the currently invoked feature extraction sub-model and the real-time data stream, and generates a stage-specific input vector that can be directly used for feature calculation.

[0035] Furthermore, through parameter initialization and state cache loading methods, the internal computation graph of the sub-model is initialized, feature calculation drift during cross-stage switching is avoided, and intermediate states that can be used for key criterion feature calculation are output.

[0036] S4.3: Apply the first-order difference algorithm to the standardized temperature curve sequence generated in S2 to calculate the temperature rise slope sequence, and combine it with the broadband spectral reflectance matrix to identify the inflection point of spectral reflectance by the second derivative detection method. Output the maximum slope value and the timestamp of the first significant inflection point as the key criterion features of the heating segment to form a quantitative index reflecting the initial response intensity of material surface activation.

[0037] Furthermore, a second derivative detection method is employed on the broadband spectral reflectance matrix to calculate the second derivative of reflectance as a function of wavelength. By searching for points on the second derivative curve where the sign changes and amplitudes exceed a set threshold, the location of the first significant inflection point is identified. Furthermore, the maximum slope calculation result is correlated and mapped with the timestamp of the first significant inflection point to achieve a combined output of key features of the heating phase.

[0038] Furthermore, through feature combination logic, the maximum temperature rise slope and inflection point timestamp obtained in the previous step are encapsulated into a quantitative index vector, which serves as the core criterion data reflecting the initial response intensity of material surface activation.

[0039] The rate of temperature change is calculated using the following formula: in, For temperature number Segment sample value, The sampling interval time. For the first Slope of the segment.

[0040] The second derivative is calculated using the following formula: in, Spectral reflectance The second derivative, For wavelength, By using wavelength stepping and combining the first-order temperature difference with the second-order spectral derivative to identify features, the quantitative determination of the material surface response during the heating stage is improved.

[0041] S4.4: Principal component analysis (PCA) is performed within the insulation section window based on the standardized spectral reflectance matrix to extract the energy concentration ratio of the first principal component. The variance of the thermal radiation signal is calculated by sliding window to obtain the thermal radiation stability index. The two are normalized and combined into the key criterion feature vector of the insulation section, which is used to characterize the stability of the adsorbed state on the zirconium wire surface and the uniformity of energy distribution.

[0042] Furthermore, the eigenvalues ​​and eigenvectors of the spectral covariance matrix are solved using the eigenvalue decomposition method, and the variance contribution value of the first principal component is obtained as the basis for calculating the energy concentration.

[0043] Furthermore, a sliding window variance calculation method is used to dynamically estimate the thermal radiation stability during the heat preservation stage, and a stability index sequence is generated.

[0044] The energy concentration ratio and thermal radiation stability index are mapped to a dimensionless unified scale by using the min-max normalization method, generating normalized feature vectors.

[0045] S4.5: Apply a linear regression model to fit the cooling rate of the standardized temperature curve of the cooling section, and perform Pearson correlation analysis on the spectral attenuation process to calculate the attenuation consistency coefficient between multiple bands. The regression slope and correlation coefficient are used together as the key criterion feature output of the cooling section to evaluate the synchronicity between the relaxation behavior and spectral response of the zirconium wire structure.

[0046] Furthermore, band separation processing is performed on the broadband spectral reflectance time series matrix within the corresponding window of the cooling section (to achieve segmented signal extraction of the multi-band spectral attenuation process). Further, Pearson correlation analysis is performed on each band attenuation curve and the reference band curve to calculate the multi-band attenuation consistency coefficient and generate a phased spectral synchronization quantification index.

[0047] In this embodiment, in step S5, the specific features of each stage are input into the multi-model output fusion module, and the features of each criterion are weighted and integrated in combination with the confidence weight of the current dominant process stage to generate a dynamic comprehensive judgment score. This score is a quantitative indicator reflecting the degree to which the activation process is close to the end point, and specifically includes: S5.1: Based on the category confidence distribution output from the previous steps and the extraction results of special features for each stage, construct the input tensor of the multi-model output fusion module. The input tensor consists of three parts: the feature vector of the heating stage, the feature vector of the heat preservation stage, and the feature vector of the cooling stage, and is supplemented with the stage confidence weight vector at the current moment to form a multi-dimensional fusion input structure with process context awareness, providing a data foundation for subsequent weighted integration. Furthermore, through the confidence weight generation algorithm, the confidence of the heating segment, heat preservation segment, and cooling segment at the current time step is used as weights to be added to the corresponding feature vectors to generate weighted stage input units. These units are then combined into a matrix to form a four-column input tensor, where the first three columns are stage feature vectors and the fourth column is the corresponding stage confidence weight.

[0048] Furthermore, by using matrix concatenation operations (to achieve a physical binding combination of multi-stage features and stage weights), the resulting input tensor has process context awareness capabilities, which can support stage priority judgment in the subsequent weighted integration process.

[0049] S5.2: Perform feature alignment processing on the multidimensional fusion input structure, use the standardized mapping function to normalize the scale of the feature vectors at each stage, eliminate the weight bias caused by the difference in the physical dimensions of the features, and generate a standardized feature matrix under a unified dimension as a prerequisite for the weighted fusion operation, so as to ensure the comparability of the criteria features at different stages in the fusion space. S5.3: Based on the confidence weight vector of the current dominant process stage, construct a dynamic weighted coefficient set, where the weighted coefficient corresponding to each stage is determined by its class confidence at the current time. Use a soft selection mechanism to weight and activate the feature components of each stage in the normalized feature matrix, suppress the feature responses of irrelevant stages, highlight the key criterion contribution of the dominant stage, and generate a stage-aware weighted feature vector. S5.4: The stage-aware weighted feature vector is input into the linear-nonlinear composite mapping unit. First, it is reduced in dimension and compressed by a learnable linear transformation matrix to extract high-contribution combined features. Then, a saturation mapping is performed by a Sigmoid nonlinear activation function to generate a preliminary comprehensive judgment score between 0 and 1. This score is an intermediate quantitative indicator representing the stage completion degree of the current activation process. S5.5: Perform time series smoothing filtering on the preliminary comprehensive judgment score, and use the moving exponential weighted average algorithm to eliminate instantaneous fluctuation interference to generate a stable dynamic comprehensive judgment score. This score is used as the final output to characterize the degree to which the zirconium wire high-temperature activation process is close to the end point, and to provide a continuous and reliable judgment basis for subsequent threshold comparison and end point confirmation.

[0050] In this embodiment, step S6 involves loading the corresponding dynamic judgment threshold range from the preset rule base based on the current dominant process stage, and using historical batch data to compensate for the drift of the upper and lower limits of the threshold range to generate an adaptive judgment threshold that fits the current operating conditions. Specifically, this includes the following steps: S6.1: Obtain the category identifier of the current dominant process stage, and retrieve the corresponding initial dynamic judgment threshold range from the preset rule base based on the stage category. The rule base stores the upper and lower limit sets of the benchmark thresholds respectively calibrated for the heating stage, the heat preservation stage and the cooling stage, so as to form the initial value of the stage-specific judgment boundary. S6.2: Obtain the comprehensive judgment score sequence of the most recent N historical batches under the same dominant process stage, calculate its statistical distribution characteristics, including the mean and standard deviation, generate a historical score distribution model, and use it as a reference benchmark to characterize process stability for subsequent threshold drift trend analysis. S6.3: Based on the deviation between the historical score distribution model and the real-time comprehensive judgment score of the current batch, calculate the threshold drift compensation coefficient. This coefficient is obtained by fitting through the sliding window regression algorithm, reflecting the systematic offset trend caused by equipment aging or environmental disturbances, and the output is a continuous adjustment factor. S6.4: Perform linear weighted operations on the upper and lower limits of the initial dynamic judgment threshold interval with the threshold drift compensation coefficient, perform boundary correction processing, and generate the intermediate threshold interval after drift compensation. This interval has included adaptive adjustments for time-varying operating conditions and is reflected as the updated high-confidence judgment boundary. S6.5: Apply upper and lower limit constraints to the intermediate threshold range, perform rationality verification based on the process safety boundary knowledge base, and truncate to the limit value if it exceeds the allowable range. Finally, generate an adaptive judgment threshold that adapts to the current working condition, which serves as the dynamic decision-making basis for the endpoint confirmation judgment in S7.

[0051] In this embodiment, step S7 determines whether the dynamic comprehensive judgment score continuously exceeds the adaptive judgment threshold and maintains it within a preset time window. If this condition is met, an activation endpoint confirmation signal is generated, triggering a process termination command. Specifically, this includes the following steps: S7.1: Obtain the dynamic comprehensive judgment score sequence generated by the multi-model output fusion module, and the adaptive judgment threshold generated by loading the rule base and after historical batch data drift compensation, as the input data for this step, providing a basic signal source for subsequent continuous discrimination.

[0052] Furthermore, the integrity of the collected judgment score sequence is verified through a data verification algorithm, and a high-confidence score data stream that has passed the verification is obtained.

[0053] Furthermore, a timestamp resampling algorithm is used to ensure the consistency of the judgment score sequence with the clock domain of the judgment link in the process stage, and to generate a judgment score signal vector under a unified time reference.

[0054] Furthermore, the adaptive judgment threshold after drift compensation in step S6 is retrieved from the rule base through a memory mapping loading method, enabling rapid reading of threshold data and generating a threshold scalar corresponding to the current operating condition.

[0055] Furthermore, a unified data container encapsulation strategy is adopted to combine the time-aligned judgment score signal vector with the adaptive judgment threshold into a judgment input pair, which is used for downstream sliding time window judgment algorithm calls.

[0056] S7.2: Based on the sliding time window mechanism, perform temporal continuity analysis on the dynamic comprehensive judgment score, calculate whether the judgment scores of all sampling points in the current time window are higher than the adaptive judgment threshold, and generate a Boolean discrimination flag sequence to quantify the continuous satisfaction of the judgment condition; S7.3: Based on the preset time window length T and system sampling frequency f, determine the minimum number of consecutive qualifying frames, and perform a cumulative counting operation on the Boolean discrimination flag sequence to determine whether its longest consecutive True segment is not less than N, so as to verify the continuous validity of the judgment condition; S7.4: If the longest consecutive True segment reaches or exceeds N, it is determined that the current activation process has stably entered the endpoint interval, and a high-confidence activation endpoint trigger request is generated; otherwise, the current running state is maintained, and the dynamic comprehensive judgment score of the next moment is received for cyclic evaluation. S7.5: After confirming that the continuity judgment condition is met, output an activation endpoint confirmation signal to the process control system, trigger the relay action or PLC logic interruption, issue a process termination command, and complete the closed-loop control response from intelligent judgment to physical execution.

[0057] In this embodiment, in step S8, while generating the activation endpoint confirmation signal, the main contribution characteristics and their weight change trends at each stage of the activation process are recorded to form an interpretable feedback log, which is used to support subsequent process optimization and model iteration. Specifically, the steps include the following: S8.1: Based on the weighted integration path of the dedicated criteria features of each stage in the multi-model output fusion module, the key feature input items corresponding to the heating stage, heat preservation stage and cooling stage are obtained, including temperature rise slope, spectral reflectance inflection point, spectral energy concentration, thermal radiation stability index, cooling rate and spectral attenuation consistency coefficient, as the original feature dependent variables for interpretability analysis. S8.2: The real-time contribution weight of the key features in each stage in the dynamic comprehensive judgment score calculation process is sampled over time. The mean weight and variance fluctuation value of each feature in each process stage are calculated using the sliding window statistical method to generate a quantitative index of the stage importance of each feature, so as to reflect the dynamic evolution law of the dominant criteria in different stages. S8.3: Based on the time series data of category confidence distribution output by the process stage identification engine, the boundary time of each stage is precisely delimited, and the change trend of each feature weight is spatiotemporally aligned with its process stage to generate a feature contribution trajectory matrix with stage labels, which serves as the core intermediate product for constructing interpretable feedback logs. S8.4: The feature contribution trajectory matrix and the process parameter context of the current batch (including adaptive judgment threshold, historical drift compensation amount, and trigger time window length) are structurally encapsulated, and an interpretable feedback log file is generated using a predefined log template, which includes the main control features of each stage, weight evolution trend map and abnormal fluctuation alarm flags. S8.5: Upload the interpretability feedback log to the process knowledge base system, and perform inter-batch comparative analysis based on the log content to identify high-contribution feature patterns or abnormal weight drift phenomena that occur repeatedly across batches, and generate model optimization suggestion instructions to guide the parameter iteration update of the feature extraction sub-model group and the fusion rule base.

[0058] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0059] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0060] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent real-time determination of the endpoint of a high-temperature activation process for a zirconium filament in an ultra-high pressure mercury lamp, characterized in that, Specifically, it includes: S1: Real-time acquisition of temperature time series data and broadband spectral data stream during the high-temperature activation process of zirconium filament in ultra-high pressure mercury lamp, and recording of time stamp information of process stages to obtain the raw dataset; S2: Denoise and normalize the original dataset to generate a standardized temperature curve sequence and spectral reflectance matrix; S3: The process stage identification engine, built using a pre-trained time series classification network, performs online inference on the standardized temperature curve sequence and outputs the category confidence distribution of the process stage at the current moment. S4: Determine the current dominant process stage based on the category confidence distribution, and calculate the key criterion features by calling the corresponding dedicated feature extraction sub-model group based on the stage category; S5: Input the special features of each stage into the multi-model output fusion module, and combine the confidence weight of the current dominant process stage to weight and integrate the features of each criterion to generate a dynamic comprehensive judgment score. S6: Based on the current dominant process stage, load the corresponding dynamic judgment threshold range from the preset rule base, and perform drift compensation on the upper and lower limits of the threshold range to generate an adaptive judgment threshold. S7: Determine whether the dynamic comprehensive judgment score continuously exceeds the adaptive judgment threshold and maintains it within a preset time window. If the condition is met, generate an activation endpoint confirmation signal and trigger a process termination command. S8: While generating the activation endpoint confirmation signal, record the main contribution characteristics and their weight change trends of each stage in this activation process.

2. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 1, characterized in that, In step S1, the temperature simulation signal output from the thermocouple sensor arranged inside the high-temperature activation furnace is obtained, and the ADC module is used to perform analog-to-digital conversion to generate a discrete temperature time-series sampling sequence.

3. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 2, characterized in that, In step S1, by acquiring the broadband reflectance spectrum simulation signal collected by the fiber optic spectrometer, wavelength calibration and photocurrent digitization are performed using the embedded spectral acquisition unit to generate a spectral reflectance vector sequence with wavelength index alignment.

4. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 3, characterized in that, In step S1: Based on the process control timing signal output by the PLC control system, the preset time window boundary information of the current operating stage is extracted to generate a process stage time stamp sequence synchronized with the physical sampling clock. The temperature time-series sampling sequence, spectral reflectance vector sequence, and process stage time stamp sequence are time-stamp aligned. An interpolation compensation algorithm is used to fill the data asynchrony gap caused by the difference in sampling frequency, and a multimodal sensing data frame set is generated. The multimodal sensor data frame set is encapsulated and stored according to batch number and furnace ID to generate an original dataset with operating condition context labels.

5. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 4, characterized in that, Step S2 specifically includes: The temperature time series data is processed by low-frequency noise suppression based on the sliding window mean filtering algorithm to obtain a smoothed temperature signal sequence. The original broadband spectral data stream is decomposed and reconstructed using a wavelet thresholding denoising algorithm. The high-frequency detail coefficients of the spectrum are extracted using the Daubechies wavelet basis function and then soft thresholding is performed. The smoothed temperature signal sequence is input into the adaptive anomaly detection module to identify anomalous sampling points that significantly deviate from the normal trend based on the local outlier factor algorithm, and then repairs them using linear interpolation to generate a clean temperature curve sequence. The clean temperature curve sequence and the denoised spectral reflectance time series matrix are subjected to scale unification processing to generate a standardized temperature curve sequence and a standardized spectral reflectance matrix; and then aligned and fused according to timestamps to generate a structured multimodal standardized dataset.

6. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 5, characterized in that, In step S3, a lightweight TFT network architecture is used as the basic model. A temporal feature extraction structure including static variable embedding, known future input encoding and multi-head attention mechanism is designed. The temperature sequence in the training dataset is trained end-to-end, the cross-entropy loss function is optimized, and a pre-trained process stage recognition engine is generated.

7. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 6, characterized in that, In step S3, the temperature sequence block is input into the process stage recognition engine that has completed offline training. The forward propagation calculation is performed through the encoder-decoder structure, and the probability distribution vectors of the three process stages corresponding to each time step, namely the category confidence distribution, are output.

8. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 7, characterized in that, In step S4, the feature extraction sub-model for the corresponding stage is loaded from the pre-configured dedicated feature extraction sub-model library: if it is the heating stage, the feature extraction sub-model for the heating stage is called; if it is the heat preservation stage, the feature extraction sub-model for the heat preservation stage is called. If it is a cooling section, then the cooling section feature extraction sub-model is invoked.

9. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 8, characterized in that, In step S5, the inputs to the multi-model output fusion module include: the feature vector of the heating section, the feature vector of the heat preservation section, and the feature vector of the cooling section; Based on the confidence weight vector of the current dominant process stage, a dynamic weighted coefficient set is constructed, where the weighted coefficient corresponding to each stage is determined by its class confidence at the current time. A soft selection mechanism is used to weight and activate the feature components of each stage in the normalized feature matrix to generate a stage-aware weighted feature vector.

10. The intelligent real-time determination method for the endpoint of the high-temperature activation process of zirconium filament in an ultra-high pressure mercury lamp according to claim 9, characterized in that, In step S5, the stage-aware weighted feature vector is input into the linear-nonlinear composite mapping unit, and the dimension reduction and compression are performed by the learnable linear transformation matrix to extract the high contribution combination features. Then, the saturation mapping is performed by the Sigmoid nonlinear activation function to generate a preliminary comprehensive judgment score between 0 and 1. The preliminary comprehensive judgment score is processed by time series smoothing filtering, and the moving exponential weighted average algorithm is used to eliminate instantaneous fluctuation interference, generating a stable dynamic comprehensive judgment score, which is used to characterize the degree to which the zirconium wire high-temperature activation process is close to the end.

Images