A method for controlling the atmosphere in a furnace for precious metal refining

Abstract

The present application relates to the technical fields of industrial process control and data processing, and discloses a kind of noble metal refining process furnace atmosphere control method, including obtaining furnace gas data set in noble metal refining process, according to tail gas component switching relationship and air supply corresponding relationship division disturbance fragment, again based on oxygen content and reducing gas content in each disturbance fragment Construction hysteresis loop and extract hysteresis phase value, then according to hysteresis phase value and furnace gas data set division boundary sample set, and according to the adjacency relationship of boundary sample set in adjacent disturbance fragment, generate boundary constraint set, finally to initial neural network model is sparsely perceived and is reformed and inputs furnace gas data set, boundary sample set and boundary constraint set generate atmosphere instruction set, so as to keep atmosphere window tracking and shorten fluctuation convergence time, realize impurity oxidation boundary stability.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and data processing technology, and more specifically, to a method for controlling the atmosphere inside a furnace during a precious metal refining process. Background Technology

[0002] In the refining process of precious metals, it is usually necessary to adjust the atmosphere in the furnace in stages to keep the oxidation reaction of impurities within the target boundary range. The relevant control is generally deployed in the industrial furnace and its supporting tail gas recirculation and emission unit, and is limited by the on-site response delay, detection link stability and continuous operation constraints. Existing technologies are mostly based on process quantity monitoring, threshold determination, correlation measurement, time smoothing and feedback correction to adjust the atmosphere state. Such methods are usually applicable under the premise that the tail gas composition changes slowly, the reaction stage in the furnace is clear, and the supply and exhaust response relationship is relatively stable.

[0003] In precious metal refining, fluctuations in exhaust gas composition and changes in furnace reaction intensity cause continuous disturbances in the atmosphere. This makes it difficult for control methods based on process quantity monitoring, threshold determination, correlation measurement, time smoothing, and feedback correction to maintain the target atmosphere range in a timely manner. Consequently, the recovery process after atmosphere deviation correction is prolonged, and the stability of the impurity oxidation boundary is affected. Therefore, the technical problem to be solved is how to achieve the stability of the impurity oxidation boundary while maintaining atmosphere window tracking and shortening the fluctuation convergence time.

[0004] In view of this, the present invention proposes a method for controlling the furnace atmosphere in the precious metal refining process to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for controlling the atmosphere inside the furnace during the refining process of precious metals.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for controlling the furnace atmosphere in a precious metal refining process is provided, including: Obtain the furnace gas dataset during the precious metal refining process, and divide the disturbance segment according to the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset. Hysteresis loops are constructed based on the oxygen and reducing gas contents in each perturbation segment, and hysteresis phase values ​​are extracted based on the component crossing positions in the hysteresis loops. The boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, and the boundary constraint set is generated based on the adjacency relationship of the boundary sample set in adjacent perturbation segments. Based on the hysteresis phase value, the boundary sample extraction path in the initial neural network model is modified by sparse sensing to obtain a neural network model that integrates sparse sensing. The furnace gas dataset, boundary sample set, and boundary constraint set are then input into the neural network model that integrates sparse sensing to generate an atmosphere instruction set.

[0007] In some embodiments, disturbance segments are divided according to the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset, including: Extract tail gas component data, gas supply parameter data, and exhaust parameter data from the furnace gas dataset, and generate the dominant component sequence and gas supply and exhaust displacement sequence according to the sampling time. A switching anchor sequence is generated based on the transposition order of adjacent sampling times in the dominant component sequence, and a set of coupling segments is generated based on the reverse transposition order around the same sampling time in the supply and exhaust displacement sequence. A fragment boundary set is generated based on the overlapping relationship between the switching anchor point sequence and the coupling segment set, and a candidate fragment set is obtained by segmenting the furnace gas dataset based on the fragment boundary set. The candidate segments are screened based on continuous sections of exhaust gas components and alternating sections of exhaust gas supply and exhaust gas supply to identify disturbance segments.

[0008] In some embodiments, generating a set of coupled segments based on the reverse displacement order around the same sampling time in the supply and exhaust displacement sequence includes: Extract the gas supply displacement segment and exhaust displacement segment corresponding to each sampling time in the gas supply and exhaust displacement sequence, and generate candidate displacement pairs based on the distribution of the gas supply displacement segment and exhaust displacement segment around the same sampling time. Based on the opposite relationship between the gas supply displacement direction and the exhaust displacement direction in each candidate displacement pair, reverse displacement pairs are selected to obtain the set of reverse displacement pairs. Coupled segments are generated based on the continuous connection relationship of the reverse displacement pairs in the supply and exhaust displacement sequence, and coupled segment sets are generated based on the beginning and end connection relationship of each coupled segment.

[0009] In some embodiments, a hysteresis loop is constructed based on the oxygen content and reducing gas content in each perturbation segment, including: Oxygen and reducing gas content data are extracted from each perturbation segment, and a series of oxygen reduction points are generated according to the sampling time. A sequence of return nodes is generated based on the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the return node sequence. A closed trajectory is generated based on the turnaround node sequence and the lap envelope sequence, and a loop skeleton is generated based on the connection relationship between the beginning and end of the closed trajectory. Hysteresis loops are extracted based on the loop skeleton and closed trajectory.

[0010] In some embodiments, a turnaround node sequence is generated based on the round-trip connection order in the reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the turnaround node sequence, including: Generate forward connection sets and backward connection sets based on the connection directions of continuous sampling times in the oxygen reduction point sequence; Candidate points for turning back are generated based on the intersection of the forward connection set and the backward connection set, and a sequence of turning back nodes is generated based on the temporal arrangement of the candidate points for turning back. Extract the trajectory extension segments corresponding to both sides of the turnaround node sequence, and generate envelope segments based on the coverage relationship of the trajectory extension segments; Generate a cyclic envelope sequence based on the connection between the beginning and end of the envelope segments.

[0011] In some embodiments, the boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, including: Based on the furnace gas dataset, tail gas component data, gas supply parameter data, and exhaust parameter data are obtained, and sample time series clusters are generated according to the perturbation segments. The hysteresis phase values ​​are then mapped to the corresponding sample time series clusters. Phase adjacency chains are generated based on the sequential connection relationship of hysteresis phase values ​​in each sample time series cluster, and boundary candidate clusters are generated based on the transposition position of exhaust gas component data in the phase adjacency chains. Boundary sample fragments are selected based on the inverse adjoint relationship between gas supply parameter data and exhaust parameter data in the boundary candidate cluster, and boundary sample clusters are generated. The boundary sample set is determined based on the sample attribution relationship of the boundary sample cluster in the furnace gas dataset.

[0012] In some embodiments, the boundary sample extraction path in the initial neural network model is sparsely sensed and modified according to the hysteresis phase value to obtain a neural network model fused with sparse sensing, including: Extract hysteresis phase values ​​and boundary sample segments from the boundary sample set, and generate a phase layering sequence based on the arrangement of hysteresis phase values ​​in the boundary sample segments; A sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set; Rewrite the boundary sample extraction path in the initial neural network model based on the sparse gating matrix to generate a path routing graph; Based on the path routing graph, the boundary sample extraction paths and non-boundary sample extraction paths in the initial neural network model are combined to obtain a neural network model that integrates sparse perception.

[0013] In some embodiments, a sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set, including: Phase access segments are generated based on the interlayer connection relationship between adjacent hysteresis phase values ​​in the phase layering sequence. Based on the adjacency relationship of the samples in the boundary sample set, the adjacent sample pairs corresponding to each phase access segment are extracted to generate sample connection clusters; Path opening and closing units are generated based on the co-occurrence relationship of boundary sample segments in each sample connection cluster; A sparse gating matrix is ​​generated based on the arrangement of path opening and closing units in each phase access segment.

[0014] In some embodiments, path opening and closing units are generated based on the co-occurrence relationship of boundary sample fragments in each sample connection cluster, including: Extract boundary sample fragments from each sample connection cluster, and generate co-occurrence fragment groups based on the common occurrence positions of the boundary sample fragments in the same phase access segment; Open path segments are generated based on the sequential relationship of boundary sample segments in the co-occurring segment group, and closed path segments are generated based on the discontinuity relationship of boundary sample segments in the co-occurring segment group. Path opening and closing units are generated based on the corresponding combination relationship between open and closed path segments in each sample connection cluster.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires a furnace gas dataset from the precious metal refining process, divides it into perturbation segments based on the tail gas component switching relationship and the supply and exhaust correspondence, constructs hysteresis loops based on the oxygen and reducing gas contents in each perturbation segment, and extracts hysteresis phase values ​​based on the component crossing positions in the hysteresis loops. Subsequently, it divides a boundary sample set based on the hysteresis phase values ​​and the furnace gas dataset, and generates a boundary constraint set based on the adjacency relationship of the boundary sample set in adjacent perturbation segments. Finally, it modifies the boundary sample extraction path in the initial neural network model using sparse sensing based on the hysteresis phase values ​​to obtain a neural network model fused with sparse sensing. The furnace gas dataset, boundary sample set, and boundary constraint set are then input into the neural network model fused with sparse sensing to generate an atmosphere instruction set, thereby maintaining atmosphere window tracking and shortening the fluctuation convergence time, achieving impurity oxidation boundary stability. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a method for controlling the atmosphere inside a furnace during a precious metal refining process according to the present invention. Figure 2 This is a schematic diagram of the furnace atmosphere control system for a precious metal refining process according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. In the following detailed description, many specific details are set forth to provide a thorough understanding of the exemplary embodiments described. However, it will be apparent to those skilled in the art that the described embodiments may be practiced without some or all of these specific details. In other exemplary embodiments, well-known structures have not been described in detail to avoid unnecessarily obscuring the concepts of this disclosure. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention. Furthermore, the various aspects described in the embodiments may be combined arbitrarily without conflict.

[0018] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0019] Example 1 Figure 1 This disclosure illustrates a method for controlling the furnace atmosphere in a precious metal refining process, provided by at least one embodiment, including: S10: Obtain the furnace gas dataset during the precious metal refining process, and divide the disturbance segment according to the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset. The disturbance segments are divided based on the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset, including: Extract tail gas component data, gas supply parameter data, and exhaust parameter data from the furnace gas dataset, and generate the dominant component sequence and gas supply and exhaust displacement sequence according to the sampling time. A switching anchor sequence is generated based on the transposition order of adjacent sampling times in the dominant component sequence, and a set of coupling segments is generated based on the reverse transposition order around the same sampling time in the supply and exhaust displacement sequence. A fragment boundary set is generated based on the overlapping relationship between the switching anchor point sequence and the coupling segment set, and a candidate fragment set is obtained by segmenting the furnace gas dataset based on the fragment boundary set. The candidate segments are screened based on continuous sections of exhaust gas components and alternating sections of exhaust gas supply and exhaust gas supply to identify disturbance segments.

[0020] It should be noted that the furnace gas dataset in this embodiment is collected by the detection devices on the furnace body and its supporting tail gas reflux and emission units. Among them, the tail gas component data can be obtained by the gas component detection device arranged on the tail gas sampling pipeline, the gas supply parameter data can be obtained by the flow detection device, pressure detection device or valve position detection device on the gas supply pipeline, and the exhaust parameter data can be obtained by the flow detection device, pressure detection device on the exhaust pipeline or the operation state detection device of the exhaust and extraction execution unit. Correspondingly, the atmosphere instruction set is used to output to the gas supply valve, exhaust valve, tail gas reflux valve, air blowing execution unit or exhaust and extraction execution unit supporting the furnace body to adjust the gas supply flow rate, exhaust intensity, reflux ratio or execution rhythm, so that the generated control result can be applied to the specific industrial execution end.

[0021] In this embodiment, a continuous data interval corresponding to the same atmosphere disturbance is identified from the continuously collected furnace gas dataset, so as to construct a hysteresis loop and extract the hysteresis phase value around the same disturbance background subsequently. Among them, the tail gas component switching relationship reflects the change position of the dominant reaction environment, and the gas supply and exhaust corresponding relationship reflects the response interval generated by the execution end around this change position. Therefore, instead of segmenting the furnace gas dataset at a fixed time length, the start and end boundaries of the disturbance segment are determined by combining the tail gas component data with the gas supply parameter data and exhaust parameter data, so as to avoid mixing data belonging to different reaction stages into the same disturbance segment.

[0022] It can be understood that the dominant component sequence is used to represent the time sequence arrangement of the tail gas components with the highest occupancy at each sampling moment, the gas supply and exhaust displacement sequence is used to represent the change positions of the gas supply parameter data and exhaust parameter data at adjacent sampling moments, the switching anchor point sequence is used to mark the sampling positions where the dominant component sequence changes, and the coupling section set is used to mark the continuous intervals where the gas supply parameter data and exhaust parameter data form reverse displacement responses around the same sampling moment. When the switching anchor point sequence and the coupling section set overlap on the time axis, this overlapping interval can be used as the basis for generating the segment boundary set. Therefore, in this embodiment, the candidate segment set is determined by the overlapping relationship between the switching anchor point sequence and the coupling section set, and then the intervals where both the continuous section of the tail gas components and the alternating section of the gas supply and exhaust are established are screened out from the candidate segment set as the disturbance segments.

[0023] For example, in a certain batch of furnaces, within six consecutive sampling times t1 to t6, the exhaust gas composition data is SO2, SO2, O2, O2, CO2, CO2, the gas supply parameter data is 12, 15, 15, 11, 11, 14, and the exhaust parameter data is 8, 8, 6, 6, 9, 9. Then the dominant component sequence corresponds to SO2, SO2, O2, O2, CO2, CO2. The dominant component transposes between t2 and t3, and between t4 and t5, resulting in the switching anchor point sequence {(t2, t3), (t4, t5)}. At the same time, the gas supply parameter data and the exhaust parameter data form reverse transpositions around t3 (15→11 and 6→9), and around t5 (11→14 and 9 remain after transposition), resulting in a coupling segment set {(t2, t4), (t4, t6)}. This set is used to limit the dominant component switching from an isolated event to a boundary of the disturbance segment together with the gas supply and exhaust responses.

[0024] A set of coupled segments is generated based on the reverse displacement order around the same sampling time in the supply and exhaust displacement sequence, including: Extract the gas supply displacement segment and exhaust displacement segment corresponding to each sampling time in the gas supply and exhaust displacement sequence, and generate candidate displacement pairs based on the distribution of the gas supply displacement segment and exhaust displacement segment around the same sampling time. Based on the opposite relationship between the gas supply displacement direction and the exhaust displacement direction in each candidate displacement pair, reverse displacement pairs are selected to obtain the set of reverse displacement pairs. Coupled segments are generated based on the continuous connection relationship of the reverse displacement pairs in the supply and exhaust displacement sequence, and coupled segment sets are generated based on the beginning and end connection relationship of each coupled segment.

[0025] Understandably, the reverse displacement sequence around the same sampling time is mainly used to identify whether the gas supply parameter data and exhaust parameter data jointly respond to the same exhaust gas component switching. Here, the gas supply displacement segment refers to the continuous interval in which the gas supply parameter data maintains the same direction of change between adjacent sampling times, the exhaust displacement segment refers to the continuous interval in which the exhaust parameter data maintains the same direction of change between adjacent sampling times, the candidate displacement pair refers to the corresponding segment extracted from the gas supply displacement segment and the exhaust displacement segment around the same sampling time, the reverse displacement pair refers to the candidate displacement pair in which the gas supply displacement direction is opposite to the exhaust displacement direction, the coupled segment refers to the time interval covered by the continuous reverse displacement pair, and the coupled segment set is the interval set formed by the combination of multiple coupled segments connected end to end.

[0026] It should be noted that the process of first generating candidate displacement pairs, then screening out reverse displacement pairs, and finally generating a set of coupled segments is because the response of gas supply parameter data and exhaust parameter data to tail gas component switching in precious metal refining sites usually has a sequential order. It is difficult to determine whether the two belong to the same disturbance process based solely on single-point data at the same sampling time. However, by examining the distribution before and after the same sampling time, we can unify the situations where gas supply parameter data changes first and exhaust parameter data changes later, or vice versa, into the same identification framework. This allows us to extract the execution actions that truly correspond to the same atmospheric disturbance from the continuous data.

[0027] For example, continuing with the previous example from t1 to t6, if the gas supply parameter data is 12, 15, 15, 11, 11, 14, and the exhaust parameter data is 8, 8, 6, 6, 9, 9, then around t3, the gas supply displacement segment {t1, t2, t3} and the exhaust displacement segment {t2, t3, t4} can be extracted to form a candidate displacement pair. Around t5, the gas supply displacement segment {t3, t4, t5} and the exhaust displacement segment {t4, t5, t6} can be extracted to form another candidate displacement pair. The gas supply parameters in the first candidate displacement pair... The data shows an initial increase followed by stabilization, while the exhaust parameter data shows an initial decrease followed by stabilization. In the next candidate displacement pair, the gas supply parameter data shows an initial decrease followed by stabilization, while the exhaust parameter data shows an initial increase followed by stabilization. Therefore, both candidate displacement pairs can be screened as reverse displacement pairs, and further coupling segments {t2,t4} and {t4,t6} are generated. These two coupling segments are then connected end to end to form a coupling segment set {(t2,t4),(t4,t6)}, which is used to define the response interval that works in conjunction with the switching anchor point sequence.

[0028] Furthermore, compared to the existing technology that determines the response interval based solely on changes in gas supply parameters or exhaust parameters, this embodiment generates reverse displacement pairs around the same sampling time and generates a set of coupled segments based on the continuous connection relationship of the reverse displacement pairs. This allows the common response of gas supply parameters and exhaust parameters to the same tail gas component switching to be completely preserved. Therefore, it helps to reduce segment offset caused by mismatch of execution end actions and provides a more stable time constraint for the subsequent generation of segment boundary sets based on the switching anchor sequence and the set of coupled segments.

[0029] S20: Construct hysteresis loops based on the oxygen and reducing gas contents in each perturbation segment, and extract hysteresis phase values ​​based on the component crossing positions in the hysteresis loops; Hysteresis loops are constructed based on the oxygen and reducing gas content in each perturbation segment, including: Oxygen and reducing gas content data are extracted from each perturbation segment, and a series of oxygen reduction points are generated according to the sampling time. A sequence of return nodes is generated based on the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the return node sequence. A closed trajectory is generated based on the turnaround node sequence and the lap envelope sequence, and a loop skeleton is generated based on the connection relationship between the beginning and end of the closed trajectory. Hysteresis loops are extracted based on the loop skeleton and closed trajectory.

[0030] In this embodiment, a hysteresis loop is constructed around the oxygen content data and reducing gas content data. This is mainly to transform the round-trip change process of the atmosphere state within the same perturbation segment from a time-axis expression to a trajectory expression. Here, the oxygen-reduction point sequence refers to the point sequence formed by arranging the oxygen content data and reducing gas content data in pairs according to the sampling time. The turnaround node sequence is used to mark the position where the round-trip change turns from forward to backward. The cyclic envelope sequence is used to limit the outer coverage range of the trajectory on both sides of the turnaround node. The loop skeleton is used to represent the main connecting framework of the closed trajectory. The hysteresis loop is a closed trajectory jointly determined by the loop skeleton and the closed trajectory, which is used to carry out the extraction of the subsequent component crossing position.

[0031] It should be noted that, to ensure the stability of hysteresis loop construction and hysteresis phase value extraction, this embodiment performs validity verification and anti-interference processing on oxygen content data and reducing gas content data before entering the oxygen-reduction point series construction. Specifically, this may include: removing abnormal sampling values ​​that exceed the preset range, screening out sudden interference points according to the continuous change relationship between adjacent sampling times, and using adjacent time-based completion or local smoothing to unify short-term missing points; for baseline drift caused by detection link contamination after long-term operation, the sampling results can be baseline corrected based on preset calibration gas samples or reference detection values ​​under stable operating conditions. Through the above processing, the oxygen content data and reducing gas content data entering the hysteresis loop construction can maintain temporal continuity and quantitative comparability, thereby reducing the impact of on-site dust, metal vapor, and detection link fluctuations on the stability of hysteresis phase value extraction.

[0032] It is understandable that the process of first generating the oxygen reduction point sequence, then the turnaround node sequence and the cyclic envelope sequence, and finally forming the closed trajectory and loop skeleton is because in the same disturbance segment of precious metal refining, the response of oxygen content data and reducing gas content data to the switching of exhaust gas components usually involves a round-trip process of first rising and then falling or first falling and then rising. If the values ​​are directly compared according to time sequence, it is difficult to distinguish between entering and leaving a certain atmosphere. However, by identifying the turnaround nodes through the round-trip connection sequence and using the outward expansion relationship on both sides of the turnaround nodes to form the cyclic envelope sequence, the forward and return paths within the same disturbance segment can be organized into the same closed trajectory, thus providing a positional basis for the hysteresis phase value.

[0033] For example, continuing with the previous example from t1 to t6, within the perturbation segment defined by {(t2,t4),(t4,t6)}, the oxygen content data, reducing gas content data, and oxygen reduction point sequence corresponding to each sampling time can be shown in Table 1: Table 1: Examples of Oxygen Reduction Points

[0034] It is understandable that when the oxygen reduction points in Table 1 are connected sequentially according to the sampling time, they can form a forward trajectory (2.3,5.6) → (3.1,4.8) → (3.8,4.1) and a return trajectory (3.8,4.1) → (3.0,4.7) → (2.6,5.2). Among them, (3.8,4.1) corresponds to the turning point of the round trip, which can form the turning point node sequence {(3.8,4.1)}. After extracting the extended segments of the front and back trajectories around this turning point, a turning envelope sequence can be obtained. This turning envelope sequence is used to limit the outer coverage of the closed trajectory, so that the closed trajectory can represent the round trip path of the atmosphere state within the same disturbance segment.

[0035] A turnaround node sequence is generated based on the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the turnaround node sequence, including: Generate forward connection sets and backward connection sets based on the connection directions of continuous sampling times in the oxygen reduction point sequence; Candidate points for turning back are generated based on the intersection of the forward connection set and the backward connection set, and a sequence of turning back nodes is generated based on the temporal arrangement of the candidate points for turning back. Extract the trajectory extension segments corresponding to both sides of the turnaround node sequence, and generate envelope segments based on the coverage relationship of the trajectory extension segments; Generate a cyclic envelope sequence based on the connection between the beginning and end of the envelope segments.

[0036] Furthermore, a reversal node sequence is generated around the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the reversal node sequence. This is mainly to separate the forward trajectory and the cyclic trajectory in the same disturbance segment before the hysteresis loop is formed. The forward connection set is used to represent the connection result of the oxygen reduction point sequence along the direction of entering the current atmosphere state, the cyclic connection set is used to represent the connection result of the oxygen reduction point sequence along the direction of leaving the current atmosphere state, the reversal candidate point is used to mark the intersection position of the forward connection set and the cyclic connection set, the reversal node sequence is used to mark the temporal position of the round-trip switching, the trajectory extension segment is used to represent the local trajectory extension range on both sides of the reversal node, and the cyclic envelope sequence is used to limit the outer coverage boundary of the subsequent closed trajectory.

[0037] Understandably, the process of first generating the forward and backward connection sets, then generating the turnaround node sequence, and finally generating the cyclic envelope sequence is because although the oxygen and reducing gas content data exhibit round-trip changes within the same perturbation segment, not all transitions correspond to actual atmospheric turnaround locations. Only the area at the intersection of the end of the forward connection set and the beginning of the backward connection set is suitable as a candidate turnaround point. Generating the turnaround node sequence through the temporal arrangement of these candidate turnaround points can distinguish local fluctuations from actual round-trip turns. Subsequently, the trajectory extension segments are extracted around both sides of the turnaround node sequence, ensuring that the cyclic envelope sequence retains the trajectory coverage near the turnaround point, rather than having the envelope boundary determined solely by a single turnaround point.

[0038] For example, the forward connection set, backward connection set, turnaround candidate points, turnaround node sequence, and trajectory extension segment formed after connecting according to the sampling time can be shown in Table 2: Table 2: Examples of Turnaround Node Sequences and Rotation Envelope Sequences

[0039] It is understandable that the forward connection set and the backward connection set in Table 2 intersect at (3.8, 4.1). Therefore, (3.8, 4.1) is taken as a candidate point for turning back and further determined as a sequence of turning back nodes. Then, the first trajectory extension segment and the second trajectory extension segment are extracted around both sides of the sequence of turning back nodes to form an envelope segment. This envelope segment is used to limit the trajectory coverage on both sides of the turning back position and to provide a basis for generating a turning envelope sequence based on the connection relationship between the beginning and end of the envelope segment.

[0040] It should be noted that, compared with the existing technology that directly determines the round-trip boundary based on the local extreme value position or a single turning point, this embodiment generates a gyratory envelope sequence by jointly using the forward connection set, the return connection set, and the turnaround node sequence. This makes the determination of the turnaround position no longer dependent on a single sampling point, but on the connection and coverage relationship of the trajectory before and after the turnaround. Therefore, it can reduce the misjudgment of local turnaround caused by the fluctuation of exhaust gas composition and the change of reaction intensity in the furnace, and provide a more stable trajectory boundary for the subsequent generation of closed trajectories based on the turnaround node sequence and the gyratory envelope sequence to extract the hysteresis loop.

[0041] S30: Divide the boundary sample set according to the hysteresis phase value and the furnace gas dataset, and generate the boundary constraint set according to the adjacency relationship of the boundary sample set in adjacent disturbance segments; The boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, including: Based on the furnace gas dataset, tail gas component data, gas supply parameter data, and exhaust parameter data are obtained, and sample time series clusters are generated according to the perturbation segments. The hysteresis phase values ​​are then mapped to the corresponding sample time series clusters. Phase adjacency chains are generated based on the sequential connection relationship of hysteresis phase values ​​in each sample time series cluster, and boundary candidate clusters are generated based on the transposition position of exhaust gas component data in the phase adjacency chains. Boundary sample fragments are selected based on the inverse adjoint relationship between gas supply parameter data and exhaust parameter data in the boundary candidate cluster, and boundary sample clusters are generated. The boundary sample set is determined based on the sample attribution relationship of the boundary sample cluster in the furnace gas dataset.

[0042] In this embodiment, a boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, and a boundary constraint set is generated based on the adjacency relationship of the boundary sample set in adjacent perturbation segments. This is mainly to further transform the obtained hysteresis phase values ​​into boundary data that can be identified and constrained by the model. Here, the sample time series cluster refers to the time series sample set formed after aggregating the exhaust gas component data, gas supply parameter data, and exhaust parameter data according to the perturbation segments. The phase adjacency chain refers to the phase connection sequence formed by the hysteresis phase values ​​in the sample time series cluster according to the connection relationship between the preceding and following phases. The boundary candidate cluster refers to the candidate sample set extracted from the corresponding position of the exhaust gas component data in the phase adjacency chain. The boundary sample cluster refers to the sample set in the boundary candidate cluster that further satisfies the reverse association relationship between the gas supply parameter data and the exhaust parameter data. The boundary sample set is the boundary sample set determined by the boundary sample cluster according to the sample belonging relationship.

[0043] Understandably, the focus here is not solely on dividing based on hysteresis phase values ​​or solely on the furnace gas dataset. Instead, it maps hysteresis phase values ​​to sample time-series clusters organized by perturbation segments, and then uses the phase adjacency chain and the transposition position of exhaust gas component data to jointly locate the candidate boundary clusters. This is because hysteresis phase values ​​can reflect the relative position of the atmosphere state within the same perturbation segment in the hysteresis loop, but cannot directly distinguish which samples are truly near the impurity oxidation boundary. While the transposition position of exhaust gas component data can reflect the switching position of the dominant reaction environment, if it is separated from the hysteresis phase values, it is easy to mistake ordinary transition samples for boundary samples. Therefore, this scheme first uses hysteresis phase values ​​to give the phase position, and then uses exhaust gas component data, gas supply parameter data, and exhaust parameter data to give the execution end association relationship, thereby separating the boundary samples from the general samples.

[0044] S40: Based on the hysteresis phase value, the boundary sample extraction path in the initial neural network model is modified by sparse sensing to obtain a neural network model that integrates sparse sensing. The furnace gas dataset, boundary sample set, and boundary constraint set are then input into the neural network model that integrates sparse sensing to generate an atmosphere instruction set.

[0045] In this embodiment, the boundary sample extraction path in the initial neural network model is sparsely perceived and modified according to the hysteresis phase value to obtain a neural network model with sparse perception. This is mainly to directly embed the boundary position information represented by the aforementioned hysteresis phase value and boundary sample set into the model structure. The boundary sample extraction path can be understood as the path in the initial neural network model that specifically receives the feature input of the boundary sample set and completes the layer-by-layer extraction. The sparse perception modification can be understood as retaining, suppressing and reorganizing the connection relationship in the boundary sample extraction path according to the difference in the arrangement of the hysteresis phase value in different boundary sample segments. This makes the initial neural network model no longer use the same extraction method for all boundary sample segments, but instead forms a differentiated data path based on the positional relationship of the boundary sample set in the hysteresis loop.

[0046] It should be noted that the initial neural network model in this embodiment is a trainable network model for processing time-series industrial process data. It is not limited to a single specific network type. Those skilled in the art can select a feedforward network structure with multi-layer feature extraction capabilities, a time-series recursive network structure, or a time-series feature extraction structure with attention computation as the initial model according to the organization of the furnace gas dataset, boundary sample set, and boundary constraint set. Correspondingly, the so-called sparse perception transformation is not to reconstruct an independent model, but to apply selective retention and suppression constraints to the connection relationship of the boundary sample extraction path in the initial neural network model, so that the path consistent with the phase hierarchical sequence and the sample adjacency relationship remains active, and the path that is unrelated to boundary propagation or crosses the disconnection relationship has its weight reduced or closed.

[0047] Understandably, the focus here is not simply on increasing the input data, but on first changing the boundary sample extraction path based on the hysteresis phase value, and then using the modified sparse sensing-integrated neural network model to receive the furnace gas dataset, boundary sample set, and boundary constraint set. This is because the furnace gas dataset reflects the current atmosphere state inside the furnace, the boundary sample set reflects the sample content related to the impurity oxidation boundary, and the boundary constraint set reflects the adjacency relationship of the boundary sample set in adjacent perturbation segments. If the unified extraction path of the initial neural network model is still used, the difference between the boundary sample set and ordinary samples is easily buried in the general features. However, after the sparse sensing modification, the sparse sensing-integrated neural network model can preferentially project the boundary information corresponding to the hysteresis phase value into the boundary sample extraction path, and then combine the boundary constraint set to restrict the feature propagation direction, thereby generating an atmosphere instruction set that matches the current perturbation background.

[0048] Based on the hysteresis phase value, the boundary sample extraction path in the initial neural network model is modified using sparse sensing to obtain a neural network model fused with sparse sensing, including: Extract hysteresis phase values ​​and boundary sample segments from the boundary sample set, and generate a phase layering sequence based on the arrangement of hysteresis phase values ​​in the boundary sample segments; A sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set; Rewrite the boundary sample extraction path in the initial neural network model based on the sparse gating matrix to generate a path routing graph; Based on the path routing graph, the boundary sample extraction paths and non-boundary sample extraction paths in the initial neural network model are combined to obtain a neural network model that integrates sparse perception.

[0049] In this embodiment, the boundary sample extraction path in the initial neural network model is sparsely perceived and modified according to the hysteresis phase value to obtain a neural network model with sparse perception fusion. This mainly involves further mapping the aforementioned hysteresis phase value, boundary sample set, and boundary constraint set into path organization rules within the model. The phase hierarchical sequence is used to characterize the hierarchical arrangement of boundary sample segments in the hysteresis loop, the sparse gating matrix is ​​used to characterize the path opening and closing relationship between boundary sample segments at different levels, and the path routing graph is used to characterize the connection structure of the boundary sample extraction path after being rewritten by the sparse gating matrix. Therefore, the processing here is not to add ordinary input to the initial neural network model, but to first reconstruct the boundary sample extraction path based on the hysteresis phase value, and then use the reconstructed path and the non-boundary sample extraction path to form a neural network model with sparse perception fusion.

[0050] Understandably, the processing sequence here has a continuous data dependency. First, extracting hysteresis phase values ​​and boundary sample fragments from the boundary sample set to generate a phase layering sequence is to transform the positional differences of the boundary sample fragments in the hysteresis loop into data that can be used for path layering. Then, generating a sparse gating matrix based on the phase layering sequence and the sample adjacency relationship in the boundary sample set is to distinguish between boundary sample extraction paths that should be retained and those that should be weakened. Subsequently, rewriting the boundary sample extraction paths in the initial neural network model based on the sparse gating matrix and generating a path routing graph is to map the aforementioned boundary information from the data layer to the model structure layer. Finally, combining the boundary sample extraction paths and non-boundary sample extraction paths based on the path routing graph yields a neural network model that integrates sparse perception. This ensures that when the furnace gas dataset, boundary sample set, and boundary constraint set are subsequently input into the model, boundary-related features propagate along the constrained path, while non-boundary-related features propagate along another type of path.

[0051] For example, continuing the processing results of the boundary sample set and hysteresis phase values ​​mentioned above, if the boundary sample segments are B1, B2, B3, B4, and the corresponding hysteresis phase values ​​are 1, 2, 2, 3, then a phase layer sequence {(1,B1),(2,B2),(2,B3),(3,B4)} can be generated. If the sample adjacency relationship in the boundary sample set is {(B1,B2),(B2,B3),(B3,B4)}, then a sparse gating matrix can be generated accordingly. The sparse gating matrix is ​​used to characterize the preserved connectivity relationship between B1 and B4 in the boundary sample extraction path. The boundary sample extraction path in the initial neural network model is then rewritten based on the sparse gating matrix to generate a path routing map. This map is then combined with the non-boundary sample extraction path to form a neural network model that integrates sparse perception. Compared with the existing technology that directly uses a unified extraction path to process the furnace gas dataset, this embodiment enables the boundary sample extraction path inside the model to correspond to the boundary position relationship represented by the hysteresis phase value. Therefore, it can reduce the interference of tail gas composition fluctuations and furnace reaction intensity changes on the propagation of boundary-related features, which is beneficial for the subsequent generation of an atmosphere instruction set that better meets the needs of atmosphere window tracking.

[0052] A sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set, including: Phase access segments are generated based on the interlayer connection relationship between adjacent hysteresis phase values ​​in the phase layering sequence. Based on the adjacency relationship of the samples in the boundary sample set, the adjacent sample pairs corresponding to each phase access segment are extracted to generate sample connection clusters; Path opening and closing units are generated based on the co-occurrence relationship of boundary sample segments in each sample connection cluster; A sparse gating matrix is ​​generated based on the arrangement of path opening and closing units in each phase access segment.

[0053] Furthermore, a sparse gating matrix is ​​generated based on the phase layer sequence and the sample adjacency relationship in the boundary sample set. This is mainly to further transform the boundary position relationship represented by the hysteresis phase value into a path control relationship that can be directly applied to the initial neural network model. Here, the phase access segment is used to characterize the inter-layer segment that can propagate continuously between adjacent hysteresis phase values, the sample connection cluster is used to characterize the set of boundary sample segments connected to each phase access segment, the path opening and closing unit is used to characterize the paths that should be retained and the paths that should be suppressed between boundary sample segments in each sample connection cluster, and the sparse gating matrix is ​​used to characterize the arrangement of each path opening and closing unit in each phase access segment. Therefore, the processing here is not to directly generate the matrix, but to first convert the hysteresis phase value and the sample adjacency relationship into a data structure that can describe the path propagation conditions, and then form a sparse gating matrix based on this.

[0054] The sparse gating matrix can be generated by combining the inter-layer accessibility relations corresponding to the phase layer sequence and the adjacency relations corresponding to the boundary sample segments. The reserved positions in the matrix are used to indicate the path connections that are allowed to propagate, and the suppressed positions in the matrix are used to indicate the path connections that should be weakened or closed. Specifically, the inter-layer accessibility range can be determined first based on whether there is a continuous connection between adjacent hysteresis phase values. Then, combined with the adjacency relations between the boundary sample segments, the path connections that simultaneously satisfy phase continuity and sample adjacency are screened out and written into the reserved positions in the matrix. For path connections that do not satisfy phase continuity, cross the discontinuity interval, or are inconsistent with the propagation direction of the boundary, they are written into the suppressed positions in the matrix. In this way, the propagation conditions between boundary sample segments are transformed into gating constraints that can be directly applied to the boundary sample extraction path.

[0055] Understandably, the processing sequence here has a layer-by-layer constraint relationship. First, phase access segments are generated based on the inter-layer connection relationship of adjacent hysteresis phase values ​​in the phase hierarchical sequence. This is to determine which boundary sample segments have continuous propagation conditions at the hysteresis loop position. Then, adjacent sample pairs corresponding to each phase access segment are extracted based on the sample adjacency relationship in the boundary sample set, and sample connection clusters are generated. This is to map the positional continuity relationship with the sample adjacency relationship. Subsequently, path opening and closing units are generated based on the co-occurrence relationship of boundary sample segments in each sample connection cluster. This is to transform the co-occurrence situation between boundary sample segments into path preservation or path suppression conditions. Finally, a sparse gating matrix is ​​generated based on the arrangement relationship of path opening and closing units in each phase access segment. This allows the sparse gating matrix to retain both the boundary hierarchical relationship expressed by the hysteresis phase values ​​and the local connectivity relationship in the boundary sample set.

[0056] For example, continuing the aforementioned phase layering sequence {(1,B1),(2,B2),(2,B3),(3,B4)} and sample adjacency relationship {(B1,B2),(B2,B3),(B3,B4)}, a phase access segment {(1,2),(2,2),(2,3)} can be generated. Then, the adjacent sample pairs {(B1,B2),(B2,B3),(B3,B4)} corresponding to {(1,2),(2,2),(2,3)} are extracted and form a sample connection cluster. Within this sample connection cluster, B2 and B3 co-occur continuously in phase layer 2, while B1 and B2, and B3 and B4 co-occur across layers. Based on this, corresponding path opening and closing units can be generated, and further, a sparse gating matrix can be generated. The sparse gating matrix is ​​used to characterize the propagable paths between boundary sample segments in each phase access segment. Compared with the existing technology that directly sets the model path based on a unified connection method, this embodiment makes the path control relationship constrained by both the hysteresis phase value and the sample adjacency relationship. Therefore, it is beneficial to reduce the propagation of paths unrelated to the boundary and improve the response of the boundary sample extraction path to changes in the atmosphere boundary.

[0057] Path opening and closing units are generated based on the co-occurrence relationships of boundary sample segments in each sample connection cluster, including: Extract boundary sample fragments from each sample connection cluster, and generate co-occurrence fragment groups based on the common occurrence positions of the boundary sample fragments in the same phase access segment; Open path segments are generated based on the sequential relationship of boundary sample segments in the co-occurring segment group, and closed path segments are generated based on the discontinuity relationship of boundary sample segments in the co-occurring segment group. Path opening and closing units are generated based on the corresponding combination relationship between open and closed path segments in each sample connection cluster.

[0058] In this embodiment, path opening and closing units are generated based on the co-occurrence relationship of boundary sample segments in each sample connection cluster. This is mainly to further transform the co-occurrence relationship of boundary sample segments within the same phase access segment into a structure that can be directly used for path control. The co-occurrence segment group is used to represent the combination of boundary sample segments that appear simultaneously and have a connection relationship within the same phase access segment. The open path segment is used to represent the path segment that continues to propagate within the same phase access segment according to the sequential relationship. The closed path segment is used to represent the path segment that stops propagating within the same phase access segment due to the interval disconnection relationship. The path opening and closing unit is used to represent the combination result of the open path segment and the closed path segment within the same sample connection cluster. Therefore, the focus of the processing here is not to directly determine whether a certain path is retained, but to first decompose the co-occurrence relationship of boundary sample segments into the sequential relationship and the disconnection relationship, and then combine them to form the path opening and closing unit.

[0059] Understandably, the processing sequence here has a progressively shrinking relationship. First, boundary sample fragments are extracted from each sample connection cluster, and co-occurrence fragment groups are generated based on the common occurrence positions of these fragments in the same phase access segment. This is to determine which boundary sample fragments are in the same propagation background. Then, open path segments are generated based on the sequential relationship of the boundary sample fragments in the co-occurrence fragment groups, and closed path segments are generated based on the interval disconnection relationship of these fragments. This is to distinguish between continuously propagating sample connections and those that should be blocked. Finally, path opening and closing units are generated based on the corresponding combination relationship of open and closed path segments in each sample connection cluster. This ensures that the path opening and closing units can reflect both the co-occurrence range of boundary sample fragments and which connections within the same range should be retained and which should be closed.

[0060] For example, continuing the aforementioned sample connection cluster {(B1,B2),(B2,B3),(B3,B4)} and phase access segment {(1,2),(2,2),(2,3)}, then the co-occurrence fragment group {B1,B2} can be obtained within the phase access segment (1,2), the co-occurrence fragment group {B2,B3} can be obtained within the phase access segment (2,2), and the co-occurrence fragment group {B3,B4} can be obtained within the phase access segment (2,3). Here, B1 and B2, B2 and B3, and B3 and B4 respectively form consecutive relationships, which can form the open path segment {(B1,B2),(B2,B3)}. If a combination of B1 and B3 or B2 and B4, separated by other boundary sample segments, appears within a certain phase access segment, a closed path segment can be formed. Then, path opening and closing units are generated according to the corresponding combination relationship of open and closed path segments in each sample connection cluster. These path opening and closing units are used to generate sparse gating matrices in the future. Compared with the existing technology that directly controls path propagation based on fixed connection methods, this embodiment can directly map the co-occurrence relationship of boundary sample segments to path opening and closing relationship, thereby reducing path propagation that is unrelated to atmospheric boundaries.

[0061] Example 2 Please see Figure 2 As shown, based on the same inventive concept, this embodiment discloses a furnace atmosphere control system for a precious metal refining process. For details not covered in this embodiment, please refer to the relevant sections of Embodiment 1. The system includes: Disturbance partitioning module: used to acquire furnace gas datasets during precious metal refining process, and partition disturbance segments based on the tail gas component switching relationship and supply and exhaust correspondence in the furnace gas dataset; Phase extraction module: used to construct hysteresis loops based on oxygen and reducing gas content in each perturbation segment, and extract hysteresis phase values ​​based on the component crossing positions in the hysteresis loops; Constraint generation module: used to divide the boundary sample set according to the hysteresis phase value and the furnace gas dataset, and generate the boundary constraint set according to the adjacency relationship of the boundary sample set in adjacent perturbation segments; The instruction generation module is used to perform sparse sensing transformation on the boundary sample extraction path in the initial neural network model based on the hysteresis phase value, so as to obtain a neural network model that integrates sparse sensing. The furnace gas dataset, boundary sample set and boundary constraint set are input into the neural network model that integrates sparse sensing to generate the atmosphere instruction set.

[0062] The detailed description above, in conjunction with the accompanying drawings, describes examples but does not represent all examples that can be implemented or fall within the scope of the claims. The terms “example” and “exemplary” are used in this specification to mean “serving as an example, instance or illustration” and do not mean “superior to or better than other examples”.

[0063] Throughout this specification, the phrase "an embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the invention. Therefore, the use of these phrases may refer to more than one embodiment. Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0064] It should also be noted that these embodiments may be described as processes depicted as flowcharts, structural diagrams, or block diagrams. Although a flowchart may describe the operations as sequential processes, many of these operations can be performed in parallel or concurrently, and the order of these operations may be rearranged.

Claims

1. A method for controlling the furnace atmosphere in a precious metal refining process, characterized in that, include: Obtain the furnace gas dataset during the precious metal refining process, and divide the disturbance segment according to the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset. Hysteresis loops are constructed based on the oxygen and reducing gas contents in each perturbation segment, and hysteresis phase values ​​are extracted based on the component crossing positions in the hysteresis loops. The boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, and the boundary constraint set is generated based on the adjacency relationship of the boundary sample set in adjacent perturbation segments. Based on the hysteresis phase value, the boundary sample extraction path in the initial neural network model is modified by sparse sensing to obtain a neural network model that integrates sparse sensing. The furnace gas dataset, boundary sample set, and boundary constraint set are then input into the neural network model that integrates sparse sensing to generate an atmosphere instruction set.

2. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 1, characterized in that, The disturbance segments are divided based on the tail gas component switching relationship and the supply and exhaust correspondence in the furnace gas dataset, including: Extract tail gas component data, gas supply parameter data, and exhaust parameter data from the furnace gas dataset, and generate the dominant component sequence and gas supply and exhaust displacement sequence according to the sampling time. A switching anchor sequence is generated based on the transposition order of adjacent sampling times in the dominant component sequence, and a set of coupling segments is generated based on the reverse transposition order around the same sampling time in the supply and exhaust displacement sequence. A fragment boundary set is generated based on the overlapping relationship between the switching anchor point sequence and the coupling segment set, and a candidate fragment set is obtained by segmenting the furnace gas dataset based on the fragment boundary set. The candidate segments are screened based on continuous sections of exhaust gas components and alternating sections of exhaust gas supply and exhaust gas supply to identify disturbance segments.

3. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 2, characterized in that, A set of coupled segments is generated based on the reverse displacement order around the same sampling time in the supply and exhaust displacement sequence, including: Extract the gas supply displacement segment and exhaust displacement segment corresponding to each sampling time in the gas supply and exhaust displacement sequence, and generate candidate displacement pairs based on the distribution of the gas supply displacement segment and exhaust displacement segment around the same sampling time. Based on the opposite relationship between the gas supply displacement direction and the exhaust displacement direction in each candidate displacement pair, reverse displacement pairs are selected to obtain the set of reverse displacement pairs. Coupled segments are generated based on the continuous connection relationship of the reverse displacement pairs in the supply and exhaust displacement sequence, and coupled segment sets are generated based on the beginning and end connection relationship of each coupled segment.

4. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 1, characterized in that, Hysteresis loops are constructed based on the oxygen and reducing gas content in each perturbation segment, including: Oxygen and reducing gas content data are extracted from each perturbation segment, and a series of oxygen reduction points are generated according to the sampling time. A sequence of return nodes is generated based on the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the return node sequence. A closed trajectory is generated based on the turnaround node sequence and the lap envelope sequence, and a loop skeleton is generated based on the connection relationship between the beginning and end of the closed trajectory. Hysteresis loops are extracted based on the loop skeleton and closed trajectory.

5. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 4, characterized in that, A turnaround node sequence is generated based on the round-trip connection order in the oxygen reduction point sequence, and a cyclic envelope sequence is generated based on the outward expansion relationship on both sides of the turnaround node sequence, including: Generate forward connection sets and backward connection sets based on the connection directions of continuous sampling times in the oxygen reduction point sequence; Candidate points for turning back are generated based on the intersection of the forward connection set and the backward connection set, and a sequence of turning back nodes is generated based on the temporal arrangement of the candidate points for turning back. Extract the trajectory extension segments corresponding to both sides of the turnaround node sequence, and generate envelope segments based on the coverage relationship of the trajectory extension segments; Generate a cyclic envelope sequence based on the connection between the beginning and end of the envelope segments.

6. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 1, characterized in that, The boundary sample set is divided based on the hysteresis phase value and the furnace gas dataset, including: Based on the furnace gas dataset, tail gas component data, gas supply parameter data, and exhaust parameter data are obtained, and sample time series clusters are generated according to the perturbation segments. The hysteresis phase values ​​are then mapped to the corresponding sample time series clusters. Phase adjacency chains are generated based on the sequential connection relationship of hysteresis phase values ​​in each sample time series cluster, and boundary candidate clusters are generated based on the transposition position of exhaust gas component data in the phase adjacency chains. Boundary sample fragments are selected based on the inverse adjoint relationship between gas supply parameter data and exhaust parameter data in the boundary candidate cluster, and boundary sample clusters are generated. The boundary sample set is determined based on the sample attribution relationship of the boundary sample cluster in the furnace gas dataset.

7. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 6, characterized in that, Based on the hysteresis phase value, the boundary sample extraction path in the initial neural network model is modified using sparse sensing to obtain a neural network model fused with sparse sensing, including: Extract hysteresis phase values ​​and boundary sample segments from the boundary sample set, and generate a phase layering sequence based on the arrangement of hysteresis phase values ​​in the boundary sample segments; A sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set; Rewrite the boundary sample extraction path in the initial neural network model based on the sparse gating matrix to generate a path routing graph; Based on the path routing graph, the boundary sample extraction paths and non-boundary sample extraction paths in the initial neural network model are combined to obtain a neural network model that integrates sparse perception.

8. The method for controlling the furnace atmosphere in a precious metal refining process according to claim 7, characterized in that, A sparse gating matrix is ​​generated based on the phase hierarchical sequence and the sample adjacency relationship in the boundary sample set, including: Phase access segments are generated based on the interlayer connection relationship between adjacent hysteresis phase values ​​in the phase layering sequence. Based on the adjacency relationship of the samples in the boundary sample set, the adjacent sample pairs corresponding to each phase access segment are extracted to generate sample connection clusters; Path opening and closing units are generated based on the co-occurrence relationship of boundary sample segments in each sample connection cluster; A sparse gating matrix is ​​generated based on the arrangement of path opening and closing units in each phase access segment.

9. A method for controlling the furnace atmosphere in a precious metal refining process according to claim 8, characterized in that, Path opening and closing units are generated based on the co-occurrence relationships of boundary sample segments in each sample connection cluster, including: Extract boundary sample fragments from each sample connection cluster, and generate co-occurrence fragment groups based on the common occurrence positions of the boundary sample fragments in the same phase access segment; Open path segments are generated based on the sequential relationship of boundary sample segments in the co-occurring segment group, and closed path segments are generated based on the discontinuity relationship of boundary sample segments in the co-occurring segment group. Path opening and closing units are generated based on the corresponding combination relationship between open and closed path segments in each sample connection cluster.

Images