DRL-based identification method for misplacement of assembly direction of corrugated sheet of heat accumulating element of air preheater

Abstract

The application provides a DRL-based air preheater heat accumulating element corrugated sheet assembly direction misplacement recognition method, and belongs to the air preheater maintenance technical field.The method comprises the following steps: obtaining the maintenance operation data of the air preheater to form operation window data; calculating the flow resistance deviation and heat exchange lag deviation according to the operation window data and the same load reference data to form the flow resistance-heat exchange lag correlation quantity; inputting the operation window data into the feature extraction unit of the DRL to obtain channel state potential data, and generating initial decomposition activation quantity from the operation window data by the weight learner of the DRL; embedding the flow resistance-heat exchange lag correlation quantity into the generation link of the initial decomposition activation quantity to obtain channel direction abnormal decomposition data, and performing direction misplacement judgment based on the channel direction abnormal decomposition data to output the corrugated sheet assembly direction misplacement state.The method can complete the corrugated sheet assembly direction misplacement state judgment under the condition that only the operation measuring point data and the same load reference data can be obtained after the air preheater is maintained.

Description

Technical Field

[0001] This invention relates to the field of air preheater maintenance technology, specifically to a method, system, equipment, medium, and program for identifying misaligned assembly direction of corrugated plates in air preheater thermal storage elements based on DRL. Background Technology

[0002] Rotary air preheaters are widely used heat exchange devices in the tail flue of thermal power boilers. Their rotors contain numerous heat storage elements. High-temperature flue gas and cold air alternately contact these elements as the rotor rotates. The elements first absorb heat from the flue gas side and then release it to the air entering the boiler, thus achieving heat recovery. The corrugated fins in the heat storage elements not only provide heat exchange area but also form directional airflow channels. The corrugation direction, interlayer arrangement, and channel contact path collectively affect the air-to-flue gas pressure drop, flow distribution, hot and cold end temperature response, and thermal inertia under load changes. During maintenance, the heat storage elements typically need to be removed, cleaned, replaced, or reassembled. The quality of the reassembly will be reflected in the air-to-flue gas resistance and heat exchange response characteristics after the unit restarts. However, due to simultaneous changes in unit load, coal quality, air volume, and air leakage, the post-maintenance operating data often exhibits multi-factor coupling characteristics, making it difficult for a single measuring point or temperature index to directly correspond to the directional state of the corrugated fins.

[0003] A similar prior art to this application is the air preheater operation performance testing method and apparatus disclosed in Chinese Patent CN109357896B. This scheme converts a three-compartment air preheater into a first heat exchanger and a second heat exchanger according to a simplified theoretical model. Operating parameters, including flue gas mass flow rate, flue gas temperature, primary air mass and temperature, secondary air mass and temperature, exhaust gas temperature, and total air leakage, are obtained. Based on these operating parameters, characteristic parameters of the two heat exchangers under test conditions are calculated. Then, performance parameters under boundary conditions are deduced from these characteristic parameters, resulting in overall air preheater performance parameters. These parameters are then compared with preset thresholds to determine the performance status. The testing process revolves around heat transfer, mixing enthalpy, the product of heat transfer coefficient and heat transfer area, boundary outlet air temperature, and exhaust gas temperature, thereby completing a quantitative judgment of the air preheater's operating performance. However, the aforementioned prior art mainly treats the air preheater as an overall heat transfer performance object for status judgment. The performance parameters are derived from the comparison between the simplified heat exchanger model and the design boundary, and the diagnostic results are more biased towards whether the overall operating performance deviates from the design value. Regarding the specific issue of misaligned corrugated sheet assembly after maintenance, the reverse use of the channel caused by this misalignment will simultaneously manifest as deviation in flow resistance on the air / smoke side and lag in temperature response at both the hot and cold ends. These two phenomena are easily masked by load fluctuations, ordinary dust accumulation, air leakage, or general heat transfer degradation. The existing technology does not establish a directional correlation between flow resistance deviation and heat transfer lag deviation under the same load and window conditions, nor does it incorporate this directional correlation into the decomposition direction generation stage within the model. Therefore, it is difficult to effectively separate directional channel anomalies from ordinary resistance and heat transfer anomalies, failing to meet the need for accurate diagnosis of the directional state of the corrugated sheets after maintenance. Summary of the Invention

[0004] Existing technologies treat the air preheater as a whole for heat exchange performance evaluation, failing to establish a directional correlation between flow resistance deviation and heat transfer hysteresis deviation under the same load and window conditions. This prevents effective separation of directional channel anomalies caused by corrugated fin misalignment from ordinary resistance and heat transfer anomalies. This method provides a DRL-based identification method for misaligned corrugated fin assembly in air preheaters. Under conditions where only operational measurement data and baseline data at the same load are available after air preheater maintenance, this method identifies directional channels caused by misaligned corrugated fin assembly and allows for reverse use. This establishes a correlation constraint between flow resistance deviation and heat transfer hysteresis deviation within the same operating window, enabling the DRL decomposition and reconstruction process to be initiated. Based on this constraint, the misalignment status of the corrugated fin assembly is determined.

[0005] To achieve the above objectives, the present invention provides the following technical solution.

[0006] In a first aspect, the present invention provides a method for identifying misalignment of the corrugated sheet assembly direction of an air preheater heat storage element based on DRL, comprising: Acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; Flow resistance deviation and heat transfer hysteresis deviation are calculated based on the operating window data and the same load baseline data to form the flow resistance-heat transfer hysteresis correlation quantity; The running window data is input into the feature extraction unit of the DRL to obtain the channel state potential data, and the weight learner of the DRL generates the initial decomposition activation based on the running window data. The flow resistance-heat transfer hysteresis correlation is embedded into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and potential decomposition coefficients are generated based on the scene decomposition activation quantity. The potential data of the channel state is decomposed and reconstructed based on the potential decomposition coefficients to obtain the channel direction anomaly decomposition data. Based on the channel direction anomaly decomposition data, the direction misalignment is determined and the corrugated sheet assembly direction misalignment status is output.

[0007] As a further improvement of the present invention, the step of acquiring the maintenance and operation data of the air preheater and forming operation window data based on the maintenance and operation data includes: Obtain the operating data of the air preheater after maintenance, and filter the operating data of the air preheater after maintenance according to the preset stable operating conditions to form effective operating data after maintenance; Maintenance operation data is generated based on effective operating data after maintenance and benchmark data under the same load. Based on the maintenance and operation data, the pressure drop, flow rate, cold and hot end temperatures, and load response are extracted on the air and flue side to form the operation window data.

[0008] As a further improvement of the present invention, the step of calculating the flow resistance deviation and heat transfer hysteresis deviation based on the operating window data and the same load reference data to form the flow resistance-heat transfer hysteresis correlation includes: Based on the load response in the running window data, a benchmark segment in the same load response range is matched in the same load benchmark data, and then truncated and aligned to form the same load comparison window data. Calculate the flow resistance deviation and temperature response hysteresis based on the data from the same load comparison window. The effective window of resistance deviation is determined based on the flow resistance deviation, and the window consistency constraint is applied to the temperature response hysteresis to form the heat transfer hysteresis deviation. The direction of resistance deviation is determined based on the flow resistance deviation, and the direction of heat transfer lag is determined based on the heat transfer lag deviation. The correlation constraint between the direction of resistance deviation and the direction of heat transfer lag is constructed within the same time window of the operating window data to form the correlation offset constraint quantity. By embedding the associated offset constraint into the fusion process of flow resistance deviation and heat transfer hysteresis deviation, the deviation of individual air and flue gas side pressure drop and individual cold and hot end temperature deviation are suppressed, forming the flow resistance-heat transfer hysteresis correlation quantity.

[0009] As a further improvement of the present invention, the step of inputting the running window data into the feature extraction unit of the DRL to obtain the channel state potential data, and having the weight learner of the DRL generate the initial decomposition activation value based on the running window data, includes: The running window data is input into the feature extraction unit of the DRL according to the same time window to form window latent mapping data; Channel state constraints are constructed based on the window potential mapping data, and potential mapping components formed by single-phase air-smoke side pressure drop deviation or single-phase cold and hot end temperature deviation are suppressed to obtain channel state potential data. The running window data is input into the weight learner of the DRL to form candidate activations for decomposition directions; Within the weight learner of the DRL, the candidate activation values ​​for decomposition directions are constrained to correspond to the potential decomposition directions of the channel state potential data, thereby generating the initial decomposition activation values.

[0010] As a further improvement of the present invention, the step of embedding the flow resistance-heat transfer hysteresis correlation quantity into the generation step of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and generating potential decomposition coefficients based on the scene decomposition activation quantity, includes: Obtain the activation generation position of the initial decomposed activation in the weight learner of the DRL, and determine the intermediate activation to be embedded according to the activation generation position and the decomposition direction. Determine the associated embedding constraint based on the flow resistance-heat transfer hysteresis correlation; The associated embedding constraint quantity is embedded into the activation generation position where the intermediate activation quantity to be embedded is located, forming the scene constraint activation intermediate quantity. The initial decomposed activation quantity is updated according to the scene constraint activation intermediate quantity to form the associated embedding activation quantity. The activation attribution of the decomposition direction is consistently adjusted according to the associated embedding activation quantity to form the scene decomposed activation quantity candidate. The candidate activation values ​​for scene decomposition are normalized to obtain the scene decomposition activation values, and potential decomposition coefficients are generated based on the scene decomposition activation values.

[0011] As a further improvement of the present invention, the step of decomposing and reconstructing the potential data of the channel state according to the potential decomposition coefficient to obtain the channel direction anomaly decomposition data, and determining the direction misalignment based on the channel direction anomaly decomposition data to output the corrugated sheet assembly direction misalignment state includes: The proportion of channel state potential data entering the DRL decomposition and reconstruction process is determined based on the potential decomposition coefficient, and the proportion of decomposition direction is used as the basis for the allocation of channel state potential data to form the decomposition direction allocation ratio. Based on the allocation ratio of the decomposition direction, the channel decomposition projection data is determined; based on the potential decomposition coefficient, the reconstruction contribution of the channel decomposition projection data is limited to form channel reconstruction contribution data; then, the channel reconstruction contribution data is combined and reconstructed to form channel state reconstruction data. The potential reconstruction difference data is obtained by calculating the difference between the potential channel state data and the reconstructed channel state data. The potential reconstruction difference data is then reconstructed based on the correlation between the flow resistance and heat transfer hysteresis carried by the potential decomposition coefficient to obtain the channel direction anomaly decomposition data. Then, the directional channel response is aggregated to form the direction misalignment judgment input data. The direction misalignment judgment input data is used to calculate the direction misalignment response data; the judgment response quantity of the direction misalignment in the direction misalignment response data is compared with the preset threshold to form the direction misalignment judgment result; based on the direction misalignment judgment result, it is determined whether there is a direction misalignment, no direction misalignment, or direction misalignment needs to be verified, and the direction misalignment status of the corrugated sheet assembly is output.

[0012] Secondly, the present invention provides a DRL-based air preheater heat storage element corrugated sheet assembly orientation misalignment identification system, comprising: Operation window data module: used to acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; The hysteresis correlation module is used to calculate the flow resistance deviation and heat transfer hysteresis deviation based on the operating window data and the same load baseline data, forming the flow resistance-heat transfer hysteresis correlation. Initial decomposition activation quantity module: It is used to input the running window data into the feature extraction unit of DRL to obtain the channel state potential data, and the weight learner of DRL generates the initial decomposition activation quantity based on the running window data. Potential decomposition coefficient module: used to embed the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and generate potential decomposition coefficients based on the scene decomposition activation quantity. Corrugated sheet assembly orientation misalignment status module: It is used to decompose and reconstruct the potential data of the channel state according to the potential decomposition coefficient, obtain the abnormal decomposition data of the channel orientation, determine the orientation misalignment based on the abnormal decomposition data of the channel orientation, and output the corrugated sheet assembly orientation misalignment status.

[0013] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for identifying misaligned assembly orientation of corrugated sheet of air preheater heat storage element based on DRL.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for identifying misaligned assembly orientation of corrugated sheets for air preheater heat storage elements based on DRL.

[0015] Fifthly, the present invention provides a computer program product, including computer instructions, which, when executed by a processor, implement the aforementioned method for identifying misaligned assembly orientation of corrugated sheets for air preheater heat storage elements based on DRL.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This method constructs a flow resistance-heat transfer hysteresis correlation by simultaneously calculating flow resistance deviation and heat transfer hysteresis deviation under the same load and window conditions. This introduces directional correlation constraints into the air preheater performance evaluation, making directional channel anomalies structurally distinguishable from ordinary anomalies in terms of data characteristics. Secondly, this correlation is embedded into the DRL weight learner's initial decomposition activation value generation stage to obtain the scene decomposition activation value. Based on this, potential decomposition coefficients are generated, enabling the DRL feature extraction unit to possess directional sensitivity when decomposing and reconstructing potential channel state data. This allows for the effective separation of channel reverse usage features caused by corrugated sheet directional misalignment from mixed anomalies, outputting a clear directional misalignment determination result. In practical applications, this method relies solely on operational measurement data and load baseline data obtainable after air preheater maintenance, without requiring additional sensors or specialized testing. It possesses strong field applicability and engineering operability, enabling rapid identification of directional misalignment within the first operational window after maintenance. This provides a precise and efficient technical means for maintenance quality verification, effectively avoiding long-term operational efficiency losses and potential safety hazards caused by undetected directional misalignment, and significantly improving the quality control level and operational reliability of air preheaters after maintenance. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for identifying misaligned assembly orientation of corrugated sheets in an air preheater based on DRL, according to the present invention. Figure 2 This is a schematic diagram of the specific process of the method for identifying the misalignment of the corrugated sheet assembly direction of the air preheater heat storage element based on DRL according to the present invention. Figure 3 This is a flowchart illustrating step S1 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 4 This is a flowchart illustrating step S2 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 5This is a flowchart illustrating step S3 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 6 This is a flowchart illustrating step S4 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 7 This is a flowchart illustrating step S5 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 8 This is a flowchart illustrating step S6 of the method for identifying misaligned assembly direction of corrugated sheet of air preheater heat storage element based on DRL according to the present invention. Figure 9 This is a schematic diagram illustrating the feature extraction, weight learning, decomposition, and reconstruction of the DRL embedding correlation quantity in this invention. Figure 10 This is a schematic diagram illustrating the calculation of deviations and correlations between the maintenance operation data and the baseline data under the same load in this invention; Figure 11 This is a schematic diagram of the orientation misalignment determination model based on orientation response aggregation of the present invention; Figure 12 This is a schematic diagram comparing the results of forward and backward orientation misalignment recognition using the present invention in an embodiment of the present invention; Figure 13 This is a schematic diagram comparing the abnormal decomposition results before and after DRL embedding correlation in an embodiment of the present invention. Detailed Implementation

[0018] Definitions: Preset stable operating conditions: These are the operating conditions under which the load, air and flue gas side pressure drop, flow rate, and hot and cold end temperatures all meet the preset fluctuation thresholds within a preset time window after the unit is connected to the grid.

[0019] Measuring point caliber consistency screening: Candidate reference data and effective operation data after maintenance are screened using the same measuring point location, the same sampling interval, the same unit, and the same time window boundary.

[0020] Same measuring point diameter: The measurement location, data unit, sampling interval, and time window boundary of the corresponding measuring point in the effective operating data after maintenance and the reference data under the same load are consistent.

[0021] Same sampling time sequence: Effective operating data after maintenance and baseline data under the same load are arranged in the same sampling time sequence.

[0022] Same load baseline data: This is the baseline data that is matched from the normal operation data of the air preheater and is in the same load range as the effective operation data after maintenance.

[0023] Synchronous measurement point segment: A continuous operating data segment composed of measurement points on the air preheater's flue gas side, cold and hot ends, and load under the same sampling time sequence.

[0024] Candidate data for pressure-flow window: Data from the same time window formed after extracting the pressure drop and flow rate on the air-smoke side from the synchronous measurement point segment.

[0025] Load response quantity: The operational response quantity is characterized by both the magnitude and duration of load change within the same time window.

[0026] Temperature load constraint window data: This is the window data formed by adding the cold and hot end temperatures and load response to the pressure flow window candidate data.

[0027] Operating window data: This includes air and flue gas side pressure drop, flow rate, hot and cold end temperatures, and load response, all arranged according to the same time window.

[0028] Same load response range: The load range in which the load response in the running window data and the load response in the same load baseline data meet the preset threshold range.

[0029] Same load comparison window data: This is data extracted from the same load baseline data and aligned with the operating window data in terms of load response, time window, and measuring point diameter.

[0030] Pressure-flow correspondence offset: This refers to the offset of the correspondence between the air and flue gas side pressure drop and flow rate in the operating window data within the same time window and the control window data under the same load.

[0031] Same-window synthesis: Synthesizes the pressure-flow corresponding offsets on the air-smoke side within the same time window of the running window data to form a pressure-flow coupled offset.

[0032] Pressure-flow coupling offset: This is the offset of the relationship between air and flue gas pressure drop and flow rate relative to the data in the same load comparison window.

[0033] Load response deduction: Based on the load response data in the same load comparison window, the offset component caused by load changes is deducted from the pressure-flow coupling offset or temperature response hysteresis.

[0034] Flow resistance deviation: This is the deviation of the resistance of the air and flue gas flowing through the heat storage element under the same load conditions, formed after subtracting the pressure-flow coupling offset from the load response.

[0035] Temperature response lag: The lag of the cold and hot end temperature response in the operating window data over the time window relative to the cold and hot end temperature response in the control window data under the same load.

[0036] Effective window for resistance deviation: This refers to the operating window data that meets the preset threshold for flow resistance deviation and maintains the same time window data as the control window data under the same load.

[0037] Window consistency constraint: to limit the calculation and use of temperature response hysteresis within the same time window corresponding to the effective window of resistance deviation.

[0038] Heat transfer hysteresis bias: This is the temperature response hysteresis deviation at the hot and cold ends formed after the temperature response hysteresis is subtracted from the window consistency constraint and the load response.

[0039] Resistance deviation direction: This refers to the direction of resistance change relative to the baseline data under the same load.

[0040] Heat transfer hysteresis direction: This refers to the direction of the hysteresis change in the temperature response at the hot and cold ends, as exhibited by the heat transfer hysteresis deviation relative to the baseline data under the same load.

[0041] Associated offset constraint: This is a fusion constraint determined by the matching relationship between the resistance deviation direction and the heat transfer hysteresis direction within the same time window.

[0042] Flow resistance-heat transfer hysteresis correlation: This is an intermediate object formed by fusing flow resistance deviation and heat transfer hysteresis deviation under the constraint of correlation offset, used to characterize the correlation offset between resistance deviation and heat transfer hysteresis when a directional channel is used in reverse.

[0043] Joint latent mapping: The feature extraction unit of DRL performs joint mapping of air and flue gas side pressure drop, flow rate, cold and hot end temperature and load response within the same time window.

[0044] Window latent mapping data: latent data formed by joint latent mapping of running window data and corresponding to the same time window.

[0045] Channel state constraint: This constraint limits the potential mapping data of the window to reflect the channel state deviation of the thermal storage element and suppresses the dominant components of single-phase flue gas side pressure drop deviation or single-phase hot and cold end temperature deviation.

[0046] Channel state potential data: This is the data formed by channel state constraints on the window potential mapping data and used in the DRL decomposition and reconstruction process.

[0047] Same window distribution relationship: This refers to the corresponding changes in air and flue gas side pressure drop, flow rate, cold and hot end temperatures, and load response within the same time window in the operating window data.

[0048] Decomposition direction candidate activations: These are the activations formed by the weight learners of the DRL based on the same window distribution relationship, and for which the decomposition direction assignment constraints have not yet been completed.

[0049] Decomposition direction attribution constraint: A constraint that ensures the candidate activation values ​​of the decomposition direction correspond to the potential decomposition directions of the channel state potential data.

[0050] Initial decomposition activation: The activation value generated by the weight learner of DRL after the candidate activation values ​​of the decomposition direction are subject to the decomposition direction assignment constraint.

[0051] Activation generation location: The existing computational location in the weight learner of DRL where the initial decomposed activation is generated from the running window data.

[0052] Intermediate activation quantity to be embedded: This refers to the intermediate activation object that has not yet formed a scene decomposition activation quantity within the activation generation location.

[0053] Correlation offset effect strength: The effect strength determined by the flow resistance-heat transfer hysteresis correlation and used to constrain the intermediate quantity to be embedded for activation.

[0054] Association embedding constraint: This is the constraint formed after mapping the intensity of the association offset to the decomposition direction.

[0055] Scene constraint activation intermediate quantity: is the activation intermediate object formed after the activation intermediate quantity to be embedded is updated by the associated embedded constraint quantity within the activation generation position.

[0056] Associated embedded activation quantity: The activation quantity formed after updating the initial decomposed activation quantity according to the scenario constraints of the intermediate activation quantity.

[0057] The flow resistance-heat transfer hysteresis correlation constraint relationship is the constraint relationship formed by the flow resistance-heat transfer hysteresis correlation on the generation basis of the initial decomposition activation quantity and the decomposition direction assignment.

[0058] Decomposition direction related to channel orientation anomalies: Decomposition direction used to satisfy the flow resistance-heat transfer hysteresis correlation constraint relationship and to allocate potential channel state data.

[0059] Ordinary resistance anomaly direction: This is the decomposition direction dominated by the deviation of the single-phase air and flue gas side pressure drop and does not satisfy the constraint relationship of flow resistance-heat transfer hysteresis.

[0060] Ordinary heat transfer anomaly direction: This is the decomposition direction dominated by a single deviation in hot and cold end temperatures and that does not satisfy the flow resistance-heat transfer hysteresis correlation constraint relationship.

[0061] Candidate activation quantities for scene decomposition: These are the activation quantities formed after the activation attribution consistency adjustment and before normalization of the associated embedded activation quantities.

[0062] Scene decomposition activation quantity: The activation quantity formed by normalizing the scene decomposition activation quantity candidates and used to generate potential decomposition coefficients.

[0063] Potential decomposition coefficient: The proportion of the decomposition direction allocated to constrain the potential data decomposition and reconstruction of the channel state, formed based on the scene decomposition activation amount.

[0064] Decomposition direction ratio: The proportion of potential data in the receiving channel state corresponding to each decomposition direction of the potential decomposition coefficient.

[0065] Decomposition direction allocation ratio: This is the basis for allocating potential channel state data into the decomposition and reconstruction process based on the proportion of the decomposition direction.

[0066] Directional channel status: This refers to the channel status characterized by the channel direction of the airflow through the heat storage element from the flue gas side and the heat exchange contact path at the hot and cold ends, as shown in the operating window data.

[0067] Directional attribution matching data: This is data formed by matching the potential components in the channel state potential data with the directional channel state related decomposition directions according to the decomposition direction allocation ratio.

[0068] Non-directional reconstruction components: These are potential components that do not conform to the directional channel state and should not be used as the main basis for determining directional misalignment.

[0069] Channel orientation reconstruction constraint: This constraint limits the potential components dominated by unidirectional wind and smoke side pressure drop deviation and the potential components dominated by unidirectional hot and cold end temperature deviation to non-directional reconstruction components.

[0070] Channel direction reconstruction constraint: This is a constraint formed by the channel direction reconstruction constraint and used to limit the decomposition projection and combined reconstruction.

[0071] Channel decomposition projection data: This is the data formed by projecting the potential channel state data onto the decomposition direction according to the allocation ratio under the constraints of the channel direction reconstruction.

[0072] Channel reconstruction contribution data: Data formed after defining the reconstruction contribution of each decomposition direction to the channel state potential data for the potential decomposition coefficients.

[0073] Channel state reconstruction data: This is the data formed by combining and reconstructing the channel reconstruction contribution data based on the channel direction reconstruction constraints.

[0074] Potential Reconstruction Difference Data: This data is obtained by calculating the difference between the potential channel state data and the reconstructed channel state data.

[0075] Correlation Consistency Reconstruction: This involves reconstructing the potential reconstruction difference data based on the flow resistance-heat transfer hysteresis correlation amount carried by the potential decomposition coefficient, so that the difference components retain the correlation offset between the flow resistance deviation and the heat transfer hysteresis deviation within the same operating window data.

[0076] Channel direction anomaly decomposition data: This is anomaly decomposition data formed after the potential reconstruction difference data is reconstructed with correlation consistency and used for direction misalignment determination.

[0077] Directional channel response aggregation: This aggregates the anomalous decomposition components that are used in reverse for the corresponding directional channels in the channel direction anomaly decomposition data.

[0078] Directional misalignment determination input data: This is the data generated after aggregating directional channel responses and used for directional misalignment determination.

[0079] Judgment response quantity: The output response corresponding to the input data for determining direction misalignment, indicating whether direction misalignment exists, does not exist, or needs to be verified.

[0080] Direction misalignment response data: This is the decision data composed of various decision response quantities.

[0081] Mutual exclusion constraint: This constraint ensures that only one type of corrugated sheet assembly orientation misalignment state can be output between the states of orientation misalignment (existing, non-existent, and orientation misalignment pending verification).

[0082] Direction misalignment determination result: The direction misalignment determination result is formed based on the direction misalignment response data and mutual exclusion constraints.

[0083] Corrugated sheet assembly orientation misalignment status: This is the status output based on the orientation misalignment judgment result, indicating whether there is an orientation misalignment, no orientation misalignment, or orientation misalignment pending verification.

[0084] Existing technologies treat the air preheater as a whole for heat transfer performance evaluation, failing to establish a directional correlation between flow resistance deviation and heat transfer hysteresis deviation under the same load and window conditions. Therefore, they cannot effectively separate directional channel anomalies caused by corrugated fin misalignment from ordinary resistance and heat transfer anomalies. This method provides a DRL-based method for identifying the misalignment of corrugated fin assembly in air preheaters, such as... Figure 1 As shown, it includes: Acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; Flow resistance deviation and heat transfer hysteresis deviation are calculated based on the operating window data and the same load baseline data to form the flow resistance-heat transfer hysteresis correlation quantity; The running window data is input into the feature extraction unit of the DRL to obtain the channel state potential data, and the weight learner of the DRL generates the initial decomposition activation based on the running window data. The flow resistance-heat transfer hysteresis correlation is embedded into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and potential decomposition coefficients are generated based on the scene decomposition activation quantity. The potential data of the channel state is decomposed and reconstructed based on the potential decomposition coefficients to obtain the channel direction anomaly decomposition data. Based on the channel direction anomaly decomposition data, the direction misalignment is determined and the corrugated sheet assembly direction misalignment status is output.

[0085] This method, under the condition that only operational measurement point data and same load baseline data can be obtained after the air preheater is overhauled, identifies the reverse use of directional channels caused by the misalignment of the corrugated sheet assembly direction of the heat storage element. This makes the flow resistance deviation and heat transfer hysteresis deviation form an associated constraint within the same operating window that can enter the DRL decomposition and reconstruction process, and completes the determination of the misalignment state of the corrugated sheet assembly direction accordingly.

[0086] The present invention will be further explained and described below with reference to the accompanying drawings.

[0087] like Figure 2 As shown, a method for identifying misaligned assembly orientation of corrugated sheets in air preheater thermal storage elements based on DRL includes: S1: As Figure 3 As shown, the operation data and load baseline data of the air preheater after maintenance are obtained to form maintenance operation data; Specifically, the operation data of the air preheater after maintenance is obtained, and the operation data after maintenance is filtered according to the preset stable operating conditions to form effective operation data after maintenance. Acquire normal operating data of the air preheater, and filter candidate baseline data from the normal operating data of the air preheater based on the load range of the effective operating data after maintenance; The candidate benchmark data and the effective operating data after maintenance are screened for consistency of measuring point diameter to form benchmark candidate data with the same measuring point diameter; The baseline candidate data are matched with the effective operating data after maintenance according to the same sampling time sequence to form the same load baseline data; Maintenance operation data is generated based on effective operating data after maintenance and benchmark data under the same load.

[0088] S2: As Figure 4 As shown, the pressure drop, flow rate, cold and hot end temperatures, and load response are extracted from the maintenance and operation data to form the operation window data; Specifically, based on the sampling time in the maintenance and operation data, the air preheater's flue gas side, cold and hot ends, and load measurement points are time-aligned, and the sampling interval is compared with a preset threshold. Measurement point segments that meet the comparison conditions are selected to form synchronous measurement point segments. Based on the pressure difference between the inlet and outlet of the flue gas side and the flow measurement value of the flue gas side in the synchronous measurement point segment, the pressure drop and flow rate of the flue gas side are extracted to form candidate data for the pressure-flow window. Based on the synchronous measurement point segments corresponding to the candidate data of the pressure-flow window, the cold and hot end temperatures are extracted, and the load response is calculated based on the load change amplitude and load change duration. The candidate data of the pressure-flow window is then updated to form the temperature load constraint window data. The temperature load constraint window data, including air and flue gas side pressure drop, flow rate, cold and hot end temperatures, and load response, are arranged according to the same time window to form the operating window data.

[0089] S3: As Figure 5 As shown, flow resistance deviation and heat transfer hysteresis deviation are calculated based on the operating window data and the same load reference data. The flow resistance deviation and heat transfer hysteresis deviation are then combined to form the flow resistance-heat transfer hysteresis correlation quantity. Specifically, based on the load response in the running window data, a benchmark segment in the same load response interval is matched in the same load benchmark data, and the benchmark segment is truncated and aligned according to the time window of the running window data to form the same load comparison window data. Based on the pressure drop and flow rate on the flue gas side in the same load comparison window data, calculate the pressure-flow corresponding offset of the operating window data relative to the same load benchmark data, and synthesize the pressure-flow corresponding offset on the flue gas side within the same window to form the pressure-flow coupling offset. Based on the load response data in the same load comparison window, the pressure-flow coupling offset is deducted by the load response to prevent the change in air and smoke side pressure drop caused by load fluctuation from being included in the resistance deviation calculation, thus forming the flow resistance deviation. Based on the cold and hot end temperatures in the same load comparison window data, calculate the time lag of the cold and hot end temperature response relative to the same load baseline data, and form the temperature response lag amount with the time window of the running window data as the boundary. The effective window of resistance deviation is determined based on the flow resistance deviation, the window consistency constraint is applied to the temperature response hysteresis, and the hysteresis component affected by load disturbance is deducted from the load response in the comparison window data under the same load to form the heat transfer hysteresis deviation. The direction of resistance deviation is determined based on the flow resistance deviation, and the direction of heat transfer lag is determined based on the heat transfer lag deviation. The correlation constraint between the direction of resistance deviation and the direction of heat transfer lag is constructed within the same time window of the operating window data to form the correlation offset constraint quantity. By embedding the associated offset constraint into the fusion process of flow resistance deviation and heat transfer hysteresis deviation, the deviation of single-phase air and flue gas pressure drop and the deviation of single-phase hot and cold end temperature are suppressed. This allows the directional channel to be used in reverse, and the corresponding resistance deviation and heat transfer hysteresis enter the same intermediate object together, forming the flow resistance-heat transfer hysteresis correlation quantity.

[0090] The flow resistance deviation and heat transfer hysteresis deviation are combined to form a flow resistance-heat transfer hysteresis correlation quantity. This includes: determining the resistance deviation direction based on the flow resistance deviation and the heat transfer hysteresis direction based on the heat transfer hysteresis deviation within the same time window of the operating window data; matching the resistance deviation direction with the heat transfer hysteresis direction; when the matching result meets the preset directional relationship, the flow resistance deviation and heat transfer hysteresis deviation are combined to form a flow resistance-heat transfer hysteresis correlation quantity; when the matching result does not meet the preset directional relationship, the effects of single-phase flue gas side pressure drop deviation and single-phase hot and cold end temperature deviation on the formation of the flow resistance-heat transfer hysteresis correlation quantity are suppressed.

[0091] S4: As Figure 6 As shown, the running window data is input into the feature extraction unit of the DRL to obtain the channel state potential data, and the weight learner of the DRL generates the initial decomposition activation based on the running window data. Specifically, the operating window data is input into the feature extraction unit of DRL according to the same time window, and the air and flue gas side pressure drop, flow rate, cold and hot end temperature and load response in the operating window data are jointly potential mapped to form window potential mapping data; Channel state constraints are constructed based on the window potential mapping data, and potential mapping components formed by single-phase air-smoke side pressure drop deviation or single-phase cold and hot end temperature deviation are suppressed to obtain channel state potential data. The running window data is input into the weight learner of the DRL. Based on the same window distribution relationship of the air and flue gas side pressure drop, flow rate, cold and hot end temperature and load response in the running window data, candidate activation quantities for decomposition direction are formed. Within the weight learner of the DRL, the candidate activation values ​​for decomposition directions are constrained to correspond to the potential decomposition directions of the channel state potential data, thereby generating the initial decomposition activation values.

[0092] DRL, short for Decomposed Representation Learning for Tabular Anomaly Detection, is a decomposed representation learning method for tabular anomaly detection. It maps the running window data to channel state latent data and decomposes and reconstructs the channel state latent data using latent decomposition coefficients.

[0093] S5: As Figure 7 As shown, the flow resistance-heat transfer hysteresis correlation is embedded into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and the potential decomposition coefficient is generated based on the scene decomposition activation quantity. Specifically, the activation generation position of the initial decomposed activation quantity in the weight learner of the DRL is obtained, and the intermediate activation quantity to be embedded is determined according to the activation generation position and the decomposition direction. The correlation offset intensity within the same operating window is determined based on the flow resistance-heat transfer hysteresis correlation, and the correlation offset intensity is mapped to the decomposition direction corresponding to the intermediate quantity to be embedded and activated, thus forming the correlation embedding constraint quantity. Embed the associated embedding constraint quantity into the activation generation position where the activation intermediate quantity to be embedded is located, so that the flow resistance-heat transfer hysteresis correlation quantity updates the activation intermediate quantity to be embedded within the activation generation position, forming the scene constraint activation intermediate quantity. The initial decomposition activation quantity is updated based on the intermediate activation quantity according to the scenario constraints. The generation basis of the initial decomposition activation quantity is updated from the distribution deviation of the running window data to the generation basis jointly limited by the distribution deviation of the running window data and the flow resistance-heat transfer hysteresis correlation quantity, forming the correlation embedded activation quantity. Based on the correlation embedded activation amount, the activation attribution of the decomposition direction is adjusted in a consistent manner. The activation components that satisfy the flow resistance-heat transfer hysteresis correlation amount constraint relationship are retained in the decomposition direction that is abnormally related to the channel direction, and the activation components of the ordinary resistance abnormal direction and the ordinary heat transfer abnormal direction that do not satisfy the flow resistance-heat transfer hysteresis correlation amount constraint relationship are suppressed, thus forming a candidate scene decomposition activation amount. The scene decomposition activation quantity candidates are normalized and adjusted according to the decomposition direction allocation relationship required for the channel state potential data decomposition and reconstruction to obtain the scene decomposition activation quantity. Based on the scene decomposition activation quantity, potential decomposition coefficients are generated, so that the potential decomposition coefficients bear the embedded constraint of the flow resistance-heat transfer hysteresis correlation quantity on the initial decomposition activation quantity.

[0094] The process of embedding the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation quantity includes: inputting the flow resistance-heat transfer hysteresis correlation into the activation generation position before normalizing the scene decomposition activation quantity candidates; updating the intermediate activation quantity to be embedded in each decomposition direction according to the flow resistance-heat transfer hysteresis correlation; forming scene constraint activation intermediate quantities according to the updated intermediate activation quantities to be embedded, and sequentially forming correlation embedding activation quantities and scene decomposition activation quantity candidates; normalizing the scene decomposition activation quantity candidates to obtain the scene decomposition activation quantity, and generating potential decomposition coefficients according to the scene decomposition activation quantity.

[0095] S6: As Figure 8 As shown, the channel state potential data is decomposed and reconstructed based on the potential decomposition coefficients to obtain the channel direction anomaly decomposition data; Specifically, the process receives potential decomposition coefficients and potential channel status data. Based on the potential decomposition coefficients, it determines the proportion of channel status potential data that will enter the DRL decomposition and reconstruction process, and uses the proportion of decomposition directions as the basis for allocating channel status potential data to form a decomposition direction allocation ratio. Based on the decomposition direction allocation ratio, the potential components in the potential data of the channel state are decomposed and matched in the direction, and the potential components that do not match the directional channel state are marked as non-directional reconstruction components, forming direction assignment matching data. Based on the direction attribution matching data, channel direction reconstruction constraints are constructed, which restrict the potential components dominated by single-phase wind and smoke side pressure drop deviation and the potential components dominated by single-phase cold and hot end temperature deviation to non-direction reconstruction components, thus forming the channel direction reconstruction constraint quantity. Based on the channel direction reconstruction constraint, the potential data of the channel state is decomposed and projected, and the projection results are allocated to the decomposition direction related to the directional channel state according to the decomposition direction allocation ratio to form channel decomposition projection data. The reconstruction contribution of the channel decomposition projection data is limited based on the potential decomposition coefficients, so that the potential decomposition coefficients directly constrain the reconstruction contribution of each decomposition direction to the channel state potential data, thus forming channel reconstruction contribution data. Based on the channel direction reconstruction constraints, the channel reconstruction contribution data is combined and reconstructed so that the retained reconstruction contribution comes from the directional channel state decomposition direction within the same running window data, forming channel state reconstruction data. The potential reconstruction difference data is obtained by calculating the difference between the potential channel state data and the reconstructed channel state data. The potential reconstruction difference data is then reconstructed based on the flow resistance-heat transfer hysteresis correlation amount carried by the potential decomposition coefficient to obtain the channel direction anomaly decomposition data.

[0096] The potential data of the channel state is decomposed and reconstructed based on the potential decomposition coefficients, including: when forming the channel reconstruction contribution data, determining the reconstruction contribution of each decomposition direction to the potential data of the channel state based on the potential decomposition coefficients; based on the flow resistance-heat transfer hysteresis correlation amount carried by the potential decomposition coefficients, limiting the reconstruction contributions corresponding to the potential components dominated by single-phase air-smoke side pressure drop deviation and the potential components dominated by single-phase cold and hot end temperature deviation to the non-directional reconstruction components; forming the channel state reconstruction data based on the reconstruction contributions not limited to the non-directional reconstruction components; and calculating the difference between the potential data of the channel state and the channel state reconstruction data to obtain the channel directional anomaly decomposition data.

[0097] S7: Based on the channel direction anomaly decomposition data, determine the direction misalignment and output the corrugated sheet assembly direction misalignment status.

[0098] Specifically, directional channel response aggregation is performed on the channel direction anomaly decomposition data, and the anomaly decomposition components used in reverse of the corresponding directional channels are extracted to form the direction misalignment judgment input data. Based on the input data for direction misalignment determination, the determination response quantities for the presence of direction misalignment, the absence of direction misalignment, and the direction misalignment pending verification are calculated to form direction misalignment response data; The judgment response quantity of the direction misalignment response data is compared with the preset threshold, and the judgment response quantities of the direction misalignment response data that do not contain direction misalignment and the direction misalignment that needs to be reviewed are used as mutually exclusive constraints to form the direction misalignment judgment result. Based on the orientation misalignment determination result, determine whether orientation misalignment exists, does not exist, or is pending verification, and output the orientation misalignment status of the corrugated sheet assembly.

[0099] This method addresses the issue of misalignment of corrugated sheet assembly direction after maintenance being easily confused by load disturbances, abnormal general resistance, and abnormal general heat transfer. It acquires post-maintenance operating data and baseline data under the same load, extracts air-to-flue gas side pressure drop, flow rate, hot and cold end temperatures, and load response, calculates flow resistance deviation and heat transfer hysteresis deviation, and merges them to form a flow resistance-heat transfer hysteresis correlation. This correlation is embedded into the DRL initial decomposition activation generation stage to generate potential decomposition coefficients. The potential channel state data is then decomposed and reconstructed to obtain channel direction anomaly decomposition data, which is used to output the corrugated sheet assembly direction misalignment status. This invention is used for identifying the assembly status of air preheaters after maintenance, reducing the interference of single operational deviations on the determination of direction misalignment.

[0100] In this method, DRL stands for Decomposed Representation Learning for Tabular Anomaly Detection, a decomposition representation learning method for tabular anomaly detection. In this approach, DRL can be understood as a main algorithm for anomaly identification in tabular data. It first maps the input tabular data to a latent space, forming a latent representation; then, using several fixed decomposition directions and corresponding decomposition coefficients, it decomposes and reconstructs the latent representation; finally, based on the difference between the latent representation and the decomposition and reconstruction results, it determines whether the sample contains an anomaly.

[0101] In this method, the DRL is not directly used for ordinary anomaly detection, but rather serves as the main basis for identifying deviations in the channel state of the air preheater's heat storage elements. Specifically, after the operating window data enters the feature extraction unit of the DRL, it forms potential channel state data; the weight learner of the DRL generates initial decomposition activation values ​​based on the operating window data; then, the flow resistance-heat transfer hysteresis correlation is embedded in the generation process of the initial decomposition activation values, so that the potential decomposition coefficients can reflect the correlation shift between flow resistance deviation and heat transfer hysteresis deviation caused by the misalignment of the corrugated sheet assembly direction.

[0102] Therefore, the focus of this method in using DRL is not simply calling DRL for anomaly detection, but rather improving the decomposition coefficient generation process within DRL so that the decomposition and reconstruction process of DRL serves the identification of misalignment in the assembly direction of the air preheater corrugated sheets.

[0103] like Figure 9This diagram illustrates the core processing flow of the DRL model in this method. The running window data first enters the feature extraction section, where it forms channel state latent values ​​through joint mapping. Simultaneously, the running window data enters the weight learning section, generating initial activation values ​​and embedding the flow resistance-heat transfer hysteresis correlation value into the activation generation process, forming scene activation and latent coefficients. Subsequently, in the decomposition and reconstruction section, the channel state latent data is decomposed and reconstructed based on the latent coefficients, suppressing common resistance anomalies and common heat transfer anomalies, and outputting anomaly decompositions. This diagram highlights that this invention does not simply use correlation values ​​as external judgment conditions, but rather embeds the flow resistance-heat transfer hysteresis correlation value into the generation stage of the initial decomposition activation value, allowing the latent decomposition coefficients to bear correlation constraints, thereby improving the support of channel direction anomaly decomposition data for direction misalignment judgment.

[0104] Figure 10 This diagram illustrates the process from input data to the formation of the flow resistance-heat transfer hysteresis correlation. The left side inputs include maintenance operation data and baseline data under the same load. The data is then organized into an operation window, containing pressure drop, flow rate, hot and cold end temperatures, and load. Deviation calculations are then performed to obtain flow resistance deviation and hysteresis deviation, further forming the flow resistance-hysteresis correlation. This diagram corresponds to the technical process in this invention of acquiring maintenance operation data, forming operation window data, calculating flow resistance deviation and heat transfer hysteresis deviation, and fusing them to form the flow resistance-heat transfer hysteresis correlation. The advantage lies in using baseline data under the same load for comparison, reducing the impact of load changes and normal operational fluctuations on the judgment results, and providing a clearer intermediate object for subsequent DRL embedding and direction misalignment determination through the correlation.

[0105] Figure 11 This diagram illustrates the judgment process after channel anomaly decomposition data enters directional channel response aggregation. The left side uses channel anomaly decomposition as input, representing the anomaly components obtained after decomposing and reconstructing the potential channel state data. The middle section aggregates the anomaly components used in reverse for the corresponding directional channels through directional aggregation and directional response aggregation, forming the judgment input. The right side generates misaligned quantities, normal quantities, and quantities to be reviewed, and outputs the misaligned state through thresholds and mutual exclusion constraints. This diagram demonstrates that the invention does not directly judge based solely on single-item pressure drop or temperature anomalies, but rather first obtains channel directional anomaly decomposition data and then performs directional misalignment judgment. The advantage is that it can distinguish between ordinary resistance anomaly directions, ordinary heat transfer anomaly directions, and channel directional anomaly directions, improving the specificity and stability of the directional misalignment state output.

[0106] In summary, this method embeds the correlation between flow resistance and heat transfer hysteresis into the initial decomposition activation generation stage, rather than performing threshold correction after model output. Existing overall performance detection or direct classification methods typically use air-to-smoke side pressure drop deviation and hot-and-cold end temperature deviation as input features or result judgment criteria, making it difficult to distinguish between directional channel reverse use and similar operational performance caused by ordinary ash accumulation, air leakage, and heat transfer degradation. This method introduces correlation embedding constraints at the activation generation position of the DRL weight learner, enabling the scene decomposition activation quantity to complete the decomposition direction assignment adjustment before normalization, retaining activation components that satisfy the correlation between flow resistance deviation and heat transfer hysteresis deviation, and suppressing activation components of ordinary resistance abnormal directions and ordinary heat transfer abnormal directions. Thus, directional misalignment related information enters the model's internal representation formation process, helping to reduce the interference of single operational deviations on assembly directional misalignment judgment. Secondly, the post-maintenance operational data is matched with the same load baseline data in terms of load range, measuring point diameter, sampling sequence, and time window. The operational window data simultaneously includes air-to-smoke side pressure drop, flow rate, hot-and-cold end temperature, and load response. This method calculates the pressure-flow offset based on a load comparison window and forms the flow resistance deviation by subtracting from the load response. Simultaneously, it constrains the lag in the cold and hot end temperature response using the effective window of the resistance deviation, forming a heat transfer lag deviation. Both are linked within the same time window according to the direction of resistance deviation and the direction of heat transfer lag, ensuring that the combined operational performance generated when directional channels are used in reverse enters the same intermediate object. Compared to judging solely based on overall heat transfer performance or a single pressure drop / temperature index, this invention better reflects the specific performance of corrugated sheet misalignment in air preheater operation. Finally, the potential decomposition coefficient in this method constrains the scene decomposition activation amount based on the flow resistance / heat transfer lag correlation amount and is used to limit the allocation ratio of potential channel state data into each fixed decomposition direction. During the decomposition and reconstruction process, this method, through direction-assigned matching data and channel direction reconstruction constraints, restricts the dominant components of single-phase air / smoke side pressure drop deviation and single-phase cold / hot end temperature deviation as non-directional reconstruction components, ensuring that abnormal channel direction decomposition data mainly originates from the directional channel state-related decomposition directions. The final judgment step is based on the channel direction anomaly decomposition data output, which indicates whether there is a direction misalignment, no direction misalignment, or a direction misalignment pending verification, thus forming a closed technical chain corresponding to the post-maintenance assembly status identification task.

[0107] The present invention will be further explained and illustrated below with reference to specific embodiments.

[0108] Example In this embodiment, step S1 specifically includes: The operating data after the air preheater overhaul is represented as a timestamped sampling sequence. ,in, This represents the sampling index of the operating data after the air preheater maintenance. This represents the total number of samples of operating data after the air preheater maintenance. Indicates the first Each sampling timestamp Indicates the first Each sampling timestamp corresponds to a multi-point record. The multi-point record includes the unit's grid connection status, load, air-side inlet pressure, air-side outlet pressure, flue gas inlet pressure, flue gas outlet pressure, air-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, and hot-end outlet temperature. Normal air preheater operation data is represented as a timestamped sampling sequence. ,in, The sampling index represents the data of the air preheater during normal operation. This represents the total number of samples taken for normal operation of the air preheater. Indicates the first Each sampling timestamp Indicates the first Each sampling timestamp corresponds to a multi-point record. The multi-point records in the air preheater's normal operation data use the same fields as the operation data after the air preheater's overhaul.

[0109] The operational data after the air preheater overhaul are arranged in ascending order according to the sampling timestamp, and a unified time axis is constructed. ,in, Indicates the start time of the unified timeline. Indicates a uniform sampling interval. , This indicates the number of sampling points on a unified time axis. Multiple measurement point records falling within the same unified sampling interval from the operational data after the air preheater overhaul are merged into the corresponding index. To form a discrete sequence For any measurement point field, if the duration of the missing data does not exceed the preset missing time threshold, the boundary is maintained using adjacent valid sample values; if the duration of the missing data exceeds the preset missing time threshold, the time window containing the missing measurement point field will not participate in the formation of valid operating data after maintenance.

[0110] The preset stable operating conditions are jointly defined by the unit's grid connection status, a preset time window, and a preset fluctuation threshold. The stability screening window is represented as... ,in, Indicates the starting index of the stable filter window. This indicates the length of the stability screening window. For each stability screening window, the maximum and minimum differences for load, flue gas side pressure drop, flow rate, and hot / cold end temperature are determined, and these differences are compared with corresponding preset fluctuation thresholds. If the unit is effectively connected to the grid, and the load difference, flue gas side pressure drop difference, flow rate difference, and hot / cold end temperature difference do not exceed the corresponding preset fluctuation thresholds, the multi-measurement records within the stability screening window are retained; otherwise, they are discarded. The retained multi-measurement records are arranged in index order along a unified time axis, forming valid post-maintenance operating data. ,in, This represents the sampling index of valid operating data after maintenance. This represents the total number of valid operational data samples collected after maintenance.

[0111] The load range is determined based on the load field of the effective operating data after maintenance. The load sequence in the effective operating data after maintenance is represented as follows: ,in, Indicates the index in the valid operating data after maintenance. The corresponding load value. Based on... The minimum load value, maximum load value, and preset load tolerance form the load range. The load values ​​of each multi-measuring point record in the air preheater's normal operation data are compared. Multi-measuring point records with load values ​​within the load range are retained, while those with load values ​​outside the load range are removed. The retained air preheater normal operation data forms the candidate baseline data. ,in, This represents the sampling index of the candidate benchmark data. This represents the total number of samples taken from the candidate benchmark data.

[0112] The candidate baseline data and the effective operational data after maintenance are screened for consistency of the measuring point caliber. The measuring point field index is represented as follows: The diameter of the measuring point is expressed as ,in, Represents measurement point field Measurement location, Represents measurement point field sampling interval, Represents measurement point field Data units Represents measurement point field The time window boundaries are defined. The caliber of each measuring point in the candidate reference data is compared item by item with the corresponding measuring point caliber in the effective operational data after maintenance. Only candidate reference data with consistent measurement location, sampling interval, data unit, and time window boundaries are retained. These retained candidate reference data form the reference candidate data with the same measuring point caliber. ,in, This represents the total number of samples taken from the baseline candidate data.

[0113] The baseline candidate data is matched with the effective operational data after maintenance, following the same sampling time sequence. For each index in the effective operational data after maintenance... The record with the smallest sampling time difference and load difference not exceeding the preset load matching threshold is selected from the baseline candidate data. The matching index is represented as follows: ,in, Indicates the index in the valid operating data after maintenance. The corresponding sampling time, Indicating the index in the benchmark candidate data The corresponding sampling time, Representation and Index The index of the matching baseline candidate data. When Furthermore, when the load difference does not exceed the preset load matching threshold, a matching relationship is established between the effective operating data after maintenance and the baseline candidate data. This indicates the preset alignment time threshold. Matching relationships are checked according to the sampling index order of the effective operating data after maintenance. If the baseline candidate data indices in adjacent matching relationships do not meet the ascending order, the matching relationships that disrupt the same sampling time sequence are eliminated. The retained baseline candidate data are arranged according to the sampling index order of the effective operating data after maintenance, forming baseline data for the same load. ,in, This represents the total number of samples after matching is complete.

[0114] Maintenance operation data is generated based on effective operational data after maintenance and baseline data under the same load. Maintenance operation data is represented as follows: ,in, This indicates the first valid operating data after maintenance. A number of retained multi-point measurement records Indicates the same load baseline data as Corresponding multi-point records. Each and each All measurements include load, wind-side inlet pressure, wind-side outlet pressure, smoke-side inlet pressure, smoke-side outlet pressure, wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, and hot-end outlet temperature, and are consistent in terms of load range, measuring point diameter, and sampling sequence.

[0115] In this embodiment, step S2 specifically includes: Post-maintenance valid operational data records from the maintenance operation data serve as the source for extracting operational window data. The timestamped sampling sequence participating in the measurement point extraction from the maintenance operation data is represented as follows: , Indicates the measurement point field. Represents the set of measurement point fields. , It includes wind-side inlet pressure, wind-side outlet pressure, smoke-side inlet pressure, smoke-side outlet pressure, wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, hot-end outlet temperature, and load measurement points; Represents measurement point field The sampling index, Represents measurement point field The total number of samples, Represents measurement point field Sampling time, Represents measurement point field At sampling time The measured values ​​at the following points.

[0116] For each measurement point, the fields are arranged in ascending order of sampling time, with a uniform sampling interval. Constructing a unified timeline , Indicates the start time of the unified timeline. Indicates a uniform sampling interval. , This indicates the number of sampling points on the unified timeline. For each unified timeline index... In each measurement point field Selecting from the sampling sequence The measurement point value with the smallest time difference is represented by the matching index as follows: , Represents measurement point field In a unified timeline index The original sampling index is bound below. This applies when the minimum time difference corresponding to all measurement point fields does not exceed the preset measurement point alignment threshold. When the time comes, write the corresponding measurement point value into the alignment measurement point record. , Represents measurement point field In a unified timeline index The alignment measurement point values ​​below.

[0117] The time difference between adjacent bound samples is checked for each measurement point field, and the time difference between adjacent bound samples is compared with the preset sampling interval threshold. Comparison; when the sampling time difference between adjacent bound samples of all measurement point fields within the continuous index range does not exceed the preset sampling interval threshold. At that time, the corresponding continuous index range is retained. The retained continuous index ranges are arranged according to the same sampling time sequence to form synchronous measurement point segments. , This indicates a multi-field record composed of the air preheater flue gas side, cold and hot ends, and load measurement points under the same sampling time sequence. This indicates the number of sampling points in the synchronized measurement point segment. Valid sample values ​​are aligned measurement point values ​​that have been time-aligned and have not been discarded due to sampling interval comparison.

[0118] Extract the air-to-smoke pressure drop and flow rate from the synchronous measurement point segments. For each synchronous sampling index... The difference between the wind-side inlet pressure and the wind-side outlet pressure is defined as the wind-side pressure drop, and the difference between the smoke-side inlet pressure and the smoke-side outlet pressure is defined as the smoke-side pressure drop. The wind-side flow rate and flue gas flow rate under the same synchronous sampling index are read. The window length parameter is expressed as... The window step size parameter is expressed as , will the Each time window is represented as , , , This indicates the number of time windows. For each time window, the arithmetic mean of the effective sampled values ​​for wind-side pressure drop, smoke-side pressure drop, wind-side flow rate, and flue gas flow rate are used as the window field value to form candidate data for the pressure-flow window. , , Indicates wind-side pressure drop. Indicates the pressure drop on the smoke side. Indicates wind-side flow rate. This indicates the flue gas flow rate.

[0119] Extract the hot and cold end temperatures from the synchronous measurement point segments corresponding to the candidate data of the pressure-flow window. For the first... Within a time window, the cold end inlet temperature, cold end outlet temperature, hot end inlet temperature, and hot end outlet temperature are read from the synchronous measurement point segment, and the arithmetic mean of the effective sampled values ​​within the time window is used as the corresponding temperature field value. The load measurement point values ​​within the time window are represented as follows: , Indicates the synchronous sampling index The load value under load. Load variation range. For the first The difference between the maximum and minimum load values ​​within a time window. Duration of load variation. For the first The duration from when the deviation of the load value from the initial load value within a time window reaches a preset load change threshold until the load value meets the preset load stabilization threshold again; the duration of load change when the load value within the time window does not reach the preset load change threshold. The duration of the load change is zero; when the load value fails to meet the preset load stabilization threshold before the end of the time window, the duration of the load change is zero. The duration from the time the preset load change threshold is reached until the end of the time window is taken.

[0120] For the magnitude of load change Duration of load change Standardize the dimensions. Define the load variation range. According to the preset load range benchmark Convert to amplitude normalization value, and convert the duration of load variation. Based on the preset duration Converted to a duration normalized value; using a preset load amplitude weight. and preset duration weight The load response is obtained by weighted summing of the normalized amplitude value and the normalized duration value. , and Preset load response weights. This involves weighting the cold end inlet temperature, cold end outlet temperature, hot end inlet temperature, hot end outlet temperature, and load response. Add candidate data for the pressure flow window to form temperature load constraint window data. .

[0121] The temperature load constraint window data, including air and flue gas side pressure drop, flow rate, cold and hot end temperatures, and load response, are arranged according to the same time window to form the operating window data. Each running window's data is a nine-dimensional vector, represented as... , Indicates wind-side pressure drop. Indicates the pressure drop on the smoke side. Indicates wind-side flow rate. Indicates flue gas flow rate, Indicates the cold end inlet temperature. Indicates the cold end outlet temperature. Indicates the hot end inlet temperature. Indicates the hot end outlet temperature. This represents the load response. The running window data is indexed according to the time window. Maintain the order and receive fixed-dimensional input as the feature extraction unit of the DRL and the weight learner of the DRL.

[0122] In this embodiment, step S3 specifically includes: Run window data is represented as , This represents the time window index that indicates the running window data. This indicates the number of time windows that run the window data. This represents a nine-dimensional real vector. The parameters are arranged in the following order: wind-side pressure drop, smoke-side pressure drop, wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, hot-end outlet temperature, and load response. The baseline data for the same load is expressed as follows: , The index represents the time window index of the load baseline data. This indicates the number of time windows for the same load baseline data. This represents a nine-dimensional real vector. The baseline data for the same load uses the same field order as the data in the running window.

[0123] Based on the load response values ​​in the running window data, match benchmark segments within the same load response range in the same load benchmark data. For each running window data... ,Will The load response and the baseline data at the same load The load response values ​​are compared, and the difference between the load response values ​​does not exceed the preset load response matching threshold. At that time, the corresponding load baseline data will be used. Candidate records within the same load response range are selected. Candidate records are filtered based on the continuity of the time window index of the same load baseline data, retaining continuous baseline segments with the same time window length, measuring point diameter, and aligned time window boundaries as the operating window data. These continuous baseline segments are then truncated and aligned according to the time window index of the operating window data to form the same load comparison window data. , Represents a nine-dimensional real vector. Indicates and A reference window record aligned with the load response, time window, and measuring point diameter.

[0124] Based on the air and flue gas pressure drop and flow rate in the same load comparison window data, calculate the pressure-flow offset of the operating window data relative to the same load baseline data. For each time window... Both the wind-side flow rate and the flue gas flow rate meet the preset minimum effective flow rate threshold. Under these conditions, the offsets in the correspondence between wind-side pressure drop and wind-side flow rate, and the offsets in the correspondence between smoke-side pressure drop and flue gas flow rate are determined separately. The offset in the correspondence between wind-side pressure drop and wind-side flow rate is determined by comparing the correspondence between wind-side pressure drop and wind-side flow rate in the operating window data with the correspondence between wind-side pressure drop and wind-side flow rate in the same load control window data. The offsets in the correspondence between wind-side pressure drop and flue gas flow rate are determined by comparing the correspondence between smoke-side pressure drop and flue gas flow rate in the operating window data with the correspondence between smoke-side pressure drop and flue gas flow rate in the same load control window data. The offsets in the wind-side pressure-flow correspondence and the offsets in the smoke-side pressure-flow correspondence are then weighted according to a preset wind-side composite weight. Combined with preset smoke side weights A pressure-flow coupling offset is generated by performing in-window synthesis. The pressure-flow coupling offset is used to characterize the deviation of the relationship between air and flue gas pressure drop and flow rate relative to the data in the same load control window within the same time window.

[0125] Based on the load response data in the same load comparison window, the pressure-flow coupling offset is subtracted from the load response. A preset flow resistance load subtraction factor is used. It is calibrated by the ratio of the change in pressure-flow correspondence to the change in load response within adjacent load response intervals in the air preheater's normal operation data. For each time window... The load response difference is determined by comparing the load response in the operation window data with the load response in the comparison window data, and a preset flow resistance load deduction factor is used. The load response difference is mapped to a load response compensation amount. The load response compensation amount is deducted from the pressure-flow coupling offset to prevent the pressure drop changes on the flue gas side caused by load fluctuations from being included in the resistance deviation calculation, thus forming the flow resistance deviation.

[0126] Based on the cold and hot end temperatures in the same load comparison window data, the time lag of the cold and hot end temperature responses relative to the same load baseline data is calculated. The cold end inlet temperature, cold end outlet temperature, hot end inlet temperature, and hot end outlet temperature in the operating window data are used as the cold and hot end temperature responses; the cold end inlet temperature, cold end outlet temperature, hot end inlet temperature, and hot end outlet temperature in the same load comparison window data are used as the baseline cold and hot end temperature responses. Within a preset lag search range, the cold and hot end temperature responses of the operating window data are matched with the cold and hot end temperature responses of the same load comparison window data at different time offsets, and the time offset with the smallest difference in cold and hot end temperature responses is selected as the time lag index. The time lag index is multiplied by the time interval between the centers of adjacent time windows to form the temperature response lag.

[0127] The effective resistance deviation window is determined based on the flow resistance deviation. The flow resistance deviation is then compared to a preset flow resistance deviation threshold. The comparison is performed when the absolute value of the flow resistance deviation is not less than the preset flow resistance deviation threshold. Furthermore, when the load comparison window data and the running window data maintain a time window correspondence, the time window... The calibration is set as the effective window for resistance deviation. A window consistency constraint is applied to the temperature response hysteresis, retaining the temperature response hysteresis only within the same time window corresponding to the effective window for resistance deviation. A preset heat transfer load deduction factor is used. The load response difference is determined by calibrating the ratio of the change in temperature response lag to the change in load response within adjacent load response intervals in the air preheater's normal operation data. A preset heat exchange load deduction factor is then used to determine the load response difference between the load response data in the operating window and the load response data in the same load comparison window. The heat transfer hysteresis bias is formed by subtracting the hysteresis component affected by load disturbance from the temperature response hysteresis.

[0128] The direction of resistance deviation is determined based on the flow resistance deviation. The flow resistance deviation exceeds a preset flow resistance direction threshold. When the resistance deviation direction indicates the direction of increased resistance; if the flow resistance deviation is less than the preset flow resistance direction threshold... When the resistance deviation is the opposite value, the direction of resistance deviation indicates the direction of resistance reduction; when the flow resistance deviation does not meet the two direction determination conditions, the direction of resistance deviation indicates that the resistance direction is undetermined. The heat transfer hysteresis direction is determined based on the heat transfer hysteresis deviation. The heat transfer hysteresis deviation is greater than the preset hysteresis direction threshold. When the heat transfer hysteresis direction indicates the direction of increasing hysteresis, the heat transfer hysteresis deviation is less than the preset hysteresis direction threshold. When the value is the opposite of the value, the direction of heat transfer lag indicates the direction of decreasing lag; when the heat transfer lag deviation does not meet the two direction determination conditions, the direction of heat transfer lag indicates the direction of undetermined lag.

[0129] Within the same time window of the running window data, construct the correlation constraint between the resistance deviation direction and the heat transfer hysteresis direction. Match the resistance deviation direction and the heat transfer hysteresis direction, and compare the matching result with a preset directional relationship. The preset directional relationship is used to limit the directional combination of resistance deviation and heat transfer hysteresis when the directional channel is used in reverse. The preset directional relationship is pre-calibrated using the corrugated sheet design flow direction and the directional response boundary in the air preheater normal operation data. The preset directional relationship is saved as a set of directional combinations. directional combination set The elements consist of the direction of increasing resistance, the direction of decreasing resistance, the direction of increasing hysteresis, and the direction of decreasing hysteresis. When the matching result satisfies the preset directional relationship, the associated offset constraint is a retention constraint; when the matching result does not satisfy the preset directional relationship, the associated offset constraint is a suppression constraint.

[0130] At the location where the flow resistance deviation and heat transfer hysteresis deviation are combined to form the flow resistance-heat transfer hysteresis correlation, the pressure-flow corresponding offset, load response deduction, time hysteresis, window consistency constraint, and correlation offset constraint are uniformly written into the same calculation rule using the following formula:

[0131] In the formula, Indicates the first The flow resistance-heat transfer hysteresis correlation for each time window; The time window index represents the data of the running window; Indicates the first The associated offset constraint for each time window; Indicates the first Flow resistance deviation within a time window; Indicates the first Heat transfer hysteresis deviation within a time window; This represents the preset suppression coefficient. It is a dimensionless constant that is not less than zero and less than one; This represents a preset directional relationship indicator function. The directional combination formed by the direction of resistance deviation and the direction of heat transfer hysteresis belongs to the set of directional combinations. When the value is one, the direction combination formed by the direction of resistance deviation and the direction of heat transfer hysteresis does not belong to the set of direction combinations. The value is zero at that time. This represents the set of direction combinations corresponding to a preset directional relationship; This represents the drag direction function, which is based on a preset flow resistance direction threshold. The direction in which output resistance increases, the direction in which resistance decreases, or the direction in which resistance is undetermined; This represents the hysteresis direction function, which is based on a preset hysteresis direction threshold. The output hysteresis can be in the direction of increasing hysteresis, decreasing hysteresis, or an undetermined hysteresis direction. Indicates the preset flow resistance fusion weight; This indicates the preset lag fusion weight; This indicates the preset flow resistance scale, and the dimension of the preset flow resistance scale is consistent with the dimension of the resistance formed by dividing the pressure drop by the flow rate. This indicates the preset lag scale, and the dimension of the preset lag scale is time. Indicates the preset wind-side composite weights; This indicates the preset smoke-side synthesis weights; This represents the wind-side pressure drop in the running window data; This represents the smoke-side pressure drop in the running window data; This represents the wind-side flow rate in the running window data; This represents the flue gas flow rate in the running window data; This indicates the load response amount in the running window data; This represents the wind-side pressure drop in the same load comparison window data; This represents the smoke-side pressure drop in the data from the same load comparison window; This represents the wind-side flow rate in the same load comparison window data; This represents the flue gas flow rate in the same load comparison window data; This represents the load response in the same load comparison window data; This indicates the preset minimum effective traffic threshold; This indicates that the larger of the two input values ​​is selected. This represents the preset flow resistance load deduction factor, the dimension of which is the dimension of resistance divided by the dimension of load response. This indicates that the resistance deviates from the effective window indicator function. Not less than the preset flow resistance deviation threshold The value is 1 at time. Less than the preset flow resistance deviation threshold The value is zero at that time. This represents the time interval between the centers of adjacent time windows; This indicates the selection of candidate indexes with the smallest time lag that minimizes the temperature response difference between the hot and cold ends within the brackets. Represents an integer-type time-lag candidate index; Indicates the first The set of effective lag candidate indices corresponding to each time window. By satisfying The candidate indexes are composed of time lags that fall within the effective time window index range of the data in the same load comparison window. This indicates the index of the hot and cold end temperature field. Corresponding to the cold end inlet temperature, Corresponding to the cold end outlet temperature, Corresponding to the hot end inlet temperature, Corresponding hot end outlet temperature; Indicates the first data in the running window One temperature field value; This indicates the candidate index with time lag in the same load comparison window data. The corresponding number One temperature field value; Indicates the first Each temperature field corresponds to a preset temperature scale, and the unit of the preset temperature scale is temperature. This represents the preset heat exchange load deduction factor, the dimension of which is time divided by the dimension of load response.

[0132] Through the calculation of the flow resistance-heat transfer hysteresis correlation, the correlation offset constraint is incorporated into the fusion process in the form of a directional relationship indicator function and a preset suppression coefficient. When the matching result satisfies the preset directional relationship, the normalized flow resistance deviation and the normalized heat transfer hysteresis deviation are jointly incorporated into the flow resistance-heat transfer hysteresis correlation. When the matching result does not satisfy the preset directional relationship, the effects of individual airflow / flue gas side pressure drop deviation and individual hot / cold end temperature deviation on the formation of the flow resistance-heat transfer hysteresis correlation are limited by the preset suppression coefficient. Flow resistance-heat transfer hysteresis correlation Arranged according to the time window index of the running window data, each It is a one-dimensional intermediate object used to characterize the correlation offset between flow resistance deviation and heat transfer hysteresis deviation within the same operating window data, and serves as a one-dimensional scene physical input received by the weight learner of DRL.

[0133] In this embodiment, step S4 specifically includes: Run window data is represented as , This represents the time window index that indicates the running window data. This indicates the number of time windows that run the window data. Indicates the first A nine-dimensional real vector corresponding to each time window. , Indicates wind-side pressure drop. Indicates the pressure drop on the smoke side. Indicates wind-side flow rate. Indicates flue gas flow rate, Indicates the cold end inlet temperature. Indicates the cold end outlet temperature. Indicates the hot end inlet temperature. Indicates the hot end outlet temperature. This indicates the load response.

[0134] The feature extraction unit and weight learner of the DRL are pre-configured networks. The model parameters of the DRL's feature extraction unit are read from the DRL's model parameter storage area, represented as... , This includes the weights and biases from the input layer to the first fully connected layer, from the first fully connected layer to the second fully connected layer, and from the second fully connected layer to the latent output layer in the feature extraction unit of the DRL. The model parameters of the DRL's weight learner are read from the DRL's model parameter storage area and are represented as follows: , The weights and biases from the input layer to the hidden layer and from the hidden layer to the activation and output layers in the weighted learner containing DRL. Indexing for each time window. Run window data A forward pass is completed as an independent input, and the recursive state is not passed between different time windows.

[0135] The running window data is input into the feature extraction unit of DRL according to the same time window. The feature extraction unit of DRL uses mapping. , Indicates the model parameters A defined forward mapping. The feature extraction unit of DRL consists of an input layer, a first fully connected layer, a second fully connected layer, and a latent output layer. The input layer contains 9 neurons, each receiving a specific input signal. to The first fully connected layer contains 32 neurons. Each neuron in the first fully connected layer receives all the outputs of the 9 neurons in the input layer and generates the first layer features through weights, biases, and corrected linear nonlinear units. The second fully connected layer contains 64 neurons. Each neuron in the second fully connected layer receives all the outputs of the 32 neurons in the first fully connected layer and generates the second layer features through weights, biases, and corrected linear nonlinear units. The latent output layer contains 64 neurons. The latent output layer receives the outputs of the 64 neurons in the second fully connected layer and generates a 64-dimensional latent representation.

[0136] Index for each time window The feature extraction unit of DRL is used for... The pressure drop and flow rate on the smoke side, the temperature at the hot and cold ends, and the load response are jointly mapped to form windowed potential mapping data. . , , , Indicates the potential output layer number 1 The neuron in the first Output of each time window The joint latent mapping simultaneously receives wind-side pressure drop, flue gas pressure drop, wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, hot-end outlet temperature, and load response at the same input layer, enabling the pressure drop, flow rate, temperature, and load fields within the same time window to participate in the latent representation generation in the fully connected path.

[0137] Construct channel state constraints based on the window's potential mapping data. Represent the input field index as... , Input field index and The field order is consistent. The index of the first fully connected layer neuron is represented as... , The index of the neurons in the second fully connected layer is represented as... , The input layer's first... The first fully connected layer contains the first neuron. The connection weights of each neuron are represented as follows: The first fully connected layer The neurons are connected to the second fully connected layer. The connection weights of each neuron are represented as follows: The second fully connected layer The first neuron to the potential output layer The connection weights of each neuron are represented as follows: For the first The potential output neuron and the first One input field, enumerate all derived from the input field After passing through the neurons of the first fully connected layer Second fully connected layer neurons Reaching the potential output neuron The fully connected paths are identified; the absolute values ​​of the three connection weights on each path are multiplied as the path contribution value; the field contribution value is obtained by summing all path contribution values. The contribution values ​​of the nine fields corresponding to the same potential output neuron are normalized to obtain the normalized field contribution ratio. When the sum of the contribution values ​​of the nine fields is zero, the contribution percentage of each of the nine normalized fields is set to zero.

[0138] The contribution ratio of normalized fields is used to determine whether the potential mapping components in the potential mapping data of the window are dominated by single-item wind and smoke side pressure drop deviations or single-item cold and hot end temperature deviations. When the first Among the potential dimensions, the sum of the contributions of the normalized fields corresponding to wind side pressure drop and smoke side pressure drop exceeds the preset pressure drop dominance threshold. Furthermore, the contribution percentages of the normalized fields corresponding to wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, hot-end outlet temperature, and load response all failed to reach the preset joint contribution threshold for the channels. At that time, the first The first potential mapping component is identified as the potential mapping component formed by the deviation of the single-phase wind and smoke side pressure drop. When the first... Among the potential dimensions, the sum of the contributions of the normalized fields corresponding to the cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, and hot-end outlet temperature exceeds the preset temperature dominance threshold. Furthermore, the contribution percentages of the normalized fields corresponding to wind-side pressure drop, smoke-side pressure drop, wind-side flow rate, flue gas flow rate, and load response all failed to reach the preset joint contribution threshold for the channels. At that time, the first Each potential mapping component is calibrated as a potential mapping component formed by a single deviation in the temperature of the hot and cold ends.

[0139] For potential mapping components identified as being caused by single-phase airflow side pressure drop deviations and potential mapping components identified as being caused by single-phase hot and cold end temperature deviations, a preset potential suppression coefficient is used. For the corresponding Amplitude suppression is performed. It is a dimensionless constant that is not less than zero and less than one. Unlabeled latent mapping components retain their amplitudes. After channel state constraint processing, the window latent mapping data forms channel state latent data. , , , Indicates the first The first time window Channel state potential components. Channel state potential data are used to characterize channel state deviations of thermal storage elements and suppress the dominant components of unidirectional flue gas side pressure drop deviations or unidirectional hot and cold end temperature deviations.

[0140] The running window data is input into the DRL's weight learner. The DRL's weight learner uses a mapping... , Indicates the model parameters A defined forward mapping. The weight learner in DRL consists of an input layer, hidden layers, and an activation output layer. The input layer contains 9 neurons, and the input layer receives... The DRL has 9 fields. The hidden layer contains 32 neurons, each receiving all the outputs of the 9 neurons in the input layer and generating a hidden representation through weights, biases, and corrected linear nonlinear units. The activated output layer contains 16 neurons, each corresponding to one of the 16 fixed decomposition directions within the DRL.

[0141] The distribution relationship of air and flue gas side pressure drop, flow rate, cold and hot end temperatures, and load response in the same window data is as follows: The nine fields are represented by a fixed order within the same time window, the field binding relationships between wind-side pressure drop and wind-side flow rate, the field binding relationship between smoke-side pressure drop and flue gas flow rate, and the field binding relationship between cold and hot end temperatures and load response. The DRL weight learner generates candidate activation values ​​for decomposition directions based on the distribution relationship within the same window. . , , Indicates activation of the output layer. The neuron in the first Candidate activation values ​​formed by time windows The candidate activation values ​​for decomposition directions have not yet been properly assigned to the potential decomposition directions of the channel state potential data.

[0142] Within the weight learner of the DRL, the candidate activations of the decomposition directions are constrained to belong to the decomposition direction. The 16 fixed decomposition directions within the DRL are represented as follows: , Indicates the first There are 64 fixed decomposition directions, with 16 fixed decomposition directions read from the fixed decomposition direction table of the DRL. This is based on the activation output layer... The first neuron and the second A fixed decomposition direction The correspondence will be used to decompose the candidate activation values ​​in the direction. The first binding to the channel state potential data One potential decomposition direction. After completing the decomposition direction assignment constraints, the initial decomposition activation quantity is generated. , , , Indicates the first The corresponding time window is the The initial activation values ​​for each fixed decomposition direction are determined. The channel state latent data maintains 64 dimensions, and the initial decomposition activation values ​​maintain 16 dimensions. The 64-dimensional dimensions of the channel state latent data are consistent with the 64-dimensional dimensions of each fixed decomposition direction.

[0143] In this embodiment, step S5 specifically includes: The flow resistance-heat transfer hysteresis correlation is expressed as follows: ,in, This represents the time window index that indicates the running window data. This indicates the number of time windows used to run the window data. Indicates the first The one-dimensional flow resistance-heat transfer hysteresis correlation corresponding to each time window. The initial decomposition activation quantity is expressed as... , , , Indicates the first The corresponding time window is the Initial activation values ​​for a fixed decomposition direction. .

[0144] The DRL's weight learner consists of an input layer, hidden layers, and an activation output layer. The input layer contains 9 neurons, receiving data from the operating window including wind-side pressure drop, smoke-side pressure drop, wind-side flow rate, flue gas flow rate, cold-end inlet temperature, cold-end outlet temperature, hot-end inlet temperature, hot-end outlet temperature, and load response. The hidden layer contains 32 neurons, each receiving all the outputs from the 9 neurons in the input layer and generating a hidden representation through weights, biases, and corrected linear / nonlinear units. The activation output layer contains 16 neurons, corresponding to the 16 fixed decomposition directions within the DRL. The activation generation position is determined by the output positions of the 16 neurons within the activation output layer of the DRL's weight learner, before the normalized exponential unit. The DRL's weight learner forms the hidden layer output when generating the initial decomposition activation values. , , , Indicates the first The hidden layer in the first time window The output of each neuron Activating the output layer is based on Generate initial decomposition activation quantity, and receive one-dimensional flow resistance-heat transfer hysteresis correlation quantity within the same activation output layer. .

[0145] Based on the activation generation position, the first The unnormalized outputs of the activated output neurons, without the embedding of the flow resistance-heat transfer hysteresis correlation, were determined as the intermediate quantities to be embedded for activation. When the initial decomposition activation value is stored as the output before normalization of the activation output layer, and Take the same value; when the initial decomposed activation value is stored as the output after the activation output layer is bound, read the same value from the activation output layer cache of the DRL weight learner. The unnormalized outputs with the same fixed decomposition direction index are used as The 16 intermediate activation variables to be embedded are arranged according to a fixed decomposition direction index. , The intermediate quantity to be embedded and the initial decomposed activation quantity correspond in the decomposition direction index. The intermediate quantity to be embedded and activated has not yet received the scenario constraints formed by the flow resistance-heat transfer hysteresis correlation quantity.

[0146] The 16 fixed decomposition directions within the DRL are represented as follows: , Indicates the first A fixed 64-dimensional decomposition direction , Indicates the first The fixed decomposition direction in the channel state potential data of the [number]th [section / direction]. Fixed directional components in each potential dimension The DRL's fixed decomposition direction table also stores the direction assignment label for each fixed decomposition direction. The direction assignment label is written into the fixed decomposition direction table during the DRL model configuration phase: the fixed decomposition direction corresponding to the training window defined by the flow resistance-heat transfer hysteresis correlation is labeled as a channel direction anomaly-related decomposition direction; the fixed decomposition direction corresponding to the training window dominated by a single air-to-smoke side pressure drop deviation is labeled as a normal drag anomaly direction; and the fixed decomposition direction corresponding to the training window dominated by a single hot-cold end temperature deviation is labeled as a normal heat transfer anomaly direction. Based on the direction assignment label, the fixed decomposition direction index is divided into a set of channel direction anomaly-related decomposition directions. Common resistance anomaly direction set and the set of normal heat transfer anomalies After reading the direction attribution flag, each fixed decomposition direction index is converted into three mutually exclusive zero-one values, representing the decomposition direction related to channel direction anomalies, the direction of ordinary resistance anomalies, and the direction of ordinary heat transfer anomalies, respectively.

[0147] When embedding the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation, the mapping from the correlation offset intensity to the correlation embedding constraint is used as the key calculation location. This calculation location is before the normalization processing of the scene decomposition activation candidate. The correlation offset intensity, correlation embedding constraint, scene constraint activation intermediate value, correlation embedding activation, and scene decomposition activation candidate are updated under the same rule using the following formula:

[0148] In the formula, Indicates the first The intensity of the associated offset effect within a time window; Indicates the first The flow resistance-heat transfer hysteresis correlation for each time window; The preset correlation scale has the same dimensions as the flow resistance-heat transfer hysteresis correlation. The preset correlation scale is determined by the median or engineering calibration value of the absolute value of the flow resistance-heat transfer hysteresis correlation formed by the normal operation data of the air preheater within the stable window. This indicates the preset upper limit of the effect intensity, which is a dimensionless positive number. This represents the truncation function, when... Greater than Time to take ,when Less than Time to take ,when lie in Take the time inside ; Indicates the first Effective labeling of association strength within a time window; This indicates a conditional indicator function; the value is 1 when the condition within the parentheses is true, and 0 when the condition within the parentheses is false. Represents the absolute value of the intensity of the associated offset effect; This indicates the preset correlation strength threshold, which is determined by the absolute value quantile boundary of the correlation offset effect intensity corresponding to the normal operation data of the air preheater; Indicates the first The first time window The associated embedding constraint quantity corresponding to each fixed decomposition direction; Indicates a fixed decomposition direction index; Indicates the first Preset association embedding direction coefficients corresponding to a fixed decomposition direction. The preset association embedding direction coefficients are dimensionless values ​​and are stored in the activation output layer parameter table of the weight learner of DRL. Indicates the first A set of fixed decomposition directions related to channel direction anomalies. The value is 1 if the condition is met, and 0 otherwise. Indicates the first The fixed decomposition direction belongs to the set of ordinary resistance anomaly directions. The value is 1 if the condition is met, and 0 otherwise. Indicates the first The fixed decomposition direction belongs to the set of ordinary heat transfer anomaly directions. The value is 1 if the condition is met, and 0 otherwise. This represents the preset suppression embedding coefficient for the abnormal direction of ordinary resistance. The preset suppression embedding coefficient for the abnormal direction of ordinary resistance is a dimensionless value. This represents the preset suppression embedding coefficient for the direction of normal heat transfer anomalies. The preset suppression embedding coefficient for the direction of normal heat transfer anomalies is a dimensionless value. Indicates the first The first time window The intermediate activation quantities to be embedded corresponding to a fixed decomposition direction; Indicates the first The first time window Activate intermediate quantities of scene constraints corresponding to a fixed decomposition direction; Indicates the first The first time window The associated embedding activation quantity corresponding to a fixed decomposition direction; Indicates the first The first time window Candidates for scene decomposition activation quantities corresponding to a fixed decomposition direction; This represents the preset activation inhibition coefficient. It is a dimensionless constant that is not less than zero and less than one.

[0149] Based on the above calculations, the flow resistance-heat transfer hysteresis correlation is first converted into a dimensionless correlation offset intensity using a preset correlation scale, and then the effective correlation intensity marker is determined by a preset correlation intensity threshold. When the effective correlation intensity marker is 1, the decomposition direction related to channel direction anomalies receives the embedding constraint of the flow resistance-heat transfer hysteresis correlation through a preset correlation embedding direction coefficient, while the ordinary resistance anomaly direction and the ordinary heat transfer anomaly direction receive the suppression constraint through a preset suppression embedding coefficient. When the effective correlation intensity marker is 0, the correlation embedding constraint does not change the intermediate quantity to be embedded activation. The scene constraint activation intermediate quantity is formed by the intermediate quantity to be embedded activation and the correlation embedding constraint. The correlation embedding activation quantity is formed by writing the scene constraint activation intermediate quantity back to the activation output layer. The scene decomposition activation quantity candidates complete the activation affiliation consistency adjustment in the same rule, retaining the activation components that satisfy the flow resistance-heat transfer hysteresis correlation constraint relationship in the decomposition direction related to channel direction anomalies, and suppressing the activation components of the ordinary resistance anomaly direction and the ordinary heat transfer anomaly direction that do not satisfy the flow resistance-heat transfer hysteresis correlation constraint relationship.

[0150] Arrange the scene decomposition activation candidates according to a fixed decomposition direction index. , The candidate activation values ​​for scene decomposition are normalized. The normalization exponent unit is then used to normalize the values. The 16 components are each subjected to an exponential transformation, and the sum of the 16 exponential transformation values ​​is used as the normalized denominator to obtain 16 non-negative components that sum to one. Following the index order of the 16 fixed decomposition directions within the DRL, the 16 normalized components are arranged as the scene decomposition activation values. , , The first activation quantity of scene decomposition The component corresponds to the first A fixed decomposition direction is used when the 64-dimensional channel state potential data enters the decomposition and reconstruction stage. The 64-dimensional fixed decomposition direction remains consistent.

[0151] Generate potential decomposition coefficients based on scene decomposition activation volume. , , The scene is decomposed into activation quantities. The i-th component is written into the latent decomposition coefficients. Each component corresponds to a proportion of 16 fixed decomposition directions for the potential decomposition coefficients. The potential decomposition coefficients inherit the embedded constraint of the flow resistance-heat transfer hysteresis correlation on the initial decomposition activation amount, and maintain the correspondence between the channel state potential data and the 16 fixed decomposition directions in 64 dimensions.

[0152] In this embodiment, step S6 specifically includes: Channel state potential data is represented as , This represents the time window index that indicates the running window data. This indicates the number of time windows that run the window data. , , Indicates the first The first time window Potential components of each channel state The potential decomposition coefficients are expressed as... , , , Indicates the first The first time window The allocation ratio corresponding to each fixed decomposition direction. The potential decomposition coefficients are generated from the scene decomposition activations and already incorporate the embedded constraints of the flow resistance-heat transfer hysteresis correlation on the initial decomposition activations.

[0153] The DRL decomposition and reconstruction stage receives 64-dimensional channel state latent data and 16-dimensional latent decomposition coefficients without adding new neural network layers. The DRL decomposition and reconstruction stage reads 16 fixed decomposition directions from within the DRL. , , , Indicates the first The fixed decomposition direction is in the... Fixed directional components in each potential dimension. Fixed decomposition directions are stored in the DRL's fixed decomposition direction table as dimensionless unit norms. The DRL's fixed decomposition direction table stores the direction assignment label for each fixed decomposition direction. The direction assignment label includes decomposition directions related to channel directional anomalies, general drag anomalies, and general heat transfer anomalies. Decomposition directions related to channel directional anomalies correspond to directional channel states; general drag anomalies correspond to potential components dominated by unidirectional airflow-smoke side pressure drop deviations; and general heat transfer anomalies correspond to potential components dominated by unidirectional hot and cold end temperature deviations. The direction assignment label is written into the DRL's fixed decomposition direction table during the DRL model configuration phase and remains unchanged during the DRL decomposition and reconstruction phase.

[0154] Based on the direction attribution label, the fixed decomposition direction index is divided into a set of decomposition directions that are related to channel direction anomalies. Common resistance anomaly direction set and the set of normal heat transfer anomalies . Includes fixed decomposition direction indices that are allowed as sources of directional channel state reconstruction; Includes a fixed decomposition direction index dominated by the deviation of the unidirectional wind and smoke side pressure drop; It contains a fixed decomposition direction index dominated by a single hot-cold end temperature deviation. All elements of the three sets are fixed decomposition direction indices. .

[0155] The proportion of potential channel state data entering the DRL decomposition and reconstruction process is determined based on the potential decomposition coefficients. Indexing is performed for each time window. ,Will Read as the first The proportion of potential data for receiving channel status in a fixed decomposition direction. According to to The fixed decomposition direction index order is used to arrange the 16 decomposition directions according to their proportions. , , , and Take the same value. The allocation ratio of the decomposition direction serves as the basis for allocating potential channel state data into the decomposition and reconstruction process.

[0156] Decomposition direction matching is performed on the latent components in the channel state potential data according to the decomposition direction allocation ratio. For the first... A time window will be used to store the potential data of the 64-dimensional channel states. The orientation consistency is matched one by one with each of the 16 fixed 64-dimensional decomposition directions. The preset minimum allocation ratio threshold is expressed as follows: The preset effective direction component threshold is expressed as: The preset threshold for the number of identical symbols is expressed as: The preset amplitude distribution threshold is expressed as For the first With a fixed decomposition direction, first select a fixed direction component whose absolute value is not less than... The latent dimensions are used as effective latent dimensions. If Less than Then the first A fixed decomposition direction does not receive latent components from the channel state latent data. If Not less than Then, a dimension-by-dimensional comparison is performed within the effective potential dimensions. and The number of latent dimensions with the same sign is taken as the sign consistency number; at the same time, the number of effective latent dimensions is... Normalize by the sum of the absolute values ​​of all effective latent dimensions, and then normalize the values ​​within the effective latent dimensions. Normalize the sum of the absolute values ​​of all effective latent dimensions, then sum the absolute differences between the two normalized results dimension by dimension to obtain the amplitude distribution bias. When the number of sign-consistent values ​​is not less than... And the amplitude distribution deviation is no greater than At that time, the corresponding potential components will be matched to the first... A fixed decomposition direction. The matching results are arranged into direction-based matching data. , , , 1 indicates the first Potential components in the potential data of the receiving channel state in a fixed decomposition direction 0 indicates the first A fixed decomposition direction does not receive potential components from the potential data of the channel state.

[0157] Channel direction reconstruction constraints are constructed based on direction attribution matching data. Indexing is performed for each time window. and index of each fixed decomposition direction Read the first A fixed decomposition direction is assigned to a directional marker. When =1 and belong When this happens, the corresponding reconstruction channel is designated as the direction reconstruction channel. =1 and belong or At that time, the corresponding potential components are labeled as non-directional reconstruction components. The channel direction reconstruction constraint is formed from the labeling results. , , , Indicates the first The fixed decomposition direction is in the... Whether it is permissible to use it as a source for directional channel state reconstruction within a given time window. A value of 1 indicates an enabled state. A value of 0 indicates a restricted state. The potential components corresponding to the directions of normal drag anomalies and normal heat transfer anomalies are obtained through... The restricted state enters the non-directional reconstruction component and is not used as the main basis for determining directional misalignment.

[0158] At the locations where decomposition projection, reconstruction contribution constraint, combined reconstruction, and correlation consistency reconstruction are performed, the potential dimensional components of the channel-direction anomaly decomposition data are directly formed using the following formula:

[0159] In the formula, Indicates the first The first time window Decompose data components with channel orientation anomalies in each potential dimension; Indicates a potential dimension index; The time window index represents the data of the running window; Indicates a fixed decomposition direction index; Indicates the potential dimension index participating in the projection calculation; Indicates the first The fixed decomposition direction is in the... Channel direction reconstruction constraint within a time window; Indicates the first The fixed decomposition direction is in the... Potential decomposition coefficients within a time window; Indicates the first The first time window Potential components of each channel state; Indicates the first The first time window Potential components of each channel state; Indicates the first The fixed decomposition direction is in the... Fixed directional components in each potential dimension; Indicates the first The fixed decomposition direction is in the... Fixed directional components in each potential dimension.

[0160] In the above calculations, the inner summation incorporates the potential data of the channel state. The 64 potential components and the first A fixed decomposition direction The 64 fixed-direction components are multiplied and summed one by one to obtain the first... The projected intensity along each fixed decomposition direction. Multiply the projected intensity by... The 64 fixed-direction components form the first Decomposed projection data for each channel , .Will Each latent dimensional component is simultaneously affected and Limitations, forming the first Contribution data for channel reconstruction , .when When it is 0, the first The reconstruction contribution of a fixed decomposition direction is restricted to non-directional reconstruction components.

[0161] The reconstruction contribution data of the 16 channels were merged item by item according to the same potential dimension to form the channel state reconstruction data. , , Channel state potential data Reconstructing data with channel status Calculate the difference item by item according to the potential dimensions to form potential reconstructed difference data. , , The summation term outside the square brackets in the formula represents the retention ratio of the potential decomposition coefficients in the directional channel state decomposition direction. This retention ratio is jointly determined by the potential decomposition coefficients and the channel direction reconstruction constraint. After being constrained by this retention ratio, the potential reconstruction difference data forms the channel direction abnormal decomposition data, ensuring that the difference components retain the correlation offset between flow resistance deviation and heat transfer hysteresis deviation within the same operating window data.

[0162] according to to Calculate in sequence , forming the first Channel direction anomaly decomposition data corresponding to each time window , , .according to to The potential data of channel status and the potential decomposition coefficients are processed sequentially to form 64-dimensional channel direction anomaly decomposition data corresponding to the time window index of the running window data.

[0163] In this embodiment, step S7 specifically includes: Channel direction anomaly decomposition data is represented as , This represents the time window index that indicates the running window data. This indicates the number of time windows used to run the window data. It is a positive integer. , , Indicates the first The first time window Anomalous decomposition components of each potential dimension The channel direction anomaly decomposition data is formed by decomposing and reconstructing the potential decomposition coefficients and the potential channel state data. The channel direction anomaly decomposition data retains the correlation offset between flow resistance deviation and heat transfer hysteresis deviation within the same operating window.

[0164] The orientation misalignment determination stage receives channel orientation anomaly decomposition data. This stage comprises a fully connected layer with 16 neurons and an output layer with 3 neurons. The 16-neuron fully connected layer is used for directional channel response aggregation. The output layer with 3 neurons corresponds to the presence of orientation misalignment, the absence of orientation misalignment, and orientation misalignment pending verification, respectively. The orientation misalignment determination stage indexes each time window. Each forward computation is performed independently, without propagating the recursive state between different time windows. The input to the direction misalignment determination stage is a 64-dimensional vector. The orientation misalignment determination process generates 16-dimensional orientation misalignment determination input data and 3-dimensional orientation misalignment response data within a single time window.

[0165] The weights of the 16-neuron fully connected layer are represented as follows: , Indicates the first The abnormal decomposition component enters the first The connection weights of each aggregate neuron The bias of the fully connected layer with 16 neurons is represented as follows: , Indicates the first Bias of each converging neuron. and Read from the model parameter storage area of ​​the orientation misalignment determination stage. The model parameter storage area of ​​the orientation misalignment determination stage is used to store the connection weights and bias values ​​that are fixed after the DRL model is configured. In this implementation, the connection weights and biases are not re-estimated in the orientation misalignment determination stage.

[0166] The preset aggregation weight threshold is expressed as: , Non-negative values All of the direction misalignment determination steps The absolute value quantile boundaries are calibrated and saved during the model configuration phase. For the ... A clustered neuron, reading an absolute value not less than... The latent dimension corresponding to the connection weight, will be on the corresponding latent dimension. Abnormal decomposition components calibrated for reverse use of directional channels. Absolute value less than The latent dimension corresponding to the connection weight is not included in the first... The directional channel response of the converging neurons converges. If the first... There are no aggregate neurons whose absolute value is not less than 1. If the connection weight is , then the ... The input contribution of the aberrant decomposition component of each aggregate neuron is set to zero, and only based on The values ​​before aggregation are formed.

[0167] When polymerizing anomalous decomposition components used in reverse of directional channels, the first Each aggregated neuron will be labeled Multiply by the corresponding And add all the multiplications and additions together. , forming the first The first set of values ​​before aggregation is determined. The corrected linear nonlinear unit sets values ​​before aggregation less than zero to zero, and leaves values ​​before aggregation not less than zero unchanged, forming the first set of values ​​before aggregation. Sixteen aggregate responses are arranged according to the aggregate neuron index, forming the input data for orientation misalignment determination. , , , Indicates the first The first time window The directional channel response aggregation results are formed by the aggregation neurons. The 16 components of the direction misalignment determination input data correspond to the aggregation results of the 16 aggregation neurons on the 64-dimensional channel direction anomaly decomposition data.

[0168] The weights of the output layer are represented as... , Indicates the first The aggregation response enters the... The connection weights of each output neuron The output layer bias is represented as... , Indicates the first Bias of each output neuron. and Read from the model parameter storage area of ​​the orientation misalignment determination stage. There is a corresponding directional misalignment. There is no directional misalignment. The corresponding direction misalignment needs to be verified. The output layer will have 16... According to Weighted and added This results in three decision response quantities. These three decision response quantities are arranged in order of output category to form the direction-misaligned response data. , , , This indicates the determination response quantity where a direction misalignment exists. This indicates the response quantity for determining that there is no orientation misalignment. This indicates the response quantity for judgments that require verification due to misalignment.

[0169] The response value indicating the presence of orientation misalignment in the orientation misalignment response data is compared with the preset orientation misalignment response threshold. The responses are compared, and those with no orientation misalignment and those with orientation misalignment pending verification are used as mutually exclusive constraints. A preset orientation misalignment response threshold is set. The quantile boundaries of the output response corresponding to the orientation misalignment category in samples with inspection and repair kernel labels are identified and saved. For the first... A time window, when Not less than ,and Greater than and At that time, the first The result of the orientation misalignment determination within each time window is marked as indicating the existence of orientation misalignment. When Greater than ,and Not less than At that time, the first The orientation misalignment determination result of the first time window is marked as no orientation misalignment. When neither the orientation misalignment marking condition nor the orientation misalignment marking condition is met, the first time window is marked as no orientation misalignment. The orientation misalignment determination results for each time window are marked as orientation misalignment pending review. Mutually exclusive constraints ensure that only one determination result is retained within the same time window: orientation misalignment exists, orientation misalignment does not exist, and orientation misalignment pending review.

[0170] The first The direction misalignment determination result for each time window is represented as follows: , The value can be either "there is a direction misalignment", "there is no direction misalignment", or "direction misalignment needs to be verified". According to... to Formed in sequence The value is expressed as the number of time windows with directional misalignment. , Through statistics The value is obtained from the time window where there is a direction misalignment. The number of time windows requiring review due to direction misalignment is expressed as... , Through statistics The value is obtained from the time window during which the orientation is misaligned and needs to be reviewed. The preset threshold for the proportion of the orientation misalignment window is expressed as follows: The preset review window percentage threshold is expressed as: , and All are dimensionless thresholds greater than zero and not greater than one, and are pre-calibrated according to the continuous window coverage ratio required for maintenance and diagnosis.

[0171] when and The ratio is not less than When the corrugated sheet assembly direction is misaligned, it is determined that there is a direction misalignment. and The ratio is less than ,and and The ratio is not less than When the corrugated sheet assembly orientation is misaligned, the state is determined to be misaligned and requires verification. When neither the overall judgment condition for misalignment nor the overall judgment condition for verification is met, the state of the corrugated sheet assembly orientation misalignment is determined to be non-misaligned.

[0172] Figure 12The graph compares the identification results before and after applying this invention, using the operating window number as the horizontal axis and the direction misalignment judgment response quantity as the vertical axis. A low threshold of 0.4 and a high threshold of 1.0 are set in the graph. A response quantity less than 0.4 is considered as not having a direction misalignment, 0.4 to 1.0 is considered as needing verification, and greater than or equal to 1.0 is considered as having a direction misalignment. On the left side, without this invention, multiple operating windows fall into the areas needing verification or where a direction misalignment exists, indicating that ordinary pressure drop anomalies easily lead to false alarms, with a false alarm rate of approximately 15%–35%. On the right side, after applying this invention, true direction misalignment windows are identified collectively, with an identification accuracy of approximately 90%–98%, and the proportion needing verification is approximately 3%–12%. It can be seen that this invention, through the joint judgment of flow resistance deviation, heat transfer hysteresis deviation, and DRL decomposition and reconstruction, can reduce false judgments caused by single anomalies.

[0173] Figure 13 This figure illustrates the changes in the contribution of anomalous decompositions before and after embedding the flow resistance-heat transfer hysteresis correlation in the DRL. The left side shows the decomposition results without embedding the flow resistance-heat transfer hysteresis correlation, where the contribution percentage for the anomalous direction of ordinary resistance is 0.25, the contribution percentage for the anomalous direction of ordinary heat transfer is 0.20, and the contribution percentage for the anomalous direction of the channel is 0.55. The right side shows the results after embedding the correlation, where the anomalous direction of ordinary resistance decreases to 0.10, the anomalous direction of ordinary heat transfer decreases to 0.08, and the anomalous direction of the channel increases to 0.82. This demonstrates that embedding the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation value makes the scene decomposition activation value and potential decomposition coefficients more likely to retain decomposition components that satisfy the directional misalignment correlation relationship and suppress ordinary single-item anomalous components, thereby improving the effectiveness of the anomalous decomposition data in the channel direction.

[0174] In summary, this invention addresses the issue of misaligned corrugated sheet assembly status verification of heat storage elements after air preheater maintenance. This misalignment leads to the reverse use of directional channels, which manifests as flow resistance deviation on the flue gas side and lag in temperature response at the hot and cold ends. This reverse use is easily confused with load fluctuations, common ash accumulation, air leakage, or general heat exchange degradation. Under conditions where only post-maintenance operational data and load-based benchmark data are available, and the field measurement point diameter and sampling sequence are constrained by the operating system, a window-based correlation characterization between flow resistance deviation and heat exchange lag deviation is established. This allows the flow resistance and heat exchange lag correlation quantity corresponding to the reverse use of directional channels to enter the decomposition direction generation process within the DRL, thus forming a closed-loop technical chain for determining the misaligned corrugated sheet assembly status, from operational data acquisition, correlation quantity construction, potential decomposition constraints to misalignment status output.

[0175] This invention directly addresses the aforementioned technical problems through a technical approach involving load comparison, correlation deviation construction, embedded decomposition activation constraints, and channel direction anomaly reconstruction. First, this invention acquires operational data and load-based benchmark data of the air preheater after maintenance, matching them across load range, measuring point diameter, sampling sequence, and time window to provide a comparable benchmark for post-maintenance performance. Subsequently, this invention extracts air-flue gas side pressure drop, flow rate, hot and cold end temperatures, and load response from the operational window data. Flow resistance deviation is generated through pressure-flow correspondence offset and load response subtraction, and the lag in hot and cold end temperature response is calculated within the effective window of resistance deviation, forming a heat transfer lag deviation. Furthermore, this invention establishes correlation constraints between the resistance deviation direction and the heat transfer lag direction within the same time window, integrating the combined operational performance caused by directional misalignment into a flow resistance-heat transfer lag correlation quantity, avoiding the direct use of single-item pressure drop deviation or single-item temperature lag as the basis for assembly direction misalignment.

[0176] Compared to existing algorithms that treat the air preheater as a whole heat exchanger for performance testing, or directly input operating measurement points into a classification model for result judgment, this invention embeds the scenario mechanism into the DRL decomposition process in its algorithm structure. First, instead of adding threshold rules after the judgment result, this invention introduces an association embedding constraint at the position where the DRL weight learner generates the initial decomposition activation value, so that the lag association value of flow resistance heat transfer affects the scenario decomposition activation value before normalization. Second, this invention generates potential decomposition coefficients through the scenario decomposition activation value, enabling directional channels to continuously transmit the corresponding association information to the decomposition and reconstruction process of the channel state potential data in reverse. Third, this invention utilizes directional attribution matching data and channel directional reconstruction constraints to restrict the abnormal directions of ordinary resistance and ordinary heat transfer to non-directional reconstruction components, ensuring that the channel directional anomaly decomposition data mainly retains the potential components related to the misalignment of the corrugated sheet assembly direction. Therefore, this invention can specifically separate the complex operational characteristics caused by directional misalignment within the range of operational data available on-site after maintenance, and output the status results of whether directional misalignment exists, does not exist, or is pending verification.

[0177] The second objective of this invention is to propose a method for identifying misalignment of the corrugated sheet assembly direction of an air preheater heat storage element based on DRL, comprising: Operation window data module: used to acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; The hysteresis correlation module is used to calculate the flow resistance deviation and heat transfer hysteresis deviation based on the operating window data and the same load baseline data, forming the flow resistance-heat transfer hysteresis correlation. Initial decomposition activation quantity module: It is used to input the running window data into the feature extraction unit of DRL to obtain the channel state potential data, and the weight learner of DRL generates the initial decomposition activation quantity based on the running window data. Potential decomposition coefficient module: used to embed the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and generate potential decomposition coefficients based on the scene decomposition activation quantity. Corrugated sheet assembly orientation misalignment status module: It is used to decompose and reconstruct the potential data of the channel state according to the potential decomposition coefficient, obtain the abnormal decomposition data of the channel orientation, determine the orientation misalignment based on the abnormal decomposition data of the channel orientation, and output the corrugated sheet assembly orientation misalignment status.

[0178] A third objective of this invention is to provide an electronic device comprising a processor, a memory, and a display screen. The memory and display screen are both connected to the processor, such as via a bus. Optionally, the electronic device may further include a transceiver. It should be noted that in practical applications, the transceiver is not limited to a single unit, and the structure of this electronic device does not constitute a limitation on the embodiments of this application.

[0179] The processor can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0180] A bus can include a pathway for transmitting information between the aforementioned components. The bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc.

[0181] The memory may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited to these.

[0182] The memory stores the application code that executes the solution of this application, and its execution is controlled by the processor. The processor executes the application code stored in the memory to implement the content shown in the foregoing method embodiments.

[0183] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the aforementioned functions. Figure 1 The illustrated method embodiments include various processes. For example, a memory may include instructions that can be executed by a processor of an electronic device to perform the described method.

[0184] A computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. A computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, a computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0185] A fifth objective of this invention is to provide a computer program product comprising computer instructions that, when executed by a processor, implement the above-described... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

Claims

1. A method for identifying misaligned assembly orientation of corrugated sheets in air preheater heat storage elements based on DRL, characterized in that, include: Acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; Flow resistance deviation and heat transfer hysteresis deviation are calculated based on the operating window data and the same load baseline data to form the flow resistance-heat transfer hysteresis correlation quantity; The running window data is input into the feature extraction unit of the DRL to obtain the channel state potential data, and the weight learner of the DRL generates the initial decomposition activation based on the running window data. The flow resistance-heat transfer hysteresis correlation is embedded into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and potential decomposition coefficients are generated based on the scene decomposition activation quantity. The potential data of the channel state is decomposed and reconstructed based on the potential decomposition coefficients to obtain the channel direction anomaly decomposition data. Based on the channel direction anomaly decomposition data, the direction misalignment is determined and the corrugated sheet assembly direction misalignment status is output.

2. The method for identifying misaligned assembly orientation of corrugated sheets in air preheater heat storage elements based on DRL as described in claim 1, characterized in that, The process of acquiring maintenance and operation data of the air preheater and generating operation window data based on this data includes: Obtain the operating data of the air preheater after maintenance, and filter the operating data of the air preheater after maintenance according to the preset stable operating conditions to form effective operating data after maintenance; Maintenance operation data is generated based on effective operating data after maintenance and benchmark data under the same load. Based on the maintenance and operation data, the pressure drop, flow rate, cold and hot end temperatures, and load response are extracted on the air and flue side to form the operation window data.

3. The method for identifying misaligned assembly orientation of corrugated sheets in air preheater heat storage elements based on DRL as described in claim 1, characterized in that, The process of calculating flow resistance deviation and heat transfer hysteresis deviation based on operating window data and load baseline data to form a flow resistance-heat transfer hysteresis correlation includes: Based on the load response in the running window data, a benchmark segment in the same load response range is matched in the same load benchmark data, and then truncated and aligned to form the same load comparison window data. Calculate the flow resistance deviation and temperature response hysteresis based on the data from the same load comparison window. The effective window of resistance deviation is determined based on the flow resistance deviation, and the window consistency constraint is applied to the temperature response hysteresis to form the heat transfer hysteresis deviation. The direction of resistance deviation is determined based on the flow resistance deviation, and the direction of heat transfer lag is determined based on the heat transfer lag deviation. The correlation constraint between the direction of resistance deviation and the direction of heat transfer lag is constructed within the same time window of the operating window data to form the correlation offset constraint quantity. By embedding the associated offset constraint into the fusion process of flow resistance deviation and heat transfer hysteresis deviation, the deviation of individual air and flue gas side pressure drop and individual cold and hot end temperature deviation are suppressed, forming the flow resistance-heat transfer hysteresis correlation quantity.

4. The method for identifying misaligned assembly orientation of corrugated sheets in air preheater heat storage elements based on DRL as described in claim 1, characterized in that, The process of inputting the running window data into the feature extraction unit of the DRL to obtain channel state potential data, and then having the weight learner of the DRL generate initial decomposition activation values ​​based on the running window data, includes: The running window data is input into the feature extraction unit of the DRL according to the same time window to form window latent mapping data; Channel state constraints are constructed based on the window potential mapping data, and potential mapping components formed by single-phase air-smoke side pressure drop deviation or single-phase cold and hot end temperature deviation are suppressed to obtain channel state potential data. The running window data is input into the weight learner of the DRL to form candidate activations for decomposition directions; Within the weight learner of the DRL, the candidate activation values ​​for decomposition directions are constrained to correspond to the potential decomposition directions of the channel state potential data, thereby generating the initial decomposition activation values.

5. The method for identifying misaligned assembly orientation of corrugated sheets in air preheater heat storage elements based on DRL according to claim 1, characterized in that, The step of embedding the flow resistance-heat transfer hysteresis correlation value into the generation of the initial decomposition activation value to obtain the scene decomposition activation value, and generating potential decomposition coefficients based on the scene decomposition activation value, includes: Obtain the activation generation position of the initial decomposed activation in the weight learner of the DRL, and determine the intermediate activation to be embedded according to the activation generation position and the decomposition direction. Determine the associated embedding constraint based on the flow resistance-heat transfer hysteresis correlation; The associated embedding constraint quantity is embedded into the activation generation position where the intermediate activation quantity to be embedded is located, forming the scene constraint activation intermediate quantity. The initial decomposed activation quantity is updated according to the scene constraint activation intermediate quantity to form the associated embedding activation quantity. The activation attribution of the decomposition direction is consistently adjusted according to the associated embedding activation quantity to form the scene decomposed activation quantity candidate. The candidate activation values ​​for scene decomposition are normalized to obtain the scene decomposition activation values, and potential decomposition coefficients are generated based on the scene decomposition activation values.

6. The method for identifying misaligned assembly orientation of corrugated sheets in air preheater thermal storage elements based on DRL according to claim 1, characterized in that, The process of decomposing and reconstructing the potential data of the channel state based on the potential decomposition coefficients to obtain the channel direction anomaly decomposition data, determining the direction misalignment based on the channel direction anomaly decomposition data, and outputting the corrugated sheet assembly direction misalignment state includes: The proportion of channel state potential data entering the DRL decomposition and reconstruction process is determined based on the potential decomposition coefficient, and the proportion of decomposition direction is used as the basis for the allocation of channel state potential data to form the decomposition direction allocation ratio. Based on the allocation ratio of the decomposition direction, the channel decomposition projection data is determined; based on the potential decomposition coefficient, the reconstruction contribution of the channel decomposition projection data is limited to form channel reconstruction contribution data; then, the channel reconstruction contribution data is combined and reconstructed to form channel state reconstruction data. The potential reconstruction difference data is obtained by calculating the difference between the potential channel state data and the reconstructed channel state data. The potential reconstruction difference data is then reconstructed based on the correlation between the flow resistance and heat transfer hysteresis carried by the potential decomposition coefficient to obtain the channel direction anomaly decomposition data. Then, the directional channel response is aggregated to form the direction misalignment judgment input data. The direction misalignment judgment input data is used to calculate the direction misalignment response data; the judgment response quantity of the direction misalignment in the direction misalignment response data is compared with the preset threshold to form the direction misalignment judgment result; based on the direction misalignment judgment result, it is determined whether there is a direction misalignment, no direction misalignment, or direction misalignment needs to be verified, and the direction misalignment status of the corrugated sheet assembly is output.

7. A system for identifying misaligned assembly orientation of corrugated sheets in an air preheater based on DRL, comprising the method for identifying misaligned assembly orientation of corrugated sheets in an air preheater based on DRL as described in any one of claims 1-6, characterized in that, include: Operation window data module: used to acquire maintenance and operation data of the air preheater, and generate operation window data based on the maintenance and operation data; The hysteresis correlation module is used to calculate the flow resistance deviation and heat transfer hysteresis deviation based on the operating window data and the same load baseline data, forming the flow resistance-heat transfer hysteresis correlation. Initial decomposition activation quantity module: It is used to input the running window data into the feature extraction unit of DRL to obtain the channel state potential data, and the weight learner of DRL generates the initial decomposition activation quantity based on the running window data. Potential decomposition coefficient module: used to embed the flow resistance-heat transfer hysteresis correlation into the generation of the initial decomposition activation quantity to obtain the scene decomposition activation quantity, and generate potential decomposition coefficients based on the scene decomposition activation quantity. Corrugated sheet assembly orientation misalignment status module: It is used to decompose and reconstruct the potential data of the channel state according to the potential decomposition coefficient, obtain the abnormal decomposition data of the channel orientation, determine the orientation misalignment based on the abnormal decomposition data of the channel orientation, and output the corrugated sheet assembly orientation misalignment status.

8. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for identifying misaligned assembly orientation of corrugated sheet of air preheater heat storage element according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for identifying misaligned assembly orientation of corrugated sheets for air preheater heat storage elements based on DRL, as described in any one of claims 1-6.

10. A computer program product, characterized in that, The method includes computer instructions that, when executed by a processor, implement a method for identifying misaligned assembly orientation of corrugated sheets for air preheater heat storage elements as described in any one of claims 1-6.

Images