A wind power flange production line equipment state remote monitoring method

Abstract

This invention belongs to the field of industrial control and monitoring technology, specifically relating to a method for remote monitoring of the status of equipment in a wind power flange production line. The method includes: initializing a baseline model, including constructing a process transfer model, an initial root cause analysis model, and an initial soft sensor model, and calculating initial process residuals; real-time monitoring of new batches, using the updated process transfer model to calculate real-time process residuals, and using the updated soft sensor model to predict real-time quality indicators; when the predicted value of the real-time quality indicator is abnormal, using the updated root cause analysis model to locate the root cause of the abnormality; obtaining the actual quality indicators of the new batch and updating the soft sensor model; when the actual quality indicators are qualified, updating the root cause analysis model. This invention achieves the tracking of chronic equipment drift, effectively distinguishes sudden failures, solves the problem of quality inspection lag, and improves the accuracy and timeliness of monitoring.

Description

Technical Field

[0001] This invention relates to the field of industrial control and monitoring technology. More specifically, this invention relates to a method for remotely monitoring the status of equipment in a wind power flange production line. Background Technology

[0002] Wind turbine flanges are key structural components of wind power generation equipment. Their production process involves multiple high-temperature and high-pressure precision machining processes, such as heating in a heating furnace, forging in a hydraulic press, precision rolling in a CNC ring rolling machine, and tempering in a pit-type heat treatment furnace. The state parameters of these processes are transmitted along the production chain and affect each other, ultimately determining the final quality of the flange.

[0003] In remote monitoring of wind power flange production lines, establishing a correlation model between process and quality is crucial for locating the root cause of faults, and partial least squares regression (PLSR) is a commonly used technique. However, standard PLSR models have technical limitations. On the one hand, they suffer from causal confusion, failing to distinguish between root causes (such as low furnace temperature) and intermediate symptoms (such as high pressure in the ring rolling mill due to low material temperature). On the other hand, they have a static baseline problem; PLSR models are typically trained once based on historical batch data, and the established normal process baseline remains fixed.

[0004] In actual production, the condition of equipment itself will slowly drift over time. For example, the rolls of a ring rolling mill will experience progressive wear, or the heating elements of a heat treatment furnace will slowly age, causing a systematic shift in the baseline of normal process parameters. Static PLSR models cannot detect this baseline drift and may diagnose normal, slow drift as a sudden fault, generating false alarms. Alternatively, when a sudden fault is superimposed on an already drifted baseline, the incorrect reference may prevent the effective identification of the true fault. Furthermore, quality inspections, such as tensile strength and impact energy, typically have a time lag of 3-7 days, making it difficult to detect quality problems in a timely manner and potentially leading to a large number of defective products.

[0005] Therefore, there is an urgent need for a method for remote monitoring of the equipment status of wind power flange production lines. Summary of the Invention

[0006] To address the technical problems of existing static models being unable to distinguish causal relationships, unable to adaptively track chronic equipment drift, and struggling to handle quality inspection lag, this invention provides a method for remote monitoring of the equipment status of a wind power flange production line, comprising:

[0007] The process involves collecting process parameters and quality indicators from initial qualified batches, dividing the process parameters into at least two process blocks according to the physical production sequence; predicting the process parameters of downstream process blocks based on the process parameters of upstream process blocks, and constructing a process transfer model; calculating the process residual between the actual process parameters of downstream process blocks and the predicted values ​​of the process transfer model; constructing an initial root cause analysis model and an initial soft sensor model based on the process parameters of upstream process blocks and the process residual; collecting real-time process parameters for new batches; predicting the process parameters of downstream processes using the updated process transfer model based on the real-time process parameters of upstream processes; calculating the real-time process residual between the real-time process parameters and the predicted values ​​of the process parameters of downstream processes; constructing a real-time feature vector, which includes the real-time process parameters and real-time process residual of upstream processes; predicting real-time quality indicators based on the real-time feature vector using the updated soft sensor model; analyzing the real-time feature vector using the updated root cause analysis model to locate the root cause of the anomaly in response to an anomaly in the predicted values ​​of the real-time quality indicators; obtaining the actual quality indicators of new batches; updating the soft sensor model using the real-time feature vector and actual quality indicators of new batches; and updating the root cause analysis model in response to the actual quality indicators being qualified.

[0008] This invention decouples the causal relationships between production processes by constructing a process transfer model and calculating process residuals, laying the foundation for distinguishing between root causes and intermediate symptoms. By adopting a differentiated model update strategy—where the soft sensor model uses all data to learn various states, while the root cause analysis model only uses qualified data—it ensures that the model can adaptively track the chronic drift of equipment while avoiding the contamination of the normal baseline by abnormal data, thus accurately distinguishing between chronic drift and sudden failures. Furthermore, the soft sensor model enables real-time quality prediction, solving the problem of quality inspection lag, and the decoupled root cause analysis model improves the accuracy of fault location.

[0009] Preferably, dividing the process parameters into at least two process blocks according to the physical production sequence includes: dividing the process parameters into a first process block matrix, a second process block matrix, and a third process block matrix; constructing the process transfer model includes: training a first initial transfer model using the first process block matrix as the independent variable and the second process block matrix as the dependent variable; using the difference between the predicted values ​​of the second process block matrix and the first initial transfer model as the second process residual matrix; training a second initial transfer model using the first and second process block matrices as independent variables and the third process block matrix as the dependent variable; and using the difference between the predicted values ​​of the third process block matrix and the second initial transfer model as the third process residual matrix.

[0010] This invention quantifies the impact of upstream processes on downstream processes by dividing process parameters into blocks according to physical order and establishing process transfer models step by step. By calculating process residuals, i.e., the difference between the actual value of the downstream process and the value predicted based on the upstream process, the pure deviation of the downstream process itself is successfully extracted. This deviation no longer includes the influence transmitted from the upstream process. This ensures that the subsequent model construction is based on the decoupled information, providing a clear model foundation for accurately distinguishing the root cause and intermediate symptoms.

[0011] Preferably, the construction of the initial root cause analysis model and the initial soft measurement model includes: constructing an initial decoupling independent variable matrix, which includes a first process block matrix, a second process residual matrix, and a third process residual matrix; and training the initial root cause analysis model and the initial soft measurement model using the initial decoupling independent variable matrix as independent variables and the quality index as dependent variables.

[0012] This invention utilizes a decoupled independent variable matrix containing the first process block and the residuals of each process level to construct an initial model. This ensures that the initial root cause analysis model and soft measurement model are trained from the outset based on the pure deviation information after process decoupling, guaranteeing the model's accurate grasp of causal relationships and providing a high-quality initial baseline for subsequent real-time monitoring.

[0013] Preferably, the real-time process parameter acquisition includes: acquiring a first process parameter vector, a second process parameter vector, and a third process parameter vector for a new batch; the construction of the real-time feature vector includes: loading the updated first process transfer model and second process transfer model from the previous batch; using the first process transfer model and the first process parameter vector to predict the predicted parameter vector for the second process; using the difference between the second process parameter vector and the predicted parameter vector for the second process as the second process residual; using the second process transfer model, the first process parameter vector, and the second process parameter vector to predict the predicted parameter vector for the third process; using the difference between the third process parameter vector and the predicted parameter vector for the third process as the third process residual; and constructing the real-time feature vector from the first process parameter vector, the second process residual, and the third process residual.

[0014] In the real-time monitoring phase, this invention utilizes the updated transmission model from the previous batch, which has adapted to the chronic drift of the equipment, to predict the expected state of the downstream processes in the current batch based on the parameters of the upstream processes. By calculating the real-time process residuals between the actual parameters of the current batch and the predicted state, the transmission influence of the upstream processes and the background of the chronic drift of the equipment are effectively separated, thereby quantifying the unexpected deviations that actually occur in the current batch and need to be diagnosed. The constructed real-time feature vector contains only the actual state of the first process and the pure deviations of the downstream processes, so that subsequent quality prediction and fault diagnosis are based entirely on the decoupled information, improving the accuracy of real-time monitoring.

[0015] Preferably, the root cause of the location anomaly includes: loading the root cause analysis model updated in the previous batch; calculating the variable importance projection score of each variable in the real-time feature vector to the predicted value of the real-time quality indicator based on the predicted value of the real-time quality indicator and the root cause analysis model; obtaining the variables whose variable importance projection scores exceed a preset root cause contribution threshold, and determining the variable with the highest score as the root cause.

[0016] When a quality warning occurs, this invention utilizes an updated root cause analysis model representing the current normal baseline. By calculating the variable importance projection score, it quantifies the contribution of each variable in the real-time feature vector to the quality anomaly. Since the feature vector is decoupled, the variable with the highest score points to the root cause of the fault. For example, is it an anomaly in the parameters of the first process itself, or a pure deviation in the second or third process? This achieves precise location of the fault source and avoids misdiagnosis caused by causal confusion.

[0017] Preferably, updating the soft measurement model includes: using real-time feature vectors and actual quality indicators to update the soft measurement model using recursive least squares method; updating the root cause analysis model includes: in response to the actual quality indicators being qualified, using real-time feature vectors and actual quality indicators to update the root cause analysis model using recursive least squares method.

[0018] This invention employs a differentiated model update strategy: the soft measurement model is updated using data from all batches, enabling it to learn and understand the process-quality relationship under various production conditions, thereby improving its accuracy in predicting future quality; while the root cause analysis model is updated only using data from batches that have passed quality standards. This specific operation ensures that the model always reflects the normal state baseline of the equipment, allowing it to adaptively track this chronic drift as the equipment gradually wears or ages, avoiding the contamination of the normal baseline by abnormal data, and ensuring that the benchmark for fault diagnosis is always the latest normal state.

[0019] Preferably, the process transfer model, the initial root cause analysis model, and the initial soft measurement model are all trained using the partial least squares regression algorithm.

[0020] Preferably, after locating the root cause of the anomaly, the method further includes: constructing a historical fault mode database, wherein the historical fault mode database includes root cause feature vectors of historical fault modes; constructing a root cause feature vector of the current fault based on the real-time feature vector, the predicted value of the real-time quality index, and the updated root cause analysis model; calculating the cosine similarity between the root cause feature vector of the current fault and the root cause feature vector of the historical fault modes in the historical fault mode database; and, in response to the maximum cosine similarity being less than a preset similarity threshold, marking the real-time quality index anomaly as a suspected false alarm.

[0021] This invention introduces a false alarm assessment mechanism, which compares the currently diagnosed fault characteristics with a historical fault mode database. If the current fault characteristics do not match any of the known historical fault modes, the system marks it as a suspected false alarm. This mechanism can effectively identify false alarms caused by inaccurate predictions or short-term fluctuations in soft measurement models, avoid overreacting to unknown or uncertain faults, reduce unnecessary process adjustments and production interruptions, and improve the robustness and practicality of the remote monitoring system.

[0022] Preferably, the method further includes: calculating the time interval between the current batch and the most recent batch that has obtained actual quality indicators, determining the model confidence of the currently used model based on the time interval; and relaxing the warning threshold used to determine whether the predicted value of the real-time quality indicator is abnormal in response to the model confidence being lower than a preset confidence threshold.

[0023] This invention addresses the uncertainty caused by the lag in quality detection leading to untimely model updates by introducing model confidence assessment. As the time interval between the last actual quality data and the current data becomes longer, the model's ability to track the current equipment status baseline becomes more uncertain, and the confidence level decreases accordingly. When the confidence level is below a threshold, the system reduces the sensitivity of monitoring by relaxing the warning threshold, thereby effectively reducing false alarms that may occur due to high model uncertainty, achieving a balance between ensuring timely warnings and avoiding excessive alarms.

[0024] Preferably, the model confidence level satisfies the expression: In the formula, Model confidence; This is the confidence decay coefficient; The time interval between the current batch and the most recent batch that has obtained actual quality indicators; It is a natural exponential function.

[0025] The beneficial effects of this invention are as follows: By constructing a process transfer model and calculating process residuals, this invention successfully decouples the causal relationship between production processes; by using only qualified batch data to update the root cause analysis model, while using all batch data to update the soft sensor model, it achieves adaptive tracking of equipment chronic drift and effectively distinguishes between chronic drift and sudden failures; at the same time, the soft sensor model solves the problem of quality inspection lag, realizing real-time quality prediction and early warning, and the root cause analysis model accurately locates the source of failure, improving the accuracy and timeliness of remote monitoring. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a method for remote monitoring of the equipment status of a wind power flange production line according to the present invention;

[0027] Figure 2 This is a schematic diagram illustrating the actual temperature change of the heating furnace in the first process of this invention.

[0028] Figure 3 This is a schematic diagram illustrating the actual value change of the rolling pressure in the second process of this invention embodiment;

[0029] Figure 4 This is a schematic diagram illustrating the decoupling result of the ring rolling pressure residual in the second process of this embodiment of the invention;

[0030] Figure 5 This is a schematic diagram illustrating the quality prediction and root cause diagnosis results in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0033] This invention discloses a method for remote monitoring of the equipment status of a wind power flange production line, referring to... Figure 1 This includes steps S1 to S4:

[0034] S1. Collect process parameters and quality indicators of the initial qualified batch, divide the process parameters into at least two process blocks according to the physical production sequence, predict the process parameters of the downstream process block based on the process parameters of the upstream process block, construct a process transfer model, calculate the process residual between the actual process parameters of the downstream process block and the predicted value of the process transfer model, and construct an initial root cause analysis model and an initial soft measurement model based on the process parameters of the upstream process block and the process residual.

[0035] Before starting adaptive monitoring, two initial baseline models that can represent the current normal state are first established, including a root cause analysis (RCA) model and a soft measurement model. Subsequent recursive updates will be based on these two models.

[0036] Specifically, collecting data from continuously produced data on the production line. Each batch of flange data, and ensure this All batches passed inspections in the metallographic testing laboratory, tensile testing laboratory, etc. The data from each batch constitutes the initial baseline set, which includes process parameters and corresponding quality index data. The process parameters include those for each step of the process, such as heating temperature, heating time, hydraulic press pressure, and hydraulic press holding time for the heating and blanking processes; rolling pressure, rolling speed, rolling feed, and spindle vibration data for the precision ring rolling process; and heat treatment temperature, holding time, cooling rate, and cooling medium pressure for the heat treatment process. Quality indicators include tensile strength, yield strength, impact energy, and hardness. The preset number of samples is used to determine the statistical significance of the sample size. This number is typically set to 50 to 100 batches to ensure sufficient representativeness of the model. In this embodiment, 50 batches are used. In other embodiments, the implementer can set the number of batches according to the actual implementation situation. The value of .

[0037] It should be noted that, due to the strict physical transmission relationship between processes in the wind power flange production line, such as the heating furnace temperature directly affecting the forging effect of the hydraulic press, the forging parameters of the hydraulic press affecting the feeding state of the ring rolling mill, and the parameters of the ring rolling mill further affecting the heat treatment effect, this invention structures and divides the collected process parameters into blocks according to the physical production sequence to distinguish the independent contributions of different processes.

[0038] In one embodiment, The data from each batch are used to construct the initial independent variable matrix. and the initial dependent variable matrix , where the initial dependent variable matrix Depend on The quality indicators consist of [number] dimensions. Dimension; Initial independent variable matrix Based on physical processes, it is divided into Each process block: First process block matrix It includes heating and blank-forming process parameters for the heating furnace and hydraulic press, such as heating furnace temperature, heating time, hydraulic press pressure, and hydraulic press holding time, etc., in the following dimensions. Dimension; Second process block matrix It includes precision rolling process parameters of CNC ring rolling machines, such as rolling pressure, rolling speed, rolling feed rate, spindle vibration data, etc., in the following dimensions. Dimension; Third process block matrix It includes heat treatment process parameters for pit-type heat treatment furnaces, such as heat treatment temperature, holding time, cooling rate, and cooling medium pressure, with the following dimensions: Dimensions. Among them, The quantity of quality indicators; The number of process parameters for the heating and blank-forming processes; The number of process parameters for the precision ring rolling process; This refers to the number of process parameters for the heat treatment process.

[0039] It should be noted that, since there is a clear physical transfer effect between processes in the wind power flange production line, for example, the state of the heating process directly affects the parameter performance of the rolling process, this invention establishes a process transfer model to reflect this dependency.

[0040] Specifically, using the first process block matrix The independent variable is the second process block matrix. Using PLSR algorithm as the dependent variable, the first initial transitive model is obtained through training. ;by The independent variable is the third process block matrix. Using the dependent variable, a second initial transfer model is obtained through training. .

[0041] It should be noted that since the standard PLSR model cannot distinguish between the root cause and intermediate symptoms, the monitoring system has difficulty in accurately identifying the source of the fault. This invention measures the pure deviation of the downstream process by calculating the process residual, that is, the difference between the actual value of the downstream process and the predicted value of the upstream process transmission model. This pure deviation is the root cause of the fault.

[0042] Specifically, calculate the initial process residual matrix:

[0043]

[0044]

[0045] in, This is the residual matrix for the second process; This is the residual matrix for the third process; This is the first process block matrix; This is the second process block matrix; This is the third process block matrix; This is the first initial transfer model; This is the second initial transfer model; The predicted values ​​of the second process parameters obtained through the first initial transfer model. .

[0046] Furthermore, the initial decoupling independent variable matrix is ​​constructed. ; to decouple the independent variable matrix The independent variable is the initial dependent variable matrix. Using PLSR algorithm as the dependent variable, an initial Root Cause Analysis (RCA) model is obtained through training. Calculate and store the RCA model The variable importance scores of all independent variables projected onto all dependent variables.

[0047] Among them, the variable importance projection score satisfies the expression:

[0048]

[0049] In the formula, For the first The independent variable is paired with the first... Projected scores of variable importance for each quality indicator; This represents the total number of independent variables, that is, the total number of process parameters for the three processes. , The number of process parameters for the heating and blank-forming processes; The number of process parameters for the precision ring rolling process; The number of process parameters for the heat treatment process; The index of the latent variable; For the first One potential variable; For the first A vector of quality indicators; For the first Among the nth latent variables, the nth The weights of the independent variables; for The transpose of the matrix; for The transpose of .

[0050] It should be noted that latent variables are the core concept of the PLSR algorithm. They are obtained by linearly combining independent and dependent variables and reflect the main change patterns in the data. In this embodiment, the number of latent variables is determined by cross-validation and does not exceed the smaller dimension of the independent and dependent variables. These are parameters already present in the PLSR algorithm, representing the first... The independent variable at the th independent variable in the th ... The weights of the individual variables reflect the degree of contribution of each independent variable to the total number of latent variables. The absolute value of the magnitude represents the first The independent variable is paired with the first... The contribution of each latent variable is considered; the larger the absolute value, the greater the contribution. The symbol represents the first The positive and negative correlation between the independent variables and the latent variables. Independent variables with larger values ​​have a significant impact on quality indicators.

[0051] Variable importance projection score By calculating the contribution of each independent variable to all latent variables, the predictive importance of the independent variable for a specific quality indicator is measured. When it is greater than 1.0, it indicates that the first... The independent variable is paired with the first... The more significant the contribution of a process parameter to the prediction of a quality indicator, the more significant its influence on the corresponding quality indicator; when When it is less than 0.8, it means that the first... The independent variable is paired with the first... The predictive contribution of each quality indicator is relatively small.

[0052] It should be noted that, due to the time lag in quality inspection, real-time quality monitoring cannot be achieved. Therefore, this invention constructs a soft measurement model to predict quality indicators based on real-time process parameters.

[0053] Specifically, to decouple the independent variable matrix The independent variable is the initial dependent variable matrix. Using PLSR algorithm as the dependent variable, an initial soft sensor model is obtained through training. .

[0054] It should be noted that the soft measurement model With RCA model Using the same decoupling independent variable matrix and dependent variable matrix However, their objectives differ. The RCA model is used for fault diagnosis, focusing on causal relationships between variables; the soft sensor model is used for quality prediction, focusing on prediction accuracy. In subsequent RCA model updates, the RCA model is trained and updated only using batches with acceptable quality indicators to ensure it reflects the normal state of the equipment. The soft sensor model, on the other hand, uses only batches with acceptable quality indicators during the initial training phase, but in subsequent updates uses quality indicators from all batches, including both acceptable and unacceptable batches, to improve prediction accuracy.

[0055] S2. For new batches, collect real-time process parameters. Based on the real-time process parameters of the upstream process, use the updated process transfer model to predict the process parameters of the downstream process. Calculate the real-time process residual between the real-time process parameters and the predicted values ​​of the process parameters of the downstream process. Construct a real-time feature vector, which includes the real-time process parameters and the real-time process residual of the upstream process. Use the updated soft measurement model to predict real-time quality indicators based on the real-time feature vector.

[0056] Specifically, from the first Starting with the first batch, the system enters a real-time adaptive monitoring phase. For each new batch entering production, its process parameters are remotely collected, forming three process parameter vectors. For example, the first batch... The three process parameter vectors for the batch are as follows: , , ,in, This is the serial number of the new batch entering production. , For the first The first process parameter vector of the batch has a dimension of The parameters include those for the heating and blank-forming processes of the furnace and hydraulic press, such as furnace temperature, heating time, hydraulic press pressure, and hydraulic press holding time. For the first The second process parameter vector of the batch has a dimension of The dimensions include precise ring rolling process parameters of the CNC ring rolling machine, such as ring rolling pressure, ring rolling speed, ring rolling feed rate, and spindle vibration data. For the first The third process parameter vector of the batch, with dimension [missing information]. The dimensions include the heat treatment process parameters of the pit-type heat treatment furnace, such as heat treatment temperature, holding time, cooling rate, and cooling medium pressure. The number of process parameters for the heating and blank-forming processes; The number of process parameters for the precision ring rolling process; This refers to the number of process parameters for the heat treatment process.

[0057] It should be noted that during the process parameter acquisition, the data acquisition frequency should satisfy the Nyquist sampling theorem. In this embodiment, it is set to 1 second / time to ensure that rapid changes in process parameters are captured.

[0058] It should be further explained that the equipment status in the wind power flange production line will slowly drift over time. For example, wear of the rolling mill rolls will gradually increase the rolling pressure, or aging of the heating elements will cause changes in the furnace temperature control characteristics. These slow drifts will cause a systematic shift in the baseline of the process parameters during normal production. Static models cannot detect this baseline drift and will misjudge normal slow drifts as sudden faults. Alternatively, when a sudden fault is superimposed on an already drifted baseline, the static model will be unable to effectively identify the true fault due to the incorrect reference. Therefore, this invention uses the normal baseline model updated from the previous batch to predict the expected parameter status of the current batch. The difference between the actual value and the predicted value is the unexpected or sudden deviation that the system needs to diagnose.

[0059] Specifically, for the current batch Load the first process transfer model updated from the previous batch. Second process transfer model Calculate the first Batch process parameter prediction values and .

[0060] Further, calculate the first Batch process residual vector:

[0061]

[0062]

[0063] In the formula, For the first Second-stage residuals in the batch; For the first The residual of the third process in the batch; For the first The second process parameter vector for the batch; For the first The predicted parameter vector for the second process of the batch; For the first The parameter vector for the third process in the batch; For the first The predicted parameter vector for the third process in a batch; this invention measures the pure deviation of the process itself by calculating the difference between the actual value and the predicted value based on the upstream process. A large absolute value indicates an abnormality in the rolling process itself, rather than being caused by an upstream process; when A large absolute value indicates an abnormality in the heat treatment process itself. For example, the residual of the second process can effectively distinguish between two fundamentally different problems: high rolling pressure caused by insufficient heating furnace temperature and high pressure caused by hydraulic system failure of the rolling mill. Similarly, the residual of the third process can effectively distinguish between two fundamentally different problems: low heat treatment temperature caused by insufficient rolling process temperature and low temperature caused by failure of the heat treatment furnace itself.

[0064] Furthermore, the first Batch first process parameter vector , No. Batch second process residual and the Batch third process residual , constitute the first Batch feature vectors .in, Directly reflects the status of the first process. This reflects the additional deviation of the second process relative to the first process. It reflects the additional deviation of the third process relative to the first two processes.

[0065] It should be noted that, in order to address the lag issue in quality inspection, this invention utilizes the updated soft measurement model from the previous batch. For the first The quality indicators of each batch are predicted in real time.

[0066] Specifically, the first Batch feature vectors Enter the number Batch updated soft measurement model In the middle, we obtained the first Predicted values ​​of batch quality indicators .

[0067] Response to soft measurement model Predicted quality indicators When the preset warning threshold is exceeded, the system sends an early warning message to the remote monitoring terminal and marks it as a soft measurement warning, prompting maintenance personnel to pay attention to potential quality issues. The initial warning threshold is set according to industry standards and quality requirements for wind turbine flange products.

[0068] S3. In response to anomalies in the predicted values ​​of real-time quality indicators, the updated root cause analysis model is used to analyze the real-time feature vectors to locate the root cause of the anomaly. Based on the root cause location information, the relevant processes are adjusted or repaired.

[0069] It should be noted that when the quality index predicted by the soft measurement model exceeds the preset warning threshold, it means that a sudden failure may have occurred. This invention uses the latest RCA model for fault location.

[0070] Specifically, for the current batch Load the RCA models updated in the previous batch. According to the Predicted values ​​of batch quality indicators The RCA model calculates the variable importance projection score for each independent variable on each quality indicator, and obtains the root cause contribution threshold when the variable importance projection score exceeds the preset threshold. The independent variable. The contribution threshold of the root cause. Based on historical data and practical experience, the value is typically set between 0.8 and 1.2. In this embodiment... In other embodiments, implementers may set the appropriate parameters according to the actual implementation situation. The value of . From these independent variables that exceed the threshold, identify the independent variable with the highest variable importance projection score to determine the most likely root cause.

[0071] Furthermore, an alarm message is sent to the remote monitoring terminal, indicating the root cause of the quality abnormality, including the specific process and parameters: in response to the parameter vector of the first process. If the variable with the highest importance projection score in the independent variable is the first process block, then the root cause is located. For example, if the variable corresponding to the furnace temperature has the highest importance projection score, then the furnace is malfunctioning. The response is to the residual of the second process. If the variable with the highest projected importance score among the independent variables is the second process block, then the root cause is located in the second process block. For example, if the variable corresponding to the ring rolling pressure has the highest projected importance score, it indicates a sudden abnormality in the ring rolling machine itself, or a failure in the ring rolling machine's hydraulic system. The response is related to the residuals of the third process. If the variable with the highest importance projection score is the third process block, then the root cause is located. For example, if the variable with the highest importance projection score is the heat treatment temperature, then the heat treatment furnace itself has experienced a sudden abnormality.

[0072] It should be noted that in this diagnostic mode, maintenance personnel adjust or repair relevant processes based on the root cause location information provided by the system to prevent potential quality problems from worsening. For example, when the root cause is located as abnormal heating furnace temperature, the heating furnace temperature parameters are adjusted; when the root cause is located as a fault in the ring rolling mill hydraulic system, the hydraulic system is inspected and repaired. This precise diagnosis based on the RCA model makes maintenance work more targeted and significantly shortens the time for troubleshooting and handling.

[0073] Preferably, to prevent process adjustments based on erroneous predictions, the present invention introduces a false alarm assessment mechanism, specifically:

[0074] A historical failure mode database was constructed using manually labeled quality anomaly events. Each historical failure mode includes a failure type, a root cause feature vector, the range of process parameters at the time of the failure, and suggested maintenance measures. The failure type includes clearly defined categories such as insufficient furnace temperature, abnormal hydraulic press pressure, hydraulic system failure of the ring rolling mill, and abnormal temperature control of the heat treatment furnace, covering the main failure modes of the wind power flange production line. The root cause feature vector includes the variable importance projection score of each process parameter for each quality indicator, the parameter vector of the first process, and the process residuals of the remaining processes. ,in This represents the total number of process parameters. This refers to the quality indicator serial number. Indicates the first The variable importance projection score of each process parameter to the q-th quality indicator. This represents the parameter vector for the first process. Indicates the residual of the second process. This indicates the residual of the third process.

[0075] When the soft measurement model predicts a quality anomaly, the RCA model diagnoses the root cause feature vector of the current fault. The root cause feature vector has the same dimensions as the root cause feature vectors of historical failure modes. Z-score standardization is applied to each dimension of the current failure root cause feature vector and each dimension of the historical failure mode root cause feature vector to ensure that each dimension has zero mean and unit variance, thus eliminating the influence of different dimensions.

[0076] Calculate the root cause feature vector after standardizing the current fault. The cosine similarity between the standardized root cause feature vectors of historical failure modes in the historical failure mode database and the standardized root cause feature vectors is related to the similarity between the root cause feature vectors and the standardized root cause feature vectors. When the maximum cosine similarity of the root cause feature vector after standardization of historical failure modes in the historical failure mode database is less than a preset similarity threshold, the system marks the soft measurement warning as a suspected false alarm and does not immediately adjust the process. Process adjustments are only recommended when subsequent data confirms a persistent trend of quality abnormalities. This mechanism effectively avoids unnecessary process adjustments caused by false alarms from the soft measurement model. In this embodiment, the similarity threshold is set to 0.6. In other embodiments, implementers can set the similarity threshold value according to the actual implementation situation.

[0077] S4. Obtain the actual quality indicators of the new batch, update the soft measurement model using the real-time feature vector and actual quality indicators of the new batch, and update the root cause analysis model in response to the actual quality indicators being qualified.

[0078] Specifically, when the After a batch of flanges is produced, its final quality indicators are obtained from metallographic testing and tensile testing laboratories, constituting the first batch of quality indicators. Batch quality vector , dimension Dimensions include key performance indicators such as tensile strength, yield strength, impact energy, and hardness. The quantity of quality indicators.

[0079] It's important to note that soft measurement models and actual quality indicators differ fundamentally in terms of both time frame and accuracy: Soft measurement models, based on real-time process parameters, provide near-instantaneous quality predictions for early warning; actual quality indicators, obtained through physical experiments, have a 3-7 day time lag and are used for final quality assessment. The relationship between the two is that soft measurement models predict quality trends, while actual quality indicators verify whether products meet standards. The predicted values ​​of soft measurement models... Compared with actual quality indicators Together they constitute a complete quality assessment system.

[0080] Furthermore, the RCA model and the soft sensor model are updated based on the quality indicators:

[0081] Regardless of the Whether a batch meets quality standards is determined using data. Recursive Least Squares (RLS) is used to apply the method to the th... Batch updated soft measurement model Update to obtain the first Batch updated soft measurement model This invention uses all data to update the soft sensor model, enabling it to learn the quality variation patterns under various production conditions, including normal and abnormal situations, thereby improving the comprehensiveness and accuracy of the soft sensor model's predictions. For the first Batch feature vectors For the first Batch quality vector.

[0082] Only when the first When the batch quality indicators are qualified, the data is used to analyze... The RLS algorithm is used to process the first... Batch updated RCA model Update to obtain the first Batch updated RCA model This invention uses batches that meet quality standards to update the RCA model, ensuring that the RCA model always reflects the normal baseline of the equipment and avoiding abnormal data contaminating the model.

[0083] It should be noted that the model update frequency is consistent with the quality indicator acquisition frequency, usually every 3-7 days per batch. After the model is updated, it is only used for subsequent batches entering production. For batches that have started production but have not yet been completed, the model before the update will continue to be used for prediction and diagnosis.

[0084] It should be further noted that the RLS algorithm updates model parameters recursively, avoiding the computational overhead of recalculating the entire dataset for each update. In this embodiment, the RLS algorithm is used for online updates of the soft measurement model and the RCA model, ensuring that the soft measurement model and the RCA model can adapt to changes in equipment status in a timely manner.

[0085] Preferably, to address the lag in quality indicators, this invention introduces a model confidence assessment mechanism, specifically:

[0086] When the When the batch quality indicators have not yet been detected, the first The model confidence level of the model used for batch prediction is based on the first... The time interval between the batch and the most recently obtained quality index is adjusted. The model confidence level satisfies the expression:

[0087]

[0088] in, For the first The model confidence level of the model used for batch prediction; This is the confidence decay coefficient, used to control the rate at which the model confidence decays over time. In this embodiment... In other embodiments, implementers may set the appropriate parameters according to the actual implementation situation. The value; For the first The time interval between a batch and the most recent batch that has obtained quality indicators, in days; It is a natural exponential function.

[0089] When model confidence Below the preset confidence threshold In this case, the system relaxes the quality anomaly warning threshold to reduce false alarms and avoid unnecessary maintenance activities caused by model uncertainties. Specifically, for quality indicators such as tensile strength, yield strength, and impact energy, where higher values ​​are better, the new warning threshold satisfies the expression:

[0090]

[0091] in, The new warning threshold, The initial warning threshold, For the first The model confidence level of the model used for batch prediction. This is the preset confidence threshold. Due to the model confidence... Less than ,therefore This makes the new warning threshold Compared to the initial warning threshold, the warning threshold is lowered, thereby relaxing the warning threshold to reduce false alarms. Furthermore, the lower the model confidence level, The smaller the value, the lower the new warning threshold. The more it decreases.

[0092] For quality indicators such as hardness that have an optimal range, the lower limit and upper limit of the new warning threshold satisfy the following expression:

[0093]

[0094]

[0095] in, This is the lower limit of the new early warning threshold. This is the upper limit of the new warning threshold. This is the initial lower limit of the warning threshold. This is the initial upper limit of the warning threshold. For the first The model confidence level of the model used for batch prediction. This is the preset confidence threshold. Due to the model confidence... Less than ,therefore This makes the lower limit of the new early warning threshold... reduce, This makes the new warning threshold upper limit The threshold is increased to relax the warning threshold and reduce false alarms. Furthermore, the lower the new warning threshold, the lower the threshold becomes as the model confidence decreases. The more it is reduced, the higher the upper limit of the new warning threshold becomes. The more it increases.

[0096] For other quality indicators where smaller is better, the new warning threshold value satisfies the expression:

[0097]

[0098] in, The new warning threshold, The initial warning threshold, For the first The model confidence level of the model used for batch prediction. This is the preset confidence threshold. Due to the model confidence... Less than ,therefore , This makes the new warning threshold The new warning threshold is increased to relax the warning threshold and reduce false alarms. Furthermore, the lower the model confidence, the lower the new warning threshold. The more it increases.

[0099] It should be noted that the confidence threshold is determined based on the relationship between the model prediction accuracy and confidence in historical data, and is usually set to 0.7-0.8. In this embodiment, it is set to 0.7. In other embodiments, implementers can set the confidence threshold value according to the actual implementation situation.

[0100] In this invention, the historical failure mode database is maintained regularly to ensure its timeliness and effectiveness. When the system detects a brand new failure mode, it automatically creates a new historical failure mode and adds it to the database, so that the historical failure mode database can be continuously improved as production experience is accumulated.

[0101] Appendix Figure 2 To be continued Figure 5 The technical effects of the method of the present invention are illustrated schematically under a simulation of multiple working conditions:

[0102] Figure 2 This is a schematic diagram showing the actual temperature changes in the heating furnace during the first process. Figure 2 The results show that the temperature of the first-stage heating furnace dropped significantly due to a sudden malfunction near batch 110.

[0103] Figure 3 This is a schematic diagram showing the actual changes in the rolling pressure during the second process. Figure 3 The original ring rolling pressure data of the second-stage ring rolling machine is displayed, which is mixed with chronic drift due to equipment wear, intermediate symptoms transmitted from the first-stage failure, and the second-stage failure itself that occurred near batch 124.

[0104] Figure 4 This is a schematic diagram showing the decoupling results of the ring rolling pressure residual in the second process. Figure 4 The invention demonstrates the residual pressure of the second process rolling pressure calculated by the present invention. This residual value successfully filters out the intermediate symptoms of chronic drift and the first transmission. The curve always stays near zero and only rises sharply when the second process itself fails, which confirms the decoupling of the fault and the adaptive tracking capability of the present method.

[0105] Figure 5This is a schematic diagram of the results of quality prediction and root cause diagnosis. Figure 5 The integrated results of quality prediction and root cause diagnosis are presented. The orange predicted tensile strength curve closely matches the gray actual tensile strength curve, achieving accurate early warning. Furthermore, the system accurately diagnosed the first process as the root cause when the quality first deteriorated. (See...) Figure 5 The red inverted triangle in the image accurately diagnosed the root cause of the second quality degradation as the second process. (See...) Figure 5 The blue inverted triangle in the image verifies the accuracy of the location of the fundamental cause of this invention.

Claims

1. A method for remote monitoring of the equipment status of a wind power flange production line, characterized in that, include: Collect process parameters and quality indicators for the initial qualified batches; The process parameters are divided into at least two process blocks according to the physical production sequence, including: dividing the process parameters into a first process block matrix, a second process block matrix, and a third process block matrix; Constructing a process transfer model includes: training a first initial transfer model with the first process block matrix as the independent variable and the second process block matrix as the dependent variable; and training a second initial transfer model with the first process block matrix and the second process block matrix as the independent variable and the third process block matrix as the dependent variable. Based on the process parameters of the upstream process block, the process parameters of the downstream process block are predicted, and the process residual between the actual process parameters of the downstream process block and the predicted values ​​of the process transfer model is calculated, including: taking the difference between the second process block matrix and the predicted values ​​of the first initial transfer model as the second process residual matrix; and taking the difference between the third process block matrix and the predicted values ​​of the second initial transfer model as the third process residual matrix. Based on the process parameters and process residuals of the upstream process blocks, an initial root cause analysis model and an initial soft measurement model are constructed, including: constructing an initial decoupling independent variable matrix, which includes the first process block matrix, the second process residual matrix, and the third process residual matrix; using the initial decoupling independent variable matrix as independent variables and the quality index as dependent variables, the initial root cause analysis model and the initial soft measurement model are trained. The process transfer model, the initial root cause analysis model, and the initial soft measurement model were all trained using the partial least squares regression algorithm. For new batches, collect real-time process parameters, including: collecting the first process parameter vector, the second process parameter vector, and the third process parameter vector for the new batch; Based on the real-time process parameters of the upstream process, the process parameters of the downstream process are predicted using the process transfer model updated in the previous batch. The real-time process residual between the real-time process parameters and the predicted values ​​of the process parameters of the downstream process is calculated, and a real-time feature vector is constructed. This includes: loading the first and second process transfer models updated in the previous batch; using the first process transfer model and the first process parameter vector to predict the predicted parameter vector of the second process; using the difference between the second process parameter vector and the predicted parameter vector of the second process as the second process residual; using the second process transfer model, the first process parameter vector, and the second process parameter vector to predict the predicted parameter vector of the third process; using the difference between the third process parameter vector and the predicted parameter vector of the third process as the third process residual; and constructing a real-time feature vector from the first process parameter vector, the second process residual, and the third process residual. The updated soft measurement model from the previous batch is used to predict the real-time quality indicators of the new batch based on the real-time feature vectors. In response to anomalies in the predicted values ​​of the real-time quality indicators, the updated root cause analysis model from the previous batch is used to analyze the real-time feature vectors of the new batch to locate the root cause of the anomalies. Obtain the actual quality indicators of the new batch. Using the real-time feature vector and actual quality indicators of the new batch, update the soft measurement model of the previous batch using the recursive least squares method to obtain the updated soft measurement model of the new batch. In response to the fact that the actual quality indicators of the new batch are qualified, use the real-time feature vector and actual quality indicators of the new batch to update the root cause analysis model of the previous batch using the recursive least squares method to obtain the updated root cause analysis model of the new batch.

2. The method for remote monitoring of the equipment status of a wind power flange production line according to claim 1, characterized in that, The root causes of the location anomaly include: Load the root cause analysis model updated in the previous batch; Based on the predicted values ​​of real-time quality indicators and the root cause analysis model, calculate the variable importance projection score of each variable in the real-time feature vector to the predicted values ​​of real-time quality indicators. Identify variables whose importance projection scores exceed a preset root cause contribution threshold, and determine the variable with the highest score as the root cause.

3. The method for remote monitoring of the equipment status of a wind power flange production line according to claim 1, characterized in that, After locating the root cause of the anomaly, the following also includes: Construct a historical failure mode database, which includes the root cause feature vectors of historical failure modes; The root cause feature vector of the current fault is constructed based on the real-time feature vector, the predicted value of the real-time quality index, and the updated root cause analysis model. The cosine similarity between the root cause feature vector of the current fault and the root cause feature vector of the historical fault mode in the historical fault mode database is calculated. In response to the maximum cosine similarity being less than the preset similarity threshold, the real-time quality index anomaly is marked as a suspected false alarm.

4. The method for remote monitoring of the equipment status of a wind power flange production line according to claim 1, characterized in that, Also includes: Calculate the time interval between the current batch and the most recent batch that has obtained actual quality indicators, and determine the model confidence of the model currently used based on the time interval; In response to the model confidence level falling below a preset confidence threshold, the warning threshold used to determine whether the predicted value of the real-time quality indicator is abnormal is relaxed.

5. The method for remote monitoring of the equipment status of a wind power flange production line according to claim 4, characterized in that, The confidence level of the model satisfies the expression: ； In the formula, Model confidence; This is the confidence decay coefficient; For the current batch; The time interval between the current batch and the most recent batch that has obtained actual quality indicators; It is a natural exponential function.

Images