SLCC converter valve state diagnosis method based on multi-parameter correlation residual error
By constructing a multi-parameter health correlation model and residual analysis, the problems of diagnostic lag and false alarm/missed alarm in converter valve condition detection were solved, enabling early warning and accurate diagnosis of sub-health conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-17
AI Technical Summary
In existing technologies, converter valve status detection relies on a fixed threshold method, which leads to high diagnostic lag, false alarm rate, and false negative rate, making it impossible to achieve early warning of sub-health conditions and easily affected by operating conditions.
A multi-parameter health correlation model (HCM) is constructed. By establishing the inherent physical laws of the converter valve, the residual between the actual measured value and the model's expected value is quantified, and multi-dimensional parameter correlation residual analysis is used for diagnosis.
It achieves highly sensitive early warning of sub-health conditions of converter valves, eliminates the influence of operating condition interference, and improves the accuracy and reliability of diagnosis.
Smart Images
Figure CN121682579A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of converter valve state monitoring, and particularly relates to a SLCC converter valve state diagnosis method based on multi-parameter associated residual. BACKGROUND
[0002] At present, the state detection of a converter valve (for example, SLCC), especially the monitoring of key parameters such as temperature and voltage, mainly relies on the "fixed threshold" method. For example, a fixed alarm threshold is set in the detection system, such as "alarm if the sub-module temperature > 90°C". This "fixed threshold" algorithm is the mainstream scheme at present, but it has serious technical defects.
[0003] Firstly, this method has significant diagnostic lag, and the system can only alarm when the fault has occurred and the parameter (such as temperature) has become very high and broken through the threshold. This "after-the-fact" alarm cannot achieve early warning of the "sub-health" state, missing the best maintenance opportunity.
[0004] Secondly, this method is easily disturbed by the operating conditions (such as load current and environmental temperature) of the converter valve, resulting in high false alarm rate and missed alarm rate. For example, when the converter valve is running under heavy load in summer, its sub-module temperature (such as 85°C) is already very high, and the traditional algorithm may "falsely alarm" frequently because it is "close to the threshold"; on the contrary, under light load in winter, the temperature (such as 60°C) of a sub-module seems "very healthy", but if the current is extremely low at this time, this 60°C temperature may actually indicate that its radiator is clogged or the fan has stopped, and the traditional algorithm will completely "miss the alarm" of this serious fault. The root cause is that the traditional algorithm has the "information silo" problem, which looks at each parameter in isolation, ignoring the inherent physical correlation between parameters. In a healthy converter valve, its "temperature" is necessarily strongly related to its "carrying current" and "cooling water temperature", and this "mismatch between operating conditions and state" is the earliest sign of failure, which cannot be identified by the prior art. SUMMARY
[0005] The purpose of the present application is to overcome the defects of the "fixed threshold" method in the prior art, such as diagnostic lag, strong operating condition interference, and high false alarm and missed alarm rate, and to provide a SLCC converter valve state diagnosis method based on multi-dimensional parameter correlation model and residual analysis. Instead of monitoring a single parameter in isolation, the present application establishes a "health correlation model" (HCM) that can reflect the inherent physical law of the converter valve, and quantifies the "residual" between the "actual measured value" and the "model expected value", to achieve high-sensitivity early warning of the sub-health state.
[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: A SLCC converter valve state diagnosis method based on multi-parameter correlation residual, comprising the following steps: (S1) constructing a multi-parameter health correlation model HCM; (S2) collecting multi-parameter operation data in real time; (S3) calculating the expected health temperature under the current working condition; (S4) calculating the real-time residual (Residual); (S5) diagnosis decision based on residual statistical process control.
[0007] Further, the step (S1) of constructing a multi-parameter health correlation model HCM specifically includes the following steps: (S1-a) health data acquisition and feature selection: when the converter valve is in the "known health" state, apply N group typical working conditions, the working conditions include different load currents I , different cooling water temperatures T water , and high-frequency synchronous acquisition of the radiator temperature of the IGBT power unit in the sub-module T sm , the different load currents I , different cooling water temperatures T water corresponding "state output parameters" radiator temperature T sm into a sample data set.
[0008] (S1-b) model structure definition: establish thermal characteristic model. The HCM model adopts a second-order multivariate nonlinear regression model, and its mathematical expression is:
[0009] wherein, is the model coefficient vector to be identified.
[0010] (S1-c) model coefficient identification (training): the solution of the coefficient vector is realized by minimizing the "cost function" , and the cost function is:
[0011] wherein, N is the total number of samples collected in (S1-a).
[0012] The minimization process uses gradient descent method for iterative solution, and the iterative update rule is:
[0013] wherein, For learning rate, For the first One coefficient, The loss function J(β) with respect to parameter β j The partial derivatives (i.e., the gradient direction, which determines the direction of parameter updates) are given.
[0014] (S1-d) After the model training (S1-c) is completed, all N sets of health condition data (i.e., different load currents) are processed. I Different cooling water temperatures T water Substitute the values into the pre-trained model to calculate N "fitted values". Calculate the difference between the original collected "actual measured values" and the "fitted values"; this difference is the "training residual".
[0015] (Note: Due to the inherent error in model fitting, the measurement points do not fall perfectly on the regression surface. Therefore, the residual objectively exists and follows a normal distribution.) Calculate the standard deviation of the training residuals for each group. This is then incorporated into the detection equipment as a "health baseline fluctuation threshold".
[0016] Furthermore, step (S2) involves real-time acquisition of multi-parameter operational data, specifically including the following steps: (S2-a) Data snapshot synchronous acquisition: at the current moment of routine inspection or online monitoring of the converter valve. t A data snapshot containing "operating condition input parameters" and "status output parameters" is synchronously acquired from the monitoring system of the converter valve; the data snapshot includes: real-time current. I now Real-time cooling water temperature and real-time submodule temperature ; (S2-b) Data Validity and Integrity Check: Perform integrity verification on the data snapshot obtained in (S2-a); if any key parameter... If data is missing (such as NaN or null values), the current diagnostic process will be terminated, and the data at that moment will be marked as "invalid" and await the next acquisition cycle. (S2-c) Data Pre-screening (Glitch Removal): To prevent erroneous measurements (such as sensor jumps or communication interference) from contaminating subsequent residual calculations, glitch removal is performed on the data that passed the (S2-b) verification. This removal uses the first-order rate of change threshold method, and its judgment rules are as follows:
[0017] in, This is the current sampled value. The value sampled at the previous moment. This represents the physically permissible maximum rate of change of the parameter (e.g., temperature). If this inequality holds, then determine... This refers to glitch data. The glitch data is filled with valid data from the previous time step, i.e. .
[0018] (S2-d) Valid data output: The clean data after processing (S2-c) The output is used as the input to the HCM model in the subsequent step (S3).
[0019] Furthermore, step (S3) calculates the expected health temperature under the current operating conditions: by applying the Health Association Model (HCM) identified and solidified in step (S1-c), the pure real-time operating condition data output in step (S2-d) is calculated to obtain the theoretical health status benchmark under the current operating conditions.
[0020] The calculation process specifically involves: using the real-time current obtained in step (S2-d) and real-time cooling water temperature Substituting these values into the mathematical expression of the thermal characteristic model defined in step (S1-b), The calculation formula is as follows:
[0021] in, It is the model coefficient vector that has been identified and solidified in step (S1-c). and It is the valid real-time operating condition input provided by step (S2-d).
[0022] The only output of the calculation This is the "expected healthy temperature," which represents the temperature that the converter valve should have under current operating conditions if it is still in a healthy state. This value will be output as a baseline value to the subsequent step (S4) for residual calculation.
[0023] Furthermore, the real-time residual calculation in step (S4) is to perform a quantitative comparison between the actual measured value and the model expected value to isolate the parameter changes caused by normal operating conditions (current, water temperature), thereby revealing the abnormal deviations caused by the deterioration of the equipment's health status.
[0024] Step (4) receives two core inputs: 1) Preprocessed real-time submodule temperature from step (S2-d) ; 2) From step (S3), the expected health temperature calculated by the HCM model based on the current operating conditions. .
[0025] The core calculation in step (4) is to obtain the "real-time residual". R now The calculation formula is as follows:
[0026] Among them, the "real-time residual" R now It is no longer an absolute temperature value, but a characteristic quantity representing the "net anomaly" after the influence of operating conditions has been eliminated. If R now A value approaching 0 indicates that the actual operating temperature of this submodule fully conforms to its health model under current operating conditions; if R now Significantly greater than 0 (e.g.) R now Exceed
[0027] This indicates that the submodule experienced an abnormal temperature rise under identical operating conditions compared to its healthy state. This deviation from the physical model suggests that early deterioration faults may have occurred inside the device, such as aging of thermal grease or increased contact thermal resistance.
[0028] The calculated "real-time residual" R now As the sole output, it is transmitted to the subsequent step (S5) for statistical process control diagnostics.
[0029] Furthermore, step (S5), based on the diagnostic decision of residual statistical process control, no longer uses a simple fixed threshold, but instead uses the "healthy baseline fluctuation threshold" obtained in (S1-d). Performing statistical process control (SPC) assessments specifically includes: 1) If If the real-time residual is within 2 standard deviations, then the submodule is considered to be in a "healthy" state.
[0030] 2) If If the real-time residual is between 2 and 3 times the standard deviation, then the submodule is judged to be in a state of "sub-health / early deterioration".
[0031] 3) If If the real-time residual exceeds 3 times the standard deviation, then the submodule is judged to be in a "fault" state.
[0032] Compared with the prior art, the beneficial effects of the present invention are: (1) Achieves "early warning" with high sensitivity: This invention calculates real-time residuals Rnow Has it broken through? This statistical threshold can isolate minute temperature drift of +3°C caused by aging thermal grease from strong operating conditions, detect potential "sub-health" problems in advance, and solve the problem of "diagnosis lag".
[0033] (2) Accurate and reliable diagnosis, eliminating operating condition interference: This method automatically removes the normal influence of operating conditions (current, frequency), and the diagnostic basis is the "unwanted" residual value. R now This avoids the false alarms and missed alarms that traditional methods may encounter under heavy or light loads.
[0034] (3) The technical solution is feasible and non-obvious: This invention provides a complete mathematical derivation and implementation path from model establishment (second-order regression), iterative solution (gradient descent) to residual calculation and statistical diagnosis (3-sigma method), which solves the problem that the "black box" model cannot be explained. Attached Figure Description
[0035] Figure 1 This is an algorithm flowchart of the method of the present invention. Detailed Implementation
[0036] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0037] Figure 1 This is a flowchart of the algorithm for the SLCC converter valve condition diagnosis method based on a multidimensional parameter correlation model and residual analysis, as described in this invention. Figure 1 As shown, the algorithm flow of this invention is mainly divided into two stages: offline model construction (S1) and online real-time diagnosis (S2-S5).
[0038] Phase 1: Offline Model Building (S1) First, during the factory testing of the converter valve, the offline model building (S1) phase is performed.
[0039] In step (S1-a) (health data acquisition and feature selection), under the "known health" state of the converter valve, 1000 different operating conditions are applied to it, and data is collected including (operating condition input: current). Water temperature (and status output: temperature) The dataset.
[0040] In step (S1-b) (model structure definition), the thermal characteristic model is defined. The mathematical expression for .
[0041] In step (S1-c) (model coefficient identification), the cost function is minimized using the gradient descent method, and regression analysis is performed on the above 1000 sets of data.
[0042] The calculated model coefficient vector is as follows: , , , , , .
[0043] At this point, the thermal characteristic model solidified in the testing equipment Confirmed.
[0044] Next, step (S1-d) (residual baseline statistics) is performed. 1000 sets of health data are substituted into the already solidified... The model calculates 1000 "training residuals". Statistical calculations show that the training residuals follow a normal distribution, with a standard deviation of [missing information]. The temperature is 0.4°C.
[0045] Should This "health baseline fluctuation threshold" is fixed and used for diagnosis in step (S5). At this point, the offline model building phase is complete.
[0046] Phase Two: Online Real-Time Diagnosis (S2-S5) After the converter valve is put into operation, the algorithm of this invention enters the online real-time diagnosis (S2-S5) stage.
[0047] Example 1: Health Status Diagnosis (1) Step (S2) Multidimensional real-time data acquisition: At time t1, step (S2-a) captures a data snapshot. The data then undergoes integrity checking (S2-b) and glitch removal (S2-c) (the judgment rules are as follows). After that, the valid data output by (S2-d) is: Real-time current
[0048] Real-time cooling water temperature
[0049] Real-time submodule temperature
[0050] (2) Step (S3) Calculate the "expected healthy temperature": Step (S3) applies the pre-defined HCM model to the above operating condition data ( , Substituting the known coefficient vector into its mathematical expression:
[0051] To simplify the presentation, the calculation results are given directly here. .
[0052] This value represents the temperature of the health valve should be 55.8°C under operating conditions of 800A and 30°C.
[0053] (3) Step (S4) Multidimensional residual calculation: Step (S4) performs residual calculation, the formula of which is:
[0054] Substitute the values: .
[0055] (4) Step (S5) Diagnostic decision based on residual statistical process control: Step (S5) received .
[0056] The cured portion in (S1-d) is known ,but .
[0057] The algorithm executes rule 1) in judgment (S5): .
[0058] because (Right now ( ), meeting the "health" requirement.
[0059] Diagnostic conclusion: Output "Healthy".
[0060] Example 2: Diagnosis of Sub-health (Early Deterioration) State This embodiment illustrates how the present invention achieves early warning.
[0061] (1) Step (S2) Multidimensional real-time data acquisition: exist At a given moment (e.g., the converter valve has been running for a year and the thermal grease has slightly aged), step (S2) captures a data snapshot. The current operating condition is exactly the same as in Example 1, but the valid data is: Real-time current
[0062] Real-time cooling water temperature
[0063] Real-time submodule temperature
[0064] (Actual temperature) (This is slightly higher than the health value of 56.1C in Example 1) (2) Step (S3) Parallel prediction of "expected healthy temperature": Step (S3) receives the exact same operating condition ( ).
[0065] Therefore, the HCM model calculates The temperature remains at 55.8°C. (The model only recognizes the operating conditions; it is unaware that the equipment has deteriorated.)
[0066] (3) Step (S4) Multidimensional residual calculation: Step (S4) performs residual calculation, the formula of which is:
[0067] Substitute the values: .
[0068] (The residual no longer approaches 0, but instead exhibits an abnormal temperature drift of +1.2°C).
[0069] (4) Step (S5) Diagnostic decision based on residual statistical process control: Step (S5) received .
[0070] Known ,but , .
[0071] The algorithm executes rule 2 in S5: .
[0072] because and (Right now ( ), which meets the criteria for "sub-health".
[0073] Diagnostic conclusion: Output "Sub-health / Early deterioration".
[0074] Example Analysis: In Example 2, the traditional "fixed threshold" method (e.g., setting the alarm threshold to 90°C) would observe 57.0°C, a value far below the threshold, thus completely "missing" the fault symptom. This invention successfully isolates the minute abnormal temperature drift (i.e., residual) of +1.2°C in step (S4), and in step (S5) achieves statistical diagnosis (reaching...). (Boundary) accurately identifies this "sub-healthy" state.
[0075] The undescribed parts involved in this invention are the same as or implemented using existing technology.
Claims
1. A method of diagnosing a state of a SLCC converter valve based on a multi-parameter associated residual error, characterized by, Comprising the following steps: (S1) constructing a multi-parameter health correlation model HCM; (S2) collecting multi-parameter operation data in real time; (S3) calculating the expected health temperature under the current working condition; (S4) calculating the real-time residual error; (S5) diagnosis decision based on residual statistical process control.
2. A method of SLCC converter valve condition diagnosis based on multi-parameter associated residuals according to claim 1, characterized in that, The step (S1) of constructing a multi-parameter health correlation model HCM specifically comprises the following steps: (S1-a) Health data collection and feature selection: While the converter valve is in a "known healthy" state, apply group typical operating conditions, including different load currents I , different cooling water temperatures T water , and high-frequency synchronous acquisition of the heat sink temperature of the IGBT power unit in the sub-module , different load currents I , different cooling water temperatures T water corresponding "state output parameters" heat sink temperature are made into a sample data set; (S1-b) Model Structure Definition: Establishment The thermal characteristic model, wherein the HCM model adopts a second-order multivariate nonlinear regression model, and its mathematical expression is: ； wherein, is the model coefficient vector to be identified; (S1-c) Model training: The solution of the coefficient vector is achieved by minimizing a "cost function" which is ； Wherein, N is the total number of samples collected in (S1-a); The minimum process uses gradient descent method for iterative solution, and its iterative updating rule is: ； wherein, is a learning rate, is a first coefficient, is a loss function J(β) derivative of the loss function β with respect to the parameter j . (S1-d) After the model training is completed, all N groups of health condition data, i.e. different load currents I , different cooling water temperatures T water are substituted into the trained model, and N "fitting calculation values" are calculated; the difference between the "actual measurement values" originally collected and the "fitting calculation values" is calculated, and the difference is the "training residual error": Calculate the standard deviation of each set of training residuals and use it as the "healthy baseline fluctuation threshold." 3. The method of claim 1, wherein, The step (S2) of collecting multi-parameter operation data in real time specifically comprises the following steps: (S2-a) Synchronous acquisition of data snapshot: at the current time of routine detection or online monitoring of the converter valve t , a data snapshot containing "working condition input parameters" and "state output parameters" is synchronously acquired from the monitoring system of the converter valve; the data snapshot includes: real-time current I now , real-time cooling water temperature , and real-time submodule temperature ; (S2-b) data validity and integrity check: the data snapshot obtained in (S2-a) is checked for integrity; if there is any missing data for any key parameter, the current diagnosis process is aborted, and the data at this time is marked as "invalid", and the next collection cycle is waited; (S2-c) data pre-screening: in order to prevent false measurement values from polluting the subsequent residual error calculation, the data passed in (S2-b) is removed from burr; (S2-d) effective data output: the pure data processed in (S2-c) is output as the input of the HCM model in the subsequent step (S3).
4. A method of SLCC converter valve condition diagnosis based on multi-parameter associated residuals according to claim 3, characterized in that, The step (S2-c) uses the first-order change rate threshold method for removal, and its judgment rule is: ； wherein, is the current sample value, is the previous sample value, and is the physically allowed maximum rate of change of the parameter; if the inequality holds, then the glitch data is identified; the glitch data is filled in with the valid data from the previous time instant, i.e. .
5. The method of claim 3, wherein, The step (S3) of calculating the expected health temperature under the current working condition is to apply the health correlation model HCM in step (S1-c) to calculate the pure real-time working condition data output in step (S2-d) to obtain the theoretical health state benchmark under the current working condition.
6. A method of SLCC converter valve condition diagnosis based on multi-parameter associated residuals according to claim 5, characterized in that, The calculation process is specifically: substituting the real-time current I now and the real-time cooling water temperature into the mathematical expression of the thermal characteristic model defined in step (S1-b) , and the calculation formula is: ； wherein, is the model coefficient vector determined in step (S1-c); I now and is the effective real-time operating condition input provided by step (S2-d); The only output of the calculation i.e. the "expected healthy temperature", which represents the temperature that the converter valve should have under the current operating conditions if it were still in a healthy state.
7. A method of SLCC converter valve condition diagnosis based on multi-parameter associated residuals according to claim 5, characterized in that, The real-time residual error calculation of the step (S4) is specifically: subtracting the real-time sub-module temperature output by the step (S2-d) from the expected health temperature under the current working condition calculated by the step (S3) to obtain a "real-time residual error", and the calculation formula is: Subtracting the real-time sub-module temperature output by the step (S2-d) from the expected health temperature under the current working condition calculated by the step (S3) to obtain a "real-time residual error", and the calculation formula is: ； If tends to 0, it means that the actual operating temperature of the submodule fully meets its health model under the current working condition; if is significantly greater than 0, it means that the submodule has an abnormal temperature rise under the same working condition as its health, and early degradation failure may occur inside the surface device.
8. A method of SLCC converter valve condition diagnosis based on multi-parameter associated residuals according to claim 7, characterized in that, The step (S5) of diagnosis decision based on residual statistical process control is: 1) If i.e. the real-time residual is within 2 standard deviations, then the sub-module status is judged as "healthy"; 2) if if the real-time residual error is between 2 and 3 times the standard deviation, the sub-module state is determined to be "sub-health / early deterioration"; 3) if i.e. the real-time residual exceeds 3 standard deviations, the sub-module status is determined as "fault".