Industrial equipment data analysis parameter adaptive configuration method based on operating condition characteristics

Abstract

The present application relates to the technical field of industrial internet data processing, in particular to an industrial equipment data analysis parameter adaptive configuration method based on operating condition characteristics; the method models the optimization process as a multi-stage state sequence decision process; the core is to construct a data analysis model processing state characteristic vector, and input the same into a parameter configuration module to calculate and select the candidate configuration parameter with the greatest value; after the parameter configuration is executed, the system dynamically corrects the parameter configuration module according to the deviation of the actual and expected processing state characteristics; the present application converts discrete static optimization into continuous dynamic decision sequence, so that the data processing control parameter can be self-adaptively adjusted according to the real-time state of the model, and the problem of fixed processing mode is solved.

Description

Technical Field

[0001] This invention relates to the field of industrial internet data processing technology, specifically to an adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics. Background Technology

[0002] With the increasing sampling frequency of sensors, data processing systems face the need for real-time analysis of massive amounts of time-series data. Existing data processing systems typically use fixed parameter configurations or static analysis models to process data under different operating conditions. However, in industrial scenarios, equipment operating conditions are highly variable and data distribution is uneven. Static processing solutions often cannot balance computational efficiency and fault identification accuracy. For example, on embedded industrial control computers with limited computing resources, using a high-complexity full-data analysis mode indefinitely can lead to computational resource overflow. On the other hand, using a low-complexity configuration may miss minor faults. Currently, there is a lack of an adaptive processing mechanism that can dynamically adjust analysis parameters and model parameters based on the current data processing status and system resource usage, resulting in low data processing efficiency and serious waste of system resources. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics. Specifically, the technical solution of this invention includes:

[0004] The optimization process of the data analysis model is modeled as a multi-stage state sequence decision-making process, specifically including the following steps:

[0005] Collect operational status data of industrial equipment as historical status data, and construct a benchmark test set containing the operating conditions of different equipment;

[0006] Construct a processing state feature vector for the data analysis model. The processing state feature vector represents the distribution of the data analysis model's accuracy in parsing different types of fault data at the current stage.

[0007] The processing state feature vector is input into the parameter configuration module to calculate the estimated performance value of each candidate configuration parameter in the processing parameter configuration space based on the performance evaluation function.

[0008] Select the candidate configuration parameter with the highest estimated performance value as the data processing control parameter for the next stage, and use the data processing control parameter to update and calibrate the data analysis model.

[0009] Based on the deviation between the actual processing state characteristics and the expected processing state characteristics generated after the execution of the next stage data processing control parameters, the parameter configuration module is dynamically corrected.

[0010] Preferably, the steps for constructing the processing state feature vector of the data analysis model include:

[0011] Select fault probe tasks for industrial equipment components from the benchmark task library, which are pre-labeled sensor data test subsets covering different fault modes and operating conditions;

[0012] The current data analysis model is used to classify and evaluate sensor data from all fault probe tasks to obtain fault identification performance scores.

[0013] The performance score is weighted according to the preset weight coefficients to generate a processing status feature vector; the weight coefficients are preset according to the importance of each fault type to the operation of the equipment.

[0014] Preferably, the performance evaluation function comprehensively considers the immediate performance evaluation value obtained after executing the configuration parameters and the potential processing value of subsequent states;

[0015] The immediate performance evaluation value is calculated from the multidimensional performance index function.

[0016] Preferably, the performance evaluation function, based on the principle of dynamic programming, determines the value of the current configuration parameter as the immediate performance evaluation value, which is the sum of the maximum potential processing value of all possible candidate configuration parameters in the new state reached after executing the current configuration parameter.

[0017] Preferably, the multidimensional performance index function includes:

[0018] The representation model analyzes the path efficiency term of the accuracy improvement in major fault identification tasks.

[0019] Knowledge retention terms that characterize the degree to which historical failure mode features are preserved;

[0020] The computational cost item for the computational resources consumed in the characterization phase.

[0021] Preferably, by introducing a preset reference cost constant, the path efficiency term and the computational cost term are processed into dimensionless values;

[0022] The path efficiency, knowledge retention, and computational cost items are weighted and summed using preset weight hyperparameters to generate an instant performance evaluation value.

[0023] Preferably, the steps for calculating the deviation include:

[0024] The system's preset state evolution prediction logic is invoked, and the expected processing state characteristics for the next stage are calculated based on historical processing state change records.

[0025] After executing the data processing control parameters for the next stage, the actual processing state characteristics are measured.

[0026] The difference between the expected processing state characteristics and the actual processing state characteristics is calculated using Euclidean distance to generate the bias.

[0027] Preferably, the dynamic correction of the parameter configuration module includes:

[0028] When the deviation exceeds the preset second threshold, the current processing status change record will be stored in the high-priority anomaly database.

[0029] When the deviation is greater than the preset first threshold but not greater than the second threshold, the parameter update step size of the parameter configuration module is reduced and the diversity range of configuration search is increased.

[0030] When the deviation is not greater than the first threshold, maintain the current parameters of the parameter configuration module;

[0031] The second threshold is greater than the first threshold.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. This method models the optimization process of the data analysis model as a multi-stage state sequence decision-making process and introduces a parameter configuration module for dynamic decision-making. It transforms a series of discrete static optimization stages into a continuous dynamic decision sequence, enabling the data processing control parameters to be adaptively adjusted according to the real-time processing state characteristics of the model, thus solving the problem of fixed data processing control parameters in the prior art.

[0034] 2. This method constructs a multi-objective real-time performance evaluation function, which can comprehensively balance the complex relationship between model performance improvement, the degree of retention of historical failure mode features, and computational resource consumption. At the same time, the performance evaluation function also considers the potential processing value of subsequent states, thereby guiding the model to evolve along the globally optimal path and effectively avoiding catastrophic forgetting and resource waste.

[0035] 3. This method introduces an auxiliary prediction model to calculate the deviation between the actual processing state characteristics and the expected processing state characteristics, thus constructing a key closed-loop feedback correction mechanism. This mechanism can dynamically correct the parameter configuration module in stages according to the magnitude of the deviation, giving the optimization framework the ability to reflect and correct errors, and significantly improving the robustness and stability of the optimization process under complex operating conditions.

[0036] 4. This method constructs a state vector that can accurately and comprehensively quantify the current capabilities of the data analysis model by selecting probe tasks from the benchmark task library and weighting the model's performance scores. This vector provides accurate and interpretable decision-making basis for subsequent parameter optimization, enabling the system to capture subtle changes in model capabilities and improving the perception accuracy and decision quality of adaptive optimization. Attached Figure Description

[0037] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0038] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0040] Example 1:

[0041] Please see Figure 1 An adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics models the optimization process of the data analysis model as a multi-stage state sequence decision-making process, specifically including the following steps:

[0042] This embodiment takes the transmission chain fault diagnosis of a wind turbine generator set as an example. The data analysis model accepts the timing signals collected by vibration acceleration sensors installed in the gearbox and main bearing, as well as the temperature and speed data recorded by the SCADA system. The output of the model is the health status label of the equipment, including various fault categories such as normal, inner ring wear, outer ring crack, and cage breakage.

[0043] Collect operational status data of industrial equipment as historical status data, and construct a benchmark test set containing the operating conditions of different equipment;

[0044] The processing state feature vector of the data analysis model is constructed. The processing state feature vector represents the distribution of the accuracy of the data analysis model in parsing different types of fault data at the current stage. The processing state feature vector is a structured numerical vector. Each dimension of the vector represents the normalized recognition accuracy of the model on different fault conditions such as high speed heavy load and low speed light load, or different fault types such as pitting and broken teeth sample sets. This enables a quantitative and interpretable representation of the model's multi-dimensional capabilities.

[0045] The processing state feature vector is input into the parameter configuration module to calculate the estimated performance value of each candidate configuration parameter in the processing parameter configuration space based on the performance evaluation function. Here, the candidate configuration parameter is not a simple combination of multiple independent parameters, but a structured parameter set that contains multiple key hyperparameters required to guide the next optimization stage, such as parameter update step size, batch size, data augmentation method, regularization coefficient, etc.

[0046] Select the candidate configuration parameter with the highest estimated performance value as the data processing control parameter for the next stage, and use the data processing control parameter to update and calibrate the data analysis model.

[0047] Based on the deviation between the actual processing state characteristics and the expected processing state characteristics generated after the execution of the next stage data processing control parameters, the parameter configuration module is dynamically corrected. It should be clarified that the dynamic correction mainly targets the internal control parameters of the parameter configuration module, such as the parameter update step size and search range, as well as the sampling strategy of historical state data, such as enabling a high-priority abnormal data retention mechanism, rather than modifying its inherent system architecture.

[0048] This embodiment provides an adaptive configuration method for data analysis parameters of industrial equipment based on operating condition characteristics. It models the optimization process of the data analysis model as a multi-stage state sequence decision-making process, thereby transforming a series of discrete static optimization stages into a continuous dynamic decision sequence. This method aims to solve the problem in the prior art where data processing control parameters are fixed and cannot adapt to the dynamic changes in the model's state. Its core lies in creating a closed-loop optimization framework that can self-perceive, self-evaluate, and self-adjust.

[0049] A processing state feature vector of the data analysis model is constructed. This processing state feature vector is a mathematical construct used to quantitatively characterize the comprehensive capability of the data analysis model at the end of the current optimization stage. Its function is to provide accurate and comprehensive decision-making basis for subsequent parameter optimization. Its source is the performance of the model on a series of preset tasks. In this embodiment, the vector is obtained by probe evaluation and weighted calculation of the model's capabilities in multiple dimensions, ensuring a comprehensive quantitative perception of the model's state.

[0050] The generated processing state feature vector is input to the parameter configuration module. This module is a decision model built on multi-layer nonlinear mapping logic, designed to learn the mapping relationship from a given processing state feature vector to the optimal candidate configuration parameters. In this embodiment, it acts as a core control unit, responsible for receiving the model's current state information and evaluating the long-term potential value of all candidate configuration schemes based on a meticulously designed performance evaluation function. Specifically, the parameter configuration module may employ a multi-layer parameter mapping topology containing at least two intermediate layers. The number of input nodes in the multi-layer parameter mapping topology is related to the processing state feature vector. The dimensions are matched, and the number of output computing nodes corresponds to the total number of all candidate configuration parameters in the configuration parameter space; the parameter configuration module receives... Perform forward logic computation to calculate each candidate configuration parameter. Corresponding performance value Before the optimization process begins, the parameter configuration module can obtain the initial configuration logic by performing offline parameter calibration on the historical optimization task dataset, or it can initialize with random weights and rely on a wide-range search for configuration optimization in the initial stage of optimization.

[0051] Based on the calculation results of the performance evaluation function, candidate configuration parameters with the maximum estimated performance value are selected. These candidate configuration parameters are a complex combination of multiple discrete and continuous parameters, including parameter update step size, batch size, data augmentation method, and regularization coefficient. Their role is to guide the specific execution of model parameter calibration in the next stage. The design of the performance evaluation function comprehensively considers the immediate performance improvement, the degree of retention of historical failure mode characteristics, and resource consumption after implementing the configuration parameters, and uses the principle of dynamic programming to predict long-term returns. The selection of the configuration scheme with the maximum value aims to ensure that the model moves along the globally optimal evolutionary path, rather than merely pursuing short-term local optima.

[0052] Perform the selected parameter iteration and optimization on the data analysis model for the next stage;

[0053] Based on the deviation between the actual processing state characteristics and the expected processing state characteristics generated after the execution of the next stage data processing control parameters, the parameter configuration module is dynamically corrected; this step constructs a key feedback loop; by establishing an auxiliary prediction model to predict the expected state after the parameter configuration is executed and comparing it with the actual state reached, the system can quantify the degree of unexpected occurrence; this deviation signal is used to dynamically adjust the learning parameters of the parameter configuration module, for example, when the deviation is too large, increasing the exploration randomness or storing unexpected samples in a high-priority anomaly database, thereby significantly improving the robustness of the optimization process and its adaptability to complex system fluctuations;

[0054] By modeling the optimization process of the data analysis model as a multi-stage state sequence decision-making process and introducing a parameter configuration module based on adaptive control logic for dynamic decision-making, this method changes the traditional optimization paradigm that relies on static and empirical rules. This method enables the optimization process to have adaptive capabilities, dynamically selecting the optimal evolution path based on the actual processing state characteristics of the model at each stage. This ensures maximum improvement in model performance while effectively avoiding catastrophic forgetting and balancing the consumption of computational resources, ultimately achieving global optimization of the entire data analysis model optimization process.

[0055] Example 2:

[0056] The steps for constructing a data analysis model to process state feature vectors include:

[0057] Select fault probe tasks for industrial equipment components from the benchmark task library, which are pre-labeled sensor data test subsets covering different fault modes and operating conditions;

[0058] The current data analysis model is used to classify and evaluate sensor data from all fault probe tasks to obtain fault identification performance scores.

[0059] The performance score is weighted according to the preset weight coefficients to generate a processing status feature vector; the weight coefficients are preset according to the importance of each fault type to the operation of the equipment.

[0060] This embodiment is a detailed explanation of the steps for constructing the processing state feature vector of the data analysis model in Embodiment 1. Its purpose is to ensure that the constructed vector can accurately, comprehensively and reproducibly reflect the real processing state characteristics of the data analysis model.

[0061] Fault probe tasks targeting industrial equipment components are selected from a benchmark task library. This benchmark task library is a collection of tasks of various types and difficulties, such as text classification, summary generation, code writing, and logical reasoning, serving as a benchmark for evaluating the multi-dimensional capabilities of the model. In this embodiment, the following tasks are selected: For example, The fault probe tasks cover the core capability dimensions required by the model; the selection principle is to ensure that the tasks are representative and distinguishable.

[0062] The current data analysis model is used to classify and evaluate sensor data from all fault probe tasks to obtain fault identification performance scores; in the current stage... At the end, the model was in the The performance score on each probe task is denoted as These scores have all been normalized, for example, mapped to... Intervals are used to eliminate the impact of inconsistent metrics for different tasks;

[0063] The performance score is weighted according to a preset weighting coefficient to generate a processing state feature vector; this weighting coefficient is a pre-defined set of parameters used to reflect the relative importance of different capability dimensions. The weighting is determined based on the business requirements of the final application scenario, and all weights must sum to 1. Through this weighting operation, the discrete performance scores are integrated into a structured processing state feature vector. Its mathematical expression is: This standardized feature engineering method provides a stable and information-rich continuous state representation for the input of the subsequent parameter configuration module.

[0064] The processing state feature vector constructed in the above manner not only quantifies the model's capabilities, but more importantly, it provides a highly structured and interpretable state representation through multi-dimensional probe tasks and a configurable weight system. Compared to a single, general evaluation metric, this vector can more accurately capture subtle changes and distributions in the model's capabilities, providing a solid data foundation for the parameter configuration module to make more precise decisions, thereby improving the perception accuracy and decision quality of the entire adaptive optimization framework. It should be noted that using the scores of a finite number of probe tasks as the processing state feature vector is a simplified representation; in practical applications, the scale and diversity of this task library can be expanded as needed, and the state vector can be further refined. The model further includes internal features such as model parameter distribution and activation value statistics to more comprehensively characterize the model state and alleviate the problem of insufficient state observation.

[0065] Example 3:

[0066] The performance evaluation function takes into account both the immediate performance evaluation value obtained after executing the configuration parameters and the potential processing value of subsequent states.

[0067] The immediate performance evaluation value is calculated from the multidimensional performance index function.

[0068] The performance evaluation function, based on the principle of dynamic programming, determines the value of the current configuration parameter as the immediate performance evaluation value, and the sum of the maximum potential processing value of all possible candidate configuration parameters in the new state reached after executing the current configuration parameter.

[0069] Multidimensional performance index functions include:

[0070] The representation model analyzes the path efficiency term of the accuracy improvement in major fault identification tasks.

[0071] Knowledge retention terms that characterize the degree to which historical failure mode features are preserved;

[0072] The computational cost item for the computational resources consumed in the characterization phase.

[0073] By introducing a preset reference cost constant, the path efficiency term and the computational cost term are treated as dimensionless values.

[0074] The path efficiency, knowledge retention, and computational cost items are weighted and summed using preset weight hyperparameters to generate an instant performance evaluation value.

[0075] This embodiment is a concentrated explanation of the core of the parameter adjustment mechanism in Embodiment 1, namely the performance evaluation function and its associated real-time performance evaluation value function; the synergistic effect of this set of features provides a rigorous mathematical basis and clear optimization guidance for the parameter configuration module to evaluate and select the optimal configuration;

[0076] This performance evaluation function comprehensively considers the immediate performance evaluation value obtained after executing the configuration parameters and the potential processing value of subsequent states. Its purpose is to overcome the limitations of short-sighted decision-making and ensure that the selected configuration parameters are optimal globally and in the long run. In this embodiment, this performance evaluation function... The recursive optimization formula, following the principles of dynamic programming, is constructed with the core idea of ​​determining the value of the current configuration parameter as its immediate performance evaluation value, and summing it with the maximum potential processing value of all possible candidate configuration parameters in the new state reached after executing the current configuration parameter. Here, dynamic programming is a core numerical calculation formula used for iteratively evaluating the long-term value of the current configuration scheme. Its special meaning in the specific technical context of this invention lies in transforming the engineering problem of optimizing a data analysis model into a dynamic programming problem for which an optimal solution can be sought mathematically. Its specific expression is:

[0077]

[0078] in, In the current processing state characteristics Configure parameters below The long-term performance expectation is a scalar value, calculated by the parameter configuration module. Is executing configuration parameters The immediate performance evaluation value obtained afterward is a scalar value, calculated by the following multi-dimensional performance index function that integrates the improvement of analytical accuracy and computational cost. It is a discount factor, a hyperparameter used to balance the importance of short-term and long-term effectiveness. The floating-point numbers in the interval are derived from preset values, such as those determined through cross-validation. ; Is executing configuration parameters The new processing state feature vector obtained by performing the fault probe task evaluation again is a vector, which is obtained by measuring by executing the steps in Example 2. In the new processing state characteristics All possible candidate configuration parameters The maximum expected value that can be brought by the middle; to solve the candidate configuration parameters This embodiment addresses the issue of including both discrete and continuous parameters by employing a parameter discretization method. Specifically, all continuous parameters are transformed into a finite set of discrete steps or ranges according to a preset step size or range. In this way, all discrete parameters and the discretized continuous parameters are combined to form a finite set of discrete configurations. ;

[0079] Therefore, in the principle of dynamic programming The calculation of the term is explicitly implemented: the state vector of the next stage is... The input parameter configuration module iterates through all candidate configuration parameters in this finite discrete configuration set. Calculate the corresponding The maximum value among the values ​​is selected as the result for that item.

[0080] To ensure that the above value assessment has a clear engineering orientation, immediate performance evaluation values ​​are required. It is calculated by a custom multidimensional performance index function; the innovative design of this function unifies the three mutually constraining core optimization objectives of path efficiency, knowledge retention and computational cost into a single evaluation signal; the multidimensional performance index function includes: a path efficiency term representing the efficiency of the model's performance improvement on the main fault identification task; a knowledge retention term representing the degree of retention of historical fault mode features; and a computational cost term representing the computational resources consumed in each stage.

[0081] To ensure that these three terms with different physical dimensions can be mathematically calculated, this embodiment introduces a preset reference cost constant to treat the path efficiency term and the computational cost term as dimensionless values; this reference cost constant... It is a benchmark value with cost dimensions, such as GPU hours serve as a normalized benchmark, derived from industry benchmarks set based on hardware limitations or typical task execution times. Through preset weighted hyperparameters, path efficiency, knowledge retention, and computational cost are weighted and summed to generate an instantaneous performance evaluation value. This weight hyperparameter , , It is a set of dimensionless coefficients used to adjust the priorities of different optimization objectives. These coefficients were determined through optimization methods such as grid search on the validation task; instantaneous performance evaluation values. The specific calculation formula is as follows:

[0082]

[0083] in, The model is divided into stages in the main fault identification task. arrive The improvement in analytical accuracy is a dimensionless value, derived from the evaluation. It is a stage The duration of data analysis and computational resource consumption is a cost-based value, measured in actual measurements; the path efficiency term is calculated by the improvement in analytical accuracy per unit of resource consumption. Multiply by the reference cost This is transformed into a standardized, dimensionless efficiency indicator; the knowledge retention item is... ,in The test score of the model on a set of historical typical fault samples is used to characterize the retention of memory of existing fault modes and prevent the model from forgetting the fault features under old conditions when learning new conditions. In order to ensure the numerical stability of the calculation and prevent division by zero error, a small positive smoothing term is added to the corresponding terms in the denominator in the above formula.

[0084] Through the collaborative design of the aforementioned performance evaluation function and multi-dimensional performance index function, this scheme achieves a deep and broad evaluation of data processing control parameters. It not only assesses the short-term effects of the current configuration scheme but also predicts its long-term impact through recursive state transition equations, effectively avoiding long-term performance damage caused by short-sighted choices. Simultaneously, it integrates the three core indicators of performance improvement, knowledge retention, and computational cost into a unified, dimensionless comprehensive performance evaluation function, providing a comprehensive, balanced, and quantifiable guiding signal for model optimization. This makes parameter tuning more scientific and robust, ultimately guiding the model to achieve optimal capability iteration under controllable costs. It should be noted that the weighted summation method used in this embodiment is a simplified multi-objective fusion mechanism, which assumes a linear relationship between different objectives. In other embodiments, the immediate performance evaluation value... The computational logic can also take a nonlinear form to more accurately characterize the complex nonlinear trade-offs between different objectives.

[0085] Example 4:

[0086] The steps for calculating deviations, in equipment fault identification scenarios, and the actual processing status characteristics. It is a vector obtained by deploying the updated model to the validation set and calculating its classification F1-Score or accuracy for various fault signals in the validation set; bias This reflects the prediction error of the parameter configuration module regarding the model's learning ability under a specific fault data distribution, including:

[0087] The system's preset state evolution prediction logic is invoked, and the expected processing state characteristics for the next stage are calculated based on historical processing state change records.

[0088] After executing the data processing control parameters for the next stage, the actual processing state characteristics are measured.

[0089] The difference between the expected processing state characteristics and the actual processing state characteristics is calculated using Euclidean distance to generate the bias.

[0090] Dynamic adjustments to the parameter configuration module include:

[0091] When the deviation exceeds the preset second threshold, the current processing status change record will be stored in the high-priority anomaly database.

[0092] When the deviation is greater than the preset first threshold but not greater than the second threshold, the parameter update step size of the parameter configuration module is reduced and the diversity range of configuration search is increased.

[0093] When the deviation is not greater than the first threshold, maintain the current parameters of the parameter configuration module;

[0094] Wherein, the second threshold is greater than the first threshold; the first threshold It is a preset upper limit for tolerance of deviation based on the statistical characteristics of historical processing deviation data, and its specific value is set according to the historical deviation distribution. Quantiles are used to distinguish between random fluctuations during normal system operation and significant deviations from the system's state; the second threshold... To identify the alarm limit for high-learning-value anomalies, its value is set to [value]. ,in The preset multiplier and .

[0095] This embodiment is a concretization of the dynamic correction mechanism in Embodiment 1. Its purpose is to improve the stability and robustness of the entire adaptive optimization method when facing complex and uncertain data flow characteristics. The mechanism forms a closed-loop feedback system through deviation identification and hierarchical correction.

[0096] The deviation calculation is precisely defined to quantify the gap between the actual optimization effect and the expected effect; the system's pre-set state evolution prediction logic is invoked, specifically employing a time series prediction model based on long short-term memory networks or gated recurrent units, to calculate the expected processing state characteristics for the next stage based on historical processing state change records; this auxiliary prediction model can specifically employ a time-series recursive analysis model with memory function to effectively capture the time-series dependencies of state transitions; its input is the current processing state characteristics. and selected configuration parameters The combined representation of these features outputs the predicted expected processing state characteristics for the next stage. This auxiliary prediction model can be calibrated synchronously with the parameter configuration module, utilizing processing status change records collected during the optimization process. Parameter correction based on historical deviation data can be performed, or offline pre-calibration can be performed using historical data before the main loop begins;

[0097] This auxiliary prediction model is based on historical data, such as The model for learning and processing state change records is designed to address the characteristics of the current processing state. and selected configuration parameters This provides an expected feature of the next stage of processing. After executing the data processing control parameters for the next stage, the actual processing state characteristics are measured. The difference between the desired processing state features and the actual processing state features is calculated using Euclidean distance to generate a bias. Here, Euclidean distance is the true distance between two points in multidimensional space, and its role in this invention is to provide a standard, dimensionless scalar to quantify the difference between two processing state feature vectors. (Bias) The calculation formula is:

[0098]

[0099] After calculating the deviation Subsequently, the dynamic correction of the parameter configuration module will trigger different response mechanisms based on the magnitude of the deviation. This is a hierarchical control mechanism based on dual thresholds; the first threshold... The first threshold is a preset, relatively small upper limit for tolerance of deviation. Its technical basis is to cover the range of random fluctuations during normal system operation, such as the 95th percentile of historical deviation data, used to distinguish between normal prediction errors and significant deviations. The second threshold is a preset, relatively large upper limit for deviation alarms. In this embodiment, its value can be set to... ,in For preset coefficients, such as This is used to identify state transition events with high learning value.

[0100] The specific correction logic is as follows: when the deviation Not greater than the first threshold At that time, that is This indicates that the actual situation matches expectations and the system is stable. At this point, the current parameters in the parameter configuration module should be maintained without further adjustments; when deviations occur... Greater than the preset first threshold And not greater than the second threshold, i.e. This indicates a deviation from the normal range, suggesting a possible misinterpretation of the environment by the parameter configuration module. In this case, the system will reduce the parameter update step size of the parameter configuration module and increase the diversity of configuration searches; when the deviation... When it exceeds the preset second threshold, that is This indicates that a serious unexpected event has occurred, and the current processing status has been changed. It contains unconventional and valuable learning information; at this time, the system stores the current processing status change record in the high-priority anomaly database; the high-priority anomaly database is a key historical data storage mechanism in the adaptive optimization process, which will prioritize the selection of samples with a high degree of unexpectedness for backtracking calibration analysis, so that the parameter configuration module can extract patterns from these rare but key historical records more efficiently.

[0101] By introducing this closed-loop correction mechanism based on prediction bias, this scheme endows the optimization framework with the ability to reflect and correct errors; it is no longer a simple execution and update, but a continuous comparison between expectations and reality; this hierarchical correction logic ensures that the system can remain efficient when stable, and can make robust and timely adjustments when deviations occur, which greatly improves the robustness and long-term stability of the entire adaptive optimization method to the evolution of unknown working conditions, and ensures that the optimization process can converge to the ideal state even in complex scenarios.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive configuration method for data analysis parameters of industrial equipment based on operating condition characteristics, wherein the data analysis model is used to process multi-dimensional sensor time-series data during the operation of industrial equipment and output the equipment fault type, characterized in that, The optimization process of the data analysis model is modeled as a multi-stage state sequence decision-making process, specifically including the following steps: Collect operational status data of industrial equipment as historical status data, and construct a benchmark test set containing the operating conditions of different equipment; A processing state feature vector is constructed for the data analysis model. The processing state feature vector represents the distribution of the data analysis model's accuracy in parsing different types of fault data at the current stage. Each dimension of the processing state feature vector represents the normalized recognition accuracy of the model on different fault conditions or different fault type sample sets. The processing state feature vector is input to the parameter configuration module to calculate the estimated performance value of each candidate configuration parameter in the processing parameter configuration space based on the performance evaluation function. The parameter configuration module adopts a multi-layer parameter mapping topology containing at least two intermediate layers. The number of input computing nodes of the multi-layer parameter mapping topology matches the dimension of the processing state feature vector, and the number of output computing nodes corresponds to the total number of all candidate configuration parameters in the configuration parameter space. Select the candidate configuration parameter with the highest estimated performance value as the data processing control parameter for the next stage, and use the data processing control parameter to update and calibrate the data analysis model. Based on the deviation between the actual processing state characteristics and the expected processing state characteristics generated after the execution of the next stage data processing control parameters, the parameter configuration module is dynamically corrected. The performance evaluation function, based on the principle of dynamic programming, determines the value of the current configuration parameter as the immediate performance evaluation value, and the sum of the maximum potential processing value of all possible candidate configuration parameters in the new state reached after executing the current configuration parameter; The immediate performance evaluation value is calculated from the multidimensional performance index function; Multidimensional performance index functions include: The representation model analyzes the path efficiency term of the accuracy improvement in major fault identification tasks. Knowledge retention terms that characterize the degree to which historical failure mode features are preserved; The computational cost of the computational resources consumed during the characterization phase; By introducing a preset reference cost constant, the path efficiency term and the computational cost term are treated as dimensionless values. The path efficiency item, knowledge retention item, and computational cost item are weighted and summed using preset weight hyperparameters to generate an instant performance evaluation value. The formula for calculating the real-time performance evaluation value is: In this formula, The representation model consists of stages in the main fault identification task. arrive The improvement in resolution accuracy; Representation stage The duration of data analysis and computational resource consumption; It belongs to the preset reference cost constant; , , These are preset weight hyperparameters; The test score of the representation model on a set of typical historical fault samples; The test score of the model on a set of historical typical failure samples after executing the configuration parameters; and It belongs to the positive value smoothing term.

2. The adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics according to claim 1, characterized in that, The steps for constructing a data analysis model to process state feature vectors include: Select fault probe tasks for industrial equipment components from the benchmark task library; The current data analysis model is used to classify and evaluate sensor data from all fault probe tasks to obtain fault identification performance scores. The performance score is weighted according to the preset weight coefficients to generate a processing state feature vector.

3. The adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics according to claim 1, characterized in that, The steps for calculating the deviation include: The system's preset state evolution prediction logic is invoked, and the expected processing state characteristics for the next stage are calculated based on historical processing state change records. After executing the data processing control parameters for the next stage, the actual processing state characteristics are measured. The difference between the expected processing state characteristics and the actual processing state characteristics is calculated using Euclidean distance to generate the bias.

4. The adaptive configuration method for industrial equipment data analysis parameters based on operating condition characteristics according to claim 3, characterized in that, Dynamic adjustments to the parameter configuration module include: When the deviation exceeds the preset second threshold, the current processing status change record will be stored in the high-priority anomaly database. When the deviation is greater than the preset first threshold but not greater than the second threshold, the parameter update step size of the parameter configuration module is reduced and the diversity range of configuration search is increased. When the deviation is not greater than the first threshold, maintain the current parameters of the parameter configuration module; The first threshold is a tolerance threshold determined based on the statistical distribution of historical processing state deviation data, used to distinguish between normal fluctuations and system deviations; the second threshold is a preset upper limit for anomaly identification, used to determine high-value training samples that need to be stored in the high-priority anomaly database, and the second threshold must be greater than the first threshold.

Images