Industrial equipment operation state evaluation method and system based on data analysis

Abstract

The application discloses an industrial equipment running state evaluation method and system based on data analysis, and the method comprises the following steps: pre-processing the collected mechanical equipment surface image data to extract wear characteristics; inputting the extracted wear characteristics into a pre-trained classification model to obtain a preliminary wear evaluation result; adjusting the model classification threshold based on the historical wear data distribution characteristics to optimize the evaluation accuracy; and recalibrating the evaluation result by using the optimized threshold to avoid overfitting or underfitting phenomenon. The application can solve the problem of how to adjust the model classification threshold based on the historical wear data distribution characteristics to solve the overfitting or underfitting problem of the detection result.

Description

Technical Field

[0001] This application relates to the field of mechanical equipment wear detection technology, specifically to a method and system for evaluating the operating status of industrial equipment based on data analysis. Background Technology

[0002] The data-driven industrial equipment operating status assessment method is an intelligent system that utilizes image recognition technology and data analysis methods to detect and assess the wear condition of mechanical equipment components. By collecting image data, extracting features, and combining this with efficiency evaluation algorithms, it can quickly and accurately determine the degree of wear on equipment, thereby reducing maintenance costs and extending equipment lifespan. However, this system faces a pressing problem: how to reasonably adjust the model's classification threshold based on the distribution characteristics of historical wear data to effectively avoid overfitting or underfitting in the detection results. Overfitting may cause the system to become overly reliant on a specific dataset, losing its universality, while underfitting will affect the detection effect due to insufficient accuracy. Therefore, it is necessary to introduce better data analysis and adjustment strategies to improve the system's stability and accuracy. Summary of the Invention

[0003] In view of this, the present invention discloses a method and system for evaluating the operating status of industrial equipment based on data analysis, which at least partially solves the problems existing in the prior art.

[0004] Data analysis-based methods for assessing the operational status of industrial equipment include:

[0005] Preprocessing is performed on the collected surface image data of mechanical equipment to extract wear features;

[0006] Preliminary wear assessment results are obtained by inputting the extracted wear features into a pre-trained classification model;

[0007] The model classification threshold is adjusted based on the distribution characteristics of historical wear data to optimize the evaluation accuracy;

[0008] The evaluation results are recalibrated using the optimized threshold to avoid overfitting or underfitting.

[0009] Preferred,

[0010] Preprocessing based on the collected surface image data of mechanical equipment includes smoothing and filtering noise components;

[0011] The intensity value I of each pixel is extracted by grayscale method, and the standardized feature is generated according to the following formula: F=(Iμ) / σ, where μ is the pixel mean of the historical image set and σ is the standard deviation;

[0012] The deviation coefficient Δ is calculated by comparing the model prediction error with the actual wear error.

[0013] If Δ < τ, where τ is an empirically set threshold, then the current classification result is considered to conform to the distribution characteristics.

[0014] Preferred,

[0015] The feature space density needs to be calculated before the extracted wear features are input into a pre-trained classification model;

[0016] The local feature density P(feature) = exp(|feature x_mean| / h) is calculated using kernel density estimation, where x_mean is the local mean and h is the bandwidth parameter.

[0017] The feature importance weight is calculated based on density: w = log(P(feature)) + β, where β represents the global adjustment factor.

[0018] Only high-confidence wear features with w>γ are retained for subsequent model use, where γ is a user-defined threshold.

[0019] Preferred,

[0020] Before optimizing the model classification threshold, the mean M and variance V of the wear degree of historical data were statistically analyzed;

[0021] The initial threshold is dynamically calculated using the following empirical formula: threshold = (M*k + α)*(V^5), where α is the calibration constant and k is the subjective scaling factor.

[0022] Compare the threshold with the current detection value;

[0023] When |difference|>λV, adjust threshold*=ratio (λ and ratio are hyperparameters) to balance overfitting or underfitting.

[0024] Preferred,

[0025] Analyze the cumulative percentage P(cdf) from the distribution characteristics of historical wear data and establish a cumulative probability model;

[0026] For a new input feature vector feature_vec, the threshold is adjusted based on the following conditions:

[0027] cdf_condition: cdf_sum = Σ(P(feature<=i)), threshold calibration is performed when P(cdf_sum p_target|feature_vec)>=δ;

[0028] The target cumulative percentage p_new = (ΣP + ∈) is updated, where ∈ is used to fine-tune the threshold;

[0029] Substitute p_new into the original threshold function to re-estimate the final evaluation criteria.

[0030] Preferred,

[0031] The optimized threshold correction includes addressing issues arising from the model's sensitivity to outlier data points (S).

[0032] After setting the initial detection threshold t_initial, the test sequence is run to obtain the sensitivity index:

[0033] Sensitivity = (1 / (t_tested_maxmin_feature))·|sum_diffs(features) / n|, where n is the number of test samples;

[0034] If the Sensitivity value exceeds the normal fluctuation range, the model threshold is adjusted to adjusted_threshold = t_original ± Δ_adjusted * S.

[0035] Preferred,

[0036] Additional environmental information, factor_env, is introduced to compensate for and correct the impact of wear level prediction on weights.

[0037] Define the compensation weight factor_θ = exp((distance(env,default) / ξ)^2), where ξ is the normalization factor;

[0038] The adjusted overall wear prediction value is:

[0039] final_predicted_wear=initial_predicted×factor_θ+β_fact or;

[0040] The classification threshold is flexibly adjusted based on whether factor_env exceeds the environmental disturbance baseline threshold Tenv. That is, if Tenv is violated, the readjustment phase is triggered to iterate again.

[0041] Preferred,

[0042] In the process of re-evaluating wear characteristics, a hybrid model decision scheme is formed by combining the output weights of multiple sub-models.

[0043] Let the output value Wi of sub-model i be defined as confidence_ci as ci = 1 / ((error(i) + γi)^p), p = 5;

[0044] The overall weight w_final = Σ(Wi*ci) / Σci is used to calculate the optimal weighted output combination, which is then input back into the optimizer.

[0045] Real-time fine-tuning based on error feedback from the sub-model set

[0046] classification_limit = limit_initial * sqrt((error_sum / max_er ror)^r), ensuring robustness of the evaluation.

[0047] Preferred,

[0048] Add dynamic verification logic for model stability to ensure that the classification model will not make significant misclassifications due to input noise;

[0049] Construct a stability test formula

[0050] delta_score = |(score_current_modelscore_reference)| / max_reference. When delta_score > threshold_validity, the alarm correction policy is triggered.

[0051] Set a reference validity value ref_val to dynamically limit the range of changes in classification boundaries;

[0052] The final adjustment was determined to be

[0053] valid_decision_boundary = current_boundary ± (reference_offs et * validity_ratio) balances accuracy and adaptability.

[0054] Preferred,

[0055] The optimization effect is further improved by using a multi-level evaluation hierarchy, and unnecessary redundant computational burden is reduced;

[0056] Set hierarchical weights

[0057] layer_weight_l = log(layer_output_sum_layer) / sum(total_output_l), where the contribution ratio of each layer is determined by recursive decomposition;

[0058] By calculating the prediction confidence intervals [lower_bound, upper_bound] for each stratum, the result closest to the target wear_estimate is selected as the main control basis;

[0059] The model learning curve slope_new = old_slope × decay_rate^(variance_level) is dynamically modified based on the convergence state of the confidence interval, so that the optimization is more inclined to the global optimum.

[0060] Preferred,

[0061] A dynamic self-feedback correction mechanism is incorporated to improve the stability and anti-interference capability of the classification process;

[0062] Calculate the real-time performance error: rate_error_current = (actual_wear_detected reference_detection_std) / std_actual;

[0063] Update the current classification threshold using the following dynamic correction rules.

[0064] boundary_corrected+=feedback_rate*tanh(error_rate_avg);

[0065] The feedback control module continuously receives error monitoring signals until the convergence accuracy convergence_goal is reached, at which point the iterative optimization process ends.

[0066] A data-driven industrial equipment operation status assessment system includes:

[0067] The preprocessing module is used to preprocess the acquired mechanical equipment surface image data to extract wear features;

[0068] The processing module is used to obtain preliminary wear assessment results by inputting the extracted wear features into a pre-trained classification model;

[0069] The optimization module is used to adjust the model classification threshold based on the distribution characteristics of historical wear data to optimize the evaluation accuracy;

[0070] The calibration module is used to recalibrate the evaluation results using optimized thresholds to avoid overfitting or underfitting.

[0071] The beneficial effects of this invention are:

[0072] This invention can solve the problem of overfitting or underfitting detection results by adjusting the model classification threshold according to the distribution characteristics of historical wear data. Attached Figure Description

[0073] Figure 1 This is a flowchart of the industrial equipment operation status assessment method based on data analysis according to an embodiment of the present invention;

[0074] Figure 2 This is a functional module structure diagram of the industrial equipment operation status assessment system based on data analysis according to an embodiment of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the embodiments of this disclosure will be further described in detail below with reference to the accompanying drawings. The illustrative implementation methods and descriptions of the embodiments of this disclosure are only used to explain the embodiments of this disclosure and are not intended to limit the embodiments of this disclosure.

[0076] Reference Appendix Figure 1 The data analysis-based industrial equipment operation status assessment method of the present invention includes the following main stages: preprocessing and extracting wear features based on mechanical equipment surface image data; analyzing the extracted features through a pre-trained classification model to obtain preliminary wear assessment results; adjusting the classification model threshold based on the distribution characteristics of historical wear data to optimize accuracy; and finally correcting the assessment results using the corrected threshold to ensure that overfitting or underfitting is reduced.

[0077] The first step involves acquiring surface images of the mechanical equipment as the raw data source, followed by preprocessing. Specifically, this step primarily includes noise removal and enhancement of the identifiability of target areas. For example, denoising algorithms such as mean filtering or Gaussian filtering are applied to the raw images to reduce image quality degradation caused by factors like lighting. Subsequently, edge detection techniques (like the Sobel operator) are applied to delineate the boundaries of mechanical components, thus more clearly identifying key areas of interest prone to wear. Morphological transformations are also used to eliminate small isolated points or fill in internal holes to further highlight the parts of interest. In one embodiment, if a gear surface exhibits corrosion due to long-term operation, the preprocessing stage will focus on enhancing the contrast between small, dark areas showing signs of corrosion and the surrounding clean areas to facilitate subsequent analysis.

[0078] After the image has been sufficiently preprocessed, the next crucial step is to extract effective feature vectors reflecting wear conditions from the new image for input into the machine learning model. This might involve using algorithms such as Local Binary Patterns (LBP) to measure the trend of fine-grained texture changes; or relying on Scale-Invariant Feature Transform (SIFT) to track significant geometric changes. All extracted features collectively construct a comprehensive representation space. For example, in the gear case mentioned above, this level might be able to quantitatively express a series of numerical indicators such as the percentage increase in tooth clearance width or the factor by which the roughness index increases.

[0079] The third task involves dynamically modifying the boundary parameter settings of the currently used classifier using a large amount of previously stored old wear data labeled with accurate status markers. Since the probability distribution of actual damage levels varies due to differences in equipment types and environmental conditions, it is necessary to appropriately set different tolerance ranges. For example, if statistics show that a small percentage of wear phenomena exceeding a certain level are present but not promptly warned of, potentially leading to serious consequences, then the early warning triggering limits can be slightly relaxed to expose more potential hazards. Alternatively, if multiple batches of similar tests reveal excessively high prediction accuracy but poor generalization performance, indicating overfitting in the training set, then the interval boundary tolerance should be appropriately increased to prevent oversensitivity and misjudgment of interference signals as genuine fault events.

[0080] The final step involves using the adjusted strategy after finalization to re-examine each of the original classification answers to determine if further revisions are needed to achieve the optimal decision-making quality in an overall balanced manner. Taking the gear example again, if the initial model tends to classify many actually minor, normal wear and tear items as major damage, then it's necessary to fine-tune the identification threshold based on empirical feedback until a balance between sensitivity and accuracy is achieved. This results in a robust and reasonable final conclusion report, which is then provided to maintenance personnel for reference, promoting the formulation of more scientific and effective maintenance plans.

[0081] Next, the present invention will be described. This system analyzes images of the surface of mechanical equipment using image processing technology to assess the wear condition of the equipment. Its specific implementation includes the following steps: preprocessing based on image acquisition data, grayscale extraction and standardized feature calculation, prediction error comparison deviation coefficient calculation, and deviation coefficient threshold determination to verify the validity of the results.

[0082] First, the preprocessing of image data of mechanical equipment surfaces includes a noise reduction process. This operation optimizes the quality of the original image by smoothing and filtering noise components, reducing the impact of random noise introduced by external interference or sensor limitations. For example, in industrial environments, cameras may capture dust or uneven lighting conditions; smoothing filtering can alleviate image distortion caused by such noise.

[0083] Second, color pixels are converted into single-channel intensity values ​​I using grayscale methods, and then standardized features are generated according to the formula F = (I - μ) / σ. In this formula, I represents the pixel intensity value of the current image, typically ranging from [0, 255]; μ is the average pixel value statistically derived from the historical image set, representing the overall grayscale characteristics; σ represents the standard deviation used to measure the width of the data distribution, and its positive value ensures the numerical stability of the formula. The significance of this standardization operation is to remove the interference caused by absolute brightness differences and enable the model to learn a more robust image representation that adapts to different acquisition conditions. For example, for a sample set of historical images showing significant changes in the operating conditions of a certain device, setting μ and σ can significantly improve the model's generalization ability, and experience shows that using normalized feature values ​​can effectively improve classification accuracy.

[0084] Third, based on the wear error obtained from the model prediction and the actual measurement, the degree of deviation is calculated using the formula Δ. The deviation coefficient Δ defines the fluctuation of the difference between the two within an acceptable range. For example, in one embodiment, when the system predicts a surface damage depth of 1.2 mm, while the actual value measured manually is 1.1 mm, the deviation coefficient can be used to quantitatively assess the level of difference between the two.

[0085] Fourth, the judgment criterion introduces an empirically set threshold τ to determine the reliability of the classification. If the Δ value is less than τ, the classification is accepted as a reasonable prediction output. Specifically, assuming that empirical values ​​indicate the maximum allowable deviation within the normal detection range should not exceed 0.15 mm, τ can be set to 0.15. This not only improves the accuracy of the results but also enhances the engineering practicality of the system. All the above steps work together to achieve the goal of automated detection.

[0086] Furthermore, the present invention also includes the following key steps: First, step one is performed, namely, calculating the feature space density. This step involves extracting data on the wear characteristics of mechanical equipment and using a kernel density estimation method to define the local feature density as the result of the formula P(feature) = exp(|feature - x_mean| / h). In this formula, x_mean is the local average value, representing the center point position within the current local feature data range; h, as the bandwidth parameter, determines the effective range of the kernel function, typically set within the range of (0.1, 1.5], with a default optimal value of 0.5. This formula quantifies the deviation of local feature values ​​from their average values ​​and obtains a probabilistic value after smoothing through an exponential function transformation.

[0087] The next step is step two, which calculates the importance weight *w* of each wear feature based on the density results. This is achieved using the formula log(P(feature)) + β. Here, the log function is used to adjust the influence of density on the weights, making the numerical distribution more balanced and stable; while β is a global adjustment factor, typically ranging from -2 to 2, depending on the feature scale and model preferences in the specific scenario. The formula's significance lies in transforming local density into an importance score that reflects information content, and then integrating the global influence to obtain the weights. This setup better adapts to the classifier's learning process.

[0088] The third step is to retain high-confidence features for model input based on the aforementioned weights. All feature weights must satisfy the condition w > γ to be considered high-quality detection criteria. Here, γ is a threshold value, generally ranging from (0,3), with a recommended optimal value of 1, used to distinguish between significant and non-significant features. This strategy effectively filters out low-confidence and noisy data portions to ensure subsequent prediction accuracy.

[0089] For example, in one embodiment, suppose a batch of industrial cameras acquire images of the surface of ball bearings. Multiple sets of wear feature data, including surface texture changes and crack widths, are extracted using image segmentation methods. In the first step described above, a bandwidth parameter h = 0.5 is used to calculate the spatial distribution density of each feature class. Then, in the second step, β = 0.8 is set to further measure and filter out the most prominent feature subset. Finally, after filtering out unreliable items according to a threshold γ = 1.2, the remaining reliable wear features are fed into a pre-trained deep learning classification model to determine whether severe wear conditions exist and their severity. This operation significantly improves the accuracy and reliability of the system.

[0090] Next, before optimizing the model's classification threshold, it's necessary to statistically analyze the mean (M) and variance (V) of wear levels from historical data. This step involves analyzing past equipment wear data to determine the average wear level and its fluctuation range, serving as the basis for subsequent dynamic adjustments. The mean (M) represents the overall level of wear on the mechanical equipment, while the variance (V) reflects the dispersion of the data points. These values ​​can be extracted from previous image processing results.

[0091] Secondly, a specific empirical formula is used to dynamically calculate the initial threshold (text{threshold}=(Mcdot k+alpha)cdot(V^5)). Here, (k) is a subjective scaling factor, set empirically between (0.8) and (1.2), with an optimal value of (1.0), representing the strength of the weight's effect on the mean. The constant (alpha) is used to calibrate the error, with a default range typically in ([-0.05,0.05]), used to fine-tune the final calculation result to avoid deviation. (V^5) in this formula represents the mechanism introduced by a higher-order nonlinear relationship, emphasizing that when the wear distribution has significant fluctuations, the threshold calculation should be more biased towards the sensitive range. The entire formula is set to make the initial threshold more consistent with the actual equipment conditions, rather than a single fixed value.

[0092] Next, (text{threshold}) is compared with the current detected value (text{difference}). If the absolute value of the current difference satisfies the condition (|text{difference}|>lambda V) (where (lambda) is a hyperparameter, usually set to a floating-point number between 1.5 and 3), then the next dynamic adjustment process begins. This comparison step determines the adjustment direction by assessing whether the detected difference significantly deviates from historical statistical characteristics. Only values ​​exceeding the reasonable fluctuation generator range will trigger the next level of operation.

[0093] Finally, when the deviation is too large, a threshold correction will be triggered: i.e. (text{threshold}*ratio), where the hyperparameter (text{ratio}) is set to ([0.95,1.05]) by default, in order to narrow or increase the classification boundary to the target value range that fits the new state, so as to avoid the risk of missed detection due to being too tight or the risk of misjudgment due to being too loose, and to balance the underfitting and overfitting problems in model performance.

[0094] In one embodiment, suppose a slight change appears on the surface of a rotating motor bearing that needs to be detected. Historical data shows that the average wear is 0.2 micrometers with small fluctuations (variance (V = 0.0003)). The selected parameters are (M = 0.2, k = 1.0, alpha = 0, lambda = 2, ratio = 1.02). Substituting these parameters, the initial threshold is approximately within a suitable range. If, after a certain run, image acquisition shows an abnormal difference of -0.06 micrometers, and the condition (|-0.06| > 2 * 0.0003 sqrt{)} is confirmed, the original threshold is updated by multiplying it by 1.02, and the new test standard is used iteratively until it converges to the normal range.

[0095] The above process realizes an automated decision-making method based on dynamic evaluation, which improves response flexibility while maintaining accuracy and effectively reduces errors that may be caused by human intervention.

[0096] Furthermore, this invention first analyzes the distribution characteristics of historical wear data and calculates the cumulative percentage P(cdf) to establish a cumulative probability model. The core of this process is to extract feature vectors from historical image data of the mechanical equipment and, based on the relationship between feature values ​​and wear status, calculate the proportion of each value to the total data and the sum of these proportions, thus forming the cumulative probability curve CDF (Cumulative Distribution Function). Determining the probability distribution of the feature vectors through statistical methods allows for a more comprehensive understanding of equipment wear trends and their changing patterns.

[0097] Then, for each new input feature vector `feature_vec`, it is determined whether threshold adjustment is needed. The specific decision is based on the `cdf_condition` formula. Here, `cdf_sum` is defined as the sum of the cumulative probabilities of all features not greater than the current `i`. When the difference between this cumulative probability and the preset target value `p_target` exceeds the error range δ (δ represents an adjustable small range of error tolerance, with the optimal value typically between 0.01 and 0.1), threshold calibration is performed. This setting ensures that the threshold accurately matches the distribution pattern of the new data within a certain range without deviating too much.

[0098] Next, update the target cumulative percentage p_new. Specifically, use the original cumulative probability value ΣP and add a correction parameter ∈, where ∈ is the fine-tuning step size, which may be set to a positive or negative value in some cases. Its optimal value depends on the data sensitivity and detection accuracy requirements, typically ranging from -0.05 to 0.05. Adjusting this parameter allows for flexible control of the threshold change rate, thus better adapting to different detection needs.

[0099] Finally, the newly estimated cumulative percentage p_new is substituted back into the original threshold function to recalculate the evaluation benchmark and generate the final result to determine whether there is abnormal wear or other changes in state.

[0100] For example, in one embodiment, assume a vision system for industrial machinery continuously captures images of gear assemblies as input. The system records gear thickness, edge sharpness, and other wear-related features to form an initial feature library, and obtains a cumulative probability curve through statistical analysis. If new images within a certain time period show a significant reduction in gear thickness (the corresponding new feature_vec shows an abnormal distribution deviation exceeding a set threshold), an automatic calibration procedure will be initiated based on the aforementioned formula to optimize the judgment conditions, and an alarm will be promptly triggered to prompt further maintenance actions. This not only improves the accuracy of equipment monitoring but also extends the lifecycle management period.

[0101] Next, the optimized threshold correction process involves several key steps to reduce the impact of abnormal data points on the accuracy of the detection system, thereby ensuring the efficiency and reliability of mechanical equipment wear detection.

[0102] The first step is to determine the initial detection threshold (t_{text{initial}}), which defines the acceptable level of wear during the detection process. This value is usually obtained through statistical analysis of historical data and set within the range of [0,1] based on the specific equipment performance parameters. The optimal value depends on the data distribution characteristics of the specific application scenario. It aims to provide a benchmark reference point for initial evaluation in actual testing.

[0103] The second step is to evaluate the test sequence by running a dataset containing multiple test samples to obtain the system's sensitivity index to each anomalous data point: (text{Sensitivity}=(1 / (t_{tested_max}}cdotmin(text{feature})))cdot|text{sum_diffs(features)} / n|). In this formula: (t_{tested_max}} refers to the maximum allowable threshold during the test; (min(text{feature})) represents the minimum value in the feature set of all test samples; (sum_diffs(features)}) is the sum of differences between different test results, used to reflect the strength of the system's response to specific conditions; (n) represents the total number of test samples. The formula is designed to measure whether the system can accurately capture subtle changes without being overly sensitive to spurious signals.

[0104] Subsequently, if the sensitivity value calculated using the above formula exceeds the preset fluctuation range, the third step is initiated: recalibrating the model threshold. The new threshold is determined by (adjusted_threshold = t_{original}}pmΔ_{adjusted}}cdot S), where the adjustment magnitude is influenced by the sensitivity level S. Δ is defined as a small but effective increment factor, adjusted between 0 and 1 (optimal range [0.1, 0.5]), aiming to achieve both rapid convergence and avoid unnecessary errors caused by excessively frequent parameter modifications.

[0105] In one embodiment, this process can be applied to monitoring the wear condition of wind turbine blades using image processing technology. First, a basic threshold for changes in blade surface texture under normal conditions is obtained through extensive experiments; for example, an initial threshold of 0.3 (corresponding to the definition of minor cracks or dirt). Second, the algorithm is run using hundreds of high-resolution images collected in the field as test samples, assuming these data contain a small number of outliers (such as crack expansion caused by bird strikes). Next, the model performance is evaluated using the aforementioned method, potentially resulting in a Sensitivity of 0.72 (i.e., excessive sensitivity requiring correction). Finally, the strategy is adjusted, appropriately reducing the tolerance limit; the updated threshold might become 0.32, thus better balancing accuracy requirements and noise resistance. This can significantly improve the long-term operational quality of the mechanical equipment condition monitoring system.

[0106] Furthermore, the present invention also includes the following process: the first step is to define an environmental compensation correction weight factor_θ; the second step is to calculate the comprehensive predicted wear value final_predicted_wear based on initial_predicted and factor_θ; the third step is to determine whether it is necessary to re-iterate the classification threshold for readjustment by comparing the environmental factor with the baseline threshold Tenv.

[0107] For each step, the specific meaning is as follows: First, calculate the environmental compensation correction weight used to adjust the accuracy of the initial prediction results. The formula is defined as factor_θ = exp((distance(env,default) / ξ)^2), where env represents the current set of environmental parameters, and default is the set of standard parameters corresponding to the ideal environment. distance(env,default) characterizes the degree of distance difference between the two, and ξ serves as a smoothing and normalization effect. A reasonable range for ξ is between 0.1 and 1, and in actual test scenarios, the optimal value is selected as approximately 0.5 based on experimental verification, thus avoiding the problem of excessively high sensitivity or excessively low sensitivity leading to biased results. The core purpose of this formula is to quantify the degree to which the environment deviates from the ideal state and to emphasize the adjustment strength when there is a large deviation by using an exponential decay method.

[0108] Then, using the initial wear prediction value `initial_predicted` and the `factor_θ` obtained in the previous step, along with another preset term `β_factor`, the overall evaluation is corrected to generate the final prediction output.

[0109] `final_predicted_wear` = `initial_predicted` × `factor_θ` + `β_fact` or . This calculation ensures that the final result depends not only on the original detection data but also incorporates optimized additional influences, improving the system's robustness and applicability.

[0110] The final step involves flexibly controlling the decision boundary value based on whether the factor_env (a quantitative indicator reflecting external disturbances) exceeds the set standard limit Tenv. For example, in one embodiment, when the temperature suddenly changes to an abnormal range during the operation of mechanical equipment, it is considered that this situation exceeds the predetermined Tenv, and the system will automatically call the readjustment program to adapt to the changing conditions.

[0111] Specifically, in image processing and efficiency evaluation systems, for example, when detecting wear in industrial ball bearings and discovering that increased dust-induced visual blurring leads to higher misclassification rates, diagnostic accuracy can be improved by adjusting the relevant `factor_env` parameter value above a threshold to activate the retraining process. These methods effectively address diverse challenges in practical applications and improve prediction accuracy.

[0112] Next, we will first list the specific steps of the data analysis-based industrial equipment operation status assessment method, and explain the meaning of each step one by one. Then, we will use examples to illustrate the working mechanism of the system.

[0113] The first step is to define the confidence weight calculation formula for each sub-model. The output value Wi of sub-model i is set as its confidence criterion. A confidence index, confidence_ci, is constructed by combining the error assessment error(i) and the regularization term γi (to avoid a denominator of zero). Specifically, ci = 1 / ((error(i) + γi)^p), where p = 5 is used to highlight the reliability of smaller errors. This parameter is chosen because it is necessary to significantly distinguish the difference between the outputs of high-precision and low-precision models.

[0114] The second step is to determine the overall weighting scheme for the hybrid model. The weights Wi and confidence levels ci of each sub-model are combined and optimized to obtain the final overall weight w_final = Σ(Wi*ci) / Σci. This weighting method ensures that decisions are tilted towards the more reliable and accurate model, thus forming a robust output. This process utilizes data input from multiple dimensions to improve the consistency of prediction results.

[0115] The third step involves dynamically fine-tuning the classification limit (classification_limit). This is adjusted in real time based on the cumulative error feedback from all sub-model sets, ensuring the system's ability to adapt to complex scenarios.

[0116] Classification boundary updates follow

[0117] In the formula limit_initial*sqrt((error_sum / max_error)^r), the parameter r is selected in the range [1.5, 2.5] for optimization experiments, with the aim of balancing overly sensitive and overly lenient conditions.

[0118] For example, in one embodiment, suppose we need to evaluate the early wear failure status shown on a set of industrial bearing images. Three different image preprocessing algorithms are used as sub-models to extract wear features. In the first stage, if a sub-model is found to have a high error of 0.2 due to excessive contrast, its influence in the overall hybrid model will decrease as the formula ci is set to a smaller value. Correspondingly, the sub-model with lower error is given a larger weight in the final weight. Furthermore, the initial classification_limit is set based on the equipment operating load, and as more data feedback accumulates over time, the system dynamically updates the classification threshold to an appropriate range.

[0119] The entire process effectively integrates image processing technology with automated evaluation logic, achieving accurate and efficient wear condition monitoring. By introducing a complex sub-model weighting and dynamic adjustment mechanism, the system exhibits higher robustness and potential for expanding its application scope.

[0120] Next, the invention will be described. First, dynamic verification logic for model stability is added. This means that a specific mechanism will be introduced into the system to continuously monitor the performance of the classification model. The purpose of this logic is to prevent misclassification due to noise input and to ensure the reliability of the system. Subsequently, a stability verification formula is constructed.

[0121] `delta_score` = |(score_current_model - score_reference)| / `max_reference`. This formula calculates the proportional relationship between the current model score and the reference model score, and standardizes it based on the maximum reference value. Key parameters include the current model score (`score_current_model`), the reference model score (`score_reference`), and the maximum value of the reference model (`max_reference`). `max_reference` is set to be greater than zero and to reflect the true distribution of the reference conditions as closely as possible, while the optimal value is determined experimentally. If the formula result exceeds the `threshold_validity` threshold, a correction strategy is triggered, which is typically set between 0.1 and 0.25.

[0122] Next, a reference validity value (ref_val) is set to dynamically limit the range of variation in the classification boundary. Specifically, to ensure the consistency and rationality of the decision rules, ref_val will be used as a key variable to control fluctuations, reflecting a standard amount of allowable variation. Finally, the balance between accuracy and adaptability is achieved by calculating the formula valid_decision_boundary = current_boundary ± (reference_offset * validity_ratio). In this formula, the selection of reference_offset and validity_ratio depends on the specific environment. The former serves as the offset baseline, typically choosing the optimal value between [0.05, 0.15]; the latter measures the impact on stability, taking a value between [0.6, 1] to maintain adaptability without sacrificing accuracy.

[0123] For example, in data analysis-based industrial equipment operating status assessment methods, a convolutional neural network is currently used as the core model for classifying surface cracks on a gear. However, it has been found to be quite sensitive to light intensity. In this environment, if dynamic stability logic is adopted, when noise is generated by changes in light intensity, it can determine whether correction is needed based on the constructed formula. Assuming that score_reference is approximately 0.93 (ideal case), while in actual testing score_current_model fluctuates slightly to 0.78, if delta_score exceeds 0.2 (threshold_validity is set to 0.2 to indicate a larger tolerance error rate), an alarm correction process is immediately implemented to avoid missed detections or misjudgments.

[0124] In one embodiment, a further adjustment is implemented as follows: Assuming the standard boundary value `current_boundary` of 0.65 represents the classification value at the boundary between normal and abnormal states, and due to long-term data accumulation, the classification boundary may drift and require recalibration, a new adjusted boundary is derived by comprehensively calculating the reference offset value `reference_offset` and the validity ratio. This effectively improves the accuracy of subsequent classifications while also considering flexibility requirements. Specifically, when `ref_val` is within a controlled range in the aforementioned gear case, the adjusted decision boundary remains within a reasonable confidence interval, while ensuring that it does not excessively favor any single aspect, leading to oversensitivity or coarse-grained errors.

[0125] Next, the system's performance is optimized and unnecessary redundant calculations are reduced based on the established multi-level evaluation hierarchy. By analyzing data at each level, resource waste is avoided, further improving detection accuracy and efficiency. For example, the feature point extraction and recognition process in images of mechanical equipment surfaces is divided into multiple stages, with each stage progressively refining the detection target.

[0126] Secondly, use the hierarchical weight formula.

[0127] The formula `layer_weight_l`, calculated as `log(layer_output_sum_layer) / sum(total_output_l)`, measures the contribution of each layer's output to the overall result. Here, `layer_output_sum_layer` represents the cumulative value of all outputs at the current layer, and `total_output_l` is the sum of all output values ​​across all layers in the system. This formula yields the specific weight for each layer, with weight values ​​ranging from 0 to 1. In mechanical equipment image processing, for example, some lower-level features (such as texture) may have smaller weights, while higher-level features (such as wear level) may have larger weights. This ensures that more important characteristics are considered to improve accuracy.

[0128] Furthermore, confidence intervals [lower_bound, upper_bound] are calculated using the prediction results from each layer, and the result closest to the estimated target wear_estimate is selected as the core reference. The range of the confidence interval is set according to the specific application; for example, a 95% confidence level corresponds to a value within two standard deviations. In one embodiment, the prediction of surface cracks on mechanical equipment may have several predicted values ​​distributed across different layers, but only the value with the highest confidence will be selected to guide the next action or decision.

[0129] Finally, dynamically adjust the slope of the learning curve.

[0130] `slope_new` = `old_slope` × `decay_rate`^(variance_level) This formula predicts the volatility state based on the current model. In this formula, `old_slope` represents the original model's learning rate; `decay_rate` is a preset decay constant, usually slightly less than 1 to prevent over-adjustment; and `variance_level` reflects the magnitude of the model's variation in recent batches of prediction data. If error fluctuations increase frequently, faster convergence is needed; otherwise, a relatively stable speed should be maintained. Specifically, for example, in the early stages of training a mechanical equipment image dataset, when large fluctuations in predictions are caused by significant differences in data quality, the range of change in each training step will be reduced accordingly to ensure that the model eventually converges to the optimal parameter combination, thereby improving overall performance.

[0131] Furthermore, the present invention elaborates on the content and specific significance of each step, and provides specific examples to aid in the explanation.

[0132] The first step is to introduce a dynamic self-feedback correction mechanism into the classification process. This is to improve the system's ability to resist interference and its overall stability when the classification equipment is worn. For example, in image processing-based detection systems, this mechanism can effectively reduce the impact of environmental noise, light intensity variations, and other factors on detection accuracy.

[0133] The second step is to calculate the real-time performance error, using the formula `rate_error_current = (actual_wear_detected - reference_detection_std) / std_actual` to measure the difference between the current classification and the standard value. The parameters are as follows: `actual_wear_detected` represents the actual wear level of the equipment currently determined by the detection system; `reference_detection_std` is the reference standard wear detection value; and `std_actual` is the standard deviation used for normalization to avoid dimensional conflicts. These parameters are generally derived from historical data analysis and the average value of multiple experiments, and their range can vary depending on the specific mechanical equipment. For example, in the bearing detection application for a specific type of machinery, the optimal value might be selected as an acceptable error range within a tolerance rate of ±5%. The logic behind the formula is to calculate the error rate to promptly identify problems and evaluate the real-time detection effect of the system.

[0134] The third step is to dynamically adjust the current classification threshold using specific rules, specifically by using the formula `boundary_corrected += feedback_rate * tanh(error_rate_avg)` to optimize the threshold. Here, `feedback_rate` defines the correction step size, typically ranging from 0.1 to 1 (the optimal value depends on the specific scenario requirements), while `tanh(error_rate_avg)` introduces an activation function characteristic to achieve non-linear correction intensity control. `error_rate_avg` refers to the cumulative average error rate over a certain period, aiming to comprehensively consider broader system performance rather than the influence of a single data point. This formula structure can smoothly adapt to various input conditions and quickly approximate the ideal state.

[0135] The fourth step involves the feedback control system continuously monitoring the aforementioned errors until the preset convergence accuracy (convergence_goal) is reached. The convergence target is typically determined by the expected performance level. For example, if the system needs to maintain a 98% accuracy rate, the corresponding target value might be set to a fluctuation range of less than 2% to be considered acceptable and end the iterative optimization process.

[0136] In one embodiment, for a specific industrial equipment production line, the above process can be specifically described as follows: After startup, the image sensor acquires image information about the condition of the rotating shaft surface and initially determines that there are signs of wear. At this point, the initially set classification threshold determines that this sample belongs to the minor damage category. However, actual inspection reveals a certain deviation from the expected benchmark, resulting in misclassification. Therefore, the formula is immediately called to calculate the current performance problem, and the decision boundary is quickly adjusted to perform the classification action again. Assuming that after several iterations, once the accuracy exceeds the design threshold, this round of automated optimization and adjustment workflow is terminated.

[0137] The data analysis-based industrial equipment operating status assessment method of the present invention includes: acquiring, preprocessing, and extracting features from surface images of mechanical equipment, combined with classification model prediction and threshold optimization correction, to achieve accurate assessment of equipment wear status. First, the system acquires image data of the surface state in the operating environment of the mechanical equipment, and uses image processing technology to denoise, normalize, and extract features from the acquired raw images to obtain effective feature vectors that reflect the actual wear status of the equipment. Subsequently, these extracted features are input into a pre-trained classification model using machine learning algorithms to generate preliminary assessment results.

[0138] A key innovation of this invention is how to adjust the classification threshold to address overfitting or underfitting during model operation. To ensure that the model's classification output matches reality, the inventors propose a method to adjust the classifier's operating point (i.e., the classification threshold) based on the distribution characteristics of historical wear data. Specifically, this step utilizes a large amount of clearly labeled historical wear data to construct a statistical distribution model and analyzes the probability boundary regions between different categories of wear data. When the detection system finds that the model output is overconfident in a certain category (i.e., high variance / overfitting), it indicates that the model may not have fully captured the true distribution of the test samples; conversely, when it finds that the output is generally too cautious for all categories (i.e., low variance / underfitting), it means that the model's adaptability is still insufficient.

[0139] Therefore, dynamically adjusting the decision boundaries based on data analysis becomes crucial to the solution. By calculating the prediction error rate and its distribution difference index for each type, and introducing an adaptive threshold mechanism to set reasonable boundary positions, the risk of misjudgment is controlled while ensuring sensitivity, ensuring that the assessment accuracy remains stable and adaptable to the needs of various actual working conditions. This not only reduces the risk of abnormal predictions by the model in complex environments but also improves the overall reliability and robustness of wear assessment.

[0140] like Figure 2As shown, the present invention also discloses an industrial equipment operation status assessment system based on data analysis. The above-mentioned functional modules of the industrial equipment operation status assessment system based on data analysis described in the embodiments of the present invention correspond to the various operation steps of the industrial equipment operation status assessment method based on data analysis of the present invention, which will not be repeated here.

[0141] The specific embodiments described above further illustrate the purpose, technical solutions, and beneficial effects of the embodiments of this disclosure. It should be understood that the above descriptions are merely specific embodiments of the embodiments of this disclosure and are not intended to limit the scope of protection of the embodiments of this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this disclosure should be included within the scope of protection of the embodiments of this disclosure.

Claims

1. A data analysis-based method for assessing the operating status of industrial equipment, characterized in that, include: Preprocessing is performed on the collected surface image data of mechanical equipment to extract wear features; Preliminary wear assessment results are obtained by inputting the extracted wear features into a pre-trained classification model; The model classification threshold is adjusted based on the distribution characteristics of historical wear data to optimize the evaluation accuracy; The evaluation results are recalibrated using the optimized threshold to avoid overfitting or underfitting.

2. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 1, characterized in that, Preprocessing based on the collected surface image data of mechanical equipment includes smoothing and filtering noise components; The intensity value I of each pixel is extracted by grayscale method, and the standardized feature is generated according to the following formula: F=(Iμ) / σ, where μ is the pixel mean of the historical image set and σ is the standard deviation; The deviation coefficient Δ is calculated by comparing the model prediction error with the actual wear error. If Δ < τ, where τ is an empirically set threshold, then the current classification result is considered to conform to the distribution characteristics.

3. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 2, characterized in that, The feature space density needs to be calculated before the extracted wear features are input into a pre-trained classification model; The local feature density P(feature) = exp(|feature x_mean| / h) is calculated using kernel density estimation, where x_mean is the local mean and h is the bandwidth parameter. The feature importance weight is calculated based on density: w = log(P(feature)) + β, where β represents the global adjustment factor. Only high-confidence wear features with w>γ are retained for subsequent model use, where γ is a user-defined threshold.

4. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 3, characterized in that, Before optimizing the model classification threshold, the mean M and variance V of the wear degree of historical data were statistically analyzed; The initial threshold is dynamically calculated using the following empirical formula: threshold = (M*k + α)*(V^5), where α is the calibration constant and k is the subjective scaling factor. Compare the threshold with the current detection value; When |difference|>λV, adjust threshold*=ratio (λ and ratio are hyperparameters) to balance overfitting or underfitting.

5. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 4, characterized in that, Analyze the cumulative percentage P(cdf) from the distribution characteristics of historical wear data and establish a cumulative probability model; For a new input feature vector feature_vec, the threshold is adjusted based on the following conditions: cdf_condition: cdf_sum = Σ(P(feature<=i)), threshold calibration is performed when P(cdf_sum p_target|feature_vec)>=δ; The target cumulative percentage p_new = (ΣP + ∈) is updated, where ∈ is used to fine-tune the threshold; Substitute p_new into the original threshold function to re-estimate the final evaluation criteria.

6. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 5, characterized in that, The optimized threshold correction includes addressing issues arising from the model's sensitivity to outlier data points (S). After setting the initial detection threshold t_initial, the test sequence is run to obtain the sensitivity index: Sensitivity = (1 / (t_tested_maxmin_feature))·|sum_diffs(features) / n|, where n is the number of test samples; If the Sensitivity value exceeds the normal fluctuation range, the model threshold is adjusted to adjusted_threshold = t_original ± Δ_adjusted * S.

7. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 6, characterized in that, Additional environmental information, factor_env, is introduced to compensate for and correct the impact of wear level prediction on weights. Define the compensation weight factor_θ = exp((distance(env,default) / ξ)^2), where ξ is the normalization factor; The adjusted overall wear prediction value is: final_predicted_wear=initial_predicted×factor_θ+β_fact or; The classification threshold is flexibly adjusted based on whether factor_env exceeds the environmental disturbance baseline threshold Tenv. That is, if Tenv is violated, the readjustment phase is triggered to iterate again.

8. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 7, characterized in that, In the process of re-evaluating wear characteristics, a hybrid model decision scheme is formed by combining the output weights of multiple sub-models. Let the output value Wi of sub-model i be defined as confidence_ci as ci = 1 / ((error(i) + γi)^p), p = 5; The overall weight w_final = Σ(Wi*ci) / Σci is used to calculate the optimal weighted output combination, which is then input back into the optimizer. Real-time fine-tuning based on error feedback from the sub-model set classification_limit = limit_initial * sqrt((error_sum / max_error)^r), ensuring robustness of the evaluation.

9. The method for evaluating the operating status of industrial equipment based on data analysis according to claim 8, characterized in that, Add dynamic verification logic for model stability to ensure that the classification model will not make significant misclassifications due to input noise; Construct a stability test formula delta_score = |(score_current_modelscore_reference)| / max_reference. When delta_score > threshold_validity, the alarm correction policy is triggered. Set a reference validity value ref_val to dynamically limit the range of changes in classification boundaries; The final adjustment was determined to be valid_decision_boundary = current_boundary ± (reference_offs et * validity_ratio) balances accuracy and adaptability.

10. A data analysis-based industrial equipment operation status assessment system, characterized in that, include: The preprocessing module is used to preprocess the acquired mechanical equipment surface image data to extract wear features; The processing module is used to obtain preliminary wear assessment results by inputting the extracted wear features into a pre-trained classification model; The optimization module is used to adjust the model classification threshold based on the distribution characteristics of historical wear data to optimize the evaluation accuracy; The calibration module is used to recalibrate the evaluation results using optimized thresholds to avoid overfitting or underfitting.

Images