A weightless health degree analysis method of an industrial equipment

Abstract

The application discloses a kind of industrial equipment's health degree analysis method without weight, belong to electric digital data processing technical field, comprising: in healthy state, according to the data index type for analyzing health degree, respectively collect at least one historical time window of industrial equipment industrial time series data, as reference data, collect the industrial time series data in current time window, as comparison data;According to reference data and comparison data, one by one through kernel density estimation, obtain probability density curve, to obtain probability distribution;According to data index type, the probability distribution of comparison data and the probability distribution of at least one reference data corresponding, obtain the JS divergence value for measuring data distribution similarity under data index type, to obtain the health degree of industrial equipment according to multiple JS divergence values.The application is based on kernel density estimation and JS divergence, quantifies distribution difference, without weight setting, improves health degree evaluation sensitivity and consistency.

Description

Technical Field

[0001] This invention belongs to the field of electrical digital data processing technology for health status, specifically relating to a weightless health status analysis method for industrial equipment. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, health analysis of industrial equipment has become one of the core technologies for ensuring production continuity and reducing operation and maintenance costs. In process industries such as water, power, and chemicals, sudden equipment failures (such as pump sets, compressors, and piping systems) can lead to production line shutdowns, causing economic losses of millions or even tens of millions of yuan. Therefore, real-time and accurate health analysis technology has become an essential requirement for the industry.

[0003] Therefore, threshold-based analysis methods have emerged, which monitor health status by setting a safety threshold for a single indicator (such as an alarm for temperature >80℃). However, this approach can only detect obvious anomalies and cannot reflect the gradual degradation process of equipment. For example, equipment in a healthy state may gradually deviate from the normal distribution due to small fluctuations, but the values ​​are still within the threshold range, making it difficult to detect potential faults in a timely manner.

[0004] To address the limitations of single-indicator threshold methods, existing technologies are gradually shifting towards comprehensive scoring methods based on multi-indicator fusion. These methods calculate device health through weighted summation, such as the analytic hierarchy process (AHP), expert scoring, or machine learning regression models.

[0005] While such methods can combine information from multiple indicators, existing methods tend to focus on the "numerical deviation" of indicators (such as the difference between real-time values ​​and the mean) rather than the "distribution deviation." The health status of industrial data is essentially a stable probability distribution (such as pressure following a normal distribution under normal operating conditions), and a single numerical deviation cannot reflect the overall degradation of the distribution pattern (such as an increase in the range of data fluctuations, peak shifts, etc.).

[0006] Furthermore, the core flaw of the comprehensive scoring method lies in its reliance on weights. Whether weights are assigned manually based on experience (e.g., vibration index weighted at 0.3, temperature weighted at 0.2) or automatically learned by algorithms (e.g., neural networks), the indicators must be weighted through "importance ranking." This approach is limited by human experience or the quality of data annotation, often lacks theoretical basis, may overlook the cumulative degradation effect of weakly correlated indicators, and different assessment subjects may give significant differences in the health scores of the same equipment, affecting the consistency of industrial operation and maintenance decisions. For example, if an indicator fluctuates little on its own but deviates from a healthy state in conjunction with other indicators, the weighting method may miss it.

[0007] Moreover, the weight settings need to correspond to the semantics of the indicators (such as the need to clearly define the names of "vibration amplitude" and "export pressure"). When the names of the indicators are ambiguous or new monitoring dimensions are added, the model needs to be readjusted, which affects the adaptability of the model.

[0008] Therefore, there is an urgent need to develop a weightless health analysis method for industrial equipment to address the problems of insufficient utilization of data distribution characteristics, unreasonable weight settings, and limited model universality in existing technologies. Summary of the Invention

[0009] This invention provides an unweighted health analysis method for industrial equipment, which solves the problems of insufficient utilization of data distribution characteristics, unreasonable weight settings, and limited model universality in the prior art.

[0010] The technical solution adopted in this invention is as follows:

[0011] A method for unweighted health analysis of industrial equipment, comprising:

[0012] In a healthy state, according to the data index type used to analyze health, industrial time-series data within at least one historical time window of the industrial equipment are collected as baseline data, and industrial time-series data within the current time window are collected as comparison data.

[0013] Based on the baseline data and the comparison data, probability density curves are obtained one by one through kernel density estimation to obtain the probability distribution;

[0014] Based on the data indicator type, the probability distribution of the comparison data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distribution under the data indicator type, so as to obtain the health of the industrial equipment based on multiple JS divergence values.

[0015] The unweighted health analysis method for industrial equipment disclosed in this invention also has the following additional technical features.

[0016] After industrial time-series data acquisition, the following are also included:

[0017] The acquired industrial time-series data is preprocessed.

[0018] Check the industrial time-series data for null values ​​and perform a deletion operation on the null values.

[0019] Linear normalization is performed on the industrial time series data after removing null values ​​to maintain the data distribution characteristics of the industrial time series data.

[0020] The historical time window and the current time window are specifically as follows:

[0021] The time window sizes of the historical event window and the current time window are determined based on the data feature frequency of the corresponding data indicator type, and must be greater than the feature period corresponding to the minimum data feature frequency.

[0022] Furthermore, the data collection time is greater than or equal to 1 hour, and the amount of data collected within the time window is greater than or equal to 500.

[0023] The time window size for each of the historical event windows and the current time window can be set independently.

[0024] At least one historical time window of industrial time-series data is used as the baseline data, specifically:

[0025] Based on the data indicator type, different historical time windows are selected, and / or,

[0026] Multiple sets of industrial time-series data from the industrial equipment under different operating modes are collected to obtain multiple benchmark data under the data indicator type.

[0027] The probability density curve is obtained through kernel density estimation, as follows:

[0028] Based on the selected kernel function, the kernel density is estimated using the benchmark data and the comparison data, resulting in a probability density curve. The kernel function includes at least one of the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel.

[0029] The selection of the kernel function is specifically as follows:

[0030] Based on the data indicator type, one of the benchmark data and the comparison data is selected, and kernel density estimation is performed using multiple kernel functions to obtain multiple probability density curves for each data indicator type.

[0031] Based on the multiple probability density curves, the asymptotic mean square error is obtained by estimating the deviation and variance. The kernel function corresponding to the smallest asymptotic mean square error is selected as the kernel function under the data index type.

[0032] The probability density curve is obtained to determine the probability distribution, specifically:

[0033] Based on the probability density curve, multiple intervals are divided, wherein the number of intervals is determined by the theoretical bandwidth obtained from the standard deviation of industrial time series data, and the number of intervals for the comparative data and the corresponding benchmark data for each data index type is set to be consistent;

[0034] The interval probability corresponding to each interval is obtained, forming a probability distribution of the comparison data and the corresponding benchmark data.

[0035] Based on the data indicator type, the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distributions under the data indicator type, specifically:

[0036] By analyzing the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data, at least one JS divergence value can be obtained.

[0037] When multiple JS divergence values ​​are obtained from multiple benchmark data, the average of the multiple JS divergence values ​​is taken to obtain the JS divergence value corresponding to the data indicator type.

[0038] The health status of the industrial equipment is obtained based on multiple JS divergence values, specifically:

[0039]

[0040] Among them, H z To obtain the health status of the industrial equipment, M is the number of data indicator types used to analyze the health status, and Djs m is the JS divergence value corresponding to the m-th data indicator type, and k is the corresponding parameter used to adjust the function curve between the JS divergence value and the health score. The value can be any integer in [9, 13].

[0041] The present invention also provides a processing apparatus, comprising:

[0042] Memory, used to store computer programs;

[0043] A processor is used to implement the steps of the unweighted health analysis method for industrial equipment when executing the computer program.

[0044] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows:

[0045] 1. In existing technologies, comprehensive scoring methods require manual experience or automatic weight allocation via algorithms (such as the analytic hierarchy process, neural networks, etc.), resulting in highly subjective and poorly interpretable results. In this invention, however, kernel density estimation transforms discrete time-series data into a continuous probability distribution, avoiding the excessive reliance on single numerical biases found in traditional methods. The JS divergence is a symmetric bounded index based on the similarity of probability distributions (range [0,1]), and its value directly reflects the degree of deviation between the distribution of the comparative data and the benchmark data.

[0046] By measuring the distribution differences between various indicators and health benchmarks, the impact of different indicators on equipment health is made consistent, eliminating the need for manual or algorithmic weighting and avoiding the influence of varying numerical magnitudes of different indicators on weight settings and health assessments. This avoids misjudgments caused by biased weight settings; for example, a weakly correlated indicator (such as "vibration amplitude") might be ignored due to insufficient weight, but could potentially cause malfunctions when it deviates from a healthy state in conjunction with other indicators. Furthermore, it improves the consistency of health scores for the same equipment from different assessment bodies, reducing the subjective risks of operational and maintenance decisions.

[0047] Furthermore, kernel density estimation transforms discrete data into a continuous probability density curve, preserving the original data distribution characteristics (such as long tails and multimodality). JS divergence captures subtle but crucial distributional differences (such as a 5% increase in data fluctuation range while the value remains within the threshold) by comparing the probability distribution of comparative data with that of benchmark data. This invention quantifies the overall degradation of distribution morphology (such as increased fluctuation range and peak shift) through kernel density estimation and JS divergence, enabling earlier detection of potential faults compared to traditional threshold or numerical bias methods. Early warnings can be issued during the progressive degradation stage of equipment (such as the early stages of bearing wear or pipeline corrosion), avoiding missed detections due to numerical values ​​not exceeding thresholds using traditional methods. It can improve the sensitivity of health scores; for example, in the temperature index of a certain piece of equipment, even if the mean temperature does not increase significantly, an increase in the fluctuation range may be identified as an anomaly through JS divergence.

[0048] Furthermore, by measuring the distribution differences between each indicator and the health benchmark, the impact of different indicators on equipment health is consistent, eliminating the need for manual or algorithmic weighting. Therefore, the utilization of indicators does not rely on indicator semantics, avoiding a direct correspondence between weight settings and indicator semantics. This allows for adaptation to inconsistent indicator naming or the need for new monitoring dimensions in industrial scenarios. For example, when adding an indicator like "equipment vibration spectrum," it is only necessary to process the indicator type according to the benchmark and comparison data collected in this method to obtain the JS divergence value, which is then used to assess the health of industrial equipment. There is no need to readjust the model or redefine the indicator weights. This not only facilitates the addition or removal of indicators but also allows for migration between multiple industry scenarios (water, power, chemical, etc.), reducing model maintenance costs. Attached Figure Description

[0049] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0050] Figure 1 This is a flowchart illustrating the unweighted health analysis method for industrial equipment according to one embodiment of the present invention.

[0051] Figure 2This is a diagram illustrating the effect of different bandwidth selections on the smoothness in one embodiment of the present invention.

[0052] Figure 3 This is a graph showing the effect of different kernel functions on the smoothness in one embodiment of the present invention;

[0053] Figure 4 This is a graph showing the function curve shape of the JS divergence value and health degree under different values ​​of k in one embodiment of the present invention. Detailed Implementation

[0054] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0056] like Figure 1 As shown, a method for unweighted health analysis of industrial equipment includes:

[0057] S100: In a healthy state, according to the data index type used to analyze health, collect industrial time-series data within at least one historical time window of the industrial equipment as baseline data, and collect industrial time-series data within the current time window as comparison data.

[0058] The core objective of this step is to establish a basic framework for benchmark data and real-time comparative data on the health status of industrial equipment, providing reliable data input for subsequent health analysis based on distribution similarity (such as JS divergence).

[0059] It should be noted that data indicators are multiple metrics used to analyze health status. This invention is applicable to a comprehensive analysis method for analyzing the health status of industrial equipment using multiple indicators. The types of data indicators can include pressure, vibration, outlet flow rate, temperature, positive and negative pressure at the outlet, and other indicators.

[0060] To collect baseline data, under healthy equipment conditions, collect industrial time-series data within at least one historical time window. It should be noted that for each type of data indicator, at least one set of industrial time-series data should be collected as baseline data. There are no restrictions on the collection time, the size of the time window, or the amount of data; multiple data indicator types can be collected with the same settings or with different settings.

[0061] Because this invention subsequently compares the forms of the comparative data and the corresponding benchmark data based on kernel density estimation and JS divergence analysis, there are no restrictions on the settings of parameters such as the acquisition time, time window size, and data volume. These parameters can be used separately to match different acquisition methods under different data indicators, thus reducing the requirements for data acquisition. For example, the acquisition frequency may differ for different types of data.

[0062] To obtain comparative data, industrial time-series data within the current time window is collected and compared with the baseline data. Similarly, for each type of data indicator, a set of industrial time-series data at the current time needs to be collected for comparison with the corresponding baseline data.

[0063] The indicators are classified and processed. For different types of monitoring indicators (such as temperature, vibration, and pressure), benchmark and comparison data are collected independently to avoid data interference across indicator types.

[0064] Furthermore, each data indicator yields at least one set of benchmark data, allowing for the use of multiple sets for comparison. This enables benchmark data collection across different historical periods and operating conditions of industrial equipment, adapting to the evolution of equipment over time and the characteristics of different operating conditions, and reducing the randomness of a single benchmark. For example, if a benchmark underestimates the risk of degradation due to the lack of fluctuation during the sampling period, or if multiple sets of benchmark data experience score fluctuations due to noise, multi-benchmark aggregation can effectively offset such errors. If the health status distribution of equipment varies significantly under different operating modes (e.g., low load, high load), using only one benchmark may lead to score distortion. Through multi-benchmark design, the health status characteristics of equipment under all operating conditions can be covered.

[0065] This step, through the design of data collection dimensions, ensures the representativeness and statistical significance of the baseline and comparison data, providing reliable input for subsequent distribution similarity analysis. By using multiple sets of historical baseline data, it reduces randomness, improves scoring stability, adapts to dynamic changes in equipment operating status, and processes different monitoring indicators independently to avoid cross-indicator interference, enhancing the versatility of the method. This provides a key data foundation for subsequent health scoring based on JS divergence.

[0066] S200: Based on the benchmark data and the comparison data, the probability density curve is obtained by kernel density estimation to obtain the probability distribution.

[0067] The core objective of this step is to transform discrete industrial time-series data into a continuous probability distribution, providing a mathematical foundation for subsequent health analysis based on distribution similarity (such as JS divergence).

[0068] To calculate the JS divergence, the probability distribution must first be obtained. Kernel density estimation transforms discrete data into a continuous probability density curve. Essentially, it involves superimposing kernel functions on each data point to generate a continuous probability density curve, thus facilitating the estimation of the data distribution.

[0069] It's important to note that kernel functions play a crucial role in KDE (Knowledge-Definition Analysis). They are primarily used to smooth discrete data points, thereby constructing a continuous probability density function. Kernel functions assign weights to the regions near each data point, with closer regions having higher weights. For example, in a Gaussian kernel, the weights decay exponentially with distance, thus allocating weights accordingly.

[0070] Kernel density estimation can identify states where numerical deviations are not significant but the distribution pattern has degenerated (e.g., the data fluctuation range has increased by 5% but the value is still within the threshold range). For example, in the temperature index of a certain device, even if the mean has not increased significantly, the expansion of the fluctuation range can be identified as an anomaly through the peak shift of kernel density estimation. The continuity of kernel density estimation avoids the discretization error of the histogram method, making subsequent JS divergence calculations more stable.

[0071] Kernel functions include the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel, and the shape of the kernel function determines the smoothness. Among them, the Gaussian kernel has higher smoothness and is suitable for scenarios with high data noise; while the Epanechnikov kernel has compact support (non-zero only in a finite interval), making it more computationally efficient.

[0072] The optimal kernel function should be selected according to the type of data indicator. Different types of data indicators have different numerical characteristics. For example, sound data indicators are easily affected by external noise and generally have a low signal-to-noise ratio. Temperature data indicators are also easily affected by changes in the external ambient temperature. Vibration data indicators are not easily affected and generally have a high signal-to-noise ratio, but when multi-directional vibration data is used, the amount of data is large.

[0073] Therefore, the optimal kernel function is selected according to the data indicator type to leverage the processing characteristics of the kernel function, balance the smoothness and processing efficiency after data processing, and further optimize the accuracy of distribution estimation.

[0074] The probability density curve's data range (e.g., the normalized [0,1] interval) is divided into N equidistant sub-intervals. The probability value for each interval is calculated using numerical integration (e.g., the trapezoidal rule or Simpson's rule), forming a discretized probability distribution. Smaller bandwidth yields richer distribution details, requiring more intervals; larger bandwidth results in a smoother distribution, reducing the number of intervals.

[0075] This step transforms discrete data into a continuous probability distribution through kernel density estimation, addressing the core issue of insufficient utilization of distribution characteristics in existing technologies. By modeling the distribution morphology, it accurately captures distribution degradation, avoiding the missed faults caused by traditional methods where values ​​do not exceed thresholds, and providing a crucial data foundation for subsequent JS divergence calculations.

[0076] S300: Based on the data indicator type, the probability distribution of the comparison data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distribution under the data indicator type, so as to obtain the health of the industrial equipment based on multiple JS divergence values.

[0077] The core objective of this step is to quantify the similarity differences in data distribution using JavaScript divergence metrics, thereby achieving unweighted industrial equipment health scores.

[0078] JS divergence is an improvement on KL divergence. JS divergence is symmetric and bounded (value range [0,1]), where 0 represents complete consistency and 1 represents complete inconsistency. It is suitable for asymmetric distributions (such as multimodal distributions and long-tailed distributions). The probability distribution obtained through the above steps is used as the input for JS divergence calculation, which can capture subtle differences in distribution shape.

[0079] It should be noted that for single-benchmark scenarios, the JS divergence between the comparison data and a single set of benchmark data is directly calculated to obtain the JS divergence value of a single data indicator. For multi-benchmark scenarios, if a data indicator has multiple sets of benchmark data (such as health data from different time periods or work modes), multiple JS divergence values ​​are obtained, and the average value is used to obtain the JS divergence value of that data indicator.

[0080] It should be noted that this invention obtains the JS divergence value by comparing the distribution similarity between benchmark data and comparative data, eliminating the influence of numerical differences between indicators and ensuring that the contribution of different indicators to equipment health is consistent. The JS divergence values ​​of multiple data indicators can be directly used, and the average value is taken to obtain a final JS divergence value. The health of the industrial equipment is then determined based on the nonlinear function between this JS divergence value and the industrial equipment's health. By directly quantifying distribution differences through JS divergence, weightless calculation is achieved, avoiding the subjectivity of weight allocation. Therefore, there is no need to set weights based on the data indicator type. This invention does not rely on indicator semantics and does not require explicit naming such as "vibration amplitude" or "outlet pressure." It evaluates solely based on distribution characteristics, making it adaptable to multiple industries and new monitoring dimensions.

[0081] JS divergence measures the difference in data distribution between comparative and baseline data. It can identify situations where the numerical deviation is not significant but the distribution pattern has deteriorated (such as a 5% increase in the data fluctuation range but the value is still within the threshold range). For example, in temperature indicators, even if the mean does not increase significantly, the expansion of the fluctuation range can be identified as an anomaly through JS divergence.

[0082] The non-linear mapping between JS divergence and health score maps JS divergence values ​​(0-1) to health scores (0-100). When the JS divergence approaches 0, the health score approaches 100 (perfectly consistent); when the JS divergence approaches 1, the health score approaches 0 (completely inconsistent). The mapping relationship from distribution comparison to health score is intuitive (JS divergence value [0,1] → health score [0,100]), facilitating understanding and decision-making for industrial operations and maintenance personnel.

[0083] This step addresses the core issues of strong weight dependence, insufficient utilization of distribution features, and limited model versatility in existing technologies by using JS divergence to quantify distribution differences.

[0084] In a preferred embodiment of the present invention, after industrial time-series data acquisition, the method further includes:

[0085] The acquired industrial time-series data is preprocessed.

[0086] Check the industrial time-series data for null values ​​and perform a deletion operation on the null values.

[0087] Linear normalization is performed on the industrial time series data after removing null values ​​to maintain the data distribution characteristics of the industrial time series data.

[0088] The core purpose of industrial time series data preprocessing is to ensure the integrity of data quality and distribution characteristics, so as to provide reliable input for subsequent health analysis based on distribution similarity (such as JS divergence calculation).

[0089] The collected industrial time-series data undergoes a null value check, and missing values ​​are directly deleted without using methods such as mean filling or interpolation. It should be noted that the industrial equipment health analysis in this invention relies on the distribution characteristics of the data (such as fluctuation range and peak offset). Filling in missing values ​​would forcibly introduce data points unrelated to the original distribution, leading to distribution distortion. For example, in the water pump group data of the water industry, if a certain indicator (such as "pressure") has missing values, direct deletion can preserve the true distribution pattern, while mean filling may smooth the distribution, masking the actual fluctuation characteristics.

[0090] Furthermore, it is understood that the industrial equipment health analysis in this invention relies on the distribution characteristics of the data. Therefore, there are no strict limitations on the difference in the amount of baseline and comparison data, and the deletion of null values ​​will not lead to errors in the comparison calculation. This embodiment preserves the natural missing data pattern by deleting null values, ensuring the authenticity of the distribution pattern.

[0091] After removing null values, normalization is performed on the data. Normalization is an operation that maps all dimensions of data to a 0-1 vector range. For an industrial time series dataset {x1, x2, ..., x...}, the normalization process is performed. i ,…,x n}, for x i Perform normalization according to the following formula.

[0092]

[0093] Where, x min x is the minimum value in this industrial time series data. max This is the maximum value in the industrial time series data.

[0094] Understandably, in industrial settings, the numerical ranges of different indicators vary significantly (e.g., pressure indicators range from 0-100 MPa, and vibration indicators range from 0-0.1 mm). Without normalization, indicators with larger magnitudes will dominate in distribution comparisons, causing the true changes in smaller magnitude indicators to be masked.

[0095] The same normalization process is applied to meet subsequent JS divergence calculation needs. JS divergence is based on the similarity of probability distributions, and normalization ensures that each indicator is compared on the same scale, avoiding interference from differences in units of measurement.

[0096] Furthermore, it should be noted that outlier deletion or correction is not required in the preprocessing of industrial time-series data. This is because outliers generally occur infrequently in industrial time-series data, thus having a smaller impact on the distribution differences in the data and therefore producing virtually no error in the assessment of the health of industrial equipment in this invention. This further simplifies the data preprocessing steps.

[0097] The preprocessing steps in this embodiment, through null value deletion and linear normalization, solve the problems of distribution distortion and scoring bias caused by unreasonable data preprocessing in existing technologies. Deleting null values ​​avoids human interference, and normalization eliminates magnitude differences, ensuring the accuracy of subsequent distribution comparisons and preserving distribution characteristics. By retaining the true distribution shape, the reliability of JS divergence calculation is improved, avoiding misjudgments due to improper preprocessing and enhancing the stability of health scores. This step provides a crucial data foundation for subsequent health analysis based on JS divergence, directly resolving the distribution distortion problem caused by unreasonable data preprocessing in existing technologies.

[0098] In a preferred embodiment of the present invention, the historical time window and the current time window are specifically as follows:

[0099] The time window sizes of the historical event window and the current time window are determined based on the data feature frequency of the corresponding data indicator type, and must be greater than the feature period corresponding to the minimum data feature frequency.

[0100] Furthermore, the data collection time is greater than or equal to 1 hour, and the amount of data collected within the time window is greater than or equal to 500.

[0101] The time window size for each of the historical event windows and the current time window can be set independently.

[0102] The core objective of this implementation method is to adapt to the dynamic characteristics of different data indicator types, ensure the sufficiency and representativeness of the collected data, and provide high-quality data input for subsequent health analysis based on kernel density estimation and JS divergence.

[0103] For each type of data indicator (such as "pressure" or "vibration"), calculate its minimum characteristic frequency (e.g., through Fourier transform or autocorrelation function analysis). The time window size must be greater than or equal to an integer multiple of the maximum characteristic period (e.g., if the maximum characteristic period of an indicator is 30 minutes, then the window must be ≥30 minutes) to ensure that the data includes complete fluctuation patterns (such as peaks and troughs) and avoid missing key features due to an excessively small window.

[0104] In addition, data volume and collection time constraints must still be met: collection time ≥ 1 hour, data volume ≥ 500 points within the time window, and statistical significance of kernel density estimation must be satisfied (usually requiring a sample size ≥ 500 to reduce estimation bias), avoiding the failure of traditional methods due to small samples. These dual constraints of data volume and time ensure the statistical stability of kernel density estimation.

[0105] The time window size for each indicator type is configured independently, rather than being set uniformly, to adapt to the dynamic characteristics of different indicators. High-frequency indicators (such as vibration signals) require shorter time windows (such as 1 hour) to avoid oversampling and redundancy; low-frequency indicators (such as temperature trends) can have longer windows (such as 6 hours) to ensure data integrity.

[0106] This implementation solves the data mismatch problem caused by fixed windows in the prior art by setting dynamic time windows. It dynamically adapts to different indicator feature frequencies to ensure data sufficiency and representativeness. The dual constraints of data volume and time improve the reliability of kernel density estimation and JS divergence calculation. The independent window setting adapts to the heterogeneity of multiple indicators and enhances the comprehensiveness of health scores.

[0107] Specifically, industrial time-series data within at least one historical time window is used as baseline data, specifically:

[0108] Based on the data indicator type, different historical time windows are selected, and / or,

[0109] Multiple sets of industrial time-series data from the industrial equipment under different operating modes are collected to obtain multiple benchmark data under the data indicator type.

[0110] The core objective of this embodiment is to construct a more representative and diverse benchmark dataset to improve the accuracy and robustness of subsequent health analysis based on JS divergence.

[0111] Multiple sets of data are randomly selected as baseline data from different time periods when the equipment is in good health (e.g., daytime, nighttime, or after the equipment has been running stably for 10 or 15 days). For example, in the case of water pump sets in the water industry, data from multiple healthy time periods (e.g., morning, afternoon, and nighttime) are selected to ensure that the baseline covers the dynamic operating characteristics of the equipment.

[0112] Furthermore, the data collected by the equipment under various operating conditions (such as low load, high load, and start-up / shutdown phases) adapts to the diversity of equipment operating states. For example, the vibration distribution of a compressor at low load may differ significantly from that at high load, and multi-condition benchmark data can more comprehensively reflect its health status.

[0113] It is understandable that relying on a single benchmark (such as a fixed time period or a single operating condition) can lead to scoring bias. For example, if the benchmark data sampling period happens to be uneven (such as when the equipment is under low load), the risk of degradation may be underestimated.

[0114] This embodiment uses data from different time periods or working modes to ensure that the benchmark data includes various distribution characteristics of the device in a healthy state, reducing the random errors of a single benchmark (such as noise interference or no fluctuation during the sampling period), improving the stability of the score, and solving the scoring deviation problem caused by a single benchmark in the prior art.

[0115] In a preferred embodiment of the present invention, the probability density curve is obtained through kernel density estimation, specifically as follows:

[0116] Based on the selected kernel function, the kernel density is estimated using the benchmark data and the comparison data, resulting in a probability density curve. The kernel function includes at least one of the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel.

[0117] The core objective of this implementation method is to transform discrete industrial time-series data into continuous probability density curves, providing a reliable foundation for subsequent health analysis based on JS divergence.

[0118] Based on preprocessed baseline and comparison data (with null values ​​removed and normalized) {X1,X2,…,X...} n The formula for estimating nuclear density is:

[0119]

[0120] Where K(·) is the kernel function, which determines the value of each sample data point X. i The contribution method to the density at position x, where h0 is the bandwidth, used to control the smoothness, such as... Figure 2 As shown, this illustrates the impact of different bandwidth selections on the smoothness, where n is the amount of data.

[0121] In kernel density estimation, the kernel function and bandwidth are two key parameters. For bandwidth selection, the theoretically optimal bandwidth (Silverman's rule) is used. This rule is applicable to scenarios where the data approximately follows a normal distribution. Specifically,

[0122] h0 = 1.06 × σ × n -1 / 5 ,

[0123] Where σ is the sample standard deviation.

[0124] In addition, kernel functions include the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel. The Epanechnikov kernel is suitable for non-normally distributed data, especially uniformly distributed, symmetrically multimodal, and other data deviating from normality. Because it only calculates the weights of neighboring points, it is computationally efficient and consumes limited computational resources. The Gaussian kernel is suitable for noisy scenarios, offering high smoothness but potentially losing details. The linear kernel is simple and efficient to compute, suitable for scenarios with uniform data distribution and no complex nonlinear relationships. The exponential kernel has strong locality, suitable for capturing local clustering features of data, but may ignore global structure. Figure 3 The figure shows the effect of different kernel functions on the smoothness.

[0125] The selection of the kernel function is specifically as follows:

[0126] Based on the data indicator type, one of the benchmark data and the comparison data is selected, and kernel density estimation is performed using multiple kernel functions to obtain multiple probability density curves for each data indicator type.

[0127] Based on the multiple probability density curves, the asymptotic mean square error is obtained by estimating the deviation and variance. The kernel function corresponding to the smallest asymptotic mean square error is selected as the kernel function under the data index type.

[0128] The core objective of this embodiment is to dynamically select the optimal kernel function by quantifying the applicability of the kernel function to minimize the kernel density estimation error, thereby improving the reliability of subsequent JS divergence calculation and the accuracy of health score.

[0129] For the same data indicator type, samples (any dataset) are randomly selected from the benchmark or comparative data, and kernel density estimation (KDE) is performed using kernel functions such as the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel to obtain the kernel density estimate. Based on the estimated deviation and estimated variance The asymptotic mean square error (AMSE) is obtained.

[0130]

[0131] The estimation bias measures the degree to which the kernel density estimate deviates from the true distribution. The estimation variance measures the volatility (smoothing) of the estimation results.

[0132] Calculate AMSE for each kernel function, compare the performance of the kernel functions, and select the kernel function with the smallest AMSE as the final choice.

[0133] Based on the data indicator type, the kernel function is dynamically selected, and the applicability of the kernel function is quantified by AMSE to ensure that the kernel density estimation is highly matched with the data distribution characteristics.

[0134] This embodiment solves the problems of insufficient kernel density estimation accuracy and blind selection of kernel functions in the prior art by dynamically selecting kernel functions and optimizing bandwidth.

[0135] As another embodiment of this implementation, a probability density curve is obtained to obtain a probability distribution, specifically as follows:

[0136] Based on the probability density curve, multiple intervals are divided, wherein the number of intervals is determined by the theoretical bandwidth obtained from the standard deviation of industrial time series data, and the number of intervals for the comparative data and the corresponding benchmark data for each data index type is set to be consistent;

[0137] The interval probability corresponding to each interval is obtained, forming a probability distribution of the comparison data and the corresponding benchmark data.

[0138] The core objective of this embodiment is to transform the continuous probability density curve obtained from kernel density estimation into a discretized probability distribution, providing operable data input for subsequent JS divergence calculation.

[0139] Similarly, based on Silverman's rule, the theoretically optimal bandwidth h (reflecting the local fluctuation scale of the data) is calculated, and the data value range (the normalized [0,1] interval) is divided into N equidistant sub-intervals.

[0140]

[0141] Then the length of each interval is,

[0142]

[0143] The interval division needs to match the bandwidth h of the kernel density estimation (which reflects the smoothness of the distribution). The smaller the bandwidth, the richer the distribution details, and more intervals are needed to capture features; the larger the bandwidth, the smoother the distribution, and the fewer intervals can be appropriately reduced.

[0144] In this invention, the interval length Δx is less than 2h to ensure that each peak or fluctuation feature is covered by at least 1-2 intervals, thus avoiding the loss of details.

[0145] It should be noted that, under the requirement of a unified data indicator type, the number of intervals N for the benchmark data and the comparison data should be set to be consistent to ensure that the two are compared on the same scale.

[0146] Calculate the interval probability based on the divided intervals. For each subinterval [x] i x i+1 ), i∈N-1, based on the probability density curve of the benchmark data The probability value p is calculated by integrating the probability density function over this interval.i (Corresponding to the P-distribution), based on the probability density curve of the comparison data Calculate the probability value q i (Corresponding to Q-distribution).

[0147]

[0148] Approximate calculations are performed using numerical integration methods (such as the trapezoidal rule and Simpson's method). This yields the discretized probability distribution P = {p1, p2, ..., p...}. i ,…,p N} and Q = {q1,q2,…,q i ,…,q N}. Calculate the average distribution W, based on We obtain W = {w1, w2, ..., w i ,…,w N}

[0149] This embodiment solves the problem of distribution modeling distortion caused by fixed interval division in the prior art by using dynamic interval division and interval probability calculation.

[0150] In a preferred embodiment of the present invention, based on the data index type, the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distributions under the data index type, specifically:

[0151] By analyzing the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data, at least one JS divergence value can be obtained.

[0152] When multiple JS divergence values ​​are obtained from multiple benchmark data, the average of the multiple JS divergence values ​​is taken to obtain the JS divergence value corresponding to the data indicator type.

[0153] The core objective of this implementation method is to improve the stability and reliability of industrial equipment health scores through JS divergence aggregation of multi-benchmark data.

[0154] By comparing the probability distribution Q of the data with the probability distribution P of at least one benchmark data, the JS divergence under this data index is obtained.

[0155] JS divergence is an important metric for measuring the similarity of probability distributions between comparative and baseline data. It is an improvement upon KL divergence, effectively addressing the asymmetry and potential divergence issues of KL divergence. JS divergence is symmetric and bounded, with a value ranging from [0,1]. A JS divergence value of 0 indicates that the two probability distributions are identical; a value of 1 indicates that the two probability distributions are completely different.

[0156]

[0157] Where KL(P||W) represents the KL divergence between probability distributions P and W. KL(Q||W) represents the KL divergence between probability distributions Q and W. i is the position number, i∈[1,n].

[0158] In this implementation, if a certain data index type has R benchmark data, the JS divergence value between the comparison data and each benchmark data is calculated separately, resulting in R results. The JS divergence values ​​of the multiple benchmarks are then aggregated into a final result using an arithmetic mean.

[0159] Based on the Bagging concept, the arithmetic mean of multiple benchmark data is used to reduce the random error of a single benchmark (such as noise interference or no fluctuation during the sampling period) and improve the stability of the score.

[0160] Specifically, taking a data indicator of a water pump set as an example, assuming there are four sets of baseline data, as shown in Table 1, if only a single baseline data (such as baseline 1) is used, the health score will be too high (99.31); if only baseline data 4 is used, the health score will be too low (96.85). However, by combining the four sets of baseline data, a health score of 97.98 is obtained, which is more stable and offsets the randomness that may exist with a single baseline data (such as baseline data 3 having a large amount of data but no fluctuation during the sampling period, or baseline data 4 having a small amount of data that may introduce noise).

[0161] Table 1 shows the health scores obtained individually for each of the four sets of baseline data.

[0162] 1 7248 0.0024 99.31 2 6452 0.0018 98.22 3 8863 0.0025 97.53 4 4501 0.0032 96.85

[0163] This implementation solves the scoring bias problem caused by a single benchmark in the prior art by using a JS divergence aggregation strategy based on multi-benchmark data, and provides high-quality JS divergence input for subsequent health scoring.

[0164] In a preferred embodiment of the present invention, the health status of the industrial equipment is obtained based on a plurality of JS divergence values, specifically as follows:

[0165]

[0166] Among them, H z To obtain the health status of the industrial equipment, M is the number of data indicator types used to analyze the health status, and Djs mis the JS divergence value corresponding to the m-th data indicator type, and k is the corresponding parameter used to adjust the function curve between the JS divergence value and the health score. The value can be any integer in [9, 13].

[0167] The core objective of this implementation method is to achieve quantitative calculation of the health status of industrial equipment through comprehensive evaluation of multiple indicators, thereby providing a scientific basis for equipment condition monitoring and maintenance.

[0168] Collect M data indicators (such as vibration, temperature, and pressure) to ensure coverage of key operational characteristics of the equipment. For each indicator, calculate its JS divergence value (Djsm) using kernel density estimation of the baseline and comparison data to quantify the distribution differences between the comparison and baseline data.

[0169] The JS divergence is symmetricized by subtracting 0.5 to create symmetry with the zero point. At this point, the closer the JS divergence is to 0.5, the more similar the distribution (healthy state); the further it deviates from 0.5, the greater the difference (higher risk of degradation). The health of the industrial equipment is obtained by taking the mean of the JS divergence of multiple data indicators and applying the Sigmoid function.

[0170] By adjusting the corresponding parameter k, the shape of the function curve can be made "S-shaped" rather than nonlinear, meaning the derivative is not constant, thus avoiding linear superposition. For example... Figure 4 As shown, the function curve shape is displayed under different values ​​of k. The value of k is any integer in the range [9, 13] to avoid the linear superposition problem when the value of k is too small, and also to avoid the problem of excessive sensitivity due to the small slope when the value of k is too large (JS divergence is close to 0).

[0171] This implementation solves the problems of assessment bias and linearity deficiency in the prior art by using multi-index JS divergence fusion and Sigmoid nonlinear mapping. It adapts the gradual nature of device degradation by using the nonlinear characteristics of the Sigmoid function and quantifies the dynamic changes in health status. In addition, by setting the k value, it takes into account the linear superposition problem and the sensitivity problem.

[0172] The present invention further provides a processing apparatus, comprising:

[0173] Memory, used to store computer programs;

[0174] A processor is used to implement the steps of the unweighted health analysis method for industrial equipment when executing the computer program.

[0175] Therefore, the method can achieve any effect of unweighted health analysis of industrial equipment, which will not be elaborated here.

[0176] For any parts not mentioned in this invention, existing technologies can be used or referenced.

[0177] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0178] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for unweighted health analysis of industrial equipment, characterized in that, include: In a healthy state, according to the data index type used to analyze health, industrial time-series data within at least one historical time window of the industrial equipment are collected as baseline data, and industrial time-series data within the current time window are collected as comparison data. The data indicators include pressure, vibration, outlet flow rate, temperature, and positive and negative pressure at the outlet. Based on the baseline data and the comparison data, probability density curves are obtained one by one through kernel density estimation to obtain the probability distribution; Based on the data indicator type, the probability distribution of the comparison data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distribution under the data indicator type, so as to obtain the health of the industrial equipment based on multiple JS divergence values. By measuring the difference in data distribution between comparative and baseline data using JS divergence values, it is possible to identify states where numerical deviations are not significant but the distribution pattern has degenerated. The historical time window and the current time window are specifically as follows: The size of the historical time window and the current time window is determined according to the data feature frequency of the corresponding data indicator type. It is greater than the feature period corresponding to the minimum data feature frequency, and the collection time is greater than or equal to 1 hour and the amount of data collected within the time window is greater than or equal to 500. The size of each historical time window and the current time window is set independently. At least one historical time window of industrial time-series data is used as the baseline data, specifically: Based on the data indicator type, different historical time windows are selected, and / or, Multiple sets of industrial time-series data from the industrial equipment under different operating modes are collected to obtain multiple benchmark data under the data indicator type. The probability density curve is obtained through kernel density estimation, as follows: Based on the selected kernel function, the kernel density is estimated using the benchmark data and the comparison data, resulting in a probability density curve. The kernel function includes at least one of the Epanechnikov kernel, Gaussian kernel, linear kernel, and exponential kernel. The selection of the kernel function is specifically as follows: Based on the data indicator type, one of the benchmark data and the comparison data is selected, and kernel density estimation is performed using multiple kernel functions to obtain multiple probability density curves for each data indicator type. Based on multiple probability density curves, the asymptotic mean square error is obtained by estimating the deviation and variance. The kernel function corresponding to the smallest asymptotic mean square error is selected as the kernel function under the data index type. Based on the data indicator type, the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data are used to obtain the JS divergence value for measuring the similarity of data distributions under the data indicator type, specifically: By analyzing the probability distribution of the comparative data and the probability distribution of at least one corresponding benchmark data, at least one JS divergence value can be obtained. When multiple JS divergence values ​​are obtained from multiple benchmark data, the average of the multiple JS divergence values ​​is taken to obtain the JS divergence value corresponding to the data indicator type. The health status of the industrial equipment is obtained based on multiple JS divergence values, specifically: ， in, For the obtained health status of the industrial equipment, M is the number of data indicator types used to analyze the health status. is the JS divergence value corresponding to the m-th data indicator type, and k is the corresponding parameter used to adjust the function curve between the JS divergence value and the health score. The value can be any integer in [9, 13].

2. The unweighted health analysis method for industrial equipment according to claim 1, characterized in that, After industrial time-series data acquisition, the following are also included: The acquired industrial time-series data is preprocessed. Check the industrial time-series data for null values ​​and perform a deletion operation on the null values. Linear normalization is performed on the industrial time series data after removing null values ​​to maintain the data distribution characteristics of the industrial time series data.

3. The unweighted health analysis method for industrial equipment according to claim 1, characterized in that, The probability density curve is obtained to determine the probability distribution, specifically: Based on the probability density curve, multiple intervals are divided, wherein the number of intervals is determined by the theoretical bandwidth obtained from the standard deviation of industrial time series data, and the number of intervals for the comparative data and the corresponding benchmark data for each data index type is set to be consistent; The interval probability corresponding to each interval is obtained, forming a probability distribution of the comparison data and the corresponding benchmark data.

4. A processing apparatus, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the unweighted health analysis method for industrial equipment as described in any one of claims 1 to 3.

Images