Method for evaluating device health

Abstract

The application discloses a kind of equipment health degree evaluation methods, the data field of object equipment is established in this method, the structure tree of object equipment is determined according to different levels;Download the historical data of all data items;According to the fault number determined by the historical alarm number or maintenance record statistics of different data, the initial weight is determined in proportion according to the same parent node under the child node;Build benchmark data feature matrix, calculate score scale, download the data of each data item, calculate the KPI score of data item, calculate KPI dynamic weight, respectively calculate the KPI score of part, equipment, system and the corresponding dynamic weight.This method takes equipment structure as model input, and is associated with data item, quantifies the initial weight, calculates dynamic weight, uses unlabeled data modeling, effectively improves the reliability of equipment health degree evaluation.

Description

Technical Field

[0002] This invention relates to the field of equipment diagnostic technology, and in particular to a method for assessing equipment health. Background Technology

[0004] Equipment health models aim to monitor, diagnose, and predict equipment status, covering the entire lifecycle of equipment, including operational health, early failures, and critical failures. Therefore, establishing effective health assessment models for equipment has significant application value.

[0005] The existing equipment health model mainly implements the following process:

[0006] 1. Collect relevant parameters of the target device;

[0007] 2. Determine the weight of each parameter based on its impact on the equipment or the frequency of failures;

[0008] 3. Set thresholds for parameters based on their performance at each stage of equipment failure, and then quantify and evaluate the parameters sequentially to obtain corresponding scores;

[0009] 4. Based on weights, the status of the target device is evaluated using a weighted method to obtain quantitative indicators.

[0010] However, in actual production scenarios within the steel industry, the varying levels of experience and knowledge among technical personnel lead to significant subjectivity in determining parameters and weights. Furthermore, due to the relatively low level of digitalization and intelligentization in the steel industry, and the relatively few equipment failure cases, there is a limited amount of fault data with corresponding fault labels, making it difficult to quantify and assess thresholds and corresponding parameter scores. Conversely, the existing collected data consists largely of unlabeled health status data.

[0011] Considering actual production scenarios, the main drawbacks of the traditional equipment health model mentioned above are:

[0012] 1. Most of the parameter types are determined based on empirical knowledge, and the determination of parameter weights and parameter thresholds lacks a quantitative process;

[0013] 2. Lack of consideration for equipment structure, and lack of consideration for parameters and the correspondence between equipment;

[0014] 3. Lack of dynamic consideration of weights.

[0015] 4. The requirement for labeled fault data is too high and does not reflect the actual situation. Summary of the Invention

[0017] The technical problem to be solved by the present invention is to provide a method for assessing equipment health. This method overcomes the defects of traditional equipment health assessment by using equipment structure as model input, associating data items, quantifying the initial weight of assessment, calculating dynamic weight, and using unlabeled data modeling, thereby effectively improving the reliability of equipment health assessment.

[0018] To solve the above-mentioned technical problems, the equipment health assessment method of the present invention includes the following steps:

[0019] Step 1: Data preparation. Establish the data domain of the target device and determine the structure tree of the target device according to different levels.

[0020] Step 2: Download historical data. Download historical data for all data items, including historical alarms, maintenance records, historical vibration signals, and time-series data.

[0021] Step 3: Calculate the initial weights. Based on the historical alarm counts or fault counts determined from maintenance records, determine the initial weights for child nodes under the same parent node proportionally.

[0022] Step 4: Construct a baseline data feature matrix. Use the K-means algorithm to divide the historical vibration data into working conditions. Perform statistical analysis on the vibration data under each working condition. Based on the calculated statistical characteristics, determine the most stable multiple vibration data under each working condition. Finally, obtain the most stable N vibration data with a length of L1. Perform statistical analysis on the data in each window after the time series data sliding window. Based on the calculated statistical characteristics, determine the most stable M data windows with a length of L2. Construct a baseline data matrix with a dimension of N×L1 or M×L2 for each parameter. Transform the baseline data matrix into the feature space and construct the baseline data feature matrices of N×L1' and M×L2'.

[0023] Step 5: Calculate the scoring scale. Traverse the historical vibration data or historical signal data window and transform it to the feature space. Use the distribution analysis method to analyze the difference between each observed data feature and each sample in the corresponding benchmark data feature matrix and calculate the average value to obtain N quantized difference values. Construct difference vectors E=[E1,E2,...,EN] respectively. Select the maximum value Emax and the minimum value Emin in the vector to construct the scoring reference interval, and then correspond to the [0,1] interval to realize quantization.

[0024] Step 6: Real-time data download. Download the data for each data item. For vibration data, acquire the waveform of length L1; for time series data, acquire the data sequence of length L2.

[0025] Step 7: Calculate the KPI score of the data item, convert the real-time data to the feature space, and perform difference analysis with the data in the corresponding benchmark data feature matrix to calculate the real-time KPI score Ereal. Determine the KPI score in the [0,1] interval based on its relationship with Emax and Emin. For KPI scores that are out of range, dynamically update the difference vector E after evaluation.

[0026] Step 8: Calculate the dynamic weights of the KPIs. Using the calculated KPI scores as independent variables, a non-linear function is used to calculate the dynamic weights based on the initial weights. The formula is as follows:

[0027] (1)

[0028] For dynamic weights The following normalization process is performed to ensure that the sum of the dynamic weights of all KPIs is 1. This dynamic adjustment mechanism automatically gives more attention to KPIs with unsatisfactory scores.

[0029] (2)

[0030] in, For the first Item dynamic weight, For the first Initial weights of items For the first Real-time KPI scores for each item;

[0031] Step 9: Combine the KPI score of each data item with its corresponding dynamic weight to calculate the KPI score of the part, and calculate the dynamic weight of the part according to Formula 1.

[0032] Step 10: Combine the KPI score of each part with its corresponding dynamic weight to calculate the KPI score of the equipment, and calculate the dynamic weight of the equipment according to Formula 1.

[0033] Step 11: Combine the KPI scores of each device with their corresponding dynamic weights to calculate the KPI score of the system.

[0034] Furthermore, the structure tree of the object device in step one includes a multi-level structure of system, device, part, and data item, and the data item is attached to each node of the structure tree.

[0035] Because the equipment health assessment method of this invention adopts the above-mentioned technical solution, namely, this method establishes a data domain for the target equipment, determines the structure tree of the target equipment according to different levels; downloads historical data of all data items; statistically determines the number of faults determined by the historical alarm count or maintenance records of different data, and proportionally determines the initial weights of child nodes under the same parent node; constructs a benchmark data feature matrix, calculates the scoring scale, downloads data for each data item, calculates the KPI score of the data item, calculates the dynamic weight of the KPI, and calculates the KPI scores and corresponding dynamic weights for parts, equipment, and systems respectively. This method uses the equipment structure as the model input, associates data items, quantifies the initial assessment weights, calculates the dynamic weights, and uses unlabeled data modeling, effectively improving the reliability of equipment health assessment. Attached Figure Description

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:

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

[0040] Implementation, for example Figure 1 As shown, the equipment health assessment method of the present invention includes the following steps:

[0041] Step 1: Data preparation. Establish the data domain of the target device and determine the structure tree of the target device according to different levels.

[0042] Step 2: Download historical data. Download historical data for all data items, including historical alarms, maintenance records, historical vibration signals, and time-series data.

[0043] Step 3: Calculate the initial weights. Based on the historical alarm counts or fault counts determined from maintenance records, determine the initial weights for child nodes under the same parent node proportionally.

[0044] Step 4: Construct a baseline data feature matrix. Use the K-means algorithm to divide the historical vibration data into working conditions. Perform statistical analysis on the vibration data under each working condition. Based on the calculated statistical characteristics, determine the most stable multiple vibration data under each working condition. Finally, obtain the most stable N vibration data with a length of L1. Perform statistical analysis on the data in each window after the time series data sliding window. Based on the calculated statistical characteristics, determine the most stable M data windows with a length of L2. Construct a baseline data matrix with a dimension of N×L1 or M×L2 for each parameter. Transform the baseline data matrix into the feature space and construct the baseline data feature matrices of N×L1' and M×L2'.

[0045] Step 5: Calculate the scoring scale. Traverse the historical vibration data or historical signal data window and transform it to the feature space. Use the distribution analysis method to analyze the difference between each observed data feature and each sample in the corresponding benchmark data feature matrix and calculate the average value to obtain N quantized difference values. Construct difference vectors E=[E1,E2,...,EN] respectively. Select the maximum value Emax and the minimum value Emin in the vector to construct the scoring reference interval, and then correspond to the [0,1] interval to realize quantization.

[0046] Step 6: Real-time data download. Download the data for each data item. For vibration data, acquire the waveform of length L1; for time series data, acquire the data sequence of length L2.

[0047] Step 7: Calculate the KPI score of the data item, convert the real-time data to the feature space, and perform difference analysis with the data in the corresponding benchmark data feature matrix to calculate the real-time KPI score Ereal. Determine the KPI score in the [0,1] interval based on its relationship with Emax and Emin. For KPI scores that are out of range, dynamically update the difference vector E after evaluation.

[0048] Step 8: Calculate the dynamic weights of the KPIs. Using the calculated KPI scores as independent variables, a non-linear function is used to calculate the dynamic weights based on the initial weights. The formula is as follows:

[0049] (1)

[0050] For dynamic weights The following normalization process is performed to ensure that the sum of the dynamic weights of all KPIs is 1. This dynamic adjustment mechanism automatically gives more attention to KPIs with unsatisfactory scores.

[0051] (2)

[0052] in, For the first Item dynamic weight, For the first Initial weights of items For the first Real-time KPI scores for each item;

[0053] Step 9: Combine the KPI score of each data item with its corresponding dynamic weight to calculate the KPI score of the part, and calculate the dynamic weight of the part according to Formula 1; where the KPI score of the data item is the health score of the data item.

[0054] Step 10: Combine the KPI scores of each part with their corresponding dynamic weights to calculate the KPI score of the equipment, and calculate the dynamic weight of the equipment according to Formula 1; where the KPI score of a part is the health score of the part.

[0055] Step 11: Combine the KPI scores of each device with their corresponding dynamic weights to calculate the system's KPI score; the KPI score of each device is its health score.

[0056] Preferably, the structure tree of the object device in step one includes a multi-level structure of system, device, part, and data item, and the data item is attached to each node of the structure tree.

[0057] This method is applied in practice using a dust collector fan as an example, employing data from the past six months, from December 20, 2024, 00:00:00 to May 6, 2025, 00:00:00, as historical data. N=M=100 is set, and the STD method is used to evaluate the stability of the data, selecting a benchmark data feature matrix for construction. For vibration signals, common time-domain, frequency-domain, and time-frequency-domain eigenvalues ​​are used to construct feature vectors; for time-series data, common statistical features are used to construct feature vectors. The equipment structure is shown in Table 1.

[0058]

[0059] Following steps 1-5 of this method, the results of the initialization phase are shown in Table 2 below:

[0060] Table 2

[0061]

[0062] In the prediction phase, a single vibration signal was used. For the time-series signal, a window of 150 points was selected. After obtaining a set of prediction data, the calculation results are shown in Table 3.

[0063] Table 3

[0064]

[0065]

[0066] The system scores are divided into 5 levels, with priority from high to low, as shown in Table 4:

[0067] Table 4

[0068]

[0069] In this method, the weight calculation is based on statistical analysis of historical data, realizing the quantification process of parameter determination. When constructing the baseline data feature matrix, vibration data is differentiated by operating condition, and the most stable vibration data under different operating conditions is selected into the baseline data to prevent the equipment score from decreasing due to changes in operating conditions, thus improving the stability of the equipment score. The calculation of dynamic weights considers the degree of influence of a certain factor on the equipment health status during the equipment deterioration stage. The process of constructing the data domain analyzes the structure of the equipment and data correlation, making the health score calculation more reasonable. The use of health data modeling reduces the need for labeled fault data, which is more in line with the needs of industrial scenarios.

Claims

1. A method for assessing equipment health, characterized in that... Includes the following steps: Step 1: Data preparation. Establish the data domain of the target device and determine the structure tree of the target device according to different levels. Step 2: Download historical data. Download historical data for all data items, including historical alarms, maintenance records, historical vibration signals, and time-series data. Step 3: Calculate the initial weights. Based on the historical alarm counts or fault counts determined from maintenance records, determine the initial weights for child nodes under the same parent node proportionally. Step 4: Construct a baseline data feature matrix. Use the K-means algorithm to divide the historical vibration data into working conditions. Perform statistical analysis on the vibration data under each working condition. Based on the calculated statistical characteristics, determine the most stable multiple vibration data under each working condition. Finally, obtain the most stable N vibration data with a length of L1. Perform statistical analysis on the data in each window after the time series data sliding window. Based on the calculated statistical characteristics, determine the most stable M data windows with a length of L2. Construct a baseline data matrix with a dimension of N×L1 or M×L2 for each parameter. Transform the baseline data matrix into the feature space and construct the baseline data feature matrices of N×L1' and M×L2'. Step 5: Calculate the scoring scale. Traverse the historical vibration data or historical signal data window and transform it to the feature space. Use the distribution analysis method to analyze the difference between each observed data feature and each sample in the corresponding benchmark data feature matrix and calculate the average value to obtain N quantized difference values. Construct difference vectors E=[E1,E2,...,EN] respectively. Select the maximum value Emax and the minimum value Emin in the vector to construct the scoring reference interval, and then correspond to the [0,1] interval to realize quantization. Step 6: Real-time data download. Download the data for each data item. For vibration data, acquire the waveform of length L1; for time series data, acquire the data sequence of length L2. Step 7: Calculate the KPI score of the data item, convert the real-time data to the feature space, and perform difference analysis with the data in the corresponding benchmark data feature matrix to calculate the real-time KPI score Ereal. Determine the KPI score in the [0,1] interval based on its relationship with Emax and Emin. For KPI scores that are out of range, dynamically update the difference vector E after evaluation. Step 8: Calculate the dynamic weights of the KPIs. Using the calculated KPI scores as independent variables, a non-linear function is used to calculate the dynamic weights based on the initial weights. The formula is as follows: （1）， For dynamic weights The following normalization process is performed to ensure that the sum of the dynamic weights of all KPIs is 1. This dynamic adjustment mechanism automatically gives more attention to KPIs with unsatisfactory scores. （2）， in, For the first Item dynamic weight, For the first Initial weights of items For the first Real-time KPI scores for each item; Step 9: Combine the KPI score of each data item with its corresponding dynamic weight to calculate the KPI score of the part, and calculate the dynamic weight of the part according to Formula 1. Step 10: Combine the KPI score of each part with its corresponding dynamic weight to calculate the KPI score of the equipment, and calculate the dynamic weight of the equipment according to Formula 1. Step 11: Combine the KPI scores of each device with their corresponding dynamic weights to calculate the KPI score of the system.

2. The equipment health assessment method according to claim 1, characterized in that: The structure tree of the object device in step one includes a multi-level structure of system, device, part, and data item, and the data item is attached to each node of the structure tree.

Images