A time-varying health baseline construction method based on long short-term deep regression observer

Abstract

This invention discloses a method for constructing a time-varying health baseline based on a long-short-term deep regression observer, relating to the field of health baseline construction under changing operating conditions. The method includes: acquiring equipment status detection data and obtaining equipment health characterization parameter combination data through parameter correlation analysis of the equipment status detection data; performing PCA-based time-varying feature mining on the equipment health characterization parameter combination data to obtain fused time-varying health feature data containing equipment operating condition information; constructing a health baseline model based on a long-short-term deep regression observer based on the fused time-varying health feature data, constructing a training dataset using a sliding window method, training the health baseline model based on the long-short-term deep regression observer to obtain a trained health baseline model based on the long-short-term deep regression observer; and conducting anomaly detection performance testing on the trained health baseline model based on the long-short-term deep regression observer using test data.

Description

Technical Field

[0001] This invention relates to the field of health baseline construction technology under changing operating conditions, and specifically to a time-varying health baseline construction method based on long-short-term deep regression observers. Background Technology

[0002] The health status of equipment during operation can often be reflected through sensor monitoring data. Therefore, accurate and effective health monitoring helps to understand its health status trend, identify potential risks during operation, and take corresponding control measures. This provides effective technical support for equipment operation decisions and management, and is of great significance for ensuring stable equipment operation, avoiding downtime accidents, and improving the stable and reliable operation of equipment. Existing data-driven health monitoring methods reduce the need for prior knowledge of equipment failure mechanisms, and due to the nonlinear feature extraction and generalization capabilities of the models, they have gradually become a mainstream research hotspot. Currently, health status monitoring technologies include probabilistic statistical methods and deep learning-based methods. However, current research mainly focuses on monitoring the status of equipment under a certain stable operating process, which has certain limitations regarding parameter changes caused by variations in operating conditions.

[0003] A health baseline is an effective means of online health monitoring that tracks the health status of continuously changing time-series monitoring data in real time. However, current health baseline models are often built based on monitoring data under specific operating conditions, failing to effectively model the health status under different conditions. When monitoring data changes due to variations in operating conditions, leading to signal changes, false alarms are easily generated. Furthermore, existing health baseline construction methods often only construct health status mapping relationships for single-point local data, failing to effectively handle parameter state changes in long-period series data. Summary of the Invention

[0004] To address the technical problems of poor fitting performance and low health monitoring capability of health baseline models constructed from time-series monitoring data under changing operating conditions, this invention proposes a time-varying health baseline construction method based on long-term and short-term deep regression observers.

[0005] This invention provides a method for constructing a time-varying health baseline based on a long-short-term deep regression observer, comprising: Acquire equipment status detection data, and obtain equipment health characterization parameter combination data by performing parameter correlation analysis on the equipment status detection data; Based on the combined data of the associated parameters of the equipment health characterization, time-varying feature mining based on PCA is performed to obtain fused time-varying health feature data containing equipment operating condition information; Based on the fused time-varying health feature data, a health baseline model based on a long-short-term deep regression observer is constructed, and a training dataset is constructed by sliding window segmentation. The health baseline model based on the long-short-term deep regression observer is trained to obtain a trained health baseline model based on the long-short-term deep regression observer. Using test data, the anomaly detection performance of the trained healthy baseline model based on long and short term deep regression observers was tested.

[0006] Preferably, the step of obtaining the combination of equipment health characterization parameters by performing parameter correlation analysis on the plurality of status monitoring parameters includes: Using the Pearson correlation coefficient algorithm, the correlation coefficients between each pair of state monitoring parameters are calculated, and the state monitoring parameters with correlation coefficients greater than the correlation threshold are used as the combination of associated parameters for equipment health characterization.

[0007] Preferably, the step of obtaining fused time-varying health feature data containing equipment operating condition information by performing PCA-based time-varying feature mining on the combination of equipment health characterization correlation parameters includes: By performing vector discretization processing based on one-hot encoding on the working condition identifier data in the combination of equipment health characterization associated parameters, working condition vector data that can reflect changes in working condition information is obtained. Principal component analysis is used to reduce the dimensionality of the condition vector data and the status monitoring data in the combination of equipment health characterization parameters to obtain fused feature data containing equipment health status information and operating condition information. By extracting time-domain features from the fused feature data, fused time-varying health feature data is obtained.

[0008] Preferably, the healthy baseline model based on the long short-term deep regression observer comprises two layers of long short-term memory (LSTM) units and four fully connected layers. The activation function for each layer of the Long Short-Term Memory (LSTM) unit is Sigmoid; the activation function for each layer of the fully connected layer is ReLU.

[0009] Preferably, after obtaining the trained healthy baseline model based on long and short-term deep regression observers, the method further includes: One parameter in the combination of device health characterization parameters is determined as the judgment parameter, and the data distribution of the judgment parameter under the health status is statistically analyzed. By utilizing the data distribution of the judgment parameters under healthy conditions, an adaptive threshold for the estimated values ​​of the judgment parameters is obtained.

[0010] Preferably, the adaptive threshold for obtaining the estimated value of the judgment parameter by utilizing the data distribution of the judgment parameter under the health state includes: By inputting the combined parameters of the device health characterization into the trained health baseline model based on long-term and short-term deep regression observers, the output result of the health baseline is obtained; Calculate the standard deviation of the data distribution of the judgment parameters based on the data distribution of the judgment parameters under the healthy state; An adaptive threshold for the estimated values ​​of the interpretation parameters is obtained based on the output of the health baseline and the standard deviation of the interpretation parameter data distribution.

[0011] Preferably, the adaptive threshold of the estimated interpretation parameter includes:

[0012] in To determine the upper threshold of the parameter estimate, To determine the lower limit threshold of the parameter estimate, The output results are for the healthy baseline. To determine the standard deviation of the parameter data distribution.

[0013] The beneficial effect of this invention is that by discretizing and encoding the operating condition data, combining it with the monitoring parameter sequence data, the fused time-varying health characteristics of the equipment are extracted, and a health baseline model of long-term and short-term regression observers is constructed, thereby realizing the learning of the law of parameter health status. Attached Figure Description

[0014] Figure 1 This is a flowchart of the time-varying health baseline construction method based on long-short-term deep regression observer provided by the present invention; Figure 2 is a correlation matrix diagram of the hydraulic system provided by the present invention; Figure 3 is a graph showing the time-varying monitoring parameters provided by the present invention under different operating conditions; Figure 4 is a working condition dimensionality reduction clustering diagram provided by the present invention; Figure 5 is a schematic diagram of the working condition-parameter fusion data provided by the present invention; Figure 6 is a schematic diagram of time-varying health characteristics provided by the present invention; Figure 7 is a structural diagram of the time-varying health baseline model based on the long-short-term deep regression observer provided by the present invention; Figure 8 This is a training process loss function curve provided by the present invention; Figure 9 is a schematic diagram of the construction result of the time-varying health baseline model based on the long-short-term deep regression observer provided by the present invention; Figure 10 is a schematic diagram of the second result of constructing the time-varying health baseline model based on the long-short-term deep regression observer provided by the present invention; Figure 11 is a schematic diagram of the construction results of the time-varying health baseline model based on the long-short-term deep regression observer provided by the present invention; Figure 12 is a schematic diagram showing the comparison results of different health baseline models provided by the present invention; Figure 13 This is a flowchart of a time-varying health baseline construction method based on a long-short-term deep regression observer provided by the present invention. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no inherent meaning. Therefore, "module," "part," or "unit" may be used interchangeably.

[0016] Figure 13 This is a flowchart of a time-varying health baseline construction method based on a long-short-term deep regression observer provided by the present invention, as follows: Figure 13 As shown, it includes: Step S101: Obtain multiple status monitoring parameters of the equipment under different operating conditions, and obtain the combination of equipment health characterization parameters by performing parameter correlation analysis on the multiple status monitoring parameters; Step S102: By performing PCA-based time-varying feature mining on the combination of the equipment health characterization associated parameters, fused time-varying health feature data containing equipment operating condition information is obtained; Step S103: Construct a health baseline model based on a long-short-term deep regression observer, and train the health baseline model based on the long-short-term deep regression observer using the fused time-varying health feature data to obtain a trained health baseline model based on the long-short-term deep regression observer. Step S104: Obtain test data and use the test data to test the anomaly detection performance of the trained healthy baseline model based on long and short term deep regression observers.

[0017] Furthermore, the step of obtaining the equipment health characterization associated parameter combination by performing parameter correlation analysis on the multiple status monitoring parameters includes: using the Pearson correlation coefficient algorithm to calculate the correlation coefficient between each pair of status monitoring parameters, and taking the pair of status monitoring parameters with a correlation coefficient greater than the correlation threshold as the equipment health characterization associated parameter combination.

[0018] Specifically, the step of obtaining fused time-varying health feature data containing equipment operating condition information by performing PCA-based time-varying feature mining on the combination of equipment health characterization associated parameters includes: performing vector discretization processing based on one-hot encoding on the operating condition identifier data in the combination of equipment health characterization associated parameters to obtain operating condition vector data that can reflect changes in operating condition information; using principal component analysis to perform dimensionality reduction processing on the operating condition vector data and the status monitoring data in the combination of equipment health characterization associated parameters to obtain fused feature data containing equipment health status information and operating condition information; and extracting time-domain features from the fused feature data to obtain fused time-varying health feature data.

[0019] Furthermore, the healthy baseline model based on the long short-term deep regression observer consists of two layers of long short-term memory units (LSTM) and four layers of fully connected layers; wherein, the activation function of each layer of the long short-term memory unit is Sigmoid; and the activation function of each layer of the fully connected layers is ReLU.

[0020] After obtaining a trained health baseline model based on a long-short-term deep regression observer, the present invention further includes: determining one parameter in the combination of device health characterization associated parameters as a judgment parameter, and statistically analyzing the data distribution of the judgment parameter under the health state; and using the data distribution of the judgment parameter under the health state to obtain an adaptive threshold for the estimated value of the judgment parameter.

[0021] Furthermore, the adaptive threshold for obtaining the estimated value of the interpretation parameters by utilizing the data distribution of the interpretation parameters under the health state includes: inputting the combination of the device health characterization correlation parameters into the trained health baseline model based on the long short-term deep regression observer to obtain the output result of the health baseline; calculating the standard deviation of the data distribution of the interpretation parameters according to the data distribution of the interpretation parameters under the health state; and obtaining the adaptive threshold of the estimated value of the interpretation parameters based on the output result of the health baseline and the standard deviation of the data distribution of the interpretation parameters.

[0022] The adaptive threshold for the estimated interpretation parameter includes:

[0023] in To determine the upper threshold of the parameter estimate, To determine the lower limit threshold of the parameter estimate, The output results are for the healthy baseline. To determine the standard deviation of the parameter data distribution.

[0024] This invention proposes a method for constructing a time-varying health baseline based on a long-short-term deep regression observer, comprising: Step 101: Based on the equipment status detection data, and by performing parameter correlation analysis on the equipment status detection data, obtain the combination of equipment health characterization correlation parameters; Furthermore, the parameter correlation analysis method described in step 101 specifically involves first calculating the correlation coefficient for different parameter sequence combinations, then setting a correlation threshold based on expert knowledge, thereby selecting parameter combinations that meet the correlation threshold requirements as equipment health characterization correlation parameter combination data.

[0025] Step 102: For the combined data of the associated parameters of the equipment health characterization, perform time-varying feature mining based on PCA to obtain fused time-varying health feature data containing equipment operating condition information; Furthermore, the PCA-based time-varying feature mining method described in step 102 first discretizes the operating condition identification data to obtain a vector space representing the operating condition information. Then, it combines principal component analysis (PCA) to fuse the operating condition vector data with the monitoring parameter data to obtain time-varying health information. Finally, it extracts time-domain features from the fused feature data to obtain time-varying health features.

[0026] Furthermore, the time-varying health information fusion method described above uses the PCA method to uniformly reduce the dimensionality of the parameter sequence and the operating condition coding vector, thereby obtaining a fused feature vector containing equipment health status information and operating condition information.

[0027] Step 103: Based on the training set data, construct a health baseline model based on long-term and short-term deep regression observers, and adaptively generate interpretation thresholds based on historical data; Furthermore, the method for constructing a health baseline model based on a long-short-term deep regression observer, as described in step 103, generates multiple training samples by using a sliding window to segment the sample data. Then, a long-short-term deep regression observer model is constructed and trained. Simultaneously, the distribution of health status input parameters is statistically analyzed to generate corresponding adaptive interpretation thresholds. In other words, based on the fused time-varying health feature data, a health baseline model based on a long-short-term deep regression observer is constructed. A training dataset is built using a sliding window segmentation method, and the health baseline model based on the long-short-term deep regression observer is trained to obtain a trained health baseline model based on the long-short-term deep regression observer.

[0028] Furthermore, the long-short-term deep regression observer model consists of two LSTM units and four fully connected layers, with each layer having a nonlinear activation function to better fit the nonlinear fluctuations in parameters caused by changes in operating conditions.

[0029] Step 104: Using the test data, perform anomaly detection performance testing on the constructed health baseline model.

[0030] Furthermore, the performance test of the health baseline model based on the long-short-term deep regression observer described in step 104 is based on the test set data of equipment status monitoring. For its associated parameter group, step two is repeated to obtain time-varying health features, which are then input into the trained health baseline model. By observing the output of the health baseline model and combining it with the adaptive interpretation threshold, the anomaly detection performance test of the health baseline model is completed.

[0031] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The flowchart of the time-varying health baseline construction method based on long-short-term deep regression observer proposed in this invention is as follows: Figure 1 As shown, it includes: Step 1: Correlation Analysis of Health Characteristic Parameters Typically, equipment has numerous condition monitoring parameters, and due to the complexity of the equipment's structure, there are intricate coupling relationships between different parameters. Therefore, there are often certain correlations between parameters. This invention uses the Pearson correlation coefficient between parameters to screen for parameter combinations with correlations. Assume parameter X has the following data sequence: ,parameter Its data sequence is The correlation coefficient between the two parameters is calculated as follows:

[0032] In the above formula, where P r Let f be the correlation coefficient, and let f be the correlation coefficient calculation function. When the correlation coefficient > 0.9, there is a strong correlation between the two parameter sequences. In this invention, the correlation threshold is set to 0.95 to select parameter combinations with strong correlations for constructing a health baseline model.

[0033] Step 2: Time-varying feature mining based on PCA Based on the selected strongly correlated parameter combinations In The sequence is used as the input parameter. As parameters to be interpreted. First, the operating condition identifier sequence is considered. Discretized vector coding based on one-hot encoding is performed to obtain the operating condition coding matrix. , where m represents the total number of working condition types in the entire working condition sequence. Then, the working condition encoding matrix is ​​concatenated with the input parameter vector to obtain the concatenated matrix. Based on this, principal component analysis (PCA) is used to reduce the dimensionality of the multidimensional vector and perform fusion.

[0034] In the above formula, F is the multidimensional information fusion vector. For the PCA algorithm, the dimension of the reduced vector is set to 1, and R is the concatenated vector of the parameter vector and the working condition code vector. Based on this, temporal feature extraction is performed. This invention selects the root mean square value as the temporal feature index. Assuming the window width is w, the samples generated by sliding the window are as follows:

[0035]

[0036] In the above formula, f represents the cut sample with length w. The number of samples generated by the segmentation is s, the window movement step size is s, and L is the length of the multi-factor information fusion vector F. Based on the segmented sample set generated by the above method, root mean square feature extraction is performed, calculated as follows:

[0037] Finally, the time-varying health feature sequence was obtained.

[0038] Step 3: Construction and Training of a Healthy Baseline Model Based on Long-Short-Term Deep Regression Observers Based on the time-varying health features mined in the above steps, a health baseline model based on a long-short-term deep regression observer is trained. A training dataset is constructed using a sliding window method, and the model is then trained. The model structure is as follows: Figure 7 As shown, the model mainly consists of Long Short-Term Memory (LSTM) units and fully connected layers. The LTM units comprise input gates, output gates, and forget gates. By introducing a gate mechanism, the flow and loss of information in long-period sequence data can be effectively controlled, enabling the memorization of long-period information and thus avoiding gradient vanishing and exploding problems. Long-period time-series feature data is extracted through LTM units, and finally, multiple fully connected layers are used to perform a deep nonlinear transformation on the feature vectors, thereby completing the construction of a healthy baseline model based on a LTM deep regression observer. Then, by selecting an appropriate optimizer and number of iterations, the model parameters are trained and updated using the backpropagation algorithm, completing the mapping from time-domain feature data to the parameter data to be interpreted.

[0039] Based on the original parameter data in the training set, the data distribution is statistically analyzed to obtain the upper and lower limits of the normal state. This invention uses the data standard deviation as a measure of data distribution, and takes 3 times the standard deviation as the limit of the normal state data distribution. The calculation method is as follows:

[0040] In the above formula, x is the sequence of parameters to be judged, and n is the data length.

[0041] Here, the original parameter data is one parameter in the associated parameter combination. For example, if AB is an associated parameter combination, during model training: the time-varying features of A are used as the model input, and B is used as the model output. Then, the statistics here are the data distribution of B.

[0042] Finally, the extracted time-varying health features are input into the health baseline model, and the output of the health baseline is: Combined with statistical data, the range of parts Through calculation The obtained interval range is used as the normal range of fluctuation of the parameter to be judged. Therefore, by controlling the output data range of the health baseline result, the judgment threshold can be generated.

[0043] Step 4: Performance Testing of the Healthy Baseline Model Based on Long-Short-Term Deep Regression Observers The test set data, after step two, extracts time-varying health features from the input parameter data of the correlation parameters, which are then used as input to the health baseline model. The output of the health baseline model is combined with the normal data distribution range obtained in step four to generate an adaptive threshold, thereby constructing a normal fluctuation range space for the judgment parameters. The parameters to be judged in the test data are compared with the normal fluctuation range space constructed by the health baseline model, and the proportion of abnormal data points is statistically analyzed. The proportion of abnormal data points is calculated as follows:

[0044] In the above formula, r represents the proportion of outlier data. The number of sample points that are outside the range. The total number of samples to be analyzed is considered normal when most data points are within the normal fluctuation range.

[0045] Example 1. Description of Experimental Data This invention uses hydraulic system condition monitoring data under different operating conditions for case verification. In this case, the hydraulic system has 348,945 sample points, and the collected parameters include a total of 12 parameters such as hydraulic oil tank volume, suction oil temperature, hydraulic system pressure, accumulator temperature, accumulator pressure, return oil filter temperature, high-pressure oil filter temperature, high-pressure oil filter differential pressure, and return oil filter differential pressure. The operating condition profile covers preparation, entry, departure, cruising, approach, landing and taxiing, and exit conditions. The health baseline construction method proposed in this invention is verified based on the hydraulic system condition monitoring data under the above-mentioned operating condition changes.

[0046] 2. Case Validation First, a parameter correlation analysis is performed. Based on hydraulic system data, this invention uses the Pearson coefficient to calculate parameter correlation, and the results are as follows. Figure 2As shown: Figure 2 To analyze the correlation of hydraulic system condition monitoring parameters, this method sets a correlation threshold of 0.95 to screen for strongly correlated parameter groups. The strongly correlated parameter groups obtained based on the correlation threshold include combinations such as "parameter 3-parameter 5" and "parameter 10-parameter 12". Considering the given data conditions, "parameter 3-parameter 5" represents "system pressure - accumulator pressure", and "parameter 10-parameter 12" represents "high-pressure oil filter temperature - oil pump housing return oil filter temperature". Therefore, "parameter 10-parameter 12" is selected as the correlation parameter group reflecting the health status of the hydraulic system's oil filters.

[0047] Then, for the time-varying state monitoring parameters that vary with operating conditions, time-varying feature mining is carried out to support the health baseline model in mining health status information from the monitoring data. The original monitoring data and operating condition identification data used to construct the health baseline model are as follows: Figure 3 As shown.

[0048] Depend on Figure 3 It is evident that the changing patterns of the monitoring parameters are subtle and complex, and exhibit a strong correlation with the trends in operating conditions. Therefore, it is necessary to perform time-varying feature extraction based on operating condition information and parameter status information to improve the data's ability to express health status. To better extract fusion features from the fused data, the operating condition identifier is discretized into a vector based on one-hot encoding, thereby obtaining an operating condition vector that reflects changes in operating condition information. TSNE is then used to reduce the dimensionality of the discretized high-dimensional operating condition vector, and the results are as follows. Figure 4 As shown: Depend on Figure 4 It can be seen that after data compression and dimensionality reduction of the discrete operating condition vectors, they can be mainly divided into 5 categories. Combined with expert knowledge, it can be known that the operating conditions during aircraft operation are mainly divided into 5 categories: preparation, entry / exit, departure / approach, cruise, and landing roll. It can be seen that by performing vector discretization processing on the operating condition identification data, the expressiveness of the operating condition data can be improved, and the working operation process of the object can be reflected more abstractly and comprehensively. It also improves the data's processability.

[0049] Then, the time-varying operating condition information and parameter state information are fused. Based on principal component analysis (PCA), the multidimensional vector composed of the monitoring data vector and the operating condition coding vector is subjected to principal component fusion extraction to obtain the time-varying feature vector, thereby realizing the fused expression of time-varying operating condition information and parameter information, such as... Figure 5 As shown: To improve the temporal feature variation information of the data, a sliding window method is used to extract the temporal features of the data within the window for the extracted fusion feature data, thereby obtaining a time-varying health feature sequence.

[0050] like Figure 6As shown, this invention extracts root mean square (RMS) time-domain features from fused data to obtain a time-varying health feature sequence. The extracted features not only retain the development trend of the original parameters, but also reduce data fluctuation interference in the peak region of operating condition transition, thus retaining effective operating condition change information. This achieves the extraction of time-varying feature sequences that combine parameter state information with time-varying operating condition information.

[0051] Based on the constructed time-varying health feature sequence, a health baseline model based on a long short-term deep regression observer is built. This paper uses a long short-term memory neural network as the basis for model construction, and the model structure is as follows. Figure 7 As shown in the table below, the parameters of the constructed health baseline model are as follows: Table 1. Parameters of the time-varying health baseline model based on long-short-term deep regression observers.

[0052] Table 2 Training Hyperparameter Configuration

[0053] Based on the above model parameter configuration and training parameter configuration, a health baseline model is trained using time-varying health feature data. The loss function during the training process is as follows: Figure 8 As shown.

[0054] Using a trained health baseline model, the effect of constructing the correlation mapping relationship between time-varying health features and target monitoring parameters in the training data is demonstrated. The health baseline construction effect is as follows: Figure 9 As shown.

[0055] Based on this, by statistically analyzing the distribution of the observer output parameters in the training data, upper and lower thresholds for the health baseline are adaptively mined and generated. This achieves adaptive updating of the health baseline threshold, avoiding false alarms or false negatives caused by traditional single-parameter thresholds. The generated adaptive health baseline thresholds are as follows: Figure 10 As shown.

[0056] Depend on Figure 10 It can be seen that the data to be interpreted in the training data is generally in a fluctuating upward phase, with sudden changes in parameters in local data regions due to changes in operating conditions. As shown in the figure, when the operating conditions change, for local data segments where the parameters change from an upward trend to a fluctuating downward trend, the time-varying health baseline model of the long-short-term deep regression observer trained based on the time-varying health features of the input parameters can effectively capture changes in operating conditions, thus accurately fitting the local parameter abrupt changes in the associated data caused by changes in operating conditions.

[0057] Then, based on the test data, performance tests were conducted on the time-varying health baseline model based on long-short-term deep regression observers. For the associated parameter groups in the test set, based on the parameter correlation analysis results, parameter 12 was used as the data to be judged. The health baseline detection results are as follows: Figure 11 As shown.

[0058] Depend on Figure 11 As can be seen, the health baseline proposed in this invention can effectively identify and utilize operating condition data, especially for monitoring data that varies with operating conditions. It also demonstrates excellent and timely tracking and fitting capabilities for parameter abrupt changes caused by changes in operating conditions, thereby enhancing the parameter correlation mapping expression ability and real-time tracking and detection capabilities of the health baseline model. Furthermore, this invention was compared with a healthy baseline model constructed using support vector regression and a healthy baseline model constructed based on the original parameters. The anomaly detection performance of the three methods on the test set data is as follows: Figure 12 As shown.

[0059] Depend on Figure 12 It can be seen that the time-varying health baseline model based on long-short-term deep regression observer proposed in this invention can well fit the data sequence of parameter -12 of the data to be judged in the test set sample data, and can also well fit and track the inflection point of parameter mutation caused by changes in working conditions.

[0060] When using the original associated parameter group data to build a health baseline model, the output of the health baseline model can only correspond well with the data to be interpreted in general trend. However, it cannot accurately identify the inflection point changes of the curve caused by changes in operating conditions. Therefore, when the equipment experiences an anomaly during the operating condition transition phase, such health baselines are prone to false detections because they cannot accurately track the inflection point of parameter shape changes.

[0061] When constructing a health baseline based on time-varying health feature data using the support vector regression algorithm, as shown in the figure, the output of the health baseline model has a relatively consistent trend with the parameters to be judged, and it can also track and fit the inflection point change trend of parameters caused by changes in operating conditions to a certain extent. However, it can be seen that the model still fits the sudden increase or decrease of parameters caused by changes in operating conditions at the beginning and end of the operating conditions. Therefore, although the health baseline model can track and fit the inflection point change trend of parameters caused by changes in operating conditions, it cannot learn the operating change pattern of long-term data and cannot effectively model the operating trend of long-term related parameter groups and the correlation between parameters. It can only reflect the parameter correlation mapping information at a single point in time. Therefore, it is easy to make false detections when parameters increase or decrease due to normal operating conditions.

[0062] This invention compares the performance of three methods by the degree of curve deviation and the proportion of false detection samples, as shown in Table 3.

[0063] Table 3 Comparison results of different types of health baselines

[0064] The degree of curve deviation reflects the accuracy of the health baseline model in fitting parameter correlations, while the anomaly detection ratio reflects the detection rate of missed or false positives for samples by the current baseline model combined with the discovered adaptive threshold. As shown in the table, the time-varying health baseline model based on long-term and short-term deep regression observations proposed in this invention can accurately discover parameter correlations and promptly track inflection point trends of parameters caused by time-varying operating conditions. Therefore, it has the lowest curve deviation, and the number of normal samples identified as abnormal samples is 10, with a false positive ratio of 5.73 × 10⁻⁶. -4 The overall detection performance of the model is relatively good, although the score is relatively low.

[0065] The baseline model built using the original parameters exhibits a relatively low degree of curve deviation (2.1 × 10⁻⁶). -3 However, it had the highest false positive rate, at 1.51 × 10⁻⁶. -2 This indicates that this type of health baseline can only fit the curve development trend well, but cannot effectively detect parameter changes caused by time-varying operating conditions.

[0066] For a support vector regression healthy baseline model based on time-varying features, the highest deviation of its curve is 8.2 × 10⁻⁶. -3 Although its false detection rate is 0, the reason is that the model cannot effectively learn the operating rules of long-period data sequences, so its anomaly detection capability is relatively poor.

[0067] In summary, this invention proposes a time-varying health baseline construction method based on a long-short-term deep regression observer. This method extracts equipment-related parameter groups through parameter correlation analysis, and then extracts time-varying fusion features of health status through discretization coding of operating conditions and multi-dimensional information fusion, thereby realizing the fusion expression of parameter health status information and equipment operating condition information. On this basis, by constructing a health baseline model using a long-short-term deep regression observer, a unified modeling of the operating law of long-period parameter sequences and the mapping relationship between parameter correlations is achieved, ultimately improving the real-time parameter tracking capability and state discrimination capability of the health baseline model.

[0068] The advantages and positive effects of this invention are as follows: (1) The present invention can perform discretization vector encoding processing on the working condition identification data, and then fuse the monitoring data information and the working condition vector information in the encoding space to obtain the health status fusion vector, thereby improving the ability of the monitoring data to express the state under the time-varying working conditions.

[0069] (2) The present invention can effectively retain and utilize the health status change information of long-term monitoring data, improve the fitting accuracy of the correlation between parameters and health characteristics, and adaptively generate the threshold of the judgment parameters. (3) The present invention can effectively utilize time-series monitoring parameter data and operating condition sequence data to complete the fusion expression of equipment operating condition information and health status information, and construct a time-varying health baseline model that changes with operating conditions and parameters. It can mine the long-term state change information of the parameter sequence itself and accurately fit the health correlation characteristics and state change law between parameters.

[0070] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the present invention.

Claims

1. A method for constructing a time-varying health baseline based on a long-short-term deep regression observer, characterized in that, include: Multiple status monitoring parameters of the equipment under different operating conditions are obtained, and a combination of equipment health characterization parameters is obtained by performing parameter correlation analysis on the multiple status monitoring parameters. By performing PCA-based time-varying feature mining on the combination of equipment health characterization parameters, fused time-varying health feature data containing equipment operating condition information is obtained, which includes: By performing vector discretization processing based on one-hot encoding on the working condition identifier data in the combination of equipment health characterization associated parameters, working condition vector data that can reflect changes in working condition information is obtained. Principal component analysis is used to reduce the dimensionality of the condition vector data and the status monitoring data in the combination of equipment health characterization parameters to obtain fused feature data containing equipment health status information and operating condition information. By extracting time-domain features from the fused feature data, fused time-varying health feature data is obtained; A health baseline model based on a long-short-term deep regression observer is constructed, and the fused time-varying health feature data is used to train the health baseline model based on the long-short-term deep regression observer to obtain a trained health baseline model based on the long-short-term deep regression observer. Acquire test data and use the test data to test the anomaly detection performance of the trained healthy baseline model based on long and short term deep regression observers.

2. The method according to claim 1, characterized in that, The method of obtaining the equipment health characterization correlation parameter combination by performing parameter correlation analysis on the multiple status monitoring parameters includes: Using the Pearson correlation coefficient algorithm, the correlation coefficients between each pair of state monitoring parameters are calculated, and the state monitoring parameters with correlation coefficients greater than the correlation threshold are used as the combination of associated parameters for equipment health characterization.

3. The method according to claim 2, characterized in that, The healthy baseline model based on the long short-term deep regression observer consists of two layers of long short-term memory (LSTM) units and four fully connected layers. The activation function for each layer of the Long Short-Term Memory (LSTM) unit is Sigmoid; the activation function for each layer of the fully connected layer is ReLU.

4. The method according to claim 3, characterized in that, After obtaining the trained healthy baseline model based on long and short-term deep regression observers, the following is also included: One parameter in the combination of device health characterization parameters is determined as the judgment parameter, and the data distribution of the judgment parameter under the health status is statistically analyzed. By utilizing the data distribution of the judgment parameters under healthy conditions, an adaptive threshold for the estimated values ​​of the judgment parameters is obtained.

5. The method according to claim 4, characterized in that, The adaptive threshold for obtaining the estimated value of the interpretation parameter by utilizing the data distribution of the interpretation parameter under the health state includes: By inputting the combined parameters of the device health characterization into the trained health baseline model based on long-term and short-term deep regression observers, the output result of the health baseline is obtained; Calculate the standard deviation of the data distribution of the judgment parameters based on the data distribution of the judgment parameters under the healthy state; An adaptive threshold for the estimated values ​​of the interpretation parameters is obtained based on the output of the health baseline and the standard deviation of the interpretation parameter data distribution.

6. The method according to claim 5, characterized in that, The adaptive threshold for the estimated interpretation parameter includes: in To determine the upper threshold of the parameter estimate, To determine the lower limit threshold of the parameter estimate, The output results are for the healthy baseline. To determine the standard deviation of the parameter data distribution.

Images