A device health state monitoring method and system based on survival analysis

Abstract

The application relates to the technical field of equipment state monitoring, and discloses a device health state monitoring method and system based on survival analysis, which comprises the following steps: training an LSTM model by using historical sensor data collected by a target device in a healthy running state to obtain a state reference model; inputting real-time sensor data in the running process of the target device into the state reference model for analysis, extracting state analysis results at each time step, fusing and analyzing corresponding state scores, and constructing a state analysis curve; combining the real-time state to calculate corresponding survival risks, simulating a device state evolution process, and constructing a state change curve; and calculating the remaining service life and the healthy period of the device to monitor the device health state; the application can be corrected in real time according to the actual degradation path of the device, and the accuracy and adaptability of the state monitoring process are improved.

Description

Technical Field

[0001] This application relates to the field of equipment condition monitoring technology, and more specifically to a method and system for monitoring equipment health status based on survival analysis. Background Technology

[0002] Currently, in the fields of modern industry and intelligent manufacturing, the continuous and reliable operation of large-scale machinery and critical infrastructure, such as wind turbines, rail transit equipment, and industrial robots, is crucial for ensuring production efficiency and economic benefits. With the development of the Internet of Things (IoT) and sensor technology, equipment condition monitoring has entered the era of big data, capable of collecting massive amounts of time-series data such as vibration, temperature, current, and pressure in real time. Traditional equipment health management relies on reactive repairs and periodic maintenance, failing to provide targeted maintenance measures.

[0003] Existing technologies suffer from the following problems: judging equipment status by setting fixed alarm thresholds is insensitive to early faults that slowly degrade, making it easy to miss alarms; a single LSTM model fits the time series and outputs a predicted value for the next moment, which is difficult to intuitively reflect the health of the equipment and cannot reflect the equipment degradation process; treating status identification and lifetime prediction as two independent modules makes it impossible to predict lifetime based on the status identification results, resulting in a lack of dynamic update capability for the prediction results, failing to reflect sudden changes in the equipment's status during operation, and low accuracy of equipment status monitoring; to solve at least one of the above problems, this application proposes a method and system for monitoring equipment health status based on survival analysis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide a method and system for monitoring equipment health status based on survival analysis, which can effectively solve the problems in the background technology. The specific technical solution of this application is as follows:

[0005] A method for monitoring device health status based on survival analysis, comprising:

[0006] The LSTM model is trained using historical sensor data collected from the target device when it is in a healthy operating state to obtain a state baseline model.

[0007] Real-time sensor data during the operation of the target device is input into the state benchmark model for analysis. The state analysis results of each time step are extracted, and the corresponding state score is calculated by fusion analysis to construct a state analysis curve.

[0008] The real-time status of the equipment is analyzed based on the status analysis curve, and the corresponding survival risk is calculated in combination with the real-time status of the equipment. The equipment status evolution process is simulated, and a status change curve is constructed.

[0009] Based on the state change curve analysis, the remaining service life and health period of the equipment are calculated to monitor the health status of the equipment.

[0010] Specifically, the step of inputting real-time sensor data during the operation of the target device into a state benchmark model for analysis, extracting the state analysis results at each time step, fusing and analyzing to calculate the corresponding state score, and constructing a state analysis curve includes:

[0011] Real-time sensor data during the operation of the target device is input into the state baseline model for analysis. The state analysis results of each time step are extracted, and the reconstruction error between the state analysis results and the input data is calculated to obtain the reconstruction error sequence.

[0012] By performing fusion analysis on the reconstructed error sequence, the corresponding state score is calculated, and a state analysis curve is constructed.

[0013] Specifically, the real-time sensor data during the operation of the target device is input into a state baseline model for analysis, the state analysis results of each time step are extracted, and the reconstruction error between the state analysis results and the input data is calculated to obtain a reconstruction error sequence, including:

[0014] Real-time sensor data during the operation of the target device is input into the state baseline model for analysis. At each time step, the state vectors generated by each layer of the model are extracted and concatenated to obtain the corresponding state analysis results.

[0015] By decoding the state analysis results to generate the corresponding data vector, the Euclidean distance and the cosine of the angle between the data vector and the input data vector are calculated to obtain the reconstruction error.

[0016] The reconstruction errors are integrated over time to obtain the reconstruction error sequence.

[0017] Specifically, the step of performing fusion analysis on the reconstructed error sequence, calculating the corresponding state score, and constructing a state analysis curve includes:

[0018] According to the preset time window and time step, slide on the reconstructed error sequence, calculate the error value in each window, form the baseline layer with the error value less than or equal to the preset error threshold, and form the anomaly layer with the error value greater than the preset error threshold.

[0019] For the baseline layer, analyze the error changes to extract error distribution characteristics and construct error fluctuation boundaries;

[0020] For the anomaly layer, analyze the error accumulation, calculate the cumulative error, and obtain the deviation.

[0021] Based on the error fluctuation boundary and deviation amount, the deviation indicators at different time scales are analyzed and integrated, the corresponding state scores are calculated, and a state analysis curve is constructed.

[0022] Specifically, the step of analyzing and fusing deviation indicators at different time scales based on error fluctuation boundaries and deviation amounts, calculating corresponding state scores, and constructing state analysis curves includes:

[0023] Calculate the ratio of the deviation at each time step to the error fluctuation boundary to obtain the first deviation index;

[0024] The first deviation index is weighted and averaged within the time window to obtain the second deviation index, where the weights are set according to the distance between the time point and the boundary of the time window;

[0025] By combining the first deviation index and the second deviation index, the deviation degree is calculated, and the corresponding state score is obtained by mapping through a preset scoring mapping rule.

[0026] The state scores are integrated in chronological order to construct a state analysis curve.

[0027] Specifically, the step of analyzing the real-time state of the equipment based on the state analysis curve, calculating the corresponding survival risk based on the real-time state of the equipment, simulating the state evolution process of the equipment, and constructing a state change curve includes:

[0028] Extract the corresponding time-series features from the state analysis curves to analyze the real-time state of the equipment;

[0029] By combining the real-time status of the equipment to calculate the corresponding survival risk, the equipment state evolution process is simulated, and a state change curve is constructed.

[0030] Specifically, the process of calculating the survival risk based on the real-time status of the equipment, simulating the equipment state evolution process, and constructing state change curves includes:

[0031] Based on the real-time status of the equipment, the corresponding risk rate is mapped in the preset equipment baseline risk function, and the survival risk is calculated by integrating the risk rate to obtain the cumulative risk value.

[0032] The device's state phase is determined based on the accumulated risk value;

[0033] By combining the risk rate, cumulative risk value, and state phase, and using a preset state evolution model, the state evolution process of the equipment is simulated, and a state change curve is constructed.

[0034] Specifically, the process of analyzing equipment status based on state change curves, calculating the remaining service life and health periods of the equipment, and monitoring the health status of the equipment includes:

[0035] Based on the state change curve analysis, the corresponding state lifetime is matched in the preset equipment state mapping library to obtain the state lifetime set.

[0036] Based on the state lifetime set, the remaining service life and health period of the equipment are calculated to monitor the health status of the equipment.

[0037] Specifically, the step of calculating the remaining service life and health period of the equipment based on the state lifetime set to monitor the health status of the equipment includes:

[0038] The matching degree between each state lifetime and the state change process is analyzed based on the state lifetime set, and corresponding fusion weights are set. The state lifetimes are weighted and averaged according to the fusion weights to obtain the remaining service life of the equipment.

[0039] The boundaries of the health period are determined based on the remaining useful life in order to monitor the health status of the equipment.

[0040] A survival analysis-based equipment health status monitoring system, used to implement the aforementioned survival analysis-based equipment health status monitoring method, includes:

[0041] The health status analysis module trains the LSTM model using historical sensor data collected from the target device when it is in a healthy operating state, thus obtaining a state baseline model.

[0042] The real-time status analysis module inputs real-time sensor data during the operation of the target device into the status benchmark model for analysis, extracts the status analysis results of each time step, fuses and analyzes to calculate the corresponding status score, and constructs a status analysis curve.

[0043] The state evolution simulation module analyzes the real-time state of the equipment based on the state analysis curve, calculates the corresponding survival risk based on the real-time state, simulates the state evolution process of the equipment, and constructs the state change curve.

[0044] The status monitoring module analyzes the equipment status based on status change curves, calculates the remaining service life and health period of the equipment, and monitors the health status of the equipment.

[0045] The beneficial effects of this application are as follows: by constructing a state baseline model to analyze the real-time state of the equipment, the state analysis results of the time steps are fused to construct a state analysis curve, the real-time state of the equipment is analyzed and the survival risk is calculated, and the dynamic process of equipment state change is analyzed by simulating the equipment state evolution process; it can effectively filter out noise, issue early warnings in the early stage of failure, perform survival analysis based on real-time state, and make real-time corrections based on the current actual degradation path of the equipment, thereby improving the accuracy and adaptability of the state monitoring process. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating a device health status monitoring method based on survival analysis, as described in an embodiment of this application.

[0047] Figure 2 This is a flowchart illustrating the process of constructing state analysis curves in an embodiment of this application.

[0048] Figure 3 This is a schematic diagram of the state analysis curves in an embodiment of this application;

[0049] Figure 4 This is a schematic diagram of a device health status monitoring system based on survival analysis in an embodiment of this application. Detailed Implementation

[0050] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0051] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0052] Hereinafter, the terms "first," "second," and other generic terms are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0053] refer to Figure 1 The image shows a specific implementation of a device health status monitoring method based on survival analysis according to this application, comprising:

[0054] S101. Train the LSTM model using historical sensor data collected by the target device in a healthy operating state to obtain a state baseline model.

[0055] S102. Input the real-time sensor data during the operation of the target equipment into the state benchmark model for analysis, extract the state analysis results of each time step, fuse and analyze to calculate the corresponding state score, and construct the state analysis curve.

[0056] S103. Analyze the real-time status of the equipment based on the status analysis curve, calculate the corresponding survival risk based on the real-time status of the equipment, simulate the equipment status evolution process, and construct the status change curve.

[0057] S104. Analyze the equipment status based on the status change curve, calculate the remaining service life and health period of the equipment, and monitor the health status of the equipment.

[0058] In the industrial sector, predictive maintenance of equipment such as wind turbines, rail transit systems, and aircraft engines can prevent major safety accidents and economic losses. In this embodiment, historical sensor data of the target equipment is collected during its initial operation or when it is confirmed by experts to be in a healthy operating state. Taking the gearbox of a certain type of wind turbine as an example, the data collected includes, but is not limited to, high-speed shaft vibration values, gearbox oil temperature, lubricating oil pressure, and rotational speed. The data collection frequency is set to once per minute, and continuous collection is carried out for three months to ensure coverage of the equipment's healthy operating conditions under different loads, such as low wind speed, rated wind speed, and cut-off wind speed.

[0059] Preferably, the collected raw data undergoes preprocessing, including outlier cleaning and normalization. Outlier cleaning uses the interquartile range (IMR) to remove occasional spike noise from the sensors. Normalization maps all sensor parameters to the interval [0, 1], eliminating the influence of different units on model training. After preprocessing, a training dataset suitable for LSTM supervised learning is constructed. A sliding window method is used, with a time window length of 60, indicating that the model predicts the data at the next moment based on the data from the past 60 minutes. For example, data from t=1 to t=60 minutes is used as input features, and data from t=61 minutes is used as the prediction target; then, by sliding one step, data from t=2 to t=61 minutes is used as input, and data from t=62 minutes is used as the target, and so on, generating corresponding input-output sample pairs.

[0060] Optionally, the LSTM model structure used in this embodiment is an encoder-decoder architecture. The encoder consists of two stacked LSTM layers. The first LSTM layer has 128 memory units and sets return_sequences=True to output the hidden state of each time step to the second layer. The second LSTM layer has 64 memory units and only returns the hidden state of the last time step as the context vector of the entire input sequence. The decoder consists of a repeating vector layer and an LSTM layer, which is responsible for decoding the context vector into the target sequence. The last layer of the model is connected to a fully connected layer with the number of neurons equal to the dimension of the input sensor. The output is the reconstructed sensor vector for the next time step.

[0061] During model training, the mean squared error loss function was used, measuring the squared difference between the model's predicted sensor vector and the actual sensor vector at the next time step. The Adam optimizer was used, with an initial learning rate of 0.001. Training data was divided into training and validation sets in an 8:2 ratio, with a batch size of 64 and 100 training epochs. An early stopping mechanism was introduced: training was stopped when the validation set loss no longer decreased for 10 consecutive epochs to prevent overfitting. After training, the trained LSTM model was saved as a baseline model for the device's health status. The model's input was a historical sensor time series window of length 60, and its output was a predicted vector of the sensor values ​​at the next time step.

[0062] It should be noted that by constructing a state baseline model, high-dimensional sensor data under the health state of the equipment is fixed in the network weights of the LSTM model. The state baseline model learns the coupling relationship and temporal evolution law between the parameters of each sensor under the health condition. The model can adapt to different operating conditions and provide an accurate normal baseline for subsequent precise identification of subtle and unhealthy pattern deviations, thereby improving the robustness and accuracy of the health state description process.

[0063] Specifically, real-time sensor data enters the system at the same frequency as during training (once per minute). The system maintains a data buffer of length 60. Whenever a new data point arrives, it is combined with the previous 59 data points to form a new input window, which is then input into the trained state baseline model. As the input window propagates forward through the LSTM network, at each time step, the hidden state vectors output by the first and second LSTM layers are extracted. The state vectors from different layers at the same time step are concatenated end-to-end to obtain a higher-dimensional composite vector, which serves as the state analysis result for that time step.

[0064] Preferably, the model decodes the state analysis results to generate a reconstructed data vector. The difference between the original input data vector and the model-reconstructed data vector is calculated as the reconstruction error. The straight-line distance between the two vectors in Euclidean space and the cosine of the angle between them are calculated. The weighted sum of the straight-line distance and the cosine of the angle is used to obtain the final reconstruction error for that time step. The reconstruction errors of each time step are integrated in chronological order to obtain a reconstruction error sequence.

[0065] Specifically, a sliding window is set to slide across the reconstruction error sequence. The sliding window can be set according to the accuracy requirements of device status monitoring, for example, setting the window width to 30 minutes and the step size to 1 minute. The average error within each window is calculated, and the error threshold is determined by statistically analyzing the 95th quantile of the reconstruction error of historical health data. Based on the preset error threshold, the window is divided into a baseline layer and an anomaly layer.

[0066] Furthermore, a first deviation index is calculated for each time step. This first deviation index is the ratio of the deviation amount at that time step to the corresponding error fluctuation boundary. This ratio eliminates the differences in the normal fluctuation amplitude across different periods and reflects the relative intensity of the deviation. A second deviation index is calculated by weighted averaging the first deviation index. The first and second deviation indices are then merged to obtain the deviation degree. This deviation degree is mapped to a scoring range of 0-100 using a preset Sigmoid mapping rule to obtain the corresponding state score, where 100 represents perfect health and lower scores indicate a worse state. The state scores are then integrated in chronological order to obtain the state analysis curve.

[0067] It should be noted that by extracting the internal state of the LSTM and employing a hierarchical, multi-scale fusion analysis method, the accuracy of the analysis results on equipment state changes can be improved. This allows for the identification of subtle changes in data analysis within the model, thereby enhancing the sensitivity of fault identification. By dividing the error sequence into a baseline layer and an anomaly layer and constructing a dynamic error fluctuation boundary, the normal operating condition fluctuations and abnormal deviations of the equipment can be effectively distinguished, reducing the false alarm rate. By fusing the first and second deviation indices, the generated state analysis curve can reflect sudden shocks and long-term degradation processes, providing accurate data for equipment state risk analysis.

[0068] Specifically, a baseline risk function for the equipment is constructed from historical data. Historical full lifecycle data of the same model of equipment is collected, including complete sensor records from healthy operation to eventual failure and the final failure time. Using this historical failure data, a parameterized Weibull distribution is used to fit a baseline risk function. This function describes the instantaneous failure probability of the equipment at a given time point in the absence of additional state information. The input to the baseline risk function is the time the equipment has been running, and the output is the base risk rate.

[0069] For the current moment and its corresponding state score, a risk multiplier is mapped to the state score using a pre-defined proportional risk model, including but not limited to the Cox model. For example, a mapping table is created using historical data: a state score of 90-100 corresponds to a risk multiplier of 1.0, 80-89 corresponds to a risk multiplier of 1.5, 70-79 corresponds to a risk multiplier of 2.0, and so on. The lower the score, the larger the multiplier, indicating a higher risk of failure.

[0070] Starting from the current moment, time is integrated forward. In each small future time interval Δt, the instantaneous risk rate equals the baseline risk rate multiplied by the risk proportional factor corresponding to the current state score. All instantaneous risk rates are summed to obtain the cumulative risk value from the current moment to future moments, reflecting the total risk accumulated by the equipment from this point in time. Based on the cumulative risk value, the state phase of the equipment is determined. For example, two risk thresholds C1 and C2 can be set: when the cumulative risk value is below C1, it is defined as a healthy stable period; between C1 and C2 is a slow degradation period; and above C2 is a rapid failure period. The calculation process starts from the current moment, and at each future time point, the corresponding instantaneous risk rate, cumulative risk value, and state phase are calculated. The instantaneous risk rate and phase information corresponding to the future time points are recorded and connected by a smooth curve to obtain a state change curve describing the future operating state of the equipment.

[0071] It should be noted that by introducing survival analysis, the static state score is made dynamic. By calculating the cumulative risk value and dividing the state phase, the state stage of the equipment can be reflected, providing reference data for operation and maintenance personnel. The constructed state change curve reflects the evolution trajectory of the dynamic update of the equipment state, providing accurate input data for subsequent calculation of the remaining service life of the equipment.

[0072] Specifically, a device status mapping library is constructed, storing a large amount of historical failure case data. For each historical failure case, the total lifespan at the time of failure and the entire process of state change curve from healthy state to failure are recorded. Each curve corresponds to a known state lifespan. The latest state change curve segment of the current device is obtained and matched in the device status mapping library to obtain multiple historical curves most similar to the current curve segment. The state lifespans corresponding to the historical curves constitute a state lifespan set. The state lifespans are weighted and fused to obtain the final predicted remaining lifespan and determine the corresponding healthy periods. The final monitoring results, including the remaining lifespan and the divided healthy periods, will be presented to maintenance personnel through a visual interface to formulate corresponding maintenance strategies.

[0073] It should be noted that calculating the remaining useful life by matching historical cases and weighted fusion can avoid the errors caused by using a single model for prediction, improve the accuracy of the prediction results, and transform the remaining useful life and health period of the equipment into key indicators required for operation and maintenance decisions, providing accurate data support for the intelligent and refined management of the entire equipment life cycle.

[0074] This application analyzes the real-time status of equipment by constructing a state baseline model, integrates the state analysis results of time steps to construct a state analysis curve, analyzes the real-time status of the equipment and calculates the survival risk, and analyzes the dynamic process of equipment state changes by simulating the equipment state evolution process; it can effectively filter out noise, issue early warnings in the early stage of fault occurrence, perform survival analysis based on real-time status, and make real-time corrections based on the current actual degradation path of the equipment, thereby improving the accuracy and adaptability of the state monitoring process.

[0075] Furthermore, real-time sensor data during the operation of the target device is input into the state baseline model for analysis. State analysis results for each time step are extracted, fused, and the corresponding state scores are calculated to construct state analysis curves, including:

[0076] S201. Input the real-time sensor data during the operation of the target device into the state reference model for analysis, extract the state analysis results of each time step, calculate the reconstruction error between the state analysis results and the input data, and obtain the reconstruction error sequence.

[0077] S202. By performing fusion analysis on the reconstructed error sequence, the corresponding state score is calculated, and a state analysis curve is constructed.

[0078] In this embodiment, real-time sensor data during the operation of the target device is input into a state baseline model for analysis. The state analysis results at each time step are extracted, and the reconstruction error between the state analysis results and the input data is calculated to obtain a reconstruction error sequence. By extracting and cascading the state vectors of each layer within the LSTM model, a state analysis result that can deeply characterize the device's operating mode is constructed. Compared to using only the final output of the model, this method can extract changes in the model's understanding of data at multiple levels, improving the sensitivity of the fault analysis process. The method of calculating the reconstruction error by combining Euclidean distance and cosine of the included angle reflects the overall deviation of sensor values ​​and the maintenance of the proportional relationship between various sensor parameters, which can more comprehensively reflect the abnormal modes of the device's state.

[0079] Specifically, by performing fusion analysis on the reconstructed error sequence, the corresponding state score is calculated, and a state analysis curve is constructed. By dividing the reconstructed error sequence into a baseline layer and an anomaly layer, and constructing dynamic error fluctuation boundaries and cumulative deviations for each, the fluctuations under normal operating conditions and the actual abnormal deviations can be effectively separated. This allows the model to adapt to changes in operating conditions. When changes in equipment load cause a shift in sensor data distribution, as long as this shift remains within the healthy mode range, it will not be misjudged as an anomaly, reducing the false alarm rate. By calculating the first and second deviation indices, the model can ensure a rapid response to sudden shocks and effectively filter out random noise, facilitating rapid analysis of equipment health status by maintenance personnel and providing accurate input data for subsequent equipment survival risk analysis.

[0080] Furthermore, real-time sensor data during the operation of the target device is input into the state baseline model for analysis. The state analysis results at each time step are extracted, and the reconstruction error between the state analysis results and the input data is calculated to obtain a reconstruction error sequence, including:

[0081] S301. Input the real-time sensor data during the operation of the target device into the state baseline model for analysis. Extract the state vector generated by each layer of the model at each time step and cascade them to obtain the corresponding state analysis results.

[0082] S302. Generate the corresponding data vector by decoding the state analysis results, calculate the Euclidean distance and cosine value of the included angle between the data vector and the input data vector, and obtain the reconstruction error.

[0083] S303. The reconstruction errors are integrated according to time to obtain the reconstruction error sequence.

[0084] In this embodiment, taking a large operating mining mill as an example, the real-time sensor data collected includes four dimensions: main motor stator temperature, large gear vibration acceleration, lubricating oil flow rate, and bearing pressure. The data acquisition frequency is set to once every 20 seconds, and the data is transmitted in real time from the field distributed control system to the monitoring server via industrial Ethernet. The server maintains a 50-bit first-in-first-out circular buffer. Whenever a new data point arrives, it is combined with the first 49 points in the buffer to form a new input time window, covering a time span of 16.7 minutes. This input time window is then input into the trained state baseline model.

[0085] As the input time window propagates forward in the model, the program performs corresponding state analysis at each time step. The input time window contains data from 50 time steps from time T-49 to T. For any time step i, the model first inputs the current time step's input vector x_i together with the hidden state passed from the previous time step into the first LSTM layer. The first LSTM layer updates the memory cell state of the current time step according to its internal gate control mechanism and generates the current time step's hidden state vector h1_i, which has a dimension of 96. h1_i is then used as the input of the second LSTM layer at the current time step. The second LSTM layer also performs internal gate control calculations to generate the current time step's hidden state vector h2_i, which has a dimension of 48.

[0086] The two state vectors are concatenated end-to-end, with the 96-dimensional hidden state vector h1_i and the 48-dimensional hidden state vector h2_i sequentially connected to obtain a 144-dimensional composite vector, which serves as the state analysis result for time step i. This process is repeated for all 50 time steps within the entire input time window to obtain a sequence of state analysis results for all time steps within that window.

[0087] It should be noted that by extracting the hidden state vectors generated by each layer of the LSTM model at each time step and concatenating them, the constructed state analysis results can deeply reflect the device's operating mode and extract the subtle changes in the model's understanding of data during the multi-level abstraction process. The first-layer state vector retains the local detailed features of the original sensor signal, while the second-layer state vector includes trend information over a longer time range. Through data fusion, the state analysis results simultaneously possess sensitivity to details and the ability to grasp the overall trend, providing richer input for subsequent calculation and reconstruction of errors, making the perception of early device degradation more sensitive and accurate.

[0088] Specifically, for each time step's state analysis result vector, the decoder part of the input state baseline model is reconstructed. This decoder consists of a repeating vector layer, an LSTM layer, and a fully connected layer. Taking the last time step within the input time window as an example, let its state analysis result be s_T, a 144-dimensional vector. This vector first enters the repeating vector layer, which copies s_T multiple times to generate a sequence with the same time step length as the original input time window. This is repeated 50 times to form a 50×144 sequence matrix. The sequence matrix is ​​then input into the LSTM layer in the decoder. This LSTM layer has 64 memory units and is configured to return the output sequence for each time step. The LSTM layer processes the input sequence step by step through its gating mechanism, generating a 64-dimensional hidden state vector at each time step. These hidden state vectors are finally fed into a fully connected layer, where the number of neurons is equal to the dimension of the original input data. The fully connected layer maps the 64-dimensional hidden states back to the 4-dimensional original data space, generating a reconstruction of the original input vector for each time step.

[0089] Given two 4D vectors at the current moment: the original input data vector and the reconstructed data vector, calculate the Euclidean distance between them. This distance reflects the straight-line distance between the two vectors in multidimensional space; a larger distance indicates a more significant overall deviation between the reconstructed and actual values. Calculate the cosine of the angle between the original and reconstructed data vectors, reflecting the directional consistency of the two vectors in multidimensional space. The closer the cosine value is to 1, the more similar the directions of the two vectors are, meaning the proportional relationship between the sensor parameters is more consistent with the health model. Weight the calculated Euclidean distance and cosine value. For example, based on historical experience, set the weight of the Euclidean distance to 0.5 and the weight of the difference between 1 and the cosine value to 0.5. After normalizing the Euclidean distance and the difference between 1 and the cosine value using maximum-minimum normalization, and then summing the weighted values, the final reconstruction error value at the current moment is obtained. This value reflects the overall deviation of the device's operating status from the health baseline model in terms of both amplitude and proportional relationship dimensions.

[0090] It should be noted that the state analysis results in the deep feature space are mapped back to the original data space through decoding operations to reconstruct the expected sensor readings in the health mode. The reconstruction error is calculated by combining Euclidean distance and cosine of the included angle, which reflects both the overall deviation of the sensor values ​​and the maintenance of the proportional relationship between the parameters of each sensor. This can more comprehensively reflect the abnormal mode of the equipment status and provide accurate input data for the subsequent error sequence analysis process.

[0091] The system processes the real-time data stream, collecting sensor data every 20 seconds. Each new data entry triggers steps S301 and S302, generating a reconstruction error value corresponding to the current moment. The system maintains a fixed-length circular array in memory to store the reconstruction error values ​​calculated over a recent period. The length of this array can be set according to actual monitoring needs. For example, if set to store data from the last 30 days, with a sampling interval of 20 seconds, the array length would be approximately 129,600 points. For real-time monitoring, which typically focuses on recent trends, the system can also store data from the last 24 hours, resulting in a length of 4,320 points.

[0092] When the system starts, the reconstruction error array is empty. As data arrives at the first sampling time, the reconstruction error is calculated and stored in the first position of the array; as data arrives at the second sampling time, the reconstruction error is calculated and stored in the second position, and so on. When the array is full, new reconstruction error values ​​overwrite the oldest values, ensuring that the array always stores the most recent, continuous data. Simultaneously with data storage, the system timestamps each reconstruction error value, recording the corresponding sampling time. Arranging the elements in the array in chronological order yields the reconstruction error sequence.

[0093] It should be noted that by integrating the reconstruction error values ​​of consecutive moments in chronological order, a reconstruction error sequence that can fully reflect the evolution of the equipment state is constructed. This allows maintenance personnel to observe the occurrence time, development speed, and changing trend of state deviations, providing data support for subsequent extraction of error distribution characteristics, calculation of fluctuation boundaries, and cumulative deviation. When equipment malfunctions, the time point at which the malfunction begins can be quickly located, and the entire process of malfunction development can be analyzed, providing important data support for fault cause analysis.

[0094] like Figure 2 As shown, by performing fusion analysis on the reconstruction error sequence, the corresponding state score is calculated, and a state analysis curve is constructed, including:

[0095] S401. According to the preset time window and time step, slide on the reconstruction error sequence, calculate the error value in each window, form the baseline layer with the error value less than or equal to the preset error threshold, and form the anomaly layer with the error value greater than the preset error threshold.

[0096] S402. For the baseline layer, analyze the error changes, extract the error distribution characteristics, and construct the error fluctuation boundary.

[0097] S403. For the abnormal layer, analyze the error accumulation, calculate the cumulative error, and obtain the deviation.

[0098] S404. Based on the error fluctuation boundary and deviation amount, analyze and integrate the deviation indicators at different time scales, calculate the corresponding state score, and construct the state analysis curve.

[0099] In this embodiment, the parameters for the sliding window analysis are set as follows: the time window width is set to 60, representing the time span of 60 sampling points, i.e., 20 minutes; the time step is set to 1, meaning one sampling point is slid at a time. Starting from the beginning of the reconstruction error sequence, the first to the 60th reconstruction error values ​​are taken to form the first window, and the arithmetic mean of all reconstruction error values ​​within this window is calculated to obtain the representative error value for this window. Then, the window is slid forward by one step, and the second to the 61st reconstruction error values ​​are taken to form the second window, and the average of the reconstruction error values ​​within this window is calculated in the same way. This process is repeated until the entire reconstruction error sequence is traversed, resulting in a series of representative error values ​​for each window.

[0100] A subset of untrained data is retained from the health dataset used to train the baseline model as a validation set. This validation set is then input into the model to calculate the reconstruction error, resulting in the error distribution under healthy conditions. The 95th quantile of this error distribution is then used as the error threshold. For each sliding window, the representative error value is compared to the error threshold. If the representative error value of the window is greater than the error threshold, the window is considered a period containing abnormal components and is designated as the anomaly layer. If the representative error value is less than or equal to the error threshold, the window is considered to be operating within the normal fluctuation range and is designated as the baseline layer. The baseline layer corresponds to healthy periods of equipment operation, while the anomaly layer corresponds to periods where equipment degradation may occur.

[0101] It should be noted that by using sliding window analysis and threshold comparison, the reconstructed error sequence is stratified, enabling subsequent analysis to be differentiated for different time periods. The identification of the baseline layer provides a clean data source for extracting normal fluctuation patterns and constructing dynamic fluctuation boundaries; the identification of the anomaly layer delineates key areas of focus for quantifying the degree of deviation and calculating the cumulative deviation. By setting error thresholds based on health data statistics, the objectivity and repeatability of the stratification criteria are ensured, allowing this method to adapt to the normal fluctuation range under different equipment and operating conditions.

[0102] For any given time window belonging to the baseline layer, this window contains 60 original reconstruction error values. The standard deviation of these 60 error values ​​is calculated; a larger standard deviation indicates more drastic error fluctuations within the window, while a smaller standard deviation indicates more stable fluctuations. As the sliding window moves along the time axis, each baseline layer window corresponds to a calculated standard deviation. These standard deviations are arranged in chronological order according to the window's start time, resulting in a sequence reflecting the variation of error fluctuation amplitude over time. Since the baseline layer windows are not continuous, the standard deviation sequence has discontinuities. Interpolation and smoothing are performed on these discontinuous standard deviation data.

[0103] For the time period occupied by the anomaly layer window, a linear interpolation method is used to estimate the standard deviation of each time point within the anomaly layer period based on the standard deviation of the adjacent baseline layer windows before and after the anomaly layer. An exponentially weighted moving average method is used to smooth the continuous standard deviation sequence; the smoothed value at the current time is obtained by weighted averaging of the original standard deviation at the current time and the smoothed value at the previous time, with more recent data having greater weight. By setting an appropriate smoothing factor, such as 0.1, the smoothed curve reflects the long-term trend of standard deviation changes while filtering out short-term random fluctuations. The smoothed curve serves as the error fluctuation boundary; this boundary line changes dynamically over time, reflecting the upper limit of error fluctuation allowed by the healthy state under the current time environment.

[0104] It should be noted that by analyzing the distribution characteristics of errors within the baseline layer, a dynamically changing error fluctuation boundary is constructed, which can adaptively adjust according to the actual operating environment of the equipment. When the operating conditions of the equipment change, causing the normal fluctuation range to expand, the error fluctuation boundary will rise accordingly to avoid misjudging normal operating condition changes as abnormalities. When the operating conditions tend to stabilize and the fluctuations decrease, the boundary will decrease accordingly to maintain high sensitivity to anomalies. It can adapt to various operating scenarios throughout the entire life cycle of the equipment, improving the robustness and adaptability of condition monitoring.

[0105] Specifically, for the time window belonging to the anomaly layer, this window covers 60 sampling points from time t to t+59. For each time point within this window, the original reconstruction error value is known, and the corresponding fluctuation boundary value can be found from the error fluctuation boundary curve. For each time point within the window, the original reconstruction error value is subtracted from the error fluctuation boundary value to obtain the instantaneous deviation at that point. The instantaneous deviations of all 60 time points within the window are summed to obtain the deviation of the anomaly layer window, reflecting the duration of the anomaly and the degree of deviation.

[0106] It should be noted that by calculating the cumulative excess of the error within the anomaly layer window relative to the dynamic fluctuation boundary, a deviation that can comprehensively reflect the severity of the anomaly is calculated, reflecting the intensity of the anomaly and the duration of the anomaly, thus providing accurate data support for subsequent analysis of equipment anomalies.

[0107] Specifically, based on the error fluctuation boundary and deviation amount, deviation indicators at different time scales are analyzed and fused to calculate the corresponding status score and construct a status analysis curve. By constructing and fusing deviation indicators, the first deviation indicator ensures the model's rapid response capability to sudden shocks, enabling it to capture abnormal signals in a timely manner at the moment a fault occurs; the second deviation indicator smooths out random noise through weighted averaging, avoiding overreaction to short-term fluctuations; and the fused status score can truly reflect the changing patterns of equipment health status.

[0108] Furthermore, based on the error fluctuation boundary and deviation amount, deviation indicators at different time scales are analyzed and fused to calculate the corresponding state scores and construct state analysis curves, including:

[0109] S501. Calculate the ratio of the deviation at each time step to the error fluctuation boundary to obtain the first deviation index;

[0110] S502. Within the time window, the first deviation index is weighted and averaged to obtain the second deviation index, wherein the weights are set according to the distance between the time point and the boundary of the time window.

[0111] S503. Combine the first deviation index and the second deviation index to calculate the deviation degree, and map the corresponding state score through the preset scoring mapping rule.

[0112] S504. Integrate the state scores in chronological order to construct a state analysis curve.

[0113] In this embodiment, the system maintains two synchronously updated time series arrays in memory: one is the deviation array, which stores the instantaneous deviation calculated at each sampling time; the other is the error fluctuation boundary array, which stores the error fluctuation boundary value at the corresponding time. These two arrays have the same time index, so for any given time, the deviation and error fluctuation boundary at that time can be obtained simultaneously.

[0114] For each time step, the deviation amount and the error fluctuation boundary value at that moment are read. The ratio of the deviation amount to the error fluctuation boundary value is calculated as the first deviation index for that time step. By calculating the ratio of the deviation amount to the error fluctuation boundary value, the resulting first deviation index is a dimensionless relative quantity, eliminating the absolute numerical differences caused by factors such as different equipment types, different sensor ranges, and changes in operating conditions. Through ratio processing, the relative severity of deviation under different operating conditions can be accurately reflected, avoiding misjudgments caused by changes in operating conditions.

[0115] Specifically, a fusion time window is set according to the accuracy requirements of equipment status monitoring, with a width of 120 sampling points corresponding to a 40-minute time span. For each current moment, a second deviation index centered on the current moment is calculated. From the first deviation index sequence, 120 sampling points centered on the current moment are extracted to form a fusion time window. Time points closer to the center point of the window have a larger weight; those farther from the window edge have a smaller weight. A triangular window function is used to set the weight of the window center point to the maximum value, linearly decreasing towards both sides, until the weight drops to zero or close to zero at the two endpoints of the window. The first deviation index of each time point within the window is multiplied by its corresponding weight, and all weighted values ​​are summed to obtain the second deviation index for the current moment. The above operation is repeated for each sampling moment to obtain the second deviation index sequence.

[0116] It should be noted that the second deviation index, constructed by weighting the first deviation index, can reflect the deviation trend. The weighted average within the window can effectively filter out random noise and instantaneous fluctuations caused by short-term disturbances, thereby improving the stability and reliability of the evaluation results. Based on the central focusing characteristics of the weight distribution, it maintains high sensitivity to recent deviation changes and can promptly capture the turning point of the degradation trend.

[0117] For the first and second deviation indices calculated at the current moment, corresponding fusion weights are set according to the characteristics of the equipment and monitoring needs. For example, for equipment prone to sudden failures, such as rolling bearings, a higher weight is set for the first deviation index to detect impact signals in a timely manner; for equipment mainly characterized by slow wear, such as gearboxes, a higher trend weight is set for the second deviation index to focus on degradation trends. The deviation degree at the current moment is obtained by weighted summation according to the corresponding weights, providing a quantitative indicator that comprehensively reflects the degree of deviation from the equipment's state.

[0118] Specifically, the deviation is mapped to a scoring range of 0 to 100, and a preset scoring mapping rule is used for conversion. The mapping rule is based on a modified design of the Sigmoid function, and its mathematical properties enable the following conversion relationship: when the deviation is close to 0, the status score is close to 100 points, indicating that the device is in perfect health; as the deviation increases, the score begins to decrease, and the rate of decrease is the fastest in the middle area, which corresponds to the rapid deterioration stage of the health status from good to bad; when the deviation exceeds a certain upper limit, the score approaches 0 points, indicating that the device is close to failure.

[0119] For example, the specific parameters of the scoring mapping rule are calibrated according to the equipment type and maintenance requirements. Based on historical fault data and expert experience, it is determined that when the deviation reaches 0.5, the equipment should be rated 80 points (minor attention); when the deviation reaches 1.5, it should be rated 60 points (warning); when the deviation reaches 3.0, it should be rated 40 points (serious warning); and when the deviation reaches 5.0, it should be rated 20 points (danger). Based on these key points, a smooth mapping curve is fitted.

[0120] It should be noted that by integrating the first and second deviation indicators, a deviation degree that comprehensively reflects the health status of the equipment is constructed, avoiding the one-sidedness of a single indicator. This makes the status assessment both sensitive and stable, and adaptable to various failure modes. Through a preset scoring mapping rule, the deviation degree is converted into a standardized score from 0 to 100, providing a unified quantitative standard for the equipment health status and improving operational efficiency.

[0121] Specifically, the status scores generated by real-time calculations are stored in an ordered manner. The system calculates a status score every 30 seconds, and this score corresponds to the current device health level. The system maintains a dedicated data table in the real-time database for storing status scores. This table contains two fields: a timestamp and the status score value. Whenever a new status score is calculated, the precise time of the current system is obtained as the timestamp; the timestamp and the status score are appended to the status score table in the real-time database as a new record. The records are then sorted in ascending order of timestamps to obtain an ordered sequence. Figure 3 As shown, the state analysis curve is obtained by plotting the timestamps in this sequence as the horizontal axis and the state scores as the vertical axis.

[0122] It's important to note that by integrating the status scores from consecutive moments in chronological order, the constructed status analysis curve visually displays the dynamic evolution of equipment health, allowing maintenance personnel to readily grasp the overall health status of the equipment. Observing the curve's trend allows for determination of whether the equipment is in a stable state, a slow degradation state, or a state of accelerated deterioration. By comparing curves from different time periods, the effectiveness of maintenance measures can be assessed, and the equipment's response to maintenance can be judged, providing accurate input data for subsequent survival risk analysis and remaining life prediction.

[0123] Furthermore, based on the state analysis curves, the real-time state of the equipment is analyzed, and the corresponding survival risk is calculated in combination with the real-time state of the equipment. The equipment state evolution process is simulated, and state change curves are constructed, including:

[0124] S601. Extract the corresponding timing features from the status analysis curve and analyze the real-time status of the equipment;

[0125] S602. Calculate the corresponding survival risk based on the real-time status of the equipment, simulate the equipment state evolution process, and construct the state change curve.

[0126] In this embodiment, considering that the degradation of mechanical components is usually a slow process, and to smooth out short-term fluctuations, the window width is set to 120 minutes. For each current moment, the system retrieves the most recent 120 rating values ​​from the queue to form a time series segment. Based on this segment, the following time series features are extracted to form the real-time state feature vector for the current moment, including the state rating at this moment, the arithmetic mean of the 120 state ratings within the window, the standard deviation of the state ratings within the window, the difference sequence of state ratings at adjacent moments within the window and the mean of the difference sequence, the difference between the highest and lowest state ratings within the window, and the time interval from the moment the highest state rating appears to the moment the lowest state rating appears within the window. The extracted time series features are integrated in a preset order to obtain the real-time state feature vector.

[0127] It should be noted that by extracting multi-dimensional time-series features from the state analysis curve, the instantaneous values, trends, and fluctuations of the equipment's health status are comprehensively quantified, and the constructed real-time state feature vector can fully reflect the real-time operating status of the equipment. Compared with using only a single score value, it can more sensitively identify early signs of equipment degradation and improve the accuracy and robustness of the state analysis results.

[0128] Specifically, by combining the real-time status of the equipment to calculate the corresponding survival risk, simulating the equipment's state evolution process, and constructing a state change curve, survival analysis integrates the real-time status of the equipment with historical failure patterns, enabling the output of a continuous risk curve. By simulating future state evolution and calculating risks point by point, the constructed state change curve not only provides a point estimate of the remaining lifespan but also displays the entire process of risk evolution. It allows for a clear view of the time points when risks accelerate, enabling targeted measures to be taken in advance, thereby improving the accuracy and adaptability of equipment condition monitoring and management.

[0129] Furthermore, by combining the real-time status calculation of the equipment with the corresponding survival risk, simulating the equipment state evolution process, and constructing state change curves, including:

[0130] S701. Based on the real-time status of the equipment, map the corresponding risk rate in the preset equipment baseline risk function, integrate the risk rate to calculate the corresponding survival risk, and obtain the cumulative risk value.

[0131] S702. Determine the state phase of the device based on the accumulated risk value;

[0132] S703. Combining the risk rate, cumulative risk value, and state phase, the state evolution process of the equipment is simulated through a preset state evolution model to construct a state change curve.

[0133] In this embodiment, a pre-defined baseline risk function for equipment is constructed. This function describes the variation of risk over operating time for similar equipment in an average sense. The construction of the baseline risk function relies on collecting a large amount of historical failure data for the same type of equipment, including the commissioning time, final failure time, and some censored data that remained operational at the end of the observation period. The survival function is estimated using the Kaplan-Meier method, and the baseline risk function is derived. The input of this function is the total operating time of the equipment, and the output is the baseline risk rate at the corresponding moment. For example, the baseline risk rate of a centrifugal compressor on its 1000th day of operation is 0.02, indicating that there is a 2% probability of failure on that day in an average sense.

[0134] Specifically, the real-time health status of each device is different. The baseline risk rate is adjusted based on the current status, using the mapping relationship between the device's real-time status and the risk multiplier. This mapping can be determined by statistically analyzing the failure frequency corresponding to different health scores in historical data. For example, the status score is divided into several intervals: when the score is 90 to 100, the risk multiplier is set to 1.0, indicating consistency with the average level; when the score is 80 to 89, the multiplier is set to 1.5; when the score is 70 to 79, the multiplier is set to 2.5; when the score is 60 to 69, the multiplier is set to 4.0; and when the score is below 60, the multiplier is set to 6.0. For the current moment, the corresponding risk multiplier is found in the above mapping table by extracting the status score from the real-time status feature vector. This multiplier is then multiplied by the baseline risk rate corresponding to the current running time to obtain the instantaneous risk rate at the current moment.

[0135] Furthermore, after obtaining the instantaneous risk rate, the cumulative risk value from the current moment to a future moment is calculated. The cumulative risk value is defined as the integral of the instantaneous risk rate over that time period, and in actual discrete data processing, it is implemented by numerical accumulation. A fixed time step is set, for example, one day. Starting from the current moment, the instantaneous risk rate for each day is accumulated to obtain the cumulative risk value from the current moment to the Nth future day. For example, if the instantaneous risk rate at the current moment is 0.05, the predicted risk rate for the first day in the future is 0.06, and the predicted risk rate for the second day in the future is 0.07, then the cumulative risk value until the second day in the future is 0.05 + 0.06 + 0.07 = 0.18. This cumulative risk value monotonically increases with the number of predicted days, quantifying the total risk accumulated from the current moment to a future point in time, and serves as the basis for subsequently determining the equipment's life cycle stage and predicting its remaining lifespan.

[0136] It should be noted that by combining the real-time status of equipment with historical statistical patterns, the risk of equipment failure is dynamically quantified. The baseline risk function provides an objective background risk, and the real-time status mapping of equipment reflects the individual differences of specific equipment. This makes the calculated instantaneous risk rate and cumulative risk value both statistically based and reflect the current actual situation, providing accurate data support for subsequent stage division and evolution simulation.

[0137] Specifically, the degradation process of equipment is not uniform; it typically goes through several distinct stages, each with a significantly different rate of risk accumulation and failure probability. By setting thresholds for cumulative risk values, the entire lifecycle of the equipment can be divided into multiple physically meaningful state phases, helping maintenance personnel to more intuitively understand the current degradation stage of the equipment. The thresholds are determined based on the analysis of cumulative risk curves from a large number of historically failed devices. This involves statistically analyzing the changes in cumulative risk values ​​throughout the entire lifecycle, from health to failure, and observing the typical range of cumulative risk values ​​at different stages of life. For example, analysis of historical data from 100 compressors of the same model revealed that in the first 80% of the equipment's lifespan, the cumulative risk value increases slowly, generally below 0.2; between 80% and 95% of the lifespan, the cumulative risk value is between 0.2 and 0.8; and in the last 5% of the lifespan, the cumulative risk value rapidly climbs above 0.8 and eventually reaches around 1.0. Based on this observation, two thresholds can be set: a first threshold of 0.2 and a second threshold of 0.8.

[0138] The equipment's status phase is divided into three intervals. When the cumulative risk value is less than or equal to the first threshold, the equipment is determined to be in a healthy and stable period. At this time, the equipment operates stably, the risk of failure is extremely low, and maintenance can be carried out according to the regular plan. When the cumulative risk value is greater than the first threshold but less than or equal to the second threshold, the equipment is determined to be in a slow degradation period. At this time, the equipment begins to show a slight performance decline, and the risk gradually accumulates. The monitoring frequency should be increased and maintenance plans should be prepared. When the cumulative risk value is greater than the second threshold, the equipment is determined to be in a rapid failure period. At this time, the equipment condition deteriorates rapidly, the risk of failure increases sharply, and shutdown maintenance should be arranged in the shortest possible time.

[0139] In actual online monitoring, the system continuously calculates the cumulative risk value at the current moment and compares it with two thresholds, updating the device's status phase in real time. This judgment result can be displayed intuitively on the monitoring interface, for example, using a green background to mark the healthy and stable period, a yellow background to mark the slow degradation period, and a red background to mark the rapid failure period, making it clear to maintenance personnel at a glance. The specific values ​​of the thresholds can be adjusted according to different device types, industry standards, or user needs. For example, for more critical devices, the thresholds can be appropriately lowered to enter a higher alert level earlier.

[0140] It should be noted that discretizing the continuously changing cumulative risk value into state phases with clear boundaries can reduce the difficulty for maintenance personnel to understand complex risk data. Phase division transforms the description of equipment health status from a single numerical value into a phased qualitative judgment, providing accurate data support for adopting differentiated maintenance strategies. Maintenance personnel can automatically trigger corresponding business processes based on the phase of the equipment, such as starting spare parts procurement during the slow degradation period and prioritizing shutdown during the rapid failure period, thereby achieving refined and intelligent equipment management.

[0141] Specifically, the future state evolution of equipment depends not only on the current risk level and cumulative risk, but also on its current state phase, with significant differences in degradation patterns across different phases. For example, during a healthy, stable period, the state remains stable over a long period; during a slow degradation period, the state declines slowly at an approximately linear rate; and during a rapid failure period, the state deteriorates exponentially. To accurately predict future state changes, a state evolution model capable of learning and reproducing the evolution patterns under different phases is constructed. The model's training relies on a large amount of historical equipment lifecycle data, collecting complete records of at least fifty devices of the same model from commissioning to eventual failure. For each device, its state features are extracted at fixed time intervals (e.g., daily), including state score, instantaneous risk rate, cumulative risk value, and the state phase calculated based on the cumulative risk value. This data is organized into training samples, with each sample taking the current feature vector as input and the actual state score sequence for a future period as output. The model can employ a Long Short-Term Memory (LSTM) network, where the input layer receives the current feature vector, extracts temporal features through two LSTM layers, and finally outputs predicted state scores for the next thirty time points through a fully connected layer. During training, mean squared error was used as the loss function to measure the difference between the predicted sequence and the true sequence. The optimizer was Adam, the initial learning rate was set to 0.001, the batch size was set to 32, and an early stopping mechanism was used to prevent overfitting.

[0142] In actual online monitoring, for the current moment, the extracted real-time state features, the calculated current risk rate and cumulative risk value, and the current state phase are combined to form a current feature vector, which is then input into the trained state evolution model. The model outputs a predicted state score sequence for the next M moments. Based on this predicted state score sequence, the predicted risk rate for each future moment can be calculated. The corresponding risk multiplier is obtained through a mapping table based on the predicted state score, and the baseline risk rate for that moment is obtained by looking up the baseline risk function. Multiplying the two yields the predicted risk rate. The predicted risk rates for each future moment are accumulated starting from the current moment to obtain the predicted cumulative risk value for each future moment. All these predicted information are plotted as a curve in chronological order to obtain the state change curve.

[0143] It's important to note that by constructing a state evolution model, we can learn the patterns of equipment degradation at different phases and accurately predict future states. The state change curves not only provide numerical estimates of future states but also graphically illustrate the timing of risk acceleration and the inflection points of degradation trends, offering forward-looking information for operational decisions. Operations personnel can use these curves to plan maintenance windows in advance, optimize spare parts inventory, and combine the prediction results with remaining life calculations, achieving a shift from reactive response to proactive prevention and significantly improving the scientific rigor and effectiveness of equipment management.

[0144] Furthermore, based on the state change curve analysis, the remaining service life and health period of the equipment are calculated to monitor the health status of the equipment, including:

[0145] S801. Analyze the equipment status based on the state change curve, match the corresponding state lifetime in the preset equipment status mapping library, and obtain the state lifetime set.

[0146] S802. Calculate the remaining service life and health period of the equipment based on the condition life set in order to monitor the health status of the equipment.

[0147] In this embodiment, the remaining lifespan of the device is closely related to its current health status and historical evolution trajectory. By matching the current device's state change curve with the complete life cycle curves of a large number of similar devices in the historical database, the most similar historical cases can be found. The time taken for these historical cases from the current similarity point to the final failure is a valid reference for the remaining lifespan of the current device.

[0148] Specifically, a pre-defined equipment state mapping library is constructed. The data in the library comes from the historical operating records of equipment of the same model. Taking a large centrifugal compressor in a petrochemical plant as an example, data on at least two hundred compressors of the same model from commissioning to eventual failure or decommissioning is collected. For each historical piece of equipment, its entire lifecycle is divided into segments with equal time intervals, such as each day as a sampling point. At each sampling point, the state characteristics at that moment are recorded. These characteristics include the state score, instantaneous risk rate, cumulative risk value, and the state phase determined based on the cumulative risk value. Simultaneously, the actual remaining time from that sampling point to the equipment's eventual failure is recorded; this remaining time is defined as the state lifetime. All the state characteristics of these sampling points are used as input to a case study, and the corresponding state lifetime is used as the output label of the case study, both stored in the equipment state mapping library.

[0149] In actual online monitoring, for the current moment, the state features at that moment are extracted from the state change curve as query conditions. These features include the current state score, instantaneous risk rate, cumulative risk value, and state phase. These four features together constitute a query vector, representing the overall state of the current device. The query vector is input into the device state mapping library for similarity matching. The matching process uses weighted Euclidean distance as the similarity measure, that is, calculating the weighted distance between the query vector and the feature vector of each historical case in the library. The weight setting needs to reflect the importance of different features to the remaining lifetime prediction, and can be determined through correlation analysis of historical data or expert experience. For example, the weight of the state score is set to 0.4, the weight of the instantaneous risk rate to 0.3, the weight of the cumulative risk value to 0.2, and the weight of the state phase to 0.1.

[0150] For each historical case in the database, the weighted Euclidean distance between it and the query vector is calculated. The smaller the distance, the more similar the states of the two cases are. To improve computational efficiency, it is not necessary to calculate the distance of all cases in the database. Instead, a nearest neighbor search algorithm is used to find only the K cases with the smallest distance. K can be set to 5 or 10. In this embodiment, K is set to 5, that is, to find the five historical cases most similar to the current state. Each of these five historical cases corresponds to a known state lifetime, which is the remaining time from the feature point of the case to the final failure of the device. Collecting these five state lifetimes constitutes the state lifetime set at the current moment. For example, the state lifetimes of these five cases are 210 days, 195 days, 185 days, 175 days, and 165 days, respectively. Each value in this set is based on real historical data and has a high similarity to the current device state, providing a reliable reference for subsequent calculation of the remaining lifetime.

[0151] It should be noted that by constructing a device status mapping library containing a large number of historical cases and employing similarity matching technology, we directly utilize real historical data, avoiding the need to build complex mathematical models to assume device degradation patterns. This makes the prediction results more realistic. The weighted distance calculation comprehensively considers information from multiple status dimensions, ensuring that the matched historical cases are highly consistent with the current equipment in key indicators. This improves the representativeness of the reference cases and enhances the robustness and reliability of the status monitoring results.

[0152] Specifically, based on the state lifetime set, the remaining lifespan and health period of the equipment are calculated to monitor the equipment's health status. By using weighted fusion to calculate the remaining lifespan, the similarity between different historical cases and the current state is fully considered, improving the accuracy of the prediction results, effectively reducing the risk of unplanned downtime, optimizing the allocation of maintenance resources, and improving the management level and economic benefits of the equipment throughout its entire life cycle.

[0153] Furthermore, based on the state-based lifetime set, the remaining lifespan and health period of the equipment are calculated to monitor the equipment's health status, including:

[0154] S901. Analyze the matching degree between each state lifetime and the state change process based on the state lifetime set and set the corresponding fusion weight. Then, perform a weighted average of the state lifetimes according to the fusion weight to obtain the remaining service life of the equipment.

[0155] S902. Determine the boundaries of the health period based on the remaining service life in order to monitor the health status of the equipment.

[0156] In this embodiment, for each case matched from the device state mapping library, the weighted Euclidean distance between it and the current device state is calculated, and the corresponding fusion weight is set according to the distance. The smaller the distance, the larger the fusion weight. After obtaining the weight of each case, a weighted average is calculated. For each case, its corresponding state lifetime is known. The state lifetime is multiplied by the corresponding weight and the average value is calculated to obtain the remaining lifespan of the current device.

[0157] It should be noted that by setting corresponding fusion weights for weighted averaging, multiple historical reference lifetimes can be intelligently fused. Compared with simple arithmetic averaging, dynamic weighted fusion can amplify the contribution of highly similar cases and suppress the interference of low-similarity cases, thereby improving the accuracy of remaining lifetime prediction results.

[0158] Specifically, the rules for dividing health periods should be determined based on factors such as equipment type, industry practices, and maintenance resources. Based on the equipment's historical failure patterns and industry maintenance standards, the following division ratios can be set: The first 20% of the remaining service life from the current moment is the high-health period. During this period, the equipment is expected to be in good condition with a low risk rate, and routine monitoring can be maintained, such as recording data once a day, without the need for special measures. The period from 20% to 50% of the remaining service life is the attention period. During this period, the equipment may begin to show slight signs of degradation, requiring enhanced monitoring. For example, the data collection frequency can be increased from once a day to once every 4 hours, and a status report can be generated weekly for engineers to conduct trend analysis. In addition, preparations for the procurement of spare parts can begin. The period from 50% to 100% of the remaining service life is the planned maintenance window. During this period, the equipment will enter a period of accelerated degradation, and the risk of failure will increase significantly. Shutdown maintenance should be completed during this period. The specific maintenance time can be optimized in conjunction with the production plan, for example, by selecting the time period when the production line is scheduled to stop to minimize the impact on production.

[0159] The beginning and end times of these three time periods are defined as the boundary points of the healthy time periods. This boundary information is presented to operations and maintenance personnel. Green blocks represent highly healthy time periods, yellow blocks represent periods of concern, and red blocks represent planned maintenance windows. The start and end dates are marked on each block. Simultaneously, a moving pointer or marker displays the current date's position on the timeline. This allows operations and maintenance personnel to easily see the current time period and how far away the next time period is.

[0160] It's important to note that by dividing the remaining lifespan into corresponding health periods, a transformation from numerical prediction to action guidelines is achieved. The division of health periods is not arbitrary but based on the proportion of remaining lifespan, reflecting the risk characteristics and maintenance needs of the equipment at different stages of its lifecycle. High-health periods correspond to low-risk routine monitoring, periods of concern correspond to medium-risk enhanced monitoring and contingency plan preparation, and planned maintenance windows correspond to high-risk mandatory maintenance. This tiered management model aligns with the concept of refined management. Through a visualized timeline, maintenance personnel can accurately grasp the future health trend and key time points of the equipment, plan maintenance resources in advance, avoid reactive responses, and achieve closed-loop management of predictive maintenance. This effectively reduces the risk of unplanned downtime, optimizes maintenance costs, and improves the overall operating efficiency of the equipment.

[0161] like Figure 4 As shown, a survival analysis-based equipment health status monitoring system is used to implement a survival analysis-based equipment health status monitoring method, including:

[0162] The health status analysis module trains the LSTM model using historical sensor data collected from the target device when it is in a healthy operating state, thus obtaining a state baseline model.

[0163] The real-time status analysis module inputs real-time sensor data during the operation of the target device into the status benchmark model for analysis, extracts the status analysis results of each time step, fuses and analyzes to calculate the corresponding status score, and constructs a status analysis curve.

[0164] The state evolution simulation module analyzes the real-time state of the equipment based on the state analysis curve, calculates the corresponding survival risk based on the real-time state, simulates the state evolution process of the equipment, and constructs the state change curve.

[0165] The status monitoring module analyzes the equipment status based on status change curves, calculates the remaining service life and health period of the equipment, and monitors the health status of the equipment.

[0166] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. A method for monitoring the health status of equipment based on survival analysis, characterized in that, include: The LSTM model is trained using historical sensor data collected from the target device when it is in a healthy operating state to obtain a state baseline model. Real-time sensor data during the operation of the target device is input into the state benchmark model for analysis. The state analysis results of each time step are extracted, and the corresponding state score is calculated by fusion analysis to construct a state analysis curve. The real-time status of the equipment is analyzed based on the status analysis curve, and the corresponding survival risk is calculated in combination with the real-time status of the equipment. The equipment status evolution process is simulated, and a status change curve is constructed. Based on the state change curve analysis, the remaining service life and health period of the equipment are calculated to monitor the health status of the equipment.

2. The method for monitoring device health status based on survival analysis according to claim 1, characterized in that, The process involves inputting real-time sensor data from the target device during operation into a state baseline model for analysis, extracting the state analysis results for each time step, fusing and analyzing the results to calculate the corresponding state score, and constructing a state analysis curve, including: Real-time sensor data during the operation of the target device is input into the state baseline model for analysis. The state analysis results of each time step are extracted, and the reconstruction error between the state analysis results and the input data is calculated to obtain the reconstruction error sequence. By performing fusion analysis on the reconstructed error sequence, the corresponding state score is calculated, and a state analysis curve is constructed.

3. The method for monitoring equipment health status based on survival analysis according to claim 2, characterized in that, The process involves inputting real-time sensor data from the target device's operation into a state baseline model for analysis, extracting the state analysis results for each time step, calculating the reconstruction error between the state analysis results and the input data, and obtaining a reconstruction error sequence, including: Real-time sensor data during the operation of the target device is input into the state baseline model for analysis. At each time step, the state vectors generated by each layer of the model are extracted and concatenated to obtain the corresponding state analysis results. By decoding the state analysis results to generate the corresponding data vector, the Euclidean distance and the cosine of the angle between the data vector and the input data vector are calculated to obtain the reconstruction error. The reconstruction errors are integrated over time to obtain the reconstruction error sequence.

4. The device health status monitoring method based on survival analysis according to claim 3, characterized in that, The process of calculating the corresponding state score and constructing a state analysis curve by performing fusion analysis on the reconstructed error sequence includes: According to the preset time window and time step, slide on the reconstructed error sequence, calculate the error value in each window, form the baseline layer with the error value less than or equal to the preset error threshold, and form the anomaly layer with the error value greater than the preset error threshold. For the baseline layer, analyze the error changes to extract error distribution characteristics and construct error fluctuation boundaries; For the anomaly layer, analyze the error accumulation, calculate the cumulative error, and obtain the deviation. Based on the error fluctuation boundary and deviation amount, the deviation indicators at different time scales are analyzed and integrated, the corresponding state scores are calculated, and a state analysis curve is constructed.

5. The device health status monitoring method based on survival analysis according to claim 4, characterized in that, The process involves analyzing and fusing deviation indicators at different time scales based on error fluctuation boundaries and deviation amounts, calculating corresponding state scores, and constructing state analysis curves, including: Calculate the ratio of the deviation at each time step to the error fluctuation boundary to obtain the first deviation index; The first deviation index is weighted and averaged within the time window to obtain the second deviation index, where the weights are set according to the distance between the time point and the boundary of the time window; By combining the first deviation index and the second deviation index, the deviation degree is calculated, and the corresponding state score is obtained by mapping through a preset scoring mapping rule. The state scores are integrated in chronological order to construct a state analysis curve.

6. The device health status monitoring method based on survival analysis according to claim 1, characterized in that, The process of analyzing the real-time state of the equipment based on the state analysis curve, calculating the corresponding survival risk based on the real-time state of the equipment, simulating the state evolution process of the equipment, and constructing a state change curve includes: Extract the corresponding time-series features from the state analysis curves to analyze the real-time state of the equipment; By combining the real-time status of the equipment to calculate the corresponding survival risk, the equipment state evolution process is simulated, and a state change curve is constructed.

7. The method for monitoring equipment health status based on survival analysis according to claim 6, characterized in that, The process of calculating the survival risk corresponding to the real-time status of the equipment, simulating the equipment state evolution process, and constructing state change curves includes: Based on the real-time status of the equipment, the corresponding risk rate is mapped in the preset equipment baseline risk function, and the survival risk is calculated by integrating the risk rate to obtain the cumulative risk value. The device's state phase is determined based on the accumulated risk value; By combining the risk rate, cumulative risk value, and state phase, and using a preset state evolution model, the state evolution process of the equipment is simulated, and a state change curve is constructed.

8. The method for monitoring equipment health status based on survival analysis according to claim 1, characterized in that, The method of analyzing equipment status based on state change curves, calculating the remaining service life and health period of the equipment, and monitoring the health status of the equipment includes: Based on the state change curve analysis, the corresponding state lifetime is matched in the preset equipment state mapping library to obtain the state lifetime set. Based on the state lifetime set, the remaining service life and health period of the equipment are calculated to monitor the health status of the equipment.

9. The method for monitoring equipment health status based on survival analysis according to claim 8, characterized in that, The step of calculating the remaining service life and health period of the equipment based on the state life set to monitor the health status of the equipment includes: The matching degree between each state lifetime and the state change process is analyzed based on the state lifetime set, and corresponding fusion weights are set. The state lifetimes are weighted and averaged according to the fusion weights to obtain the remaining service life of the equipment. The boundaries of the health period are determined based on the remaining useful life in order to monitor the health status of the equipment.

10. A device health status monitoring system based on survival analysis, characterized in that, A method for monitoring device health status based on survival analysis as described in any one of claims 1 to 9, comprising: The health status analysis module trains the LSTM model using historical sensor data collected from the target device when it is in a healthy operating state, thus obtaining a state baseline model. The real-time status analysis module inputs real-time sensor data during the operation of the target device into the status benchmark model for analysis, extracts the status analysis results of each time step, fuses and analyzes to calculate the corresponding status score, and constructs a status analysis curve. The state evolution simulation module analyzes the real-time state of the equipment based on the state analysis curve, calculates the corresponding survival risk based on the real-time state, simulates the state evolution process of the equipment, and constructs the state change curve. The status monitoring module analyzes the equipment status based on status change curves, calculates the remaining service life and health period of the equipment, and monitors the health status of the equipment.

Images