Energy station equipment life prediction method, device, equipment and medium

Abstract

The application discloses an energy field station equipment life prediction method and device, equipment and medium, relates to the field of equipment life prediction, and the method comprises the steps of obtaining multi-source monitoring data in the current operation process of the energy field station equipment and preprocessing to obtain standardized time series data; extracting a to-be-analyzed subsequence from the standardized time series data, inputting the to-be-analyzed subsequence into a pre-trained reconstruction model based on a diffusion mechanism, and identifying a change point of the equipment; extracting a trend feature vector at the change point, and calculating the ratio of the amplitude of the trend feature vector deviating from a health reference trend to the amplitude of a failure reference trend deviating from the health reference trend as a change point severity quantification index; determining the severity stage of the equipment according to the change point severity quantification index, selecting a prediction model corresponding to the severity stage to perform residual life prediction, and outputting a residual life prediction result. The application improves the accuracy of residual life prediction and the adaptability to different degradation modes.

Description

Technical Field

[0001] This application relates to the field of equipment life prediction technology, and in particular to a method, apparatus, equipment and medium for predicting the life of equipment in energy plants. Background Technology

[0002] With the continuous improvement of the intelligence level of energy power plants, key equipment such as wind turbines, transformers, and gas turbines have been widely equipped with various types of sensors, including vibration, temperature, and pressure sensors, as well as video monitoring equipment, providing a rich data foundation for equipment status perception and health management. During the operation of energy power plant equipment, in order to prevent sudden failures, extend equipment life, and optimize maintenance resource allocation, research on the remaining life prediction of key equipment is particularly important.

[0003] Currently, related technologies divide the entire life cycle of equipment into a health stage and a degradation stage through a stage division mechanism. Corresponding life prediction models are used to perform fault warning and life prediction for the degradation characteristics of different stages. However, this method relies on human experience or fixed thresholds for stage division, which is difficult to adapt to the dynamic changes of degradation process under different equipment and different operating conditions. This results in poor accuracy in identifying degradation starting points, making it difficult for life prediction models to perform differentiated modeling and prediction based on the actual degradation state of the equipment. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, equipment and medium for predicting the lifespan of energy station equipment, so as to solve the technical problem that lifespan prediction models are difficult to differentiate and predict based on the actual degradation state of the equipment.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting the lifespan of equipment in energy plants, including: Acquire multi-source monitoring data during the current operation of energy station equipment and preprocess it to obtain standardized time-series data; The subsequence to be analyzed is extracted from the standardized time series data, and the subsequence to be analyzed is input into a pre-trained reconstruction model based on a diffusion mechanism to identify the turning point of the device; the turning point is the critical moment when the device transitions from the healthy stage to the degradation stage. The trend feature vector at the change point is extracted, and the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend is calculated as a quantitative indicator of the severity of the change point. The trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state. The severity stage of the device is determined based on the severity quantification index of the change point, the prediction model corresponding to the severity stage is selected to predict the remaining lifetime, and the remaining lifetime prediction result of the device is output.

[0006] Secondly, this application provides an energy station equipment life prediction device, the device comprising: The acquisition module is used to acquire multi-source monitoring data during the current operation of energy station equipment and preprocess it to obtain standardized time-series data. The change point identification module is used to extract the subsequence to be analyzed from the standardized time series data, input the subsequence to be analyzed into a pre-trained reconstruction model based on the diffusion mechanism, and identify the change point of the device; the change point is the critical moment when the device transitions from the healthy stage to the degradation stage. The severity quantification module is used to extract the trend feature vector at the change point and calculate the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend, as a severity quantification index of the change point; the trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state. The lifespan prediction module is used to determine the current severity stage of the device based on the severity quantification index of the change point, select the prediction model corresponding to the severity stage to predict the remaining lifespan, and output the remaining lifespan prediction result of the device.

[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for predicting the lifespan of energy station equipment.

[0008] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for predicting the lifespan of energy station equipment.

[0009] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, apparatus, equipment, and medium for predicting the lifespan of energy plant equipment. Compared with existing technologies, this solution preprocesses the acquired multi-source monitoring data into standardized time-series data, unifying different data standards and providing stable input data for subsequent change point monitoring. Furthermore, it employs a diffusion-based reconstruction model to perform in-depth analysis on the subsequences extracted from the standardized time-series data to automatically identify change points. This avoids the subjectivity and uncertainty introduced by traditional methods that rely on manual experience or fixed thresholds to define degradation initiation points, allowing the determination of change points to directly stem from changes in the distribution of actual equipment operating data, thereby effectively improving the accuracy of degradation initiation point identification. After detecting a change point, a trend feature vector is further extracted, and the magnitude of this vector's deviation from the healthy reference trend is calculated. The ratio of the degree of degradation to the deviation of the failure reference trend from the health reference trend quantifies the severity of degradation at the turning point into a continuous numerical index. This overcomes the shortcomings of existing technologies that treat the turning point as only a binary trigger event and ignore its degradation intensity information, providing a more granular basis for the selection of subsequent prediction models. Based on this, the current severity stage of the equipment is determined according to the quantified severity index, and the corresponding prediction model is selected for remaining lifetime prediction. This allows the degradation evolution law of different stages to be characterized by their respective specialized models, avoiding the problem that a single model cannot simultaneously adapt to different degradation paths such as slight gradual degradation and severe rapid degradation. This reduces prediction bias in complex degradation scenarios and improves the accuracy of remaining lifetime prediction and adaptability to different degradation modes. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the application environment for the energy station equipment life prediction method provided in one embodiment of this application; Figure 2 A flowchart illustrating a method for predicting the lifespan of equipment at energy plants, provided in an embodiment of this application; Figure 3 A schematic flowchart illustrating a method for changing the dot of an identification device according to an embodiment of this application; Figure 4 A flowchart illustrating a method for predicting the remaining lifetime of an output device according to an embodiment of this application; Figure 5 A flowchart illustrating a method for predicting the lifespan of energy plant equipment provided in another embodiment of this application; Figure 6 A schematic diagram of the functional modules of an energy station equipment life prediction device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] To more clearly explain this application, the following is a definition of the corresponding terms: Transition point: refers to the critical time point at which the operating state of equipment changes from a healthy state to a deteriorating state. This time point is usually manifested by a significant shift in the distribution characteristics, trend of change, or reconstruction error of sensor monitoring data, and can be used to indicate the critical moment when the equipment's healthy state transitions to a deteriorating state.

[0015] Quantification: After detecting a change point, it refers not only to determine whether the equipment status has changed, but also to further calculate the degree of degradation deviation corresponding to the change point and convert it into a continuous numerical index that can be used for model recognition and processing.

[0016] Multi-stage life prediction refers to a method of estimating remaining life by employing different prediction strategies or models based on different stages or severity categories of equipment degradation. Each stage's pre-model learns the evolutionary patterns within its corresponding degradation interval, unlike the approach of using a single model for the entire life cycle.

[0017] Reconstruction error: a measure of the difference between the model's input sequence and its reconstructed output. It remains at a low level during the healthy phase and increases significantly when the input deviates from the healthy distribution; it is used for change point determination.

[0018] One approach in related technologies is to use a unified deep model, such as Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and attention mechanisms, to automatically extract degradation features from high-dimensional, multi-source monitoring data. This approach has been widely used in life prediction for rotating machinery, electrical equipment, and other fields. However, in actual inspection and complex equipment maintenance scenarios, energy plants are affected by factors such as operating condition fluctuations, load changes, environmental disturbances, and individual differences. The rate and evolution trend of their transition from a healthy state to a degraded state often vary significantly. A unified prediction model cannot simultaneously and accurately characterize the detailed features of different degradation paths, such as "mild, gradual degradation" and "rapid, severe degradation," leading to significant prediction biases in complex degradation scenarios.

[0019] Another approach introduces a phase division mechanism to divide the entire equipment lifecycle into a health phase and multiple degradation phases. The model can then employ more targeted modeling strategies based on the degradation characteristics of different phases, enabling fault warning and lifespan prediction. For example, the degradation process of typical equipment such as rolling bearings typically exhibits multi-stage characteristics, including early slow degradation, mid-term moderate degradation, and late rapid degradation. Accurately identifying the turning points between each stage is crucial for subsequent lifespan prediction. Existing turning point detection methods mainly include likelihood ratio tests, Bayesian tests, and cumulative sum tests. However, some methods rely on manual experience or fixed thresholds for phase division, making it difficult to adapt to the dynamic changes in the degradation process under different equipment and operating conditions. They are prone to misjudging short-term anomalies caused by operating condition switching, load fluctuations, etc., as the starting point of degradation. Other methods, while capable of detecting turning points, treat them merely as simple binary trigger events, i.e., only judging "whether a change has occurred," ignoring the degradation intensity and severity information contained at the moment the turning point occurs.

[0020] In summary, existing technologies suffer from insufficient accuracy in identifying degradation initiation points and underutilization of the degradation intensity and severity information reflected by change points, making it difficult for lifetime prediction models to differentiate and predict based on the actual degradation state of the equipment.

[0021] To address the aforementioned shortcomings, this application provides a method for predicting the lifespan of energy plant equipment. Compared to existing technologies, this method preprocesses the acquired multi-source monitoring data into standardized time-series data, unifying different data standards and providing stable input data for subsequent change point monitoring. Furthermore, it employs a diffusion-based reconstruction model to perform in-depth analysis on the subsequences extracted from the standardized time-series data, automatically identifying change points. This avoids the subjectivity and uncertainty introduced by traditional methods that rely on manual experience or fixed thresholds to define degradation initiation points, ensuring that change point determination directly stems from changes in the distribution of actual equipment operating data, thereby effectively improving the accuracy of degradation initiation point identification. Moreover, after detecting a change point, a trend feature vector is further extracted, and the magnitude of this vector's deviation from the healthy reference trend is calculated, along with the deviation of the failure reference trend from the healthy reference trend. The ratio of the magnitude of the potential quantifies the severity of degradation at the change point into a continuous numerical index, overcoming the shortcomings of existing technologies that treat change points merely as binary trigger events and ignore their degradation intensity information. This provides a more granular basis for the selection of subsequent prediction models. Based on this, the current severity stage of the device is determined according to the quantified severity index, and the corresponding prediction model is selected for remaining lifetime prediction. This allows the degradation evolution patterns of different stages to be characterized by their respective specialized models, avoiding the problem that a single model cannot simultaneously adapt to different degradation paths such as mild, gradual degradation and severe, rapid degradation. This reduces prediction bias in complex degradation scenarios and improves the accuracy of remaining lifetime prediction and adaptability to different degradation modes.

[0022] The lifespan prediction method for energy plant equipment provided in this application embodiment can be applied to, for example... Figure 1 The application environment of the energy plant equipment life prediction method is shown. This application environment includes a terminal 102, a server 104, and a data storage system. The terminal 102 communicates with the server 104 via a network. The data storage system can store multi-source monitoring data acquired during the current operation of the energy plant equipment. The data storage system can be set up independently, integrated into the server 104, or placed in the cloud or on another server. After acquiring the multi-source monitoring data during the current operation of the energy plant equipment, the terminal 102 sends it to the server 104. The server 104, after acquiring the multi-source monitoring data, performs a remaining life prediction operation to obtain the remaining life prediction result of the equipment. Furthermore, in some embodiments, the energy plant equipment life prediction method can also be implemented independently by the server 104 or the terminal 102; for example, the terminal 102 can directly perform the remaining life prediction operation to obtain the remaining life prediction result of the equipment.

[0023] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0024] In one exemplary embodiment, such as Figure 2 As shown, a method for predicting the lifespan of energy plant equipment is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S201 to S204. Wherein: Step S201: Obtain multi-source monitoring data of the energy station equipment during its current operation and preprocess it to obtain standardized time-series data.

[0025] It should be noted that the aforementioned energy station equipment refers to key mechanical and electrical equipment deployed in energy production and distribution sites (power plants, substations, converter stations, etc.) that undertake energy conversion, transmission, and control functions. Examples include wind power plant equipment, thermal power and gas-steam combined cycle power plant equipment, and energy storage and power transmission / transformation station equipment. Wind power plant equipment includes equipment such as wind turbine main shaft bearings, gearboxes, generators, and pitch systems; thermal power and gas-steam combined cycle power plant equipment includes key equipment such as steam turbines, gas turbines, generator rotors, main transformers, feedwater pump sets, and coal mills; and energy storage and power transmission / transformation station equipment includes pumped storage units, compressed air energy storage expanders, high-voltage DC converter valves, and GIS high-voltage switchgear.

[0026] The aforementioned multi-source monitoring data refers to time-series data collected by various sensors deployed on energy station equipment during equipment operation, reflecting the equipment's operating status and degradation trends. Optionally, this multi-source monitoring data can be imported from external devices, obtained from blockchain or databases, or acquired in real time by various different sensors. This embodiment does not impose any limitations on the acquisition method of multi-source monitoring data.

[0027] In one embodiment, multi-source monitoring data of the energy station equipment during its current operation is acquired and preprocessed to obtain standardized time-series data, including: Acquire multi-source monitoring data of the energy station equipment during its current operation. The multi-source monitoring data includes at least: vibration signals collected by vibration sensors, temperature signals collected by temperature sensors, pressure signals collected by pressure sensors, current signals collected by current sensors, and speed signals collected by speed sensors. Identify and correct outliers for each sensor variable in the multi-source monitoring data, and calculate the standardized deviation of the monitored value of each sensor variable at each moment relative to the mean of the sensor variable. Identify and correct monitored values ​​whose standardized deviation exceeds a preset outlier threshold, obtaining corrected sensor variables. Normalize each corrected sensor variable to linearly map the monitored values ​​of each sensor variable to a unified numerical range, obtaining standardized time-series data.

[0028] Specifically, after acquiring multi-source monitoring data during equipment operation, the data needs to be preprocessed to eliminate differences in sampling frequency, dimension range, and data quality among different sensors, forming standardized time-series data that meets the input requirements of the subsequent diffusion-based reconstruction model. This multi-source monitoring data includes, but is not limited to, time-series data reflecting the equipment's operating status, such as vibration signals from vibration sensors, temperature signals from temperature sensors, pressure signals from pressure sensors, current signals from current sensors, and speed signals from speed sensors.

[0029] The aforementioned multi-source monitoring data includes multiple sensor variables. Each sensor variable is discretely sampled along the time axis to obtain a series of monitoring values. The arrangement of these monitoring values ​​over time constitutes the time series of that sensor variable. Assume the multi-source monitoring data collected by the energy station equipment at time T is represented as follows: ; Where n represents the total number of sensor variables, i.e., the number of dimensions of multi-source monitoring data; This represents the monitoring value of the nth sensor at time T; T is the total number of times, i.e., the number of sampling points on the time axis.

[0030] The preprocessing of multi-source monitoring data described above includes outlier handling and normalization. These processes improve the stability and consistency of the model's input data.

[0031] It is understandable that, due to factors such as electromagnetic interference, instantaneous mechanical shock, communication packet loss, or noise in the acquisition circuit, the original multi-source monitoring data may contain outliers that significantly deviate from the normal range during actual data acquisition. If such outliers are directly used for model training, they will interfere with the reconstructed model's learning of the data distribution characteristics during the healthy phase, reducing the accuracy of change point detection. Therefore, this application identifies and corrects outliers for each sensor variable in the multi-source monitoring data separately.

[0032] For the nth sensor variable in the multi-source monitoring data, first calculate the mean based on all its monitoring values ​​on the time axis. and standard deviation Then, calculate the sensor's monitored value at time t according to the following formula. Standardization bias : ; Compare the standardized deviation of the monitored value with the preset anomaly threshold. At that time, determine the monitoring value If it is an outlier, then the monitored value is considered an anomaly; otherwise, it is determined to be an outlier. This is a normal value. The preset anomaly threshold can be pre-set by engineers based on the sensor type and data quality requirements. A typical value is between 3 and 5, and this embodiment may preferably use a value between 3 and 5. The value is 3.

[0033] For monitoring values ​​identified as outliers, they are removed from subsequent processing, or replaced and corrected by linear interpolation results of monitoring values ​​at adjacent time points before and after the sensor variable, in order to maintain the continuity of time series data and thus obtain the corrected sensor variable.

[0034] Furthermore, after outlier handling, the corrected sensor variables need to be normalized to eliminate the impact of differences in dimensions and orders of magnitude between different sensor variables on model training. For example, the amplitude range of vibration signals is typically from zero to tens of gravitational accelerations (g), the numerical range of temperature signals is typically from zero to hundreds of degrees Celsius, and the numerical range of current signals may reach thousands of amperes. If data with these different dimensions are directly concatenated and input into the model, variables with larger numerical ranges will dominate gradient updates, causing model parameter updates to favor feature dimensions with larger numerical ranges, thus weakening the impact of variables with smaller numerical ranges but also carrying crucial degradation information on model training.

[0035] Therefore, this application normalizes each modified sensor variable separately. Specifically, for the monitored value of the nth sensor variable at time t... Its normalized value It can be calculated using the following formula: ; in, and These represent the maximum and minimum values ​​of the nth sensor variable across all monitored values, which can be obtained by iterating through all monitored values ​​of the sensor variable. The extremely small positive number introduced to prevent the denominator from being zero.

[0036] In this step, the normalization process described above enables the data of each sensor variable to be linearly mapped to a unified numerical range of [0, 1], thereby eliminating the dimensional differences between different sensors and placing all monitoring data under a relatively uniform numerical scale. This provides a stable and consistent standardized input for the subsequent training of the reconstructed model. All time-series data after normalization are referred to as standardized time-series data.

[0037] Optionally, to further reduce the impact of data noise on subsequent modeling, filtering can be performed after outlier handling and normalization. That is, for high-frequency sampled data such as vibration signals, wavelet threshold denoising or bandpass filtering methods are used to remove environmental noise and high-frequency interference components; for slowly varying signals such as temperature and pressure, moving average filtering methods are used to smooth short-term fluctuations while preserving long-term trends. The appropriate filtering method can be selected based on the specific sensor type and signal characteristics.

[0038] In this embodiment, by preprocessing the multi-source monitoring data, different sensor variables can be mapped to a relatively uniform numerical scale, thereby reducing the impact of the variable's numerical range on model training and providing stable input data for subsequent change point detection based on reconstruction error.

[0039] The above-mentioned reconstruction model is constructed through the following steps: Historical multi-source monitoring data is acquired, and data segments of the equipment health operation stage are extracted from the historical multi-source monitoring data as health stage samples; forward diffusion processing is performed on the health stage samples to obtain noise-injected health samples; the noise-injected health samples are input into a preset model to obtain the reconstruction output; the preset model is iteratively trained with the goal of minimizing the residual between the reconstruction output and the health stage samples until the convergence condition is met to obtain the reconstruction model.

[0040] It is understandable that the sensor monitoring data generated by energy plant equipment during the healthy operation phase follows a stable distribution pattern, which can be learned and represented by a reconstruction model trained on the healthy dataset. During the healthy phase, i.e., under normal equipment operating conditions, the model can reconstruct the input data with low error. However, as the equipment begins to degrade, the monitoring data gradually deviates from the data distribution corresponding to the healthy state, and the model cannot accurately reconstruct the input that deviates from the healthy distribution, resulting in a significant increase in reconstruction error. Therefore, this application can monitor the trend of reconstruction error over time, and when the reconstruction error exceeds a preset reconstruction error threshold, it can be determined that the equipment has entered the degradation phase at that corresponding moment.

[0041] It should be noted that increased reconstruction error may also be caused by changes in operating conditions, rather than by degradation of the equipment itself. To distinguish between these two situations, this application has eliminated differences in data dimensions and numerical ranges under different operating conditions through normalization during the data preprocessing stage. At the same time, a noise injection strategy is adopted in the training of the reconstruction model, forcing the model to learn the most essential stable distribution characteristics of healthy data rather than superficial fluctuations, thereby effectively reducing the interference of changes in operating conditions on reconstruction error.

[0042] In this embodiment, when constructing the reconstruction model based on the diffusion mechanism, historical multi-source monitoring data is first acquired and preprocessed to form historical standardized time-series data. Then, data segments corresponding to the known healthy operating stages of the equipment are extracted from the historical standardized time-series data as health stage samples. Assume that all historical standardized time-series data collected by the equipment at time T are: ; Where n is the number of sensor variables and T is the total number of time points.

[0043] Based on the equipment's prior operation records or historical maintenance files, from The time interval during which the intercepted device is in a normal and stable operating state constitutes the health phase sample, which is represented by the set: ; in, The healthy phase samples are represented by a matrix; f is the total number of time steps in the healthy phase. Indicates the first a time step Dimensional sensor data. It should be noted that... This is an important node in the data trend change. It is the monitored value of the nth sensor variable at the last time step f of the health stage. This moment corresponds to the boundary between the health stage and the subsequent degradation stage of the equipment. After this moment, the equipment status begins to show an unstable trend.

[0044] Within the framework of the diffusion mechanism, the core strategy for model training is to progressively add forward noise diffusion contamination to the healthy phase samples, constructing a degradation path from clean data to noisy data, and then training the model to learn the inverse process of restoring clean data from noisy data. Specifically, to improve the anti-interference ability and generalization performance of the reconstruction model under non-stationary conditions, it is necessary to perform forward noise diffusion contamination on the aforementioned healthy phase samples to obtain noisy versions of the healthy samples. The forward diffusion process is performed according to the following formula: ; in, The cumulative noise retention factor. and A predefined noise injection rate is used to control the proportion of the original signal retained at each step of the diffusion process; The scaled clean health monitoring data item, i.e. the stable component, represents the portion of effective information retained from the original healthy sample; The time-varying Gaussian noise matrix is ​​sampled independently at each time step, with each element sampled independently from the standard normal distribution N(0,1), used to inject random perturbations into healthy samples; The estimated trend matrix, derived from the trend decomposition module in the reconstructed model, characterizes the long-term evolution direction of the monitoring data along the time axis. A trend term is superimposed during the diffusion process, enabling the model to learn the degraded data distribution that integrates noise and trend information, thus enhancing its sensitivity to slow-drifting degradation. (Coefficients) This is used to control the contribution ratio of the trend item, so that the trend information is gradually incorporated into the sample during the diffusion process.

[0045] The diffusion process described above yields... It can be viewed as a "controlled perturbation" version of the health sample, which contains the core features (stable components) of the original health data, and is superimposed with random noise (to enhance robustness) and trend information (to simulate the directionality of degradation).

[0046] In this application, the Transformer model is used as the pre-defined model, which forms the backbone network of the reconstructed model. During training, healthy samples from noisy versions of the sequence are injected. As input to the preset model, the original, un-noisy health stage samples are used. As the reconstruction target, the parameters of the preset model are iteratively updated using the backpropagation algorithm. The optimization objective is to minimize the mean squared error between the reconstructed output and the reconstruction target. Training continues until a convergence condition is met, resulting in a well-trained reconstructed model. This convergence condition can be reaching a preset number of convergences or the residual between the reconstructed output and the original, un-noisy healthy phase sample being less than a preset residual value. The Transformer model uses a multi-head self-attention mechanism and a feedforward network to perform deep feature extraction and reconstruction of the input sequence, obtaining the reconstructed output. The reconstructed model represents the model whose parameters are optimal for the preset model.

[0047] In model training, the optimization objective is to minimize the residual between the reconstructed output and the original, un-noisy healthy stage samples. A mean squared error loss function is constructed based on the reconstructed output and the healthy stage samples. End-to-end iterative training is performed using the mean squared error loss function until the preset model meets the convergence condition, thus obtaining the reconstructed model. Through the above training process, the model can learn the distribution characteristics of the healthy stage monitoring data and has a high-precision reconstruction capability for healthy state samples.

[0048] Step S202: Extract the subsequence to be analyzed from the standardized time series data, input the subsequence to be analyzed into a pre-trained reconstruction model based on the diffusion mechanism, and identify the turning points of the device; the turning points are the key moments when the device transitions from the healthy stage to the degradation stage.

[0049] The subsequence to be analyzed is a local time series data segment extracted from standardized time series data using a sliding window method, used as input to the reconstruction model for change point detection.

[0050] In one embodiment, a method is also provided for extracting the subsequence to be analyzed from standardized time-series data, inputting the subsequence to be analyzed into a pre-trained reconstruction model based on a diffusion mechanism, and identifying the device's change points. For details on this method, please refer to [link to relevant documentation]. Figure 3 As shown, it includes: Step S301: Using a sliding window method, extract multiple subsequences to be analyzed one by one from the standardized time series data according to a preset window length; each subsequence to be analyzed includes the monitoring values ​​of all sensor variables that trace back the preset window length of consecutive time steps from the current time.

[0051] Step S302: Input each subsequence to be analyzed into the reconstruction model to obtain the reconstruction sequence corresponding to each subsequence to be analyzed.

[0052] Step S303: Calculate the root mean square error between each subsequence to be analyzed and the corresponding reconstructed sequence, use the root mean square error as the reconstruction error at each time step, and form a reconstruction error sequence based on the reconstruction error at each time step.

[0053] Step S304: Compare the reconstruction error at each time step in the reconstruction error sequence with the preset reconstruction error threshold.

[0054] Step S305: The moment when the reconstruction error exceeds the reconstruction error threshold is taken as the change point.

[0055] Specifically, after training, the trained reconstruction model is used to detect change points in the standardized time-series data to be analyzed. First, multiple subsequences to be analyzed are extracted one by one from the standardized time-series data using a sliding window approach. Let the sliding window length be w (i.e., the number of time steps contained within the window), and the number of sensor variables be f (i.e., f data dimensions). Then, the subsequence to be analyzed extracted at time t can be represented as: ; in, Let f be the f-dimensional sensor data vector at time T. The sliding window slides sequentially according to a preset window length, moving one time step forward each time, thereby generating a set of overlapping subsequences covering the entire time series data. This overlapping subsequence is used as the subsequence to be analyzed.

[0056] Each subsequence to be analyzed The model is input as a pre-trained reconstruction model, and it outputs the corresponding reconstruction sequence. The root mean square error (RMSE) is used to quantify the difference between the subsequence to be analyzed and its reconstructed sequence, resulting in the reconstruction error. Each subsequence to be analyzed corresponds to a reconstruction error at a specific time step. Calculated using the following formula: ; in, This represents the monitored value of the sensor variable at the i-th time step within the t-th window. This represents the reconstructed value corresponding to the i-th time step and the j-th sensor variable within the t-th window.

[0057] Through the above steps, the reconstruction error at each moment can be monitored over time, thus forming a reconstruction error sequence. When the equipment is in a healthy state, the reconstruction error fluctuates at a low level; when the equipment begins to degrade, the monitoring data deviates from the healthy distribution, and the reconstruction error increases significantly. The reconstruction error at each time point in the reconstruction error sequence is compared with a preset reconstruction error threshold. The comparison is performed, where the reconstruction error threshold is determined based on the reconstruction error distribution observed during the healthy phase, and is typically set as a multiple of the average reconstruction error during the healthy phase (e.g., the mean plus three standard deviations). The rule for determining change points is as follows: ; It should be noted that at a certain moment Reconstruction error Exceeding the preset reconstruction error threshold At that time, the judgment The time point is the turning point, indicating that the device has transitioned from the healthy stage to the degradation stage at that time. If the reconstruction error does not exceed the preset reconstruction error threshold, the device is determined to be still in a healthy state, and the reconstruction error changes at subsequent times will continue to be monitored.

[0058] Optionally, to avoid false alarms caused by single-point noise interference, a continuous judgment condition is set: the first moment exceeding the preset reconstruction error threshold is determined as a change point only when the reconstruction error of k consecutive moments exceeds the preset reconstruction error threshold. The value of k is determined based on the sensor sampling frequency, for example, it can be between 3 and 5. By setting the judgment condition, false alarms caused by occasional noise or transient disturbances can be effectively suppressed, improving the reliability of change point detection.

[0059] Existing methods for segmenting degradation stages often rely on manual experience, fixed thresholds, or single statistical indicators, making it difficult to adapt to the dynamic differences in degradation processes across different equipment and operating conditions in complex scenarios. This can easily lead to delays or misjudgments in identifying degradation initiation points. To address this issue, this application trains a diffusion-based reconstruction model using historical multi-source monitoring data. This model learns the distribution patterns of sensor data from various sensors during the healthy phase, thereby establishing a high-dimensional feature reconstruction benchmark under normal equipment operation. Furthermore, the reconstruction error is used to assess changes in the state of energy station equipment. When the equipment is in normal operation, the reconstruction model outputs a low error; when degradation begins, the monitoring data deviates from the healthy state characteristic distribution, and the reconstruction model output error increases significantly. Based on this principle, the reconstruction model analyzes the sub-sequence to be analyzed, accurately determining the corresponding reconstruction sequence. Then, the reconstruction error is calculated based on the reconstruction sequence and the sub-sequence to be analyzed, and a threshold judgment is made. This allows for the automatic capture of the critical moment when the equipment transitions from the healthy phase to the degradation phase, i.e., the turning point. Compared to methods that rely on human experience, fixed cycles, or single statistical indicators to determine the starting point of degradation, the degradation point identified by the method provided in this application is derived from changes in the distribution of actual equipment operating data, which can better reflect the time position when the equipment truly enters the degradation stage, thereby greatly improving the reliability of subsequent stage division and remaining life prediction.

[0060] Step S203: Extract the trend feature vector at the change point, and calculate the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend, as a quantitative indicator of the severity of the change point; the trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state.

[0061] Among them, the multi-source monitoring data at the change point refers to the data slice of all multi-source monitoring data at the change point, that is, the monitoring value vector of all sensor variables corresponding to the change point, which is a subset of all multi-source monitoring data in the time dimension.

[0062] After identifying change points based on the trained reconstruction model, it is necessary to further determine the severity of degradation corresponding to each change point. Different change points may represent different degrees of degradation. For example, some change points correspond to slight degradation, and the equipment can still operate safely for a long time; while others correspond to severe degradation, and the equipment is close to failure. Therefore, it is necessary to quantitatively assess the severity of change points to provide an accurate basis for subsequent segmented lifetime prediction.

[0063] In one embodiment, this application also provides a specific implementation method for calculating the quantitative index of change point severity, the method comprising: Based on the estimated trend matrix or the latent space representation of the model in the reconstructed model, the trend feature vector at the change point is obtained. Obtain the health reference trend vector and the failure reference trend vector; the health reference trend vector is the average value of the trend components extracted by the reconstruction model on the health phase samples; the failure reference trend vector is the trend component of the time window near the equipment failure time selected from historical multi-source monitoring data. Calculate the first magnitude of the deviation of the trend feature vector from the healthy reference trend vector, and the second magnitude of the deviation of the failure reference trend vector from the healthy reference trend vector; the first magnitude is represented by the first distance between the trend feature vector and the healthy reference trend vector; the second magnitude is represented by the second distance between the failure reference trend vector and the healthy reference trend vector. The ratio of the first amplitude to the second amplitude is used as a quantitative indicator of the severity of the change point.

[0064] Specifically, when a change point is detected Next, the trend feature vector corresponding to the change point is extracted from the reconstructed model. This trend feature vector originates from the estimated trend matrix output by the trend decomposition module in the reconstructed model. Alternatively, the model's latent space representation allows the estimated trend matrix to be extracted from processed historical standardized time-series data during model training via a decomposition module. This trend component reflects the long-term evolution direction of the data and is used to characterize the gradual shift in equipment state over time. Compared to relying solely on the location of change points, the trend component filters out random fluctuations and high-frequency noise in the data, providing a more stable and fundamental reflection of the degradation strength corresponding to the change point. The advantage of the trend component lies in its ability to filter out random fluctuations and high-frequency noise from the original monitoring data, preserving the essential evolutionary information of the equipment degradation process. Compared to directly using original sensor values ​​or reconstruction error values, the trend component offers greater robustness and interpretability in representing the degree of degradation. The change point... The trend feature vector at point is represented as: .

[0065] in, For the first Trend feature vector at any given time.

[0066] Understandably, to quantitatively assess the degree of degradation at the turning point, two reference anchors—a health benchmark and a failure benchmark—need to be established. The health reference trend vector is obtained by averaging the trend feature vectors extracted from all healthy phase samples using the trained reconstruction model along the sensor dimension. This represents the baseline trend value for the device when it is in a healthy state at the i-th sensor variable. This health reference trend vector represents the long-term evolution benchmark that the sensor monitoring data should follow under normal operating conditions, and can be used as a reference standard to determine whether the device deviates from its healthy state. Failure Reference Trend Vector This represents the trend value of the device when it approaches failure at the i-th sensor variable. It can be obtained by selecting samples from historical multi-source monitoring data where the device has already failed, taking the trend feature vector within the last time window before the failure time, and averaging it across sensor dimensions. This failure reference trend vector represents the trend characteristics of the device at the failure boundary and can serve as a reference standard for judging whether the device degradation is close to failure.

[0067] After obtaining the trend feature vector, health reference trend vector, and failure reference trend vector at the change point, the severity score of the change point is calculated. The Euclidean distance between the current trend vector and the health reference trend vector is also calculated. This distance reflects the degree to which the equipment degradation trend at the change point deviates from the health baseline and is called the first distance. This first distance quantifies the first magnitude of the trend feature vector's deviation from the health reference trend vector. The Euclidean distance between the failure reference trend vector and the health reference trend vector is then calculated. This distance reflects the maximum possible degradation magnitude from the health baseline to the failure baseline and is called the second distance. This second distance quantifies the second magnitude of the failure reference trend vector's deviation from the health reference trend vector. Finally, the first distance is divided by the second distance for normalization to obtain the severity quantification index. Its value ranges from 0 to 1. This index represents the position of the degradation trend at the current turning point within the complete interval from the health baseline to the failure baseline. The larger the index value, the closer the degradation trend at the turning point is to the failure state, and the more severe the equipment degradation; the smaller the index value, the less severe the equipment degradation, and the further away from failure. It is expressed by the following formula: ; in, For the i-th sensor in The trend feature vector at any given time; Let i be the health reference trend vector of the i-th sensor. This is the failure reference trend vector for the i-th sensor.

[0068] It should be noted that the above severity quantification indicators can be understood as severity scores, which actually mean: the degree of deviation of the degradation trend at the change point from the healthy baseline, as a percentage of the total deviation from the healthy baseline to the failure baseline. The closer it is to 0, the more likely it is to be a turning point. The closer the degradation trend is to the healthy baseline, the less severe the degradation. The closer it is to 1, the more likely it is to be a turning point. The closer the degradation trend is to the failure baseline, the more severe the degradation. The values ​​range continuously from 0 to 1, reflecting the complete gradual process from slight degradation to severe degradation.

[0069] In this embodiment, the normalization process described above enables the severity of change points under different devices and sensor configurations to be mapped to a unified [0,1] numerical range. This eliminates the influence of differences in device model, sensor type, and dimensions on the assessment of degradation, allowing the severity of change points to be compared on the same scale and providing a unified quantitative basis for the subsequent division of severity stages.

[0070] Step S204: Determine the current severity stage of the equipment based on the severity quantification index of the change point, select the prediction model corresponding to the severity stage to predict the remaining lifetime, and output the remaining lifetime prediction result of the equipment.

[0071] After completing change point detection and severity score calculation, it is necessary to further determine the discrete severity stage corresponding to the index. Since the normalized change point severity quantification index itself is a continuous value in the interval [0,1], it cannot be directly used as a discrete stage label such as "low severity", "medium severity" or "high severity". Therefore, it is necessary to divide it into different severity categories by setting a preset classification threshold to adapt to the subsequent segmented prediction model.

[0072] It should be noted that existing change point detection methods typically only determine whether the equipment state has changed, outputting a binary judgment result, but cannot further characterize the degree of degradation corresponding to the change point. For equipment in actual engineering, different change points may represent drastically different risk levels. For example, some change points may only correspond to slight degradation, and the equipment can still operate safely for a long time; while other change points may indicate that the equipment has entered a significant degradation stage, requiring timely maintenance measures. Therefore, it is necessary to quantitatively assess the severity of change points.

[0073] In one embodiment, this application also provides a specific implementation method for determining the current severity stage of the device based on a change point severity quantification index, selecting a prediction model corresponding to the severity stage to predict the remaining lifetime, and outputting the remaining lifetime prediction result of the device. Please refer to [link to relevant documentation]. Figure 4 As shown, the method includes: Step S401: Obtain at least two preset classification thresholds, compare the change point severity quantification index with each classification threshold respectively, and determine the current severity stage of the device; the severity stage includes a first severity stage, a second severity stage and a third severity stage with the severity increasing sequentially.

[0074] Step S402: Select the corresponding prediction model from multiple prediction models pre-trained for each severity stage, based on the severity stage.

[0075] Step S403: Input the multi-source monitoring data within the preset prediction window with the variable point as the endpoint to the selected prediction model, and output the remaining life prediction result.

[0076] In this embodiment, to determine the optimal classification threshold, it is first necessary to construct realistic degradation stage labels based on the degradation rate and true remaining lifespan of historical multi-source monitoring data. Specifically, taking a three-segment division as an example, the labels are based on the known true remaining lifespan values ​​of the equipment from historical operational data. Define two lifetime thresholds and ( < The historical samples were divided into three severity categories: When the true remaining lifetime value of the sample When the time is right, it is marked as the first severity level, i.e., low severity, indicating that the equipment is far from failure and the current degree of degradation is relatively mild; when At this point, it is marked as the second severity level, i.e., medium severity, indicating that the equipment has already undergone a certain degree of degradation and is moderately close to failure; when When the severity level is reached, it is marked as the third severity level, i.e., high severity, indicating that the equipment is close to failure and is about to malfunction. The above classification rules are expressed by the following formula: ; Among them, the above-mentioned lifetime threshold and This can be pre-set by engineers based on equipment type and historical failure data distribution; for example, a specific percentile of the equipment's average lifespan can be selected. =60% of the average equipment lifespan =30% of the average equipment lifespan; the specific value can be adjusted according to the actual characteristics of the equipment.

[0077] Furthermore, after obtaining the true stage labels of historical samples, it is necessary to determine the classification threshold that can optimally map the continuous NED values ​​to the above three categories. and Specifically, for each variable detected in the historical data, a severity quantification index, namely the severity score (NED value), is calculated. By traversing different candidate threshold combinations, the NED value of each sample is classified into predicted severity levels according to the following rules: ; The predicted severity levels are compared with the actual severity levels constructed based on the actual remaining lifespan, and the threshold combination with the highest classification accuracy is selected. , This serves as the final severity classification criterion. The above classification threshold is a global threshold calibrated based on the NED values ​​of historical samples and their corresponding true severity levels.

[0078] It should be noted that when When it is determined to be in the first severity stage (low severity), it means that the equipment degradation is relatively minor and is far from failure; when When the condition is determined to be in the second severity stage (moderate severity), it indicates that the equipment has undergone a certain degree of degradation and is moderately close to failure; when At this point, it is classified as the third severity stage (high severity), indicating that the equipment has severely degraded and is close to failure. Classification threshold. and The method involves pre-calibrating and optimizing using historical operational data: For each variable detected in the historical operational data, its NED value and a true severity level label constructed based on the actual remaining lifetime are calculated. Different candidate threshold combinations are iterated through, and the threshold combination with the highest classification accuracy that maps the NED value to each severity level is selected as the final threshold. and .

[0079] Furthermore, after obtaining the severity level, the predicted labels are compared with the derived severity categories, and the accuracy for each category is calculated to evaluate the degree of match between the predicted severity and the actual degradation level, i.e., the classification accuracy is calculated using the following formula: ; in, This represents the adaptive classification calibration accuracy for a specific severity level s. ; (True cases) are the number of samples with a true severity level of s that are correctly predicted as s; (True negative examples) are the number of samples whose true severity level is not s and which are correctly predicted to be not s; (False positives) is the number of samples whose true severity level is not s but is incorrectly predicted as s; (False negatives) are the number of samples whose true severity level is s but which are incorrectly predicted as not being s.

[0080] Through the above calibration process, the global classification threshold is determined. and This determination ensures that NED values ​​have a consistent severity classification standard across different devices and operating conditions.

[0081] After obtaining severity stage labels, the severity information is incorporated into the remaining lifespan prediction process. Specifically, all historical change point samples are divided into K severity levels (K can be 2, 3, or more; in this embodiment, K=3 is used as an example) based on the calibrated NED threshold. Then, a prediction model is trained for each severity level, which can be a Long Short-Term Memory (LSTM) network model. Different severity stages correspond to different prediction models, and each prediction model is trained independently using historical samples for the corresponding severity stage, with no parameter sharing. After training, each severity level corresponds to a dedicated prediction model.

[0082] In this embodiment, during the application phase, when the device is identified as a change point at a certain moment and the NED value is calculated, the severity stage to which the change point belongs is first determined according to the calibrated classification threshold, including the first severity stage (low severity), the second severity stage (medium severity), and the third severity stage (high severity). Then, the long short-term memory network model corresponding to the stage is selected to predict the remaining lifetime. The multi-source monitoring data within the preset prediction window with the change point as the endpoint is input into the corresponding prediction model, and the remaining lifetime prediction result is output.

[0083] The severity stages mentioned above are determined by comparing the change point severity quantification index with a preset classification threshold; the preset classification threshold is calibrated and optimized based on the classification accuracy between the change point severity quantification index in historical samples and the stage labels constructed based on the actual remaining lifetime.

[0084] It should be noted that, compared with the approach of using a single model for the entire life cycle, this application adopts a segmented modeling approach, which means that the prediction model for each stage only needs to learn the degradation evolution pattern within a specific severity stage, without having to take into account the differences between different degradation patterns such as the healthy stage, the early degradation stage, and the severe degradation stage. Therefore, each model has higher fitting accuracy and predictive specificity within its corresponding stage.

[0085] Optionally, to further reflect severity information during model training, this application employs an exponentially weighted mean squared error loss function instead of the traditional mean squared error loss function. The traditional mean squared error loss function treats the prediction errors of all training samples equally, without distinguishing the degradation stage of the samples. For samples in the early degradation stage (low severity), their remaining lifetime is longer and the degradation trend is slower, resulting in a larger timescale for the impact of prediction bias; for samples in the late degradation stage (high severity), their remaining lifetime is shorter and the degradation trend is more severe, leading to more urgent maintenance decision errors due to prediction bias. Therefore, it is necessary to apply differentiated weights to the prediction errors of samples at different stages.

[0086] The prediction model is trained using an exponentially weighted mean squared error loss function. Different weights are assigned to the prediction error based on the actual remaining lifetime of the sample, with samples having a larger remaining lifetime having a higher weight.

[0087] As an feasible approach, during the offline training phase, training samples corresponding to each severity stage are first extracted from historical multi-source monitoring data. For each severity stage, a training sample set is constructed from the historical operational data of that stage. The true remaining lifetime label of each sample in the training sample set corresponds to the time of the input sequence; that is, the input sequence with a certain time as the end point of the window has its label as the labeled value of the device's true remaining lifetime at that time. During training, the input sequence is used as the model input, and the true remaining lifetime percentage is used as the fitting target. The model parameters are iteratively updated using the backpropagation algorithm. By constructing an exponentially weighted mean squared error loss function, iterative optimization is performed according to minimizing the exponentially weighted mean squared error loss function to obtain the prediction model for the corresponding severity stage. The calculation formula of the exponentially weighted mean squared error loss function is as follows: ; Where N is the number of samples in the training batch (batch size); The hyperparameter for controlling the exponential scaling rate typically ranges from 0.5 to 2 and can be determined by engineers through cross-validation based on the equipment degradation characteristics. The true remaining lifetime of the t-th sample is expressed as a percentage, ranging from 0 to 1, where 1 indicates that the equipment has just been put into operation and the remaining lifetime is 100%, and 0 indicates that the equipment has failed and the remaining lifetime is 0%. This represents the percentage of remaining lifetime predicted by the model for the t-th sample. This is the weighted term for the index.

[0088] Understandably, the weighting mechanism of the loss function described above is as follows: for samples with a large actual remaining lifetime (corresponding to the early degradation stage and low severity), the value of the exponential weighting term is larger, and the contribution of the prediction error of this sample to the loss function is amplified. During training, the model will pay more attention to the prediction accuracy in the early degradation stage, encouraging the model to make more conservative and robust predictions in this stage. For samples with a small actual remaining lifetime (corresponding to the late degradation stage and high severity), the value of the exponential weighting term is smaller, and the contribution of the prediction error of this sample to the loss function is relatively reduced. The model is allowed to flexibly adjust its prediction behavior according to rapidly changing trends in the near-failure stage.

[0089] This embodiment introduces an exponentially weighted loss function, enabling the prediction model to automatically adjust its learning focus based on the different degradation stages of the samples during training. This avoids the problem of insufficient model accuracy at any stage due to the equalization of contributions between early-stage slow degradation samples and late-stage rapid degradation samples in the loss function, further improving the accuracy of remaining lifetime prediction throughout the entire lifecycle. Furthermore, the solution in this application not only determines whether the device has entered a degradation stage but also utilizes information on the strength of degradation for differentiated remaining lifetime prediction. This allows the prediction model to adaptively adjust its prediction strategy based on the device's current actual degradation state, effectively improving the accuracy of remaining lifetime prediction and its adaptability to different degradation scenarios.

[0090] Furthermore, the multi-source monitoring data within the preset prediction window ending at the change point refers to a segment of multi-source monitoring data truncated forward according to a preset time window length, with the change point as the end time of the data segment. This data segment is used as input to the prediction model for remaining lifetime estimation. The preset prediction window is the required input sequence length for the prediction model, which can be predetermined during model training.

[0091] When the device is at time Once a change point is identified and its severity is quantified, the current degradation state needs to be input into the prediction model for remaining lifetime estimation. The prediction model requires a continuous historical monitoring data segment, starting from the change point time. As the endpoint, the window length L remains consistent with that during model training. Specifically, time series data is extracted from the standardized time series data. A data subsequence with a length of L time steps and an endpoint is input into a prediction model corresponding to the current severity stage. The model outputs a prediction of the device's remaining lifetime. This remaining lifetime prediction can refer to the length of time the device is expected to operate normally from the current moment until it loses its specified function or reaches the failure criteria. It is usually expressed in time units (seconds, minutes, hours, days) or as a percentage of remaining lifetime.

[0092] For example, the main shaft bearing of a wind turbine is at a certain time It was identified as a change point, with a severity quantification index (NED) of 0.42, classifying it as a moderate severity stage. The corresponding predictive model was then input into the data (using...). Input the historical window data (as the endpoint) into the prediction model, and the model outputs... =0.25 (percentage form). Given that the bearing's total lifespan is approximately 50,000 hours, the remaining lifespan prediction is 25%, which translates to 50,000 × 25% = 12,500 hours. This means that the bearing is predicted to continue operating normally for approximately 12,500 hours without changing its operating conditions.

[0093] Traditional deep learning prediction models typically employ a uniform modeling approach for data across the entire lifecycle, making it difficult to simultaneously adapt to the different evolutionary patterns of the healthy, early degradation, and severe degradation stages. To address this issue, this application utilizes reconstruction error to identify the degradation initiation point and combines this with a severity quantification index to classify equipment degradation stages, incorporating stage information into the subsequent remaining life prediction process. When the severity index is low, the equipment is in the early degradation stage, with a relatively gentle degradation trend. The model can focus on the long-term stable trend of the data along the time axis to avoid misjudgments due to short-term fluctuations. When the severity index rises to a moderate level, the equipment enters the moderate degradation stage, with the degradation trend gradually becoming apparent and abnormal features increasing. The model needs to enhance its response weights to degradation trends and abnormal features while focusing on long-term trends to capture early signals of accelerated degradation. When the severity index reaches a high level, the equipment is nearing failure, with a drastic and rapidly changing degradation trend. The model should improve its ability to respond to rapid short-term changes, promptly capturing sudden changes in the degradation trend to provide a basis for emergency maintenance decisions.

[0094] Through the above mechanism, this application can adaptively adjust the prediction strategy according to the current degradation state of the equipment, so that the prediction model in the low severity stage focuses on long-term tracking of stable trends, and the prediction model in the high severity stage focuses on timely response to rapid changes. This effectively improves the model's comprehensive adaptability to complex degradation paths such as slight gradual degradation and severe rapid degradation, reduces prediction bias under extreme operating conditions and large-span degradation scenarios, and improves the timeliness of fault warning and the accuracy of remaining life prediction.

[0095] Please see Figure 5 As shown, the solution in this application includes four parts: "multi-source data preprocessing," "reconstruction model training based on diffusion mechanism," "severity quantification," and "segmented remaining lifetime prediction." First, multi-source monitoring data collected by various types of sensors deployed on energy plant equipment is acquired. This multi-source monitoring data is then preprocessed, including outlier handling and normalization, to eliminate differences in sampling frequency, unit range, and data quality among different sensors, thereby obtaining processed samples, i.e., standardized time-series data. The aforementioned reconstruction model can be pre-built or built in real-time during the application phase.

[0096] In the training process of the diffusion-based reconstruction model, historical standardized time-series data can be extracted from historical standardized time-series data, and data segments corresponding to the equipment's healthy operation stage can be extracted from the standardized time-series data as health stage samples. The health stage samples are input into the dynamic decomposition module of the preset model to separate stable and trend components. The stable component undergoes forward noise diffusion processing to obtain a noise diffusion perturbation sequence, which is used to train the diffusion-based reconstruction model, enabling the model to learn the data distribution characteristics under the equipment's healthy state. The trend component is used to determine the health reference trend and set a reconstruction error threshold. After model training is complete, the processed samples are input into the trained reconstruction model, and the reconstruction error is calculated time-by-time and compared with the preset threshold. When the reconstruction error exceeds the threshold, that time is determined to be a change point. After detecting a change point, the trend component at that change point is extracted, and the severity index of the change point, i.e., the severity score, is calculated in conjunction with the health reference trend. This severity score is used to characterize the severity of degradation corresponding to the change point.

[0097] In the process of performing segmented remaining lifetime prediction, the severity index (severity score) of the change point and the processed standardized time series data are obtained. Based on the comparison result of the severity score and the preset classification threshold, the current severity stage of the device is determined and the corresponding severity level label is generated. The historical samples are divided into multiple severity stages according to the severity level. A Long Short-Term Memory (LSTM) prediction model is trained for each stage. The corresponding prediction model is selected according to the severity stage of the current device. The monitoring data within the preset prediction window with the change point as the endpoint is input into the model to obtain the prediction result of the corresponding stage. For example, taking n severity stages, the prediction model corresponding to the low severity stage is LSTM (1), and the corresponding output result is the prediction result of stage 1; the prediction model corresponding to the medium severity stage is LSTM (2), and the corresponding output result is the prediction result of stage 2; the prediction model corresponding to the high severity stage is LSTM (3), and the corresponding output result is the prediction result of stage 3, ..., the prediction model corresponding to the higher severity stage is LSTM (n), and the corresponding output result is the prediction result of stage n. The prediction results of each stage are fused to output the remaining lifetime prediction result.

[0098] This application provides a method for predicting the lifespan of energy plant equipment. Compared with existing technologies, this method preprocesses the acquired multi-source monitoring data into standardized time-series data, unifying different data standards and providing stable input data for subsequent change point monitoring. Furthermore, it employs a diffusion-based reconstruction model to perform in-depth analysis on the subsequences extracted from the standardized time-series data to automatically identify change points. This avoids the subjectivity and uncertainty introduced by traditional methods that rely on manual experience or fixed thresholds to define degradation initiation points. The determination of change points directly stems from changes in the distribution of actual equipment operating data, effectively improving the accuracy of degradation initiation point identification. After detecting a change point, a trend feature vector is further extracted, and the magnitude of this vector's deviation from the healthy reference trend is calculated in relation to the failure parameter. By measuring the deviation of the degradation trend from the healthy reference trend, the severity of degradation at the turning point is quantified into a continuous numerical index. This overcomes the shortcomings of existing technologies that treat turning points merely as binary triggering events while ignoring degradation intensity information, providing a more granular basis for the selection of subsequent prediction models. Based on this, the current severity stage of the device is determined according to the quantified severity index, and the corresponding prediction model is selected for remaining lifetime prediction. This allows the degradation evolution patterns at different stages to be characterized by their respective specialized models, avoiding the problem that a single model cannot simultaneously adapt to different degradation paths such as mild, gradual degradation and severe, rapid degradation. This reduces prediction bias in complex degradation scenarios and improves the accuracy of remaining lifetime prediction and adaptability to different degradation modes.

[0099] Based on the same inventive concept, this application also provides an embodiment for implementing the energy station equipment life prediction device described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more energy station equipment life prediction device embodiments provided below can be found in the limitations of the energy station equipment life prediction method described above, and will not be repeated here.

[0100] In one exemplary embodiment, such as Figure 6 As shown, an energy plant equipment life prediction device is provided, the device comprising: The acquisition module 510 is used to acquire multi-source monitoring data during the current operation of energy station equipment and preprocess it to obtain standardized time-series data. The change point identification module 520 is used to extract the subsequence to be analyzed from standardized time series data, input the subsequence to be analyzed into a pre-trained reconstruction model based on diffusion mechanism, and identify the change point of the device; the change point is the critical moment when the device transitions from the healthy stage to the degradation stage. The severity quantification module 530 is used to extract the trend feature vector at the change point and calculate the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend, as a quantification index of the severity of the change point; the trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state. The life prediction module 540 is used to determine the current severity stage of the equipment based on the quantitative index of the severity of the change point, select the prediction model corresponding to the severity stage to predict the remaining life, and output the remaining life prediction result of the equipment.

[0101] As an optional implementation, the reconstructed model is built through the following steps: Acquire historical multi-source monitoring data and extract data segments from the historical multi-source monitoring data for the equipment health operation phase as health phase samples; The healthy samples are subjected to forward diffusion to obtain a noise-injected version of the healthy samples. Input a noisy version of a healthy sample into a preset model to obtain a reconstructed output; The optimization objective is to minimize the residual between the reconstructed output and the healthy stage samples. The parameters of the preset model are iteratively trained until the convergence condition is met, and the reconstructed model is obtained.

[0102] As an optional implementation, the change point recognition module 520 is specifically used for: Using a sliding window approach, multiple subsequences to be analyzed are extracted one by one from the standardized time series data according to a preset window length; each subsequence to be analyzed includes the monitoring values ​​of all sensor variables that have been traced back to the current time for a preset window length of consecutive time steps. Each subsequence to be analyzed is input into the reconstruction model to obtain the reconstruction sequence corresponding to each subsequence to be analyzed. Calculate the root mean square error between each subsequence to be analyzed and the corresponding reconstructed sequence, use the root mean square error as the reconstruction error at each time step, and form a reconstruction error sequence based on the reconstruction error at each time step. The reconstruction error at each time step in the reconstruction error sequence is compared with a preset reconstruction error threshold. The moment when the reconstruction error exceeds the reconstruction error threshold is taken as the variable point.

[0103] As an optional implementation, the severity quantification module 530 is specifically used for: Based on the trend matrix or latent space representation in the reconstructed model, the trend feature vector at the change point is obtained. Obtain the health reference trend vector and the failure reference trend vector; the health reference trend vector is the average value of the trend components extracted by the reconstruction model on the health phase samples; the failure reference trend vector is the trend component of the time window near the equipment failure time selected from historical multi-source monitoring data. Calculate the first magnitude of the deviation of the trend feature vector from the healthy reference trend vector, and the second magnitude of the deviation of the failure reference trend vector from the healthy reference trend vector; the first magnitude is represented by the first distance between the trend feature vector and the healthy reference trend vector; the second magnitude is represented by the second distance between the failure reference trend vector and the healthy reference trend vector. The ratio of the first amplitude to the second amplitude is used as a quantitative indicator of the severity of the change point.

[0104] As an optional implementation, the lifetime prediction module 540 is specifically used for: At least two preset classification thresholds are obtained, and the severity quantification index of the change point is compared with each classification threshold to determine the current severity stage of the device; the severity stage includes a first severity stage, a second severity stage, and a third severity stage with the severity increasing sequentially; Based on the severity stage, the corresponding prediction model is selected from multiple prediction models pre-trained for each severity stage; Input the multi-source monitoring data within the preset prediction window with the variable point as the endpoint into the selected prediction model, and output the remaining life prediction results.

[0105] As an optional implementation method, the prediction model is a long short-term memory network model; different prediction models correspond to different severity stages, and each prediction model is independently trained using historical samples of the corresponding severity stage.

[0106] As an optional implementation, the acquisition module 510 is specifically used for: Acquire multi-source monitoring data during the current operation of energy station equipment; the multi-source monitoring data includes at least: vibration signals collected by vibration sensors, temperature signals collected by temperature sensors, pressure signals collected by pressure sensors, current signals collected by current sensors, and speed signals collected by speed sensors; Outlier identification and correction are performed on the sensor variables in the multi-source monitoring data, and the standardized deviation of the monitoring value of each sensor variable at each time point relative to the mean of the sensor variable is calculated. Monitoring values ​​whose standardized deviation exceeds the preset abnormal threshold are identified as outliers and corrected to obtain the corrected sensor variables. The modified sensor variables are normalized to linearly map the monitored values ​​of each sensor variable to a unified numerical range, thus obtaining standardized time series data.

[0107] The energy station equipment life prediction device provided in this application preprocesses the acquired multi-source monitoring data into standardized time-series data, unifying different data standards and providing stable input data for subsequent change point monitoring. It employs a diffusion-based reconstruction model to perform in-depth analysis on the subsequences extracted from the standardized time-series data to automatically identify change points. This avoids the subjectivity and uncertainty introduced by traditional methods that rely on manual experience or fixed thresholds to define degradation initiation points, allowing the determination of change points to directly stem from changes in the distribution of actual equipment operating data, thereby effectively improving the accuracy of degradation initiation point identification. Furthermore, after detecting a change point, it further extracts a trend feature vector and calculates the magnitude of this vector's deviation from the healthy reference trend and the failure reference trend. The ratio of the deviation of the trend from the healthy reference trend quantifies the severity of degradation at the turning point into a continuous numerical index. This overcomes the shortcomings of existing technologies that treat turning points as merely binary trigger events and ignore their degradation intensity information, providing a more granular basis for the selection of subsequent prediction models. Based on this, the current severity stage of the device is determined according to the quantified severity index, and the corresponding prediction model is selected for remaining lifetime prediction. This allows the degradation evolution patterns of different stages to be characterized by their respective specialized models, avoiding the problem that a single model cannot simultaneously adapt to different degradation paths such as mild, gradual degradation and severe, rapid degradation. This reduces prediction bias in complex degradation scenarios and improves the accuracy of remaining lifetime prediction and adaptability to different degradation modes.

[0108] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting the lifespan of energy plant equipment.

[0109] Those skilled in the art will understand that Figure 7The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0110] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0111] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0112] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0113] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0114] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0115] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0116] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0117] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the lifespan of energy plant equipment, characterized in that, The method for predicting the lifespan of energy station equipment includes: Acquire multi-source monitoring data during the current operation of energy station equipment and preprocess it to obtain standardized time-series data; The subsequence to be analyzed is extracted from the standardized time series data, and the subsequence to be analyzed is input into a pre-trained reconstruction model based on a diffusion mechanism to identify the turning point of the device; the turning point is the critical moment when the device transitions from the healthy stage to the degradation stage. The trend feature vector at the change point is extracted, and the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend is calculated as a quantitative indicator of the severity of the change point. The trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state. The severity stage of the device is determined based on the severity quantification index of the change point, the prediction model corresponding to the severity stage is selected to predict the remaining lifetime, and the remaining lifetime prediction result of the device is output.

2. The method for predicting the lifespan of energy station equipment according to claim 1, characterized in that, The reconstruction model is constructed through the following steps: Acquire historical multi-source monitoring data and extract data segments of the equipment health operation stage from the historical multi-source monitoring data as health stage samples; The healthy phase samples are subjected to forward diffusion processing to obtain a noise-injected version of the healthy samples; The pre-defined model is input with the injected noise version of the healthy sample to obtain the reconstructed output; The parameters of the preset model are iteratively trained with the optimization objective of minimizing the residual between the reconstructed output and the healthy stage sample until the convergence condition is met, thus obtaining the reconstructed model.

3. The method for predicting the lifespan of energy station equipment according to claim 1, characterized in that, Extracting the subsequence to be analyzed from the standardized time-series data, inputting the subsequence to be analyzed into a pre-trained reconstruction model based on a diffusion mechanism, and identifying the changing points of the device, including: Using a sliding window approach, multiple subsequences to be analyzed are extracted one by one from the standardized time-series data according to a preset window length; each subsequence to be analyzed includes the monitoring values ​​of all sensor variables that trace back the preset window length of consecutive time steps from the current moment. Each subsequence to be analyzed is input into the reconstruction model to obtain the reconstruction sequence corresponding to each subsequence to be analyzed. Calculate the root mean square error between each subsequence to be analyzed and the corresponding reconstructed sequence, and use the root mean square error as the reconstruction error at each time step. A reconstruction error sequence is formed based on the reconstruction errors at each time step. Each subsequence to be analyzed corresponds to the reconstruction error at one time step. The reconstruction error at each time step in the reconstruction error sequence is compared with a preset reconstruction error threshold. The moment when the reconstruction error exceeds the reconstruction error threshold is taken as the change point.

4. The method for predicting the lifespan of energy station equipment according to claim 2, characterized in that, Extract the trend feature vector at the change point, and calculate the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failed reference trend from the healthy reference trend, as a quantitative indicator of the severity of the change point, including: Based on the estimated trend matrix or the latent space representation of the model in the reconstructed model, the trend feature vector at the change point is obtained; Obtain a health reference trend vector and a failure reference trend vector; the health reference trend vector is the average value of the trend components extracted by the reconstruction model on the health stage sample; the failure reference trend vector is the trend component of the time window near the equipment failure time selected from the historical multi-source monitoring data; Calculate a first magnitude of the deviation of the trend feature vector from the healthy reference trend vector, and a second magnitude of the deviation of the failure reference trend vector from the healthy reference trend vector; the first magnitude is represented by a first distance between the trend feature vector and the healthy reference trend vector; the second magnitude is represented by a second distance between the failure reference trend vector and the healthy reference trend vector. The ratio of the first amplitude to the second amplitude is used as a quantitative indicator of the severity of the change point.

5. The method for predicting the lifespan of energy station equipment according to claim 1, characterized in that, Based on the severity quantification index of the change point, the current severity stage of the device is determined. A prediction model corresponding to the severity stage is selected to predict the remaining lifetime, and the remaining lifetime prediction result of the device is output, including: At least two preset classification thresholds are obtained, and the severity quantification index of the change point is compared with each classification threshold to determine the current severity stage of the device; the severity stage includes a first severity stage, a second severity stage, and a third severity stage with progressively increasing severity. Based on the severity stage, a corresponding prediction model is selected from multiple prediction models pre-trained for each severity stage; The multi-source monitoring data within the preset prediction window with the change point as the endpoint are input into the selected prediction model, and the remaining life prediction result is output.

6. The method for predicting the lifespan of energy station equipment according to claim 5, characterized in that, The prediction model is a long short-term memory network model; different prediction models correspond to different severity stages, and each prediction model is independently trained using historical samples of the corresponding severity stage.

7. The method for predicting the lifespan of energy station equipment according to claim 1, characterized in that, Acquire multi-source monitoring data of energy plant equipment during its current operation and preprocess it to obtain standardized time-series data, including: Acquire multi-source monitoring data during the current operation of energy station equipment; the multi-source monitoring data includes at least: vibration signals collected by vibration sensors, temperature signals collected by temperature sensors, pressure signals collected by pressure sensors, current signals collected by current sensors, and speed signals collected by speed sensors; Outlier identification and correction are performed on each sensor variable in the multi-source monitoring data, and the standardized deviation of the monitoring value of each sensor variable at each time point relative to the mean of the sensor variable is calculated. Monitoring values ​​whose standardized deviation exceeds the preset abnormal threshold are identified as outliers and corrected to obtain the corrected sensor variables. The modified sensor variables are normalized to linearly map the monitored values ​​of each sensor variable to a unified numerical range, thus obtaining standardized time series data.

8. A device for predicting the lifespan of energy station equipment, characterized in that, The energy station equipment life prediction device includes: The acquisition module is used to acquire multi-source monitoring data during the current operation of energy station equipment and preprocess it to obtain standardized time-series data. The change point identification module is used to extract the subsequence to be analyzed from the standardized time series data, input the subsequence to be analyzed into a pre-trained reconstruction model based on the diffusion mechanism, and identify the change point of the device; the change point is the critical moment when the device transitions from the healthy stage to the degradation stage. The severity quantification module is used to extract the trend feature vector at the change point and calculate the ratio of the deviation of the trend feature vector from the healthy reference trend to the deviation of the failure reference trend from the healthy reference trend, as a severity quantification index of the change point; the trend feature vector is used to characterize the degree of deviation of the long-term evolution trend of the multi-source monitoring data of the device at the change point along the time axis from the data evolution benchmark under the healthy state. The lifespan prediction module is used to determine the current severity stage of the device based on the severity quantification index of the change point, select the prediction model corresponding to the severity stage to predict the remaining lifespan, and output the remaining lifespan prediction result of the device.

9. A computer device, comprising: The memory and processor contain a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the energy plant equipment life prediction method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting the lifespan of energy station equipment as described in any one of claims 1-7.

Images