Complex equipment health state prediction method and device based on adaptive position coding
By combining adaptive position encoding and dual-stream feature extraction networks, the perception and cognitive confusion problem of deep neural networks in equipment health status prediction is solved, and more accurate equipment health status prediction, especially the prediction of remaining service life, is achieved.
Patent Information
- Application Number
- CN202510867321.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-09-26
AI Technical Summary
Existing deep neural network methods suffer from perceptual confusion and cognitive confusion in equipment health status prediction, making it difficult to effectively utilize limited monitoring data for accurate predictions, especially during long-term continuous equipment degradation, when changes in the external environment and working settings lead to data label confusion and health status cognitive confusion.
A complex equipment health status prediction method based on adaptive position coding is adopted. By obtaining multi-dimensional sensor data for preprocessing, a "state-state gradient" dual-stream feature extraction network is used. Real-time and global feature extraction are combined, and feature dot product is used for fusion. A learnable equipment degradation trend description model is constructed, which is trained with a physical constraint loss function.
The accuracy and stability of equipment health status prediction are improved, and the global and local information of the equipment can be extracted more effectively, achieving more accurate remaining service life prediction.
Smart Images
Figure CN120705553A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of industrial equipment management, and in particular to a method and device for predicting the health status of complex equipment based on adaptive position coding. Background Art
[0002] In recent years, deep neural networks have been widely used for predicting equipment health status. Limited by the model's limited receptive field, existing methods are often only able to predict equipment health status based on a limited set of state records. However, equipment health degradation is often a long-term, continuous process and is easily affected by numerous factors, such as the external environment and operating settings. Due to a lack of global understanding of the equipment, existing methods often struggle to identify equipment health status using limited but similar monitoring data. Firstly, within a limited receptive field, different equipment operations may be recorded as similar samples, making it difficult to establish a one-to-one correspondence between data and labels (health status), leading to label confusion during model training. Secondly, due to differences in initialization settings and usage processes across different equipment, similar operating states can indicate diverse health states, leading to confusion in health status perception during model training and application. These issues severely limit the predictive performance of these methods. Summary of the Invention
[0003] Based on this, it is necessary to provide a method and device for predicting the health status of complex equipment based on adaptive position coding, which can effectively improve the prediction performance, in order to address the above technical problems.
[0004] A method for predicting the health status of complex equipment based on adaptive position coding, the method comprising: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
[0005] In one embodiment, preprocessing the observation data includes: After compensating and standardizing the observation data, preliminary processed data are obtained; After filtering the preliminary processed data with a low-pass filter, the filtered data is enhanced by taking the difference between the filtered data and the original data to obtain the status data.
[0006] In one embodiment, a sliding window of a preset size is used to extract local data from the state data as the real-time state sample; The state data is compressed using an equal-interval sampling strategy to obtain a global state sample with the same length as the real-time state sample.
[0007] In one embodiment, when the real-time feature extraction unit is used to extract features based on the real-time state sample: In the state feature extraction branch, a temporal convolution unit including multiple temporal convolution layers is used to perform multi-level feature extraction on the real-time state sample, and then a fully connected unit is used to generate real-time state features based on the features corresponding to the latest moment in the deep-level features; In the state gradient feature extraction branch, the difference matrix construction unit is used to construct a corresponding real-time difference matrix based on the real-time state sample, the 2D convolution unit is used to perform multi-level feature extraction on the real-time difference matrix, the extracted features are subjected to global mean pooling using the global mean pooling unit, and expanded into feature vectors, and the fully connected unit obtains real-time differential features based on the feature vectors; The real-time status feature is combined with the real-time differential feature to obtain the real-time feature.
[0008] In one embodiment, the difference matrix is constructed using the following formula:
[0009] In the above formula, Indicates the real-time status sample j The difference matrix corresponding to the sensor variables, Indicates the j The sensor collects m signal value.
[0010] In one embodiment, in the global feature extraction unit: respectively using a state feature extraction branch and a state gradient feature extraction branch to extract the global state features and the global state difference matrix of the global state sample; Constructing a time-coded feature sequence according to the sampling time series of the observation data, the global state difference matrix, and the global state sample; Based on two temporal convolutional layers, a learnable description model is constructed, and the temporal encoding feature sequence is used as input of the learnable description model to obtain a first attention vector and a second attention vector; Adopting a channel attention mechanism, using the first attention vector to select important variables in the global state sample to obtain a weighted global sample, constructing a weighted global state difference matrix based on the weighted global sample, and then calculating the weighted global state difference feature according to the weighted global state difference matrix; The second attention vector is combined with the weighted global state differential feature to obtain a combined feature, a mixed attention vector is calculated based on the combined feature, and the global state feature is weighted according to the mixed attention vector through the channel attention mechanism to obtain the final global feature.
[0011] In one embodiment, the time coding feature sequence is represented as:
[0012] In the above formula, and Calculated based on the global state difference matrix and the global state sample, k = 1,2, …, d p / 2, d p The feature dimension representing the time-coded feature sequence.
[0013] In one embodiment, a prediction loss function and a physical constraint loss function are used when training the real-time feature extraction unit, the global feature extraction unit and the fully connected network.
[0014] In one embodiment, the physical constraint loss function is expressed as:
[0015] in,
[0016]
[0017] In the above formula, represents an identity matrix, and Represent the sample prediction results and actual health status of a training batch respectively, and yes and The corresponding elements, Represents the batch that meets the conditions The number of
[0018] The present application also provides a complex equipment health status prediction device based on adaptive position coding, the device comprising: An observation data acquisition module is used to acquire observation data related to the device to be health predicted, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; A data preprocessing module is used to preprocess the observation data to obtain state data, select a segment of data close to the current time in the state data as a real-time state sample, and downsample the state data to obtain a global state sample with the same length as the real-time state sample; A real-time and global feature extraction module, configured to input the real-time state samples and the global state samples into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, wherein the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and wherein the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The state prediction module is used to fuse the real-time features and global features by using feature dot product to obtain fused features, and then use a fully connected network to obtain the health prediction results of complex equipment based on the fused features.
[0019] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
[0020] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
[0021] The above-mentioned complex equipment health status prediction method and device based on adaptive position coding preprocesses observation data obtained by monitoring multiple sensors in chronological order to obtain status data. Based on the status data, real-time status samples and global status samples are generated. These two status data are respectively input into a real-time feature extraction unit and a global feature extraction unit to obtain real-time features and global features. Both the real-time feature extraction unit and the global feature extraction unit utilize a "state-state gradient" dual-stream feature extraction network. The real-time features and global features are fused using a feature dot product method to obtain fused features. A fully connected network is then used to obtain the health prediction results of the complex equipment based on the fused features. This sampling method can more effectively extract features from measured data, thereby improving the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 1 is a flow chart of a method for predicting the health status of complex equipment based on adaptive position coding in one embodiment; Figure 2 Schematic diagram of the structure of a "state-state gradient" dual-stream feature extraction network in one embodiment; Figure 3 Schematic diagram of the specific process from observing data to obtaining the final fusion features in one embodiment; Figure 4 1 is a structural block diagram of a complex equipment health status prediction device based on adaptive position coding in one embodiment; Figure 5 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0023] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.
[0024] In the existing technology, the perceptual confusion and cognitive confusion existing in the data-driven deep neural network (DNN) method lead to inaccurate prediction problems, such as Figure 1 As shown, a complex equipment health status prediction method based on adaptive position coding is provided, which specifically includes the following steps: Step S100: Obtain observation data related to the device to be used for health prediction. The observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order.
[0025] Step S110 , preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment in the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample.
[0026] In step S120, the real-time state samples and the global state samples are respectively input into the real-time feature extraction unit and the global feature extraction unit to obtain real-time features and global features, wherein the real-time feature extraction unit and the global feature extraction unit both adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch. The state feature extraction branch includes a temporal convolution unit and a fully connected unit. The state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence.
[0027] In step S130 , the real-time features and the global features are fused by using a feature dot product method to obtain fused features, and then a fully connected network is used to obtain a health prediction result of the complex equipment based on the fused features.
[0028] In this example, a novel adaptive position encoding-based method for predicting the health status of machinery is proposed. This method integrates global and local information about the equipment degradation process to achieve more accurate RUL (Remaining Useful Life) prediction. Furthermore, a "state-state gradient" dual-stream feature extraction network is proposed in this method to provide more efficient feature extraction and improve prediction accuracy.
[0029] The method proposed in this paper is applicable to the remaining useful life prediction of complex equipment, such as vehicles, engines, or their key components. When continuously monitoring the condition of complex equipment, the observation data in different dimensions can be real-time continuous data received by temperature sensors and vibration frequency sensors installed at different locations.
[0030] In step S100, multiple sensors of varying types, functions, and signals are used to monitor the target complex equipment in real time, thereby obtaining observation data of various dimensions. Each observation data is time-domain data and may include temperature, pressure, vibration, and other data. It is understood that there may be one or more sensors of various types, located in different spatial locations or at different locations on the target equipment.
[0031] In this embodiment, the observation data can be represented as a multi-dimensional time series ,in Representatives in t Collected at all times mdimensional sensor signal.
[0032] In step S110, the main function of preprocessing is to Extract real-time samples describing the real-time status of complex equipment That is, real-time status samples and global samples used to describe the global status of the device That is, the global state sample, which specifically includes five steps: data compensation, data normalization, data filtering, data enhancement, and data sampling.
[0033] In this embodiment, when preprocessing the observation data, the observation data is first compensated and standardized to obtain preliminary processed data, and then the preliminary processed data is filtered using a low-pass filter. The difference between the filtered data and the original data is then used to enhance the filtered data to obtain status data.
[0034] Specifically, considering that modern equipment often has multiple operating modes, the switching of operating modes will lead to the switching of monitoring data distribution. Therefore, when conducting equipment health status assessment, it is necessary to first eliminate the impact of equipment operating condition switching on the status assessment task. Given the subordinate working mode i Observation data obtained from monitoring , where an operating mode represents a set of operating parameters. For example, a car has multiple gears, and each gear is an operating mode.
[0035] The corresponding health data can be represented as , then the data compensation operation can be characterized as:
[0036] Furthermore, the data after compensation is standardized. Perform normalization operation and get , and its normalization process is expressed as:
[0037] In the above formula, represent The j sensor variables, and They represent the upper and lower bounds of the sensor value respectively.
[0038] Next, considering the complexity of the equipment's operating environment, the actual collected data often contains varying degrees of noise. Since the frequency of data noise is often significantly higher than the frequency of normal signals for equipment health status prediction tasks, a low-pass filter can be used to filter the data.
[0039] In one embodiment, a Butterworth filter is used to perform data filtering, which is expressed as:
[0040] In the above formula, Represents the function of the Butterworth filter, and T is the corresponding threshold.
[0041] Furthermore, considering the possibility that there is effective information in the noise signal, this method still retains the noise signal as the compensation input , enhance the original data. The enhanced signal can be expressed as:
[0042]
[0043] Finally, a sliding window of preset size is used to extract local data from the state data as real-time state samples. The state data is compressed using an equal-interval sampling strategy to obtain a global state sample with the same length as the real-time state sample.
[0044] Specifically, in the state data obtained after preprocessing, the sliding window sampling is used to obtain the state data describing the current state of the machine. t Data samples of working status at all times , as real-time state samples. And adopt equal interval sampling strategy to record global Compress to the same size as the sample , used to describe the sample of the macro degradation state of the device, that is, the global state sample.
[0045] like Figure 2 As shown in the figure, a "state-state gradient" dual-stream feature extraction network is proposed to extract the real-time features of equipment at the micro scale from the real-time state samples and global state samples obtained in step S110. and global characteristics at the macro scale Since both the real-time feature extraction unit and the global feature extraction unit are constructed based on the “state-state gradient” dual-stream feature extraction network, the feature extraction process of the real-time feature extraction unit is taken as an example to illustrate the structure of the dual-stream feature extraction network.
[0046] In this embodiment, when a real-time feature extraction unit is used to perform feature extraction based on real-time state samples: in the state feature extraction branch, a temporal convolution unit including multiple temporal convolution layers is used to perform multi-level feature extraction on the real-time state samples, and then a fully connected unit is used to generate real-time state features based on the features corresponding to the latest moment in the deep-level features.
[0047] Specifically, the state feature extraction branch is used to extract the distribution characteristics of the state monitoring data. Given the input real-time state sample This branch will use multiple temporal convolutional network (TCN) modules to extract multi-level features and obtain deep features. On this basis, a fully connected network is adopted, and At the latest moment t The characteristic vector on As input, calculate the state features .
[0048] In this embodiment, in the state gradient feature extraction branch, a differential matrix construction unit is used to construct a corresponding real-time differential matrix based on the real-time state sample, and a 2D convolution unit is used to perform multi-level feature extraction on the real-time differential matrix. The extracted features are globally mean pooled using a global mean pooling unit and expanded into a feature vector. The fully connected unit obtains the real-time differential features based on the feature vector.
[0049] Specifically, the difference matrix is constructed using the following formula:
[0050] In the above formula, Indicates the real-time status sample j The difference matrix corresponding to the sensor variables, Indicates the j The sensor collects m signal value, i.e. The mth variable in No. j A portion.
[0051] Furthermore, a 2D convolution module is used to extract multi-level features from the difference matrix, and the output features of the last convolution layer are Perform global mean pooling and expand it into a feature vector, and finally use this as input to calculate real-time differential features using a fully connected network Finally, the real-time differential feature Combined with real-time state features to obtain real-time features .
[0052] In this embodiment, the basic function of the global feature extraction unit is to extract the complex equipment based on the historical records. Evaluate the macro-changing trend of its state. Its backbone network composition and structural parameters are basically the same as those of the real-time feature extraction module, both of which are a "state-state gradient" dual-stream feature extraction network to extract the state from the Extract the global state characteristics of the device And the global state gradient characteristics .
[0053] Furthermore, the degradation trends of complex equipment of the same model often follow certain patterns. In this method, a learnable model describing equipment health degradation trends (hereinafter referred to as a learnable description model) is constructed, drawing on the positional encoding strategy of the transformer. This model uses the degradation records of a large number of similar equipment to learn a general function (template) module that describes the macroscopic degradation process of the health status of that type of equipment. Based on this, this module is adaptively adjusted based on the real-time status of the monitored equipment to adapt to the degradation process of the individual equipment. Finally, the attention vector generated by this function module guides feature extraction in the global feature extraction module. This method utilizes the degradation records of a large number of similar equipment to discover the degradation patterns of that type of equipment and effectively incorporates this information as prior information into the extraction of the health status features of the current equipment, thereby achieving more stable and reliable feature extraction.
[0054] In this embodiment, a state feature extraction branch and a state gradient feature extraction branch are respectively used in the global feature extraction unit to extract the global state features and global state difference matrix of the global state sample. A time-coded feature sequence is constructed based on the sampling time series of the observation data, the global state difference matrix, and the global state sample. A learnable description model is constructed based on two temporal convolutional layers. The time-coded feature sequence is used as the input of the learnable description model to obtain a first attention vector and a second attention vector. A channel attention mechanism is adopted, and the first attention vector is used to select important variables in the global state sample to obtain a weighted global sample. A weighted global state difference matrix is constructed based on the weighted global sample. A weighted global state difference feature is then calculated based on the weighted global state difference matrix. The second attention vector is combined with the weighted global state difference feature to obtain a combined feature. A mixed attention vector is calculated based on the combined feature. The global state feature is weighted according to the mixed attention vector using the channel attention mechanism to obtain the final global feature.
[0055] Specifically, the global feature extraction unit consists of two parts: a time coding feature sequence construction module and an attention vector calculation module.
[0056] Furthermore, the time coding feature sequence construction module is mainly based on the time series Compute time-coded feature sequences , expressed as:
[0057] In the above formula, and Calculated based on the global state difference matrix and the global state sample, k = 1,2, …, d p / 2, d p Represents the feature dimension of the time-coded feature sequence. and The network modules are all combinations of two-layer convolutional networks and two-layer fully connected networks.
[0058] Furthermore, the attention vector calculation module includes CNN Ga , CNN Gwa Two time-coded feature sequences is the input convolution module. Each module is composed of two temporal convolution layers and a two-layer fully connected network. Ga Used to calculate the first attention vector , CNN Gwa It is used to calculate the second attention vector .
[0059] Each element in will be compared with the global sample Multiply the data channels in the channel and implement global sample weighting through channel weighting The selection of key variables in the Based on this, we further calculate the weighted global state difference matrix , and based on this, calculate the weighted global state differential feature ; will with Combined and calculated through a two-layer fully connected network to form a hybrid attention vector , used to combine the macro status information of the monitored equipment ( ) and prior knowledge ( ). Each element in will be combined with the global feature Multiply the feature channels in the matrix to achieve global features Further selection of key variables in the process and finally obtaining the characteristics used to characterize the macroscopic state of the device.
[0060] In this embodiment, the global features and real-time features are fused by feature dot product to obtain the final state features. Finally, a fully connected network is used to Based on the prediction of the health status of complex equipment.
[0061] like Figure 3As shown in the figure, it is the data processing flow of the above-mentioned data preprocessing process, real-time feature extraction unit, and global feature extraction unit. Figure 3 The specific process from observing data to obtaining the final fusion features is described in detail.
[0062] In this embodiment, the above-mentioned real-time feature extraction unit, global feature extraction unit and fully connected network can be used to construct a complex equipment health prediction model. When training the health prediction model, in order to improve the training stability of the model, the training adopts a multi-task learning strategy for model training, and its training loss is specifically determined by the prediction loss function. L p With physical constraint loss function L k It consists of two parts.
[0063] Specifically, the prediction loss function L p The prediction error of the model for the health status of the equipment. y p and the true value y r , L p It can be calculated as:
[0064] In the above formula, n b Represents the size of the training batch.
[0065] Furthermore, considering that the degradation process of mechanical equipment is irreversible, the health status of the equipment at the end of operation cannot be better than the initial state. Based on this common sense, a physical constraint loss function can be constructed: L k , expressed as:
[0066] in,
[0067]
[0068] In the above formula, represents an identity matrix, and Represent the sample prediction results and actual health status of a training batch respectively, and yes and The corresponding elements, Represents the batch that meets the conditions The number of
[0069] Finally, the total loss function is expressed as:
[0070] In the above-mentioned complex equipment health status prediction method based on adaptive position coding, the observation data is preprocessed to obtain status data, a section of data close to the current moment is selected from the status data as the real-time status sample, and the status data is downsampled to obtain a global status sample with the same length as the real-time status sample. The real-time status sample and the global status sample are respectively input into the real-time feature extraction unit and the global feature extraction unit to obtain real-time features and global features. Among them, the real-time feature extraction unit and the global feature extraction unit both adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch. The state feature extraction branch includes a time convolution unit and a fully connected unit. The state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit and a fully connected unit connected in sequence. The real-time features and the global features are fused by feature dot product to obtain fused features, and then the fully connected network is used to obtain the health prediction results of the complex equipment based on the fused features. A state-trend two-stream convolutional network is proposed to provide more effective feature extraction capabilities. An adaptive position encoding strategy is also proposed in the global feature extraction unit, enabling the model to more efficiently adapt to devices with individual differences. During training, multi-task training constraints based on physical information constraints are combined to achieve more efficient model training. This method integrates global and local information about the device degradation process to achieve more accurate RUL prediction performance.
[0071] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0072] In one embodiment, Figure 4 As shown, a complex equipment health status prediction device based on adaptive position coding is provided, comprising: an observation data acquisition module 200, a data preprocessing module 210, a real-time and global feature extraction module 220 and a status prediction module 230, wherein: The observation data acquisition module 200 is used to acquire observation data related to the device to be health predicted, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; A data preprocessing module 210 is configured to preprocess the observation data to obtain state data, select a segment of data close to the current time in the state data as a real-time state sample, and downsample the state data to obtain a global state sample having the same length as the real-time state sample; A real-time and global feature extraction module 220 is configured to input the real-time state samples and the global state samples into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, wherein the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and wherein the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The state prediction module 230 is used to fuse the real-time features and global features by using feature dot product to obtain fused features, and then use a fully connected network to obtain health prediction results of complex equipment based on the fused features.
[0073] Regarding the specific limitations of the complex equipment health status prediction device based on adaptive position coding, please refer to the limitations of the complex equipment health status prediction method based on adaptive position coding above, which will not be repeated here. The various modules in the above-mentioned complex equipment health status prediction device based on adaptive position coding can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0074] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for predicting the health status of complex equipment based on adaptive position coding is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0075] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.
[0076] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
[0077] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
[0078] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0079] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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.
[0080] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.
Claims
1. A method for predicting the health status of complex equipment based on adaptive position coding, characterized in that: The method comprises: Acquire observation data related to the device to be used for health prediction, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; Preprocessing the observation data to obtain state data, selecting a segment of data close to the current moment from the state data as a real-time state sample, and downsampling the state data to obtain a global state sample having the same length as the real-time state sample; Inputting the real-time state sample and the global state sample into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The real-time features and global features are fused by feature dot product to obtain fused features, and then a fully connected network is used to obtain the health prediction results of complex equipment based on the fused features.
2. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 1, characterized in that: Preprocessing the observation data includes: After compensating and standardizing the observation data, preliminary processed data are obtained; After filtering the preliminary processed data with a low-pass filter, the filtered data is enhanced by taking the difference between the filtered data and the original data to obtain the status data.
3. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 1, characterized in that: Preprocessing the observation data includes: After compensating and standardizing the observation data, preliminary processed data are obtained; After filtering the preliminary processed data with a low-pass filter, the filtered data is enhanced by taking the difference between the filtered data and the original data to obtain the status data.
4. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 3, characterized in that: When the real-time feature extraction unit is used to extract features based on the real-time state sample: In the state feature extraction branch, a temporal convolution unit including multiple temporal convolution layers is used to perform multi-level feature extraction on the real-time state sample, and then a fully connected unit is used to generate real-time state features based on the features corresponding to the latest moment in the deep-level features; In the state gradient feature extraction branch, the difference matrix construction unit is used to construct a corresponding real-time difference matrix based on the real-time state sample, the 2D convolution unit is used to perform multi-level feature extraction on the real-time difference matrix, the extracted features are subjected to global mean pooling using the global mean pooling unit, and expanded into feature vectors, and the fully connected unit obtains real-time differential features based on the feature vectors; The real-time status feature is combined with the real-time differential feature to obtain the real-time feature.
5. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 4, characterized in that: The difference matrix is constructed using the following formula: In the above formula, Indicates the real-time status sample j The difference matrix corresponding to the sensor variables, express Middle j The sensor collects m signal value.
6. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 5, characterized in that: In the global feature extraction unit: respectively adopting a state feature extraction branch and a state gradient feature extraction branch to extract a global state feature and a global state difference matrix of the global state sample; Constructing a time-coded feature sequence according to the sampling time series of the observation data, the global state difference matrix, and the global state sample; Based on two temporal convolutional layers, a learnable description model is constructed, and the temporal encoding feature sequence is used as input of the learnable description model to obtain a first attention vector and a second attention vector; Adopting a channel attention mechanism, using the first attention vector to select important variables in the global state sample to obtain a weighted global sample, constructing a weighted global state difference matrix based on the weighted global sample, and then calculating the weighted global state difference feature according to the weighted global state difference matrix; The second attention vector is combined with the weighted global state differential feature to obtain a combined feature, a mixed attention vector is calculated based on the combined feature, and the global state feature is weighted according to the mixed attention vector through the channel attention mechanism to obtain the final global feature.
7. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 6, characterized in that: The time-coded feature sequence is expressed as: In the above formula, and Calculated based on the global state difference matrix and the global state sample, k = 1, 2,…, d p / 2, d p The feature dimension representing the time-coded feature sequence.
8. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 7, characterized in that: When training the real-time feature extraction unit, the global feature extraction unit, and the fully connected network, a prediction loss function and a physical constraint loss function are used.
9. The method for predicting the health status of complex equipment based on adaptive position coding according to claim 7, characterized in that: The physical constraint loss function is expressed as: in, In the above formula, represents an identity matrix, and Represent the sample prediction results and actual health status of a training batch respectively, and yes and The corresponding elements, Represents the batch that meets the conditions The number of 10. A complex equipment health status prediction device based on adaptive position coding, characterized in that: The device comprises: An observation data acquisition module is used to acquire observation data related to the device to be health predicted, wherein the observation data is multi-dimensional data obtained by monitoring multiple sensors and arranged in chronological order; A data preprocessing module is used to preprocess the observation data to obtain state data, select a segment of data close to the current time in the state data as a real-time state sample, and downsample the state data to obtain a global state sample with the same length as the real-time state sample; A real-time and global feature extraction module, configured to input the real-time state samples and the global state samples into a real-time feature extraction unit and a global feature extraction unit, respectively, to obtain real-time features and global features, wherein both the real-time feature extraction unit and the global feature extraction unit adopt a "state-state gradient" dual-stream feature extraction network, which includes a state feature extraction branch and a state gradient feature extraction branch, wherein the state feature extraction branch includes a temporal convolution unit and a fully connected unit, and wherein the state gradient feature extraction branch includes a differential matrix construction unit, a 2D convolution unit, a global mean pooling unit, and a fully connected unit connected in sequence; The state prediction module is used to fuse the real-time features and global features by using feature dot product to obtain fused features, and then use a fully connected network to obtain the health prediction results of complex equipment based on the fused features.