Electronic oil pump health state and residual life prediction method based on cloud edge-end cooperation

Abstract

The present application relates to the technical field of intelligent manufacturing and equipment health management, and particularly relates to an electronic oil pump health state and residual life prediction method based on cloud edge end cooperation, comprising: S1: training a degradation perception composite coding network and a channel adaptive time series prediction network through a cloud end, and delivering to an edge node; S2: collecting vibration acceleration signals of a target component X axis and Y axis through a device end, and transmitting to the edge node; S3: outputting health indexes of each time step of the target component through the degradation perception composite coding network of the edge node; S4: outputting health index prediction values of future several time steps of the target component through the channel adaptive time series prediction network of the edge node as residual life prediction results; and S5: adjusting production task allocation of an electronic oil pump production line based on residual life prediction results of each target component on the electronic oil pump production line. The present application realizes high-precision residual life prediction and dynamic optimization scheduling of production tasks.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and equipment health management technology, specifically to a method for predicting the health status and remaining life of electronic oil pumps based on cloud-edge-device collaboration. Background Technology

[0002] As a core component of high-precision manufacturing equipment, the production of electronic oil pumps involves multiple stakeholders, multi-task collaboration, and autonomous execution processes. On the assembly and testing production line of electronic oil pumps, a large number of rotating bearing components operate under variable loads and high temperatures for extended periods, and their health status directly affects production cycle time, product quality, and equipment safety.

[0003] Currently, the manufacturing process of electronic oil pumps generally suffers from low levels of intelligent collaboration between the cloud, edge, and device: centralized processing in the cloud leads to large data transmission latency and poor real-time response; edge nodes lack effective lightweight health assessment models; and raw vibration data from the equipment cannot be efficiently utilized. Therefore, building a cloud-edge-device collaborative health management architecture that supports real-time data edge execution feedback has become crucial for improving the production and operation efficiency of electronic oil pumps.

[0004] In current technologies, research on cloud-edge-device collaboration for equipment health management is still in its early stages. Some literature has proposed equipment status monitoring schemes based on cloud platforms, but most adopt an "edge acquisition-cloud processing" model, where edge nodes only forward data and lack independent reasoning capabilities. For example, Li et al. proposed a bearing life prediction method based on a cloud server in their paper "RUL prediction of rollingbearings across working conditions based on multi-scale convolutional parallel memory domain adaptation network". All vibration data is uploaded to the cloud for feature extraction and model inference, resulting in high network bandwidth consumption and response latency exceeding seconds, which is difficult to meet the real-time early warning requirements of production lines.

[0005] In constructing health indicators, existing technologies (such as deep autoencoder methods like DCAE and CAE) typically rely on manual feature engineering or indirect signal transformation. For example, the quadratic function deep convolutional autoencoder proposed by Chen et al. in their paper "A novel deep learning method based on attention mechanism for bearing remaining useful life prediction" requires pre-calculation of statistical quantities such as root mean square and kurtosis of the vibration signal as input. When operating conditions change, the effectiveness of the features decreases, and it does not embed physical degradation priors. The constructed health indicators often exhibit non-monotonic fluctuations, which do not match the actual wear process.

[0006] In remaining lifetime prediction, existing technologies widely employ recurrent neural networks such as LSTM and BiLSTM. For example, the BiLSTM-based prediction method proposed by Huang et al. in their paper "A bidirectional LSTM prognostics method under multiple operational conditions" uses the same nonlinear transformation for all feature channels, which cannot adapt to the different degradation rates between different channels, resulting in underfitting of fast degradation channels and overfitting of slow degradation channels. In addition, the evaluation of existing prediction models focuses mainly on error metrics (RMSE, MAE), with insufficient attention to goodness of fit (R²), making it difficult to comprehensively assess the degree of agreement between the predicted curve and the actual degradation trajectory. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration. On the basis of cloud-edge-device collaboration, a monotonic health index is constructed using a degradation-aware composite coding network with physical constraints, and a channel-adaptive temporal prediction network is combined to accurately fit the nonlinear degradation trend, thereby achieving high-precision remaining life prediction and dynamic optimization scheduling of production tasks.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A cloud-edge-device collaborative method for predicting the health status and remaining life of electronic oil pumps includes:

[0010] S1: Train a degradation-aware composite coding network for extracting health indicators and a channel-adaptive temporal prediction network for predicting remaining lifespan in the cloud, and then distribute the results to edge nodes.

[0011] S2: Collect vibration acceleration signals of the target components on the X and Y axes of the electronic oil pump production line through the equipment end and transmit them to the edge node;

[0012] S3: By inputting the vibration acceleration signals of the target component on the X-axis and Y-axis at several time steps into the trained degradation sensing composite coding network through the edge nodes, the network outputs the health indicators of the target component at each time step.

[0013] S4: Arrange the health indicators of the target component at each time step according to the time step sequence through the edge nodes to obtain the health indicator sequence, and input the health indicator sequence into the trained channel adaptive temporal prediction network to output the predicted health indicators of the target component at several future time steps as the remaining life prediction result.

[0014] S5: Adjust the production task allocation of the electronic oil pump production line based on the remaining life prediction results of each target component on the electronic oil pump production line.

[0015] Preferably, in step S2, the vibration acceleration signals of the target component on the X-axis and Y-axis of the electronic oil pump production line are collected by the triaxial accelerometer at the equipment end, and the vibration acceleration signals of the target component on the X-axis and Y-axis are preprocessed including low-pass filtering, high-frequency noise removal and data compression.

[0016] Preferably, in step S3, the processing steps of the degradation-aware composite coding network on the edge nodes include:

[0017] S301: Input the vibration acceleration signals of the target component on the X-axis and Y-axis at the corresponding time step into the two independent branches of the degradation sensing composite coding network respectively;

[0018] S302: Two independent branches sequentially extract deep temporal features and degradation modes from the corresponding vibration acceleration signals through a one-dimensional convolutional layer, a Transformer layer, a global average pooling layer, and a fully connected layer to obtain the X-axis health index and the Y-axis health index.

[0019] S303: The X-axis health index and the Y-axis health index are fused to obtain the health index for the corresponding time step.

[0020] Preferably, in step S302, the processing steps for a single independent branch include:

[0021] S3021: Acquire the single-axis vibration acceleration signal of the independent branch input;

[0022] S3022: Perform one-dimensional convolution and ReLU activation on the uniaxial vibration acceleration signal to obtain a convolution feature map;

[0023] The formula is expressed as:

[0024] ;

[0025] In the formula: For convolutional feature maps; It is a uniaxial vibration acceleration signal; These are one-dimensional convolution kernel weights; This is the convolution bias term; It is the ReLU activation function;

[0026] S3023: Convolutional feature maps After layer normalization, a linear positional coding matrix is ​​then superimposed. This yields the normalized features after superimposed position encoding;

[0027] The formula is expressed as:

[0028] ;

[0029] In the formula: These are normalized features after overlaying positional encoding; For layer normalization; It is a linear positional encoding matrix;

[0030] S3024: Normalize features The long-range temporal dependencies within a single-axis vibration acceleration signal are modeled using the Transformer layer to obtain temporal features. ;

[0031] S3025: Timing characteristics Perform global average pooling and full connectivity to obtain the corresponding single-axis health metrics;

[0032] The formula is expressed as:

[0033] ;

[0034] In the formula: It is a single-axis health indicator; Temporal characteristics The Middle Features of each time step; This is the result of global average pooling along the time dimension. Temporal characteristics Number of time steps; These are the weights for the fully connected layer; This is used for biasing the fully connected layer.

[0035] Preferably, in step S303, the formula for calculating the health indicators is:

[0036] ;

[0037] In the formula: The integrated health indicators; For X-axis health indicators; For Y-axis health indicators; and These are learnable fusion weights.

[0038] Preferably, in step S4, the processing steps of the channel adaptive temporal prediction network on the edge node include:

[0039] S401: Obtain the health indicator sequence of the target component arranged according to the time step sequence. , For the first Health indicators at each time step The number of time steps in the health indicator sequence;

[0040] S402: Extract a fixed-length data segment from the health indicator sequence using a sliding window as a prediction data window;

[0041] S403: Input the prediction data window into a six-layer bidirectional long short-term neural network to extract deep temporal features containing past information and future context information, and input the deep temporal features into the DropOut layer to obtain a bidirectional temporal feature vector;

[0042] S404: Input the bidirectional time-series feature vector into a three-layer fully connected layer in series to obtain the predicted health index values ​​of the target component at several future time steps.

[0043] Preferably, in step S404, each of the three fully connected layers includes a linear connected layer for linear mapping and a channel adaptive activation function for nonlinear activation.

[0044] The first fully connected layer performs the first-level transformation on the bidirectional temporal feature vector to obtain the first-level activation output;

[0045] The formula is expressed as:

[0046] ;

[0047] In the formula: Activate the output for the first layer; This is the channel adaptive activation function for the first fully connected layer; It is a bidirectional temporal feature vector; This is the weight matrix of the first fully connected layer; The bias is set for the first fully connected layer;

[0048] The second fully connected layer performs a second-level transformation on the activation output of the first layer to obtain the activation output of the second layer.

[0049] The formula is expressed as:

[0050] ;

[0051] In the formula: Activate the output for the second layer; This is the channel-adaptive activation function for the second fully connected layer; This is the weight matrix for the second fully connected layer; This is the bias for the second fully connected layer;

[0052] The third fully connected layer maps the activation output of the second layer to obtain the predicted values ​​of health indicators for several future time steps.

[0053] The formula is expressed as:

[0054] ;

[0055] In the formula: For the future Predicted values ​​of health indicators at each time step; This is the channel-adaptive activation function for the third fully connected layer. This is the weight matrix of the third fully connected layer; This is the bias for the third fully connected layer.

[0056] Preferably, in step S404, the formula for calculating the channel adaptive activation function is:

[0057] ;

[0058] In the formula: For the first The activation output value of each feature channel. For feature channel index; For the first Input activation values ​​for each feature channel; For the first Slope adjustment factor for each feature channel; For the first Scale scaling factor for each feature channel; For the first Translation bias factor for each feature channel; It is the hyperbolic tangent function.

[0059] Preferred options also include:

[0060] S6: Using edge nodes, the vibration acceleration signals of the X and Y axes, along with the corresponding health indicator sequences and predicted health indicator values, are treated as a new set of data. The confidence level of the predicted health indicator values ​​is calculated, and when the confidence level of the predicted health indicator values ​​is higher than a preset threshold, they are marked as pseudo-labels. New data containing pseudo-labels are periodically uploaded to the cloud.

[0061] S7: Incremental training of the degradation-aware composite coding network and the channel adaptive temporal prediction network is performed on the cloud based on new data. The model parameters of the degradation-aware composite coding network and the channel adaptive temporal prediction network are updated and distributed to the edge nodes.

[0062] Preferably, in step S7, the degenerate sensing composite coding network is incrementally trained using a combined loss function that incorporates physical monotonicity.

[0063] The formula for calculating the combined loss function is:

[0064] ;

[0065] In the formula: This represents the total loss value of the combined loss function; Weighting coefficients for data fidelity items; The weight coefficient for the monotonicity penalty term; Losses due to data fidelity issues; Loss due to monotonicity penalty term;

[0066] The formula for calculating the loss of data fidelity items is:

[0067] ;

[0068] In the formula: The total number of time steps in the health indicator sequence; For the first Health indicators at each time step; For the first To the Changes in health indicators at each time step; This is a smoothness penalty term for the health indicator sequence; The regularization coefficient is used. This is a logarithmic magnitude regularization term;

[0069] The formula for calculating the loss of the monotonicity penalty term is:

[0070] ;

[0071] In the formula: This is the tolerance threshold; For the first To the Changes in health indicators at each time step; This is the function for finding the maximum value.

[0072] Compared with existing technologies, the cloud-edge-device collaborative method for predicting the health status and remaining life of electronic oil pumps in this invention has the following advantages:

[0073] This invention directly uses raw multi-axis vibration acceleration signals as model input, abandoning the traditional method's reliance on manual feature extraction and selection. Through an end-to-end data-driven approach, it maximizes the preservation of deep temporal features and subtle information related to equipment degradation in the vibration signals, avoiding information loss due to improper manual feature selection. Furthermore, by automatically learning and extracting these high-dimensional features through a subsequent degradation-aware composite coding network, it can construct health indicators that more accurately reflect the physical degradation process of components. This lays a solid data foundation for subsequent high-precision remaining life prediction, thereby improving the overall performance ceiling of the prediction model.

[0074] This invention utilizes a degradation-aware composite coding network to construct high-quality health indicators. By introducing a combined loss function that incorporates physical monotonicity during model training, a strict monotonically decreasing constraint is imposed on the health indicators output by the network. This physical information-guided deep learning strategy effectively solves the problems of large fluctuations and unclear trends in health indicators constructed by traditional data-driven methods. As a result, the final health indicators exhibit excellent monotonicity and robustness, significantly improving the accuracy and reliability of health status assessment.

[0075] This invention employs a channel-adaptive temporal prediction network for remaining lifetime prediction. It innovatively introduces a channel-adaptive activation function (CA-Tanh), which assigns independently learnable parameters to each feature channel. This allows the model to adaptively adjust the response strength and pattern to different degradation features, significantly enhancing the network's ability to fit complex, nonlinear, and heterogeneous degradation patterns. Compared to traditional prediction models, this invention can more accurately capture the long-term dependencies and evolutionary trends of health indicator sequences, thereby making more precise predictions of future degradation paths. Ultimately, this results in remaining lifetime predictions with extremely high goodness of fit and prediction accuracy.

[0076] This invention directly applies high-precision remaining life prediction results to production scheduling decisions. By using the predicted health status as a scheduling constraint, the manufacturing task scheduler can intelligently transfer tasks from workstations with poor health status to backup workstations with better health status before component failure occurs. Through a dynamic scheduling strategy based on the actual health status of equipment, it can not only effectively avoid production line downtime caused by sudden failures and ensure the continuity and stability of production tasks, but also maximize equipment utilization while minimizing task delay costs. This achieves synergistic optimization of equipment maintenance and production scheduling, improving the overall intelligence level and operational efficiency of the production line.

[0077] This invention constructs a closed-loop self-evolving system with cloud-edge collaboration. Edge nodes feed back new data generated during actual operation to the cloud. The cloud uses this incremental data to continuously train and update the parameters of the core models (DAC-Net and CAT-Net), enabling the models to continuously learn and adapt to new degradation patterns that occur under different operating conditions and at different lifecycle stages. This effectively overcomes the performance degradation problem caused by data distribution drift in traditional static models. Furthermore, through online iterative optimization of the models, the accuracy and goodness of fit of health status assessment and remaining life prediction are ensured to remain at a high level over a long period, giving the entire system the ability to continuously learn and self-improve. Attached Figure Description

[0078] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0079] Figure 1 This is a logic block diagram of a cloud-edge-device collaborative method for predicting the health status and remaining life of an electronic oil pump.

[0080] Figure 2 This is a structural diagram of a degenerate sensing composite coding network.

[0081] Figure 3 The structure diagram of the channel adaptive temporal prediction network is shown.

[0082] Figure 4 Health metrics built for different tasks: Figure 4 (a) to Figure 4 (e) represents the health indicators of bearings 1_1, 1_3, 2_5, 2_6, and 3_3 in the FEMTO-ST dataset. Figure 4 (f) to Figure 4 (h) represents the health indicators of bearings 1_3, 2_5, and 3_4 in the XJTU dataset.

[0083] Figure 5Predicting health indicators for different tasks: Figure 5 (a) to Figure 5 (e) represents the predicted health indicators for bearings 1_1, 1_3, 2_5, 2_6, and 3_3 in the FEMTO-ST dataset. Figure 5 (f) to Figure 5 (h) represents the predicted health indicators for bearings 1_3, 2_5, and 3_4 in the XJTU dataset. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0085] The following detailed explanation illustrates the specific implementation methods:

[0086] Example:

[0087] This embodiment discloses a method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration.

[0088] like Figure 1 As shown, the method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration includes:

[0089] S1: Train the Degradation-Aware Composite Coding Network (DAC-Net) for extracting health indicators and the Channel Adaptive Temporal Prediction Network (CAT-Net) for predicting remaining lifespan in the cloud, and then distribute the training to the edge nodes.

[0090] S2: Collect vibration acceleration signals of the target components on the X and Y axes of the electronic oil pump production line through the equipment end and transmit them to the edge node;

[0091] S3: By inputting the vibration acceleration signals of the target component on the X-axis and Y-axis at several time steps into the trained degradation sensing composite coding network through the edge nodes, the network outputs the health indicators of the target component at each time step.

[0092] S4: Arrange the health indicators of the target component at each time step according to the time step sequence through the edge nodes to obtain the health indicator sequence, and input the health indicator sequence into the trained channel adaptive temporal prediction network to output the predicted health indicators of the target component at several future time steps as the remaining life prediction result.

[0093] S5: Adjust the production task allocation of the electronic oil pump production line based on the remaining life prediction results of each target component on the electronic oil pump production line.

[0094] In this embodiment, the predicted health indicators (future P-step trends) output by the edge nodes for several future time periods are reported in real time to the manufacturing task scheduler in the cloud. The scheduler dynamically adjusts the task allocation of the electronic oil pump production line: when the predicted remaining life of a bearing at a certain workstation is lower than a set threshold, its task is automatically transferred to a standby workstation with better health, and a maintenance work order is triggered simultaneously. The scheduling objective function is to minimize task delay costs and maximize equipment utilization, and the constraints include task priority, process sequence, and equipment health status.

[0095] S6: Using edge nodes, the vibration acceleration signals of the X and Y axes, along with the corresponding health indicator sequences and predicted health indicator values, are treated as a new set of data. The confidence level of the predicted health indicator values ​​is calculated, and when the confidence level of the predicted health indicator values ​​is higher than a preset threshold, they are marked as pseudo-labels. New data containing pseudo-labels are periodically uploaded to the cloud.

[0096] S7: Incremental training of the degradation-aware composite coding network and the channel adaptive temporal prediction network is performed on the cloud based on new data. The model parameters of the degradation-aware composite coding network and the channel adaptive temporal prediction network are updated and distributed to the edge nodes.

[0097] Specifically, the cloud-based system updates the degradation-aware composite coding network using vibration acceleration signals (original vibration signals) along the X and Y axes and a linear degradation benchmark (ensuring feature extraction capabilities do not deviate from physical laws). Pseudo-labels are used as supervisory signals to update the channel-adaptive temporal prediction network (ensuring trend prediction capabilities adapt to new operating conditions).

[0098] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.

[0099] I. Equipment side

[0100] In the specific implementation process, the vibration acceleration signals of the target component on the electronic oil pump production line are collected in real time at a sampling frequency of 25.6kHz by the triaxial accelerometer at the equipment end. The vibration acceleration signals of the X-axis and Y-axis (that is, the X-axis and Y-axis here correspond to the X-axis and Y-axis of the triaxial accelerometer) are then preprocessed, including low-pass filtering, high-frequency noise removal and (lightweight) data compression.

[0101] II. Edge Nodes

[0102] In practice, edge nodes deploy pre-trained DAC-Net and CAT-Net to receive and process vibration acceleration signals from the device, calculate health indicators online, and predict future degradation trends. When the predicted value falls below a safety threshold, the edge nodes directly issue warnings or shutdown commands to the actuators.

[0103] like Figure 2 As shown, the processing steps of the degradation-aware composite coding network on edge nodes include:

[0104] S301: Input the vibration acceleration signals of the target component on the X-axis and Y-axis at the corresponding time step into the two independent branches of the degradation sensing composite coding network respectively;

[0105] S302: Two independent branches sequentially extract deep temporal features and degradation modes from the corresponding vibration acceleration signals through a one-dimensional convolutional layer, a Transformer layer, a global average pooling layer, and a fully connected layer to obtain the X-axis health index and the Y-axis health index.

[0106] In step S302, the processing steps for a single independent branch include:

[0107] S3021: Acquire single-axis (X-axis and Y-axis) vibration acceleration signals from independent branch inputs;

[0108] S3022: Perform one-dimensional convolution and ReLU activation on the uniaxial vibration acceleration signal to obtain a convolution feature map;

[0109] The formula is expressed as:

[0110] ;

[0111] In the formula: For convolutional feature maps; It is a uniaxial vibration acceleration signal; These are one-dimensional convolution kernel weights; This is the convolution bias term; It is the ReLU activation function;

[0112] S3023: Convolutional feature maps After layer normalization, a linear positional coding matrix is ​​then superimposed. The normalized features after superimposed positional encoding are obtained; where the linear positional encoding matrix is... The Middle row value .

[0113] The formula is expressed as:

[0114] ;

[0115] In the formula: These are normalized features after overlaying positional encoding; For layer normalization, the mean and variance of each channel of each feature are calculated independently and then standardized. It is a linear positional encoding matrix;

[0116] S3024: Normalize features The long-range temporal dependencies within a single-axis vibration acceleration signal are modeled using the Transformer layer to obtain temporal features. ;

[0117] S3025: Timing characteristics Perform global average pooling and full connectivity to obtain the corresponding single-axis health metrics;

[0118] The formula is expressed as:

[0119] ;

[0120] In the formula: This is a single-axis (X-axis and Y-axis) health indicator, with a value range of [0,1]. The closer it is to 0, the more severe the degradation. Temporal characteristics The Middle Features of each time step; This is the result of global average pooling along the time dimension. Temporal characteristics Number of time steps; These are the weights for the fully connected layer; This is used for biasing the fully connected layer.

[0121] It should be noted that the global average pooling operation in this invention targets the temporal features after processing by a one-dimensional convolutional layer and a Transformer layer. That is, before entering the global average pooling, the original vibration acceleration signal has been processed by extracting local temporal patterns through a one-dimensional convolutional layer and explicitly modeling the long-distance temporal dependencies within the single-axis vibration acceleration signal through the self-attention mechanism of the Transformer layer.

[0122] The core function of the Transformer layer is to capture the dependencies and sequence information between time steps in temporal data. Therefore, by the time the data reaches the global average pooling layer, the temporal sequence information has already been fully encoded into the temporal features through the self-attention mechanism of the Transformer layer. In this context, global average pooling is used to reduce the dimensionality of features that already contain temporal information, rather than ignoring the temporal order at the level of the original vibration acceleration signal.

[0123] S303: The X-axis health index and the Y-axis health index are fused to obtain the health index for the corresponding time step.

[0124] In step S303, the formula for calculating health indicators is:

[0125] ;

[0126] In the formula: The integrated health indicators; For X-axis health indicators; For Y-axis health indicators; and For learnable fusion weights, .

[0127] It should be noted that, and The optimal weights are obtained automatically through network training, rather than pre-fixed empirical values. They are data-driven and adaptively learned during the training process. This learnable mechanism itself can capture the mutual influence between the X-axis and Y-axis.

[0128] Specifically, during the training phase of DAC-Net, a linearly degenerate benchmark encoding is introduced as a physically interpretable prior constraint to provide a supervisory signal for model training. The linearly degenerate benchmark encoding generates a linearly decreasing benchmark health curve under ideal conditions by processing the cumulative runtime of the training samples.

[0129] During training, the baseline curve is not used as a true label of equipment degradation, but rather as a trend guide. The model's optimization objective is twofold: firstly, through a data fidelity loss term, the network's output health indicators closely fit the actual equipment state fluctuations reflected by the vibration acceleration signal; secondly, through a monotonicity penalty loss term, the overall trend of the health indicators is constrained to remain consistent with the linear baseline curve, i.e., monotonically decreasing. In this way, the degradation-aware composite coding network is forced to learn a deep mapping relationship between vibration signal characteristics and the degree of equipment degradation, while ensuring that the extracted health indicators conform to the common sense of degradation in the physical world.

[0130] Therefore, during the model inference stage, the trained network already has the ability to map from "vibration signals" to "health indicators that conform to the degradation trend". It only needs to input vibration acceleration signals to output health indicators, without needing to input the cumulative running time.

[0131] like Figure 3 As shown, the processing steps of the channel adaptive temporal prediction network on the edge node include:

[0132] S401: Obtain the health indicator sequence of the target component arranged according to the time step sequence. , For the first Health indicators at each time step The number of time steps in the health indicator sequence;

[0133] S402: A fixed-length data segment is extracted from the health indicator sequence using a sliding window as the prediction data window; where the window length of the sliding window is... Predicting step size Forecast data window The corresponding predicted values ​​of health indicators ;

[0134] S403: Input the prediction data window into a six-layer bidirectional long short-term neural network to extract deep temporal features containing past information and future contextual information, and input the deep temporal features into the DropOut layer to obtain a bidirectional temporal feature vector containing rich contextual information;

[0135] S404: Input the bidirectional time-series feature vector into a three-layer fully connected layer in series to obtain the predicted health index values ​​of the target component at several future time steps.

[0136] In step S404, the first three fully connected layers each include a linear connected layer for linear mapping and a channel adaptive activation function for nonlinear activation.

[0137] The first fully connected layer performs the first-level transformation on the bidirectional temporal feature vector to obtain the first-level activation output;

[0138] The formula is expressed as:

[0139] ;

[0140] In the formula: Activate the output for the first layer; This is the channel adaptive activation function for the first fully connected layer; It is a bidirectional temporal feature vector; This is the weight matrix of the first fully connected layer; The bias is set for the first fully connected layer;

[0141] The second fully connected layer performs a second-level transformation on the activation output of the first layer to obtain the activation output of the second layer.

[0142] The formula is expressed as:

[0143] ;

[0144] In the formula: Activate the output for the second layer; This is the channel-adaptive activation function for the second fully connected layer; This is the weight matrix for the second fully connected layer; This is the bias for the second fully connected layer;

[0145] The third fully connected layer maps the activation output of the second layer to obtain the predicted values ​​of health indicators for several future time steps.

[0146] The formula is expressed as:

[0147] ;

[0148] In the formula: For the future Predicted values ​​of health indicators at each time step; This is the channel-adaptive activation function for the third fully connected layer. This is the weight matrix of the third fully connected layer; This is the bias for the third fully connected layer.

[0149] In step S404, the channel adaptive activation function is applied to each feature channel. Distribution slope ,scale Translation Three independent learnable parameters;

[0150] The formula for calculating the channel adaptive activation function is:

[0151] ;

[0152] In the formula: For the first The activation output value of each feature channel. This function indexes the feature channels and allows each channel to adaptively adjust its nonlinear saturation region to accommodate channels with different degradation rates. For the first Input activation values ​​for each feature channel; For the first The slope adjustment factor for each feature channel controls the nonlinear saturation. For the first Scale scaling factor for each feature channel; For the first Translation bias factor for each feature channel; It is the hyperbolic tangent function.

[0153] III. Cloud

[0154] In the specific implementation process, the cloud uses historical full life cycle datasets to train DAC-Net and CAT-Net offline; new data is collected periodically from edge nodes to trigger incremental learning and update model parameters; and the updated model is pushed to edge nodes through the model distribution channel. Specifically, the cloud will use historical full life cycle data to train DAC-Net with the original vibration signal as input and a monotonically decreasing health indicator as the optimization objective, and use the health indicator sequence output by DAC-Net as input and minimize the prediction error as the objective to train CAT-Net. After training, the model is distributed to the edge nodes.

[0155] Specifically, a combined loss function that incorporates physical monotonicity is used to incrementally train the degenerate sensing composite coding network;

[0156] The formula for calculating the combined loss function is:

[0157] ;

[0158] In the formula: This represents the total loss value of the combined loss function; Weighting coefficients for data fidelity items; The weight coefficient for the monotonicity penalty term; Losses due to data fidelity issues; Loss due to monotonicity penalty term;

[0159] The formula for calculating the loss of data fidelity items is:

[0160] ;

[0161] In the formula: The total number of time steps in the health indicator sequence; For the first Predictive health indicators at each time step; For the first To the Changes in health indicators at each time step; A smoothness penalty term is used for the health indicator sequence to suppress drastic fluctuations between adjacent steps; This is the regularization coefficient, which controls the weight of the logarithmic term; This is a logarithmic amplitude regularization term to prevent health indicators from being too small and causing numerical instability.

[0162] The formula for calculating the loss of the monotonicity penalty term is:

[0163] ;

[0164] In the formula: Tolerance threshold, allowing for minor noise fluctuations; For the first To the Changes in health indicators at each time step; To maximize the function, only positive changes exceeding the tolerance threshold are penalized; This is a monotonic loss.

[0165] In this study, the mean square error function was used to incrementally train the channel adaptive time-series prediction network.

[0166] IV. Experimental Instructions

[0167] This embodiment was validated on the XJTU dataset and the FEMTO-ST dataset. The task partitioning of the datasets is shown in Tables 1 and 2.

[0168] Figure 4 and Figure 5 The invention presents the health metric construction curves and prediction results for all tasks using the method proposed in this invention.

[0169] Figure 4 Health metrics built for different tasks: Figure 4 (a) to Figure 4 (e) represents the health indicators of bearings 1_1, 1_3, 2_5, 2_6, and 3_3 in the FEMTO-ST dataset. Figure 4 (f) to Figure 4 (h) represents the health indicators of bearings 1_3, 2_5, and 3_4 in the XJTU dataset.

[0170] Figure 5 Predicting health indicators for different tasks: Figure 5 (a) to Figure 5 (e) represents the predicted health indicators for bearings 1_1, 1_3, 2_5, 2_6, and 3_3 in the FEMTO-ST dataset. Figure 5 (f) to Figure 5 (h) represents the predicted health indicators for bearings 1_3, 2_5, and 3_4 in the XJTU dataset.

[0171] Tables 3 and 5 present ablation experimental data for DAC-Net and CAT-Net, verifying the effectiveness of the improvements proposed in this invention. Tables 4 and 6 show the comparison results with existing methods. The monotonicity index of DAC-Net in this invention can approach 1.0, and the R2 index of CAT-Net is improved from 0.34 of the baseline BiLSTM to more than 0.99, with RMSE reduced by 1-2 orders of magnitude.

[0172] Table 1 Task partitioning of the XJTU-SY dataset

[0173]

[0174] Table 2 Task Division of FEMTO-ST Dataset

[0175]

[0176] Table 3. DAC-Net ablation experimental data

[0177]

[0178] Table 4 Comparison Experimental Data for DAC-Net

[0179]

[0180] Table 5. Experimental data of CA-Tanh ablation

[0181]

[0182] Table 6 Comparative Experimental Data of CA-Tanh Model

[0183]

[0184] The above experiments show that:

[0185] (1) The monotonicity and reliability of the health indicators in this invention are higher: Compared with autoencoder methods such as DCAE and CAE, this invention improves the monotonicity of the health indicators by more than 30% by using linear baseline encoding and monotonicity constraint loss to approach 1.0.

[0186] (2) This invention achieves a dual breakthrough in prediction accuracy and goodness of fit: Compared with baseline models such as BiLSTM and LSTM, the CA-Tanh activation function proposed in this invention improves R² from less than 0.6 to more than 0.95 and reduces RMSE by 1-2 orders of magnitude, proving that the prediction curve is highly consistent with the actual degradation trajectory.

[0187] (3) The present invention achieves deep integration of manufacturing scheduling and equipment health: Most existing methods separate equipment predictive maintenance from production scheduling. The present invention triggers the adjustment of scheduling strategy in real time through the health prediction value output by the edge node, forming a closed loop of "state perception-prediction-scheduling", which can effectively reduce unplanned downtime losses.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration, characterized in that, include: S1: Train a degradation-aware composite coding network for extracting health indicators and a channel-adaptive temporal prediction network for predicting remaining lifespan in the cloud, and then distribute the results to edge nodes. S2: Collect vibration acceleration signals of the target components on the X and Y axes of the electronic oil pump production line through the equipment end and transmit them to the edge node; S3: By inputting the vibration acceleration signals of the target component on the X-axis and Y-axis at several time steps into the trained degradation sensing composite coding network through the edge nodes, the network outputs the health indicators of the target component at each time step. S4: Arrange the health indicators of the target component at each time step according to the time step sequence through the edge nodes to obtain the health indicator sequence, and input the health indicator sequence into the trained channel adaptive temporal prediction network to output the predicted health indicators of the target component at several future time steps as the remaining life prediction result. S5: Adjust the production task allocation of the electronic oil pump production line based on the remaining life prediction results of each target component on the electronic oil pump production line.

2. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 1, characterized in that: In step S2, the vibration acceleration signals of the target component on the X-axis and Y-axis of the electronic oil pump production line are collected by the triaxial accelerometer on the equipment end, and the vibration acceleration signals of the target component on the X-axis and Y-axis are preprocessed, including low-pass filtering, high-frequency noise removal and data compression.

3. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 1, characterized in that: In step S3, the processing steps of the degradation-aware composite coding network on the edge nodes include: S301: Input the vibration acceleration signals of the target component on the X-axis and Y-axis at the corresponding time step into the two independent branches of the degradation sensing composite coding network respectively; S302: Two independent branches sequentially extract deep temporal features and degradation modes from the corresponding vibration acceleration signals through a one-dimensional convolutional layer, a Transformer layer, a global average pooling layer, and a fully connected layer to obtain the X-axis health index and the Y-axis health index. S303: The X-axis health index and the Y-axis health index are fused to obtain the health index for the corresponding time step.

4. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 3, characterized in that: In step S302, the processing steps for a single independent branch include: S3021: Acquire the single-axis vibration acceleration signal of the independent branch input; S3022: Perform one-dimensional convolution and ReLU activation on the uniaxial vibration acceleration signal to obtain a convolution feature map; The formula is expressed as: ； In the formula: For convolutional feature maps; It is a uniaxial vibration acceleration signal; These are one-dimensional convolution kernel weights; This is the convolution bias term; It is the ReLU activation function; S3023: Convolutional feature maps After layer normalization, a linear positional coding matrix is ​​then superimposed. This yields the normalized features after superimposed position encoding; The formula is expressed as: ； In the formula: These are normalized features after overlaying positional encoding; For layer normalization; It is a linear positional encoding matrix; S3024: Normalize features The long-range temporal dependencies within a single-axis vibration acceleration signal are modeled using the Transformer layer to obtain temporal features. ; S3025: Timing characteristics Perform global average pooling and full connectivity to obtain the corresponding single-axis health metrics; The formula is expressed as: ； In the formula: It is a single-axis health indicator; Temporal characteristics The Middle Features of each time step; This is the result of global average pooling along the time dimension. Temporal characteristics Number of time steps; These are the weights for the fully connected layer; This is used for biasing the fully connected layer.

5. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 4, characterized in that: In step S303, the formula for calculating health indicators is: ； In the formula: The integrated health indicators; For X-axis health indicators; For Y-axis health indicators; and These are learnable fusion weights.

6. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 4, characterized in that: In step S4, the processing steps of the channel adaptive temporal prediction network on the edge nodes include: S401: Obtain the health indicator sequence of the target component arranged according to the time step sequence. , For the first Health indicators at each time step The number of time steps in the health indicator sequence; S402: Extract a fixed-length data segment from the health indicator sequence using a sliding window as a prediction data window; S403: Input the prediction data window into a six-layer bidirectional long short-term neural network to extract deep temporal features containing past information and future context information, and input the deep temporal features into the DropOut layer to obtain a bidirectional temporal feature vector; S404: Input the bidirectional time-series feature vector into a three-layer fully connected layer in series to obtain the predicted health index values ​​of the target component at several future time steps.

7. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 6, characterized in that: In step S404, each of the three fully connected layers includes a linear connected layer for linear mapping and a channel adaptive activation function for nonlinear activation. The first fully connected layer performs the first-level transformation on the bidirectional temporal feature vector to obtain the first-level activation output; The formula is expressed as: ； In the formula: Activate the output for the first layer; This is the channel adaptive activation function for the first fully connected layer; It is a bidirectional temporal feature vector; This is the weight matrix of the first fully connected layer; The bias is set for the first fully connected layer; The second fully connected layer performs a second-level transformation on the activation output of the first layer to obtain the activation output of the second layer. The formula is expressed as: ； In the formula: Activate the output for the second layer; This is the channel-adaptive activation function for the second fully connected layer; This is the weight matrix for the second fully connected layer; This is the bias for the second fully connected layer; The third fully connected layer maps the activation output of the second layer to obtain the predicted values ​​of health indicators for several future time steps. The formula is expressed as: ； In the formula: For the future Predicted values ​​of health indicators at each time step; This is the channel-adaptive activation function for the third fully connected layer. This is the weight matrix of the third fully connected layer; This is the bias for the third fully connected layer.

8. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 7, characterized in that: In step S404, the formula for calculating the channel adaptive activation function is: ； In the formula: For the first The activation output value of each feature channel. For feature channel index; For the first Input activation values ​​for each feature channel; For the first Slope adjustment factor for each feature channel; For the first Scale scaling factor for each feature channel; For the first Translation bias factor for each feature channel; It is the hyperbolic tangent function.

9. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 1, characterized in that, Also includes: S6: Using edge nodes, the vibration acceleration signals of the X and Y axes, along with the corresponding health indicator sequences and predicted health indicator values, are treated as a new set of data. The confidence level of the predicted health indicator values ​​is calculated, and when the confidence level of the predicted health indicator values ​​is higher than a preset threshold, they are marked as pseudo-labels. New data containing pseudo-labels are periodically uploaded to the cloud. S7: Incremental training of the degradation-aware composite coding network and the channel adaptive temporal prediction network is performed on the cloud based on new data. The model parameters of the degradation-aware composite coding network and the channel adaptive temporal prediction network are updated and distributed to the edge nodes.

10. The method for predicting the health status and remaining life of an electronic oil pump based on cloud-edge-device collaboration as described in claim 9, characterized in that: In step S7, the degenerate sensing composite coding network is incrementally trained using a combined loss function that incorporates physical monotonicity. The formula for calculating the combined loss function is: ； In the formula: This represents the total loss value of the combined loss function; Weighting coefficients for data fidelity items; The weight coefficient for the monotonicity penalty term; Losses due to data fidelity issues; Loss due to monotonicity penalty term; The formula for calculating the loss of data fidelity items is: ； In the formula: The total number of time steps in the health indicator sequence; For the first Health indicators at each time step; For the first To the Changes in health indicators at each time step; This is a smoothness penalty term for the health indicator sequence; The regularization coefficient is used. This is a logarithmic magnitude regularization term; The formula for calculating the loss of the monotonicity penalty term is: ； In the formula: This is the tolerance threshold; For the first To the Changes in health indicators at each time step; This is the function for finding the maximum value.

Images