Intelligent weighing equipment residual life prediction method based on multi-stage wiener process

Abstract

This invention belongs to the field of remaining life prediction technology for intelligent weighing equipment, specifically involving a method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process. First, sensor data from the equipment is acquired and normalized to obtain normalized data. Second, a fusion decision framework combining mutual information and correlation analysis is constructed to filter effective sensor features from the normalized data and perform principal component analysis to construct a one-dimensional health index. Third, the following steps are executed in parallel: first, an LSTM network is constructed and trained; second, kernel density estimation is performed on the one-dimensional health index to obtain a comprehensive health index, which is then divided into degradation stages using the CUSUM algorithm to construct a multi-stage Wiener model to predict the degradation trajectory. Finally, the LSTM network is optimized based on the degradation trajectory predicted by the multi-stage Wiener model to predict the remaining life of the intelligent weighing equipment. Experimental results show that this method can effectively improve the prediction accuracy of the remaining life of intelligent weighing equipment, reducing the correlation error to below 5%.

Description

Technical Field

[0001] This invention relates to the field of intelligent weighing equipment remaining life prediction technology, and is particularly applicable to an intelligent weighing equipment remaining life prediction method based on a multi-stage Wiener process. Background Technology

[0002] With the rapid development of technology in various fields in my country, the intelligent upgrading of weighing equipment has played a vital role in improving production and living efficiency and ensuring the operational safety of key areas. However, the ever-increasing performance requirements have led to increasingly complex internal structures, and related components are often in complex and variable working environments, significantly increasing the risk of failure. Therefore, research on RUL (Remaining Useful Life) prediction technology for intelligent weighing equipment has significant theoretical and engineering value. In the field of remaining useful life prediction, the Wiener process is often used to describe the performance degradation of equipment due to its good applicability to non-monotonic degradation processes and strong model interpretability. At the same time, neural network methods, represented by LSTM (Long Short-Term Memory), with their powerful nonlinear modeling capabilities, adaptive feature learning advantages, and ability to capture long-term temporal dependencies, also provide an effective way to improve the accuracy and generalization performance of remaining useful life prediction.

[0003] The patent document with publication number "CN112683535B" discloses a "bearing life prediction method based on multi-stage Wiener process" for predicting the remaining life of bearings. However, since its model construction usually depends on a fixed multi-stage Wiener model, its generalization ability is limited, which restricts its applicability in complex environments.

[0004] The patent document with publication number "CN114297910B" discloses "a method for predicting the life of aero-engines based on an improved LSTM". It introduces SDAE (Sparse Deep AutoEncoder) for feature extraction and combines it with an LSTM model to predict the life of aero-engines. However, it is essentially a data-driven black box model. The extracted high-dimensional features lack a clear and traceable correlation with the physical degradation mechanism of the engine, and the model's decision-making logic is not easy to explain.

[0005] The problems with the aforementioned documents are as follows: First, although the model parameters (drift coefficient, diffusion coefficient, etc.) can be updated online, once the stage is determined, the mathematical model form is fixed. For equipment with large differences in degradation trajectories and variable operating conditions, the model's generalization ability is limited. Second, although deep learning methods such as LSTM have strong nonlinear fitting capabilities, their "black box" characteristics lead to poor model interpretability, resulting in prediction results lacking clear physical meaning support. Summary of the Invention

[0006] To address the issues of insufficient generalization ability and poor model interpretability in existing technologies, this invention proposes a method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process, aiming to solve the problem of difficulty in predicting the remaining life of intelligent weighing equipment.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process, comprising the following steps:

[0008] Step 1: Acquire the device sensor data and normalize it to obtain normalized data;

[0009] Step 2: Construct a decision-making framework that integrates mutual information and correlation analysis, select effective sensor features from normalized data, perform principal component analysis on them, and construct a one-dimensional health indicator.

[0010] Step 3: Perform the following steps in parallel: First, construct and train an LSTM network; second, perform kernel density estimation on the one-dimensional health index to obtain a comprehensive health index, divide its degradation stage using the CUSUM algorithm, and construct a multi-stage Wiener model to predict the degradation trajectory.

[0011] Step 4: Optimize the LSTM network based on the degradation trajectory predicted by the multi-stage Wiener model to predict the remaining lifespan of the intelligent weighing equipment.

[0012] Furthermore, the fusion decision framework in step two above involves performing a logical AND operation on the screening results of mutual information and correlation analysis, specifically:

[0013] （2）

[0014] in, It is the final set of sensors. It is a set of sensors selected through correlation. It is a set of sensors filtered through mutual information.

[0015] Furthermore, the LSTM network in step three above is a two-layer stacked LSTM network.

[0016] Furthermore, in step three above, the degradation stages are divided using the CUSUM algorithm to construct a multi-stage Wiener degradation model as follows:

[0017] （4）

[0018] In the above formula, , And so on. .

[0019] Furthermore, the specific sub-steps of using the multi-stage Wiener model to predict the degradation trajectory as a guide for optimizing the LSTM network in step four above are as follows:

[0020] (1) Initial input construction:

[0021] Set the input time step to 50% of the length of the comprehensive health index; predict the degradation trajectory based on the multi-stage Wiener model to obtain the Wiener sequence generated by the corresponding device; set the LSTM input format to meet the input requirements of LSTM;

[0022] (2) Iterative prediction loop:

[0023] The forecasting process employs an iterative rolling forecasting strategy, executing a complete forecast-update loop at each time step; for each forecast time step, the following operations are performed:

[0024] ①LSTM prediction: Input the current sequence into the LSTM network and output the predicted value for the next time step;

[0025] ② Input sequence update: Discard the oldest element in the sequence to make room for the new prediction; after the shift, the last position of the sequence is temporarily empty and waiting to be filled; fill the empty position with the degenerate trajectory prediction value generated by the multi-stage Wiener model to obtain the updated input sequence, which is used for LSTM prediction in the next prediction step.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention proposes a dual feature selection framework that combines mutual information and correlation analysis. It quantifies the nonlinear dependence of features on degradation states by setting a mutual information threshold, while simultaneously selecting features based on low correlation principles to suppress linear redundancy. A dual-path constraint mechanism is constructed, and the results are intersected. This mechanism effectively solves the problem of coordination between heterogeneous indicators. The extracted feature subset possesses complementary linear-nonlinear correlations, high robustness, and strong interpretability, enabling effective extraction of key degradation-sensitive features under complex operating conditions.

[0028] 2. Based on the comprehensive health indicators of intelligent weighing equipment, this invention proposes a multi-stage Wiener process model based on cumulative deviation. Employing a multi-stage degradation trajectory modeling method, it can accurately characterize the full-cycle degradation characteristics of the equipment from slow decline to accelerated failure. Compared to traditional single-stage or two-stage models, this method has a more realistic physical description of the degradation path, providing a more accurate degradation modeling foundation for life prediction.

[0029] 3. This invention designs a fusion mechanism that uses a multi-stage Wiener process to predict trajectories and guides an LSTM network. It employs a dynamic correction method for degenerate trajectories and constructs a double-layer stacked LSTM network to learn the long-term dependence and nonlinearity of sequences, adaptively adjusting the trajectory with fixed parameters. Compared to the multi-stage Wiener model, this mechanism enhances the model's generalization ability, adapts to individual differences and complex working conditions, and significantly improves the robustness of the prediction system.

[0030] 4. This invention proposes a deep fusion framework of multi-stage Wiener processes and LSTM networks, employing a design approach that synergistically enhances interpretability and generalization ability, thereby improving model adaptability while maintaining physical interpretability. Compared to the multi-stage Wiener model, this framework offers the combined advantages of high prediction accuracy and strong engineering applicability.

[0031] 5. This invention constructs an integrated technical framework from feature selection, degradation modeling to lifetime prediction. First, a dual constraint mechanism combining mutual information and correlation analysis is used to extract low-redundancy, highly sensitive degradation features. Then, a multi-stage Wiener model is constructed based on cumulative bias to characterize the full-cycle degradation trajectory while maintaining physical interpretability. Finally, a prediction mechanism is designed to guide the LSTM network with the multi-stage Wiener prediction trajectory, enhancing the model's generalization ability. This framework overcomes the challenge of coordination among heterogeneous indicators, effectively integrating physical mechanisms and data-driven learning, and significantly improving the accuracy and robustness of the lifetime prediction system. Experimental verification shows that this method can effectively improve the accuracy of remaining lifetime prediction for intelligent weighing equipment, reducing the correlation error to below 5%. Attached Figure Description

[0032] Figure 1 This is an overall flowchart of the present invention;

[0033] Figure 2 This is a diagram showing the effect of the method of the present invention on predicting the degradation trend of intelligent weighing equipment. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0035] See Figure 1The design concept of this invention is as follows: First, acquire and normalize the sensor data of the equipment to obtain normalized data. Second, construct a fusion decision framework of mutual information and correlation analysis, filter out effective sensor features from the normalized data, and perform principal component analysis to construct a one-dimensional health index. Third, execute the following steps in parallel: first, construct and train an LSTM network; second, perform kernel density estimation on the one-dimensional health index to obtain a comprehensive health index, divide its degradation stages using the CUSUM algorithm, and construct a multi-stage Wiener model to predict the degradation trajectory. Finally, optimize the LSTM network based on the degradation trajectory prediction of the multi-stage Wiener model to predict the remaining lifespan of the intelligent weighing equipment.

[0036] Example: A method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process, the specific implementation steps are as follows:

[0037] Step 1: Acquire and normalize the device sensor data to obtain normalized data. This includes the following sub-steps:

[0038] 1.1 Acquiring sensor data from the device:

[0039] The sensor data used in the experiment came from a self-built "Multi-channel Intelligent Weighing Equipment Accelerated Degradation Experimental Platform." This platform consists of 100 modified intelligent warehouse units, each equipped with multi-source monitoring sensors, including: 4 load cells, 2 temperature sensors (monitoring ambient and internal temperatures respectively), 1 current sensor, 1 voltage sensor, and 1 humidity sensor. The experiment simulated accelerated degradation in a real-world industrial setting by continuously operating under overload conditions, with the weight of the item exceeding 50% of its weighing capacity (e.g., a 2kg capacity with an item weight of 4kg). Over 200 days of continuous monitoring data were collected, forming a complete degradation dataset encompassing the entire lifecycle from normal to abnormal and failure states.

[0040] 1.2. Normalize the sensor data from the equipment to obtain normalized data:

[0041] The dataset used in this embodiment includes sensor data from multiple smart weighing devices. The status of each smart weighing device is monitored by multiple sensors. Since some sensor values ​​are constant and do not contain degradation information, only the sensors containing degradation information are processed subsequently.

[0042] The sensor data is normalized using the maximum-minimum method, and the corresponding formula is shown below:

[0043] （1）

[0044] In the formula Represents a set of data, Represents each data element therein. This represents the minimum value in the set of data. This represents the minimum value in the set of data. These are the normalized data values.

[0045] After normalization, multiple intelligent weighing devices were divided into training and testing sets in a 7:3 ratio. The training set was used during subsequent sensor selection to avoid data leakage from the testing set.

[0046] Step 2: Construct a fusion decision-making framework combining mutual information and correlation analysis. Select effective sensor features from normalized data and perform principal component analysis on them to construct a one-dimensional health indicator. This includes the following sub-steps:

[0047] 2.1 For the sensor data of the training set devices, construct a fusion decision framework based on mutual information and correlation analysis to filter out effective sensor features from the normalized data. The specific sub-steps are as follows:

[0048] 2.1.1 The mutual information method is used to filter the training set data.

[0049] To ensure that the selected sensors are directly related to the device degradation process, the mutual information method is first used to assess the statistical dependence of each sensor on the remaining lifespan of the device.

[0050] Mutual information is a measure of the interdependence between two variables; it represents the amount of information shared between them. The greater the mutual information between variables, the stronger their correlation. Selecting features with high mutual information with the target variable helps improve the predictive ability of the algorithm.

[0051] In the mutual information screening, the mean mutual information value of each sensor and remaining lifetime in the training set, 0.402, was used as the judgment threshold. The mean mutual information value reflects the overall information contribution level of the sensor to lifetime prediction. Sensors with a value higher than this threshold are considered to have a significant nonlinear correlation with the degradation process and are retained; sensors with a value lower than this threshold are considered to have a weak information contribution and are eliminated. By measuring the magnitude of the mutual information between different sensor data and remaining lifetime, sensor data with a more significant impact on remaining lifetime are selected to complete the subsequent work.

[0052] 2.1.2 The training set data was filtered using correlation analysis.

[0053] Considering that multicollinearity among sensors can lead to information redundancy and reduce model efficiency, Pearson correlation coefficient is further used to analyze the correlation between sensors.

[0054] The correlation coefficient is a measure of the closeness between different physical quantities. The Pearson correlation coefficient is the most commonly used correlation coefficient. It focuses on studying the linear correlation between variables and is characterized by multiplying the difference between each variable and its respective mean. The Pearson correlation coefficient ranges from -1 to 1, and the larger the absolute value, the stronger the correlation.

[0055] When using the Pearson correlation coefficient for sensor data screening, a threshold of 0.7 is set. A correlation coefficient greater than 0.7 indicates a strong correlation between variables, potentially introducing information redundancy. In this case, the sensor with the largest variance in the original data is retained, as signals with larger variance typically contain richer dynamic information. By measuring the correlation between different sensor data, sensor data with low correlation are selected to construct a subset of sensors with low redundancy and high information content.

[0056] 2.1.3 Determination of the optimal sensor set under dual constraints:

[0057] To simultaneously meet the dual requirements of feature independence and target relevance, this embodiment adopts a hybrid screening strategy, performing a logical AND operation on the screening results of 2.1.1 and 2.1.2.

[0058] （2）

[0059] in, It is the final set of sensors. It is a set of sensors selected through correlation screening (ensuring low redundancy). It is a set of sensors selected through mutual information (ensuring high correlation).

[0060] This invention proposes a dual-constraint feature selection mechanism. By implementing dual screening of sensor data with correlation constraints and mutual information constraints, this mechanism ensures that the final sensor set simultaneously possesses the following characteristics: low correlation between sensors to avoid information duplication; strong correlation between sensors and remaining lifetime to ensure degradation characterization capability; and a balance between information retention and dimensionality reduction.

[0061] Based on this dual constraint mechanism, the method of this invention comprehensively selects the final effective sensor set by fusing correlation analysis between sensors with mutual information assessment of sensors and remaining lifetime. This method retains sensor data closely related to equipment degradation while eliminating redundant and irrelevant sensor data, thereby improving the efficiency of subsequent analysis while ensuring the accuracy of health status characterization. The selected sensor data will provide a high-quality and representative input foundation for subsequent principal component analysis dimensionality reduction, health indicator construction, and remaining lifetime prediction models.

[0062] 2.2 Principal component analysis was used to reduce the dimensionality of the screened multidimensional data to obtain a one-dimensional health indicator:

[0063] Principal component analysis (PCA) is an effective representation learning method, primarily used to reduce the computational cost of machine learning models by reducing correlated features. This method first calculates the eigenvalues ​​and eigenvectors of the data covariance matrix, sorts them according to eigenvalue size, and selects the eigenvectors corresponding to the k largest eigenvalues ​​to construct a projection matrix, thereby compressing the data from a high-dimensional space to a low-dimensional subspace.

[0064] In actual operation, intelligent weighing equipment generates high-dimensional feature data every second, resulting in a surge in the dimensionality of the machine learning feature matrix, necessitating dimensionality reduction. Principal component analysis (PCA) can significantly reduce dimensionality while preserving the main effective information, making it particularly suitable for feature compression of sensor data.

[0065] The cumulative contribution rate of the principal components is set to 85%. The principal components are obtained and fused to obtain the one-dimensional health index of each device, which is used for the construction of the subsequent multi-stage Wiener model and the training of the LSTM network.

[0066] Step 3: Perform the following steps in parallel: First, construct and train an LSTM network; second, perform kernel density estimation on the one-dimensional health index to obtain a comprehensive health index, divide its degradation stage using the CUSUM algorithm, and construct a multi-stage Wiener model to predict the degradation trajectory.

[0067] 3.1 Construct an LSTM network and train it using the training set.

[0068] LSTM neural networks are a special type of recurrent neural network that introduces three gating mechanisms: forget gate, input gate, and output gate. This overcomes the long-term dependence of the model on data while preserving the influence of previous input features on the output results, thus avoiding gradient vanishing and gradient exploding. Therefore, it has strong predictive power.

[0069] This embodiment constructs a two-layer stacked LSTM network, with the input being a one-dimensional health indicator sequence obtained through principal component analysis. The LSTM part contains two hidden layers, each with a hidden state dimension of 64. A random deactivation mechanism with a dropout rate of 0.2 is used between layers to prevent overfitting. Training employs the mean squared error loss function and the Adam optimizer, with an initial learning rate of 0.001, which is dynamically adjusted. During training, the batch size is set to 32, the maximum number of training epochs is 100, and an early stopping mechanism is introduced: if the validation set loss does not decrease for 20 consecutive epochs, training is terminated early.

[0070] 3.2 The distribution density of one-dimensional health indicators is estimated using the kernel density estimation method to obtain a comprehensive health indicator. The degradation stages are then divided using the CUSUM algorithm, and a multi-stage Wiener model is constructed to predict the degradation trajectory.

[0071] 3.2.1 Obtain the comprehensive health index of the training set using the kernel density estimation method.

[0072] Kernel density estimation is a nonparametric statistical method used to estimate unknown probability density functions. It generates a continuous, smooth density curve by placing a smooth kernel function at each observation point and superimposing the effects of all kernels. Kernel density estimation can accurately reflect the true shape of the data distribution. In this embodiment, a Gaussian kernel is chosen as the kernel function for kernel density estimation.

[0073] This study focuses on a one-dimensional health index for intelligent weighing equipment in a training set. A kernel density estimation method is employed, setting a series of fixed intervals on both the horizontal axis (time) and the vertical axis (health index values). By estimating the distribution density of data near each reference point, the location of the most concentrated distribution is determined. Then, an exponential fitting method is used to fit the data, yielding a comprehensive health index.

[0074] 3.2.2 The CUSUM algorithm is used to divide the degradation stages, and a multi-stage Wiener model is constructed to predict the degradation trajectory;

[0075] (1) Stage division of CUSUM algorithm

[0076] The CUSUM algorithm is a statistical concept based on likelihood ratio, designed to detect points of change in a set of data. In principle, it is an analysis technique for sequence. Because degradation is actually a cumulative process, even minute fluctuations during degradation can lead to an increase or decrease in the cumulative bias value. Therefore, this algorithm is highly sensitive to small changes in the data and is well-suited for capturing points at different stages.

[0077] Therefore, the CUSUM algorithm is used to identify and divide the transition points of the comprehensive health index. Then, based on the degradation characteristics of each stage, Wiener process modeling and parameter estimation are performed to construct a multi-stage Wiener model.

[0078] (2) Constructing a multi-stage Wiener model

[0079] ①Wiener model

[0080] Wiener process It is not a strictly monotonic random process; rather, it describes a degenerative process with subtle fluctuations. The mathematical expression of the Wiener model is:

[0081] （3）

[0082] in Indicates that intelligent weighing equipment is in Performance degradation value at any given moment; This is the initial value for performance degradation; It is the drift coefficient representing the degradation process of intelligent weighing equipment, reflecting the average rate of degradation of intelligent weighing equipment; The diffusion coefficient represents the uncertainty of the equipment performance degradation process; Describes standard Brownian motion and satisfies ; and For the unknown parameters of the Wiener degradation model, estimation calculations are required.

[0083] ② Multi-stage Wiener model

[0084] Assuming the constructed health indicators are divided into a total of 100 categories. There are several stages, namely, a total of If there are several transformation points, then the multi-stage Wiener degradation model is:

[0085] （4）

[0086] In the above formula, , And so on. .

[0087] Although intelligent weighing equipment follows different degradation processes at different stages, regardless of the stage of degradation, when the degradation of intelligent weighing equipment reaches a threshold, it is considered that the intelligent weighing equipment has malfunctioned.

[0088] After completing the phase division of health indicators, a linear Wiener model is established for the degradation process of each phase to obtain a multi-stage Wiener process model. Next, the model parameters of each phase are determined by the maximum likelihood estimation method to obtain the final multi-stage Wiener model, which is used to predict the degradation trajectory of intelligent weighing equipment.

[0089] ③ Using a multi-stage Wiener model to predict the degradation trajectory of intelligent weighing equipment

[0090] First, the one-dimensional health index value of each smart weighing device in the training set at the time of failure is extracted. Then, the minimum value among these values ​​is selected as a unified failure threshold to ensure that this threshold represents the most stringent failure standard for all smart weighing devices. When analyzing multiple smart weighing devices in the test set, 50% of the length of the comprehensive health index is taken as the input data length of the model. Before predicting the remaining lifetime, it is necessary to determine which stage of the multi-stage Wiener process these input data belong to. The stage is determined by matching the input data with the data characteristics of each stage in the multi-stage Wiener model. Once the current stage is determined, the model parameters of that stage and subsequent stages can be used to predict the degradation trajectory of the device.

[0091] Step 4: Optimize the LSTM network based on the degradation trajectory predicted by the multi-stage Wiener model to predict the remaining lifespan of the intelligent weighing equipment.

[0092] To overcome the insufficient generalization ability of the fixed-parameter multi-stage Wiener model under complex conditions, this invention proposes an LSTM prediction mechanism guided by the degradation trajectory of the multi-stage Wiener model. Its core lies in designing a progressive input replacement strategy: during the prediction process, historical observations in the LSTM input window are gradually replaced with Wiener predicted values, thereby achieving a smooth transition from relying entirely on real data to predictions based entirely on the physical model. Specifically, it includes the following sub-steps:

[0093] 4.1 Initial Input Construction:

[0094] Initial Input Sequence: The input time step is set to 50% of the length of the comprehensive health index, consistent with the predicted input length of the multi-stage Wiener model. For each test device, the observation data of the first time_steps time steps are extracted from the actual health index sequence to form the initial input sequence.

[0095] Multi-stage Wiener model degradation trajectory: Obtain the Wiener sequence generated by the corresponding device.

[0096] LSTM input format: The sequence is reshaped into a three-dimensional tensor format of (1, time_steps, 1), which meets the input requirements of LSTM.

[0097] 4.2 Iterative Prediction Loop:

[0098] The prediction process employs an iterative rolling prediction strategy, executing a complete prediction-update loop at each time step. The entire prediction cycle covers all time steps from the current moment until the preset maximum prediction length or the failure threshold is reached. For each prediction time step, the following operations are performed:

[0099] 4.2.1 LSTM Prediction

[0100] The current input sequence is fed into an LSTM network. The network extracts temporal features through a two-layer LSTM structure with a hidden layer dimension of 64, expanding the 1-dimensional input into a 64-dimensional feature representation. Subsequently, a three-layer fully connected network (64→32→16→1) is used for dimensionality reduction mapping, and finally, a single numerical value is output as the predicted value of the health indicator for the next time step.

[0101] During the training phase, Dropout (0.2) is used to randomly deactivate some neurons, forcing the network to learn redundant features, reducing the risk of overfitting, and enhancing the model's generalization ability.

[0102] 4.2.2 Input Sequence Update

[0103] After the prediction is completed, the system performs a sequence update operation, achieving deep coupling between the physical model and the data-driven model:

[0104] (1) Time-series rolling update: Discard the oldest element in the sequence to make room for new predicted values. After the shift, the last element of the sequence is temporarily vacant, waiting to be filled;

[0105] (2) Wiener-guided value injection: Degenerate trajectory prediction values ​​generated by the multi-stage Wiener model are used to fill in the missing positions. This realizes the direct injection of physical model prediction information into the data-driven model, so that the input of LSTM gradually transitions from pure observation data to physical-guided trajectories;

[0106] (3) Update the input sequence: After completing the above operations, the updated input sequence is obtained and used for LSTM prediction in the next prediction step.

[0107] This update mechanism creates a gradual information fusion process: in the early stage, the input sequence mainly consists of real observations, and LSTM mainly relies on historical data patterns; in the middle stage, the sequence is a mixture of real and predicted values, and LSTM learns data patterns and physical guidance at the same time; in the later stage, the sequence is gradually dominated by Wiener predictions, and the physical degradation mechanism plays a major guiding role.

[0108] Combining the multi-stage Wiener process with an LSTM neural network offers the advantages of both the interpretability of a physical model and the flexibility of a data-driven model. On the one hand, the multi-stage Wiener process provides a clear physical and statistical basis and stage division criteria for degradation modeling; on the other hand, LSTM effectively compensates for the shortcomings of fixed-parameter stochastic processes in individual adaptability and nonlinear temporal dynamics modeling by capturing complex dependencies in the sequence. Experiments show that this method significantly enhances the model's adaptability to different equipment degradation modes while improving the accuracy of remaining lifetime prediction.

[0109] To evaluate the performance of the method of the present invention, the following evaluation indicators are used:

[0110] ① Root Mean Square Error (RMSE): refers to the square root of the mean of the squares of the differences between the predicted and actual values. It can be used to evaluate the difference between the predicted and actual values ​​and is calculated using equation (5):

[0111] （5）

[0112] in, Indicates the prediction result. This represents the true value. The smaller the RMSE value, the smaller the difference between the model's prediction and the true value.

[0113] ②Relative Error (RE): refers to the ratio of the absolute error to the true value multiplied by 100%. It can be used to evaluate the error between the predicted value and the true value, and is calculated using equation (6):

[0114] （6）

[0115] in, Indicates the prediction result. Represents the true value. RE reflects the error between the predicted value and the true value, and is displayed as a percentage. The smaller the value, the smaller the error in the model's prediction.

[0116] Mean Absolute Error (MAE): refers to the average of the absolute values ​​of the differences between the predicted and actual values. It can be used to evaluate the error between the predicted and actual values ​​and is calculated using equation (7):

[0117] （7）

[0118] in, Indicates the prediction result. This represents the true value. The smaller the MAE value, the smaller the error in the model's prediction.

[0119] Table 1 Comparison of Results of Various Indicators in the Method of the Invention

[0120]

[0121] As can be seen from Table 1, the method of the present invention is significantly better than the single multi-stage Wiener model in terms of degradation trend and RUL prediction effect. The error is significantly reduced. After the LSTM model is improved, the RMSE of degradation trajectory is reduced by 0.0794, the relative error of RUL is reduced by 15.65%, and the RUL prediction effect is more accurate. The prediction accuracy is significantly improved, which verifies the effectiveness of the method of the present invention.

[0122] Figure 2 This figure shows the degradation trend prediction results of the method of this invention. The horizontal axis represents time / day, and the vertical axis represents the Health Index (HI). The red dashed line represents the failure threshold, the blue curve represents the degradation trajectory prediction result of the multi-stage Wiener model, the purple curve represents the degradation trajectory prediction result after the LSTM model is optimized by the multi-stage Wiener model degradation trajectory guidance, and the black curve represents the actual degradation trajectory of the equipment. It can be seen that the predicted trajectory of the multi-stage Wiener model deviates significantly from the actual degradation data, indicating that it is difficult to fully capture the individual degradation specificity of intelligent weighing equipment. In contrast, the method of this invention exhibits excellent prediction performance, and its prediction results are closer to the actual data.

[0123] The above description is a specific illustration of the present invention, and not a limitation thereof. Those skilled in the art can make various equivalent technical solutions without departing from the scope of the present invention; therefore, all equivalent technical solutions should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process, characterized in that: Includes the following steps: Step 1: Acquire the device sensor data and normalize it to obtain normalized data; Step 2: Construct a decision-making framework that integrates mutual information and correlation analysis, select effective sensor features from normalized data, perform principal component analysis on them, and construct a one-dimensional health indicator. Step 3: Perform the following steps in parallel: First, construct and train an LSTM network; second, perform kernel density estimation on the one-dimensional health index to obtain a comprehensive health index, divide its degradation stage using the CUSUM algorithm, and construct a multi-stage Wiener model to predict the degradation trajectory. Step 4: Optimize the LSTM network based on the degradation trajectory predicted by the multi-stage Wiener model to predict the remaining lifespan of the intelligent weighing equipment.

2. The method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process according to claim 1, characterized in that: The fusion decision framework in step two involves performing a logical AND operation on the screening results of mutual information and correlation analysis, specifically: in, It is the final set of sensors. It is a set of sensors selected through correlation. It is a set of sensors filtered through mutual information.

3. The method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process according to claim 2, characterized in that: The LSTM network in step three is a two-layer stacked LSTM network.

4. The method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process according to claim 3, characterized in that: In step three, the degradation stages are divided using the CUSUM algorithm to construct a multi-stage Wiener degradation model as follows: In the above formula, , #imgpt6#, and so on, #imgpt7#.

5. The method for predicting the remaining life of intelligent weighing equipment based on a multi-stage Wiener process according to claim 4, characterized in that: Step four, which uses the multi-stage Wiener model to predict the degradation trajectory as a guide for optimizing the LSTM network, includes the following sub-steps: (1) Initial input construction: Set the input time step to 50% of the length of the comprehensive health index; predict the degradation trajectory based on the multi-stage Wiener model to obtain the Wiener sequence generated by the corresponding device; set the LSTM input format to meet the input requirements of LSTM; (2) Iterative prediction loop: The forecasting process employs an iterative rolling forecasting strategy, executing a complete forecast-update loop at each time step; for each forecast time step, the following operations are performed: ①LSTM prediction: Input the current sequence into the LSTM network and output the predicted value for the next time step; ② Input sequence update: Discard the oldest element in the sequence to make room for the new prediction; after the shift, the last position of the sequence is temporarily empty and waiting to be filled; fill the empty position with the degenerate trajectory prediction value generated by the multi-stage Wiener model to obtain the updated input sequence, which is used for LSTM prediction in the next prediction step.

Images