Method for predicting residual life of weighing sensor based on outer edge data reinforcement

Abstract

The invention relates to a weighing sensor residual life prediction method based on outer edge data reinforcement. The method comprises the following steps: firstly, acquiring weight data and temperature data of a constant load of a weighing sensor at normal temperature, and preprocessing the weight data and the temperature data; secondly, LSTM network construction and related parameter setting are carried out, processed data are input into a network, and training of a network model is carried out; then, a temperature test set and predicted weight data are used as input of a trained model, and multi-step testing is carried out; and finally, performing RUL calculation of different test points according to the prediction results of the different test points. According to the method, the degradation trend of the weighing sensor can be well predicted, and the accuracy of RUL prediction can reach 95.6126%; according to the method, the model training stability and convergence speed can be remarkably improved, the prediction efficiency is remarkably improved, the method can effectively adapt to the influence of various environmental factors on the system performance, and the stability and response capability of the system under different working conditions are improved by adopting a comprehensive strategy.

Description

Technical Field

[0001] The present invention relates to the fields of fault prediction and health management, and in particular to a method for predicting the remaining life of a weighing sensor based on peripheral data enhanced drive. Background Art

[0002] A load cell is a device used to measure the mass of an object, usually consisting of a sensitive element and a signal processing circuit. According to different working principles, load cells are mainly divided into strain type, piezoelectric type, capacitive type, magnetic induction type, vibrating string type and other types, and are widely used in various fields. Among them, strain type load cells are particularly common in smart warehousing, with the characteristics of long continuous use time and high measurement accuracy. However, with the extension of use time and the increase of frequency of use, strain type load cells may gradually degrade, resulting in performance degradation or even failure. Therefore, it is particularly important to predict the remaining life of this type of sensor to ensure its long-term stable and reliable operation.

[0003] Currently, there are many methods that can predict degradation trends and calculate remaining life, including LSTM method, ARIMA method, etc., all of which use external data enhancement. External data enhancement refers to the use of external data to enhance the learning ability and prediction performance of the model. It is mainly used to solve the problem of insufficient data or the performance bottleneck of the model.

[0004] The document with publication number "CN202311658215.8" proposes "a bearing life prediction method and system based on LSTM", which mainly uses bearing operating status data, such as vibration data, temperature data, sound data and speed data, and uses the LSTM model to predict bearing life. The problems are as follows: 1. The generalization ability of the model is insufficient and it is difficult to adapt to different working conditions; 2. Computational resources are wasted and the calculation speed is slow; 3. LSTM single-step prediction is updated each time with the observed value. Another document with publication number "CN202111420929.6" proposes "a method for predicting the life of an aircraft engine based on an improved LSTM", which mainly uses a deep sparse autoencoder (SDAE) for unsupervised feature extraction and uses an LSTM model for aircraft generators.

[0005] The problems with life prediction are as follows: 1. The introduction of SDAE increases the complexity of the model and the training time; 2. LSTM single-step prediction is updated with observed values ​​each time; 3. The model's generalization ability is insufficient and it is difficult to adapt to different working conditions.

[0006] The common problem of the above two files is that they both use the LSTM model for single-step prediction. Single-step prediction can only be predicted step by step. Each input requires an observation value. The observation value can only predict one step, and the long-term trend cannot be obtained at one time.

[0007] The classic LSTM model method generally only contains historical weight data, and the historical time series of the weight of the weighing sensor is used as input to predict future weight data. Due to error accumulation and the limited ability of the model to capture complex relationships, it is impossible to obtain long-term trends at one time, and the predicted values ​​are prone to tend to constant values. Summary of the invention

[0008] The present invention proposes a method for predicting the remaining life of a weighing sensor based on peripheral data enhancement, so as to overcome the problem in the prior art that the long-term trend prediction tends to a constant value and the remaining life cannot be predicted.

[0009] To achieve the above object, the technical solution of the present invention is as follows: A method for predicting the remaining life of a weighing sensor based on peripheral data enhancement, comprising the following steps:

[0010] Step 1, obtaining weight data and temperature data of a constant load of a weighing sensor at room temperature, and preprocessing them respectively, wherein the preprocessing includes filtering, dividing and normalizing, and dividing the weight data and temperature data into a weight training set, a weight test set, a temperature training set and a temperature test set;

[0011] Step 2: Input the temperature data into the LSTM model network, and use the weight training set and temperature training set to train the model to obtain multiple different LSTM prediction models;

[0012] Step 3: Use the temperature test set and the predicted weight data as the input of multiple trained LSTM prediction models, perform multi-step prediction to obtain the predicted weight data, average the multi-step prediction results of multiple LSTMs, and obtain the predicted degradation trend of the weighing sensor;

[0013] Step 4: Select a test point i (i=1, ..., n), obtain the degradation trend of the weighing sensor at different test points with the help of step 3, and calculate and predict the remaining life RUL.

[0014] Furthermore, in the above step 1, the normalized formula is shown as (3):

[0015]

[0016] in, represents the smoothed result at x, is the minimum value in the smoothed data, is the maximum value in the smoothed data, and Y is the normalized data.

[0017] Furthermore, in the above step 2, the formula of the LSTM model network structure unit is shown in (4)-(9):

[0018] f t =σ[W f ·(h t-1 ,X t ,Y t )+b f ] (4)

[0019] i t =σ[W i ·(h t-1 ,X t ,Y t )+b i ] (5)

[0020]

[0021] o t =σ[W o ·(h t-1 ,X t ,Y t )+b o ] (8)

[0022]

[0023] Among them, W f , W i , W c and W o is the weight matrix, b f 、b i 、b c and b o is the bias term, (h t-1 ,X t ,Y t ) represents the output vector (h t-1 ) and (X t ,Y t ) are connected to form a longer vector, σ is the sigmoid function, f t is the output of the forget gate at time step t, X t ,Y t is the input of the current time step, i t is the output of the input gate at time step t, is the candidate memory vector, c t is the cell state update at time step t, represents element-wise multiplication, o t is the output of the output gate at time step t, h t is the hidden state at the next time step.

[0024] Furthermore, in the above step 4, the specific formulas for calculating RUL are shown in (12) and (13):

[0025] RUL 真实 =EOL 真实 -T t (12)

[0026] RUL 预测 =EOL 预测 -T t (13)

[0027] Among them, EOL 真实 Is the true end of life, EOL 预测 is the predicted end-of-life, T t is the time used.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The present invention adds a normalization step in the preprocessing process, converts the experimentally obtained temperature data and weight data into a range of -1 to 1 through normalization operations, solves the problem of unstable performance of the classic LSTM model when processing high dynamic range data, and significantly improves the stability and convergence speed of model training. ,

[0030] 2. In view of the fact that weighing sensors are affected by environmental factors, the present invention adopts an improved LSTM model, taking both temperature and weight data as input to predict future weight data, and taking temperature as the introduced peripheral data. Driven by peripheral data, the input of the model not only includes historical weight data and historical temperature data, but also uses the temperature data of the previous step and the predicted weight data of the previous step as input predictions at each step. This method solves the problem in the classic LSTM model that the predicted value tends to a constant value due to error accumulation, resulting in the result not being able to follow the actual change trend, and significantly improves the prediction efficiency.

[0031] 3. The method of the present invention adopts a multi-step process, has strong applicability and scalability, and can effectively adapt to the impact of various environmental factors on system performance. This comprehensive strategy improves the stability and responsiveness of the system under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is the overall flow chart of the present invention;

[0033] Figure 2 This is the internal structure diagram of the LSTM network repeating unit;

[0034] Figure 3 Single-step prediction graphs for the classic LSTM network and the improved LSTM network;

[0035] Figure 4Multi-step prediction graphs for the classic LSTM network and the improved LSTM network;

[0036] Figure 5 This is the remaining life prediction diagram of the weighing sensor;

[0037] Figure 6 This is the relative error diagram of the remaining life of the weighing sensor; DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0039] See also Figure 1 The present invention provides a method for predicting the remaining life of a weighing sensor based on peripheral data enhancement, comprising the following steps:

[0040] First, the constant load weight and temperature data of the weighing sensor at room temperature are preprocessed; secondly, the LSTM network is constructed and related parameters are set, and the processed data containing the edge data temperature is input into the network to train the network model. The model is tested using the edge data temperature test set to obtain the degradation trend of the weighing sensor; finally, RUL calculation is performed at different test points to predict the remaining life of the weighing sensor.

[0041] The specific implementation steps are as follows:

[0042] Step 1: Obtain the weight data and temperature data of the weighing sensor under constant load at room temperature, and preprocess them respectively. The preprocessing includes filtering, division and normalization. The weight data and temperature data are divided into a weight training set, a weight test set, a temperature training set and a temperature test set. The temperature test set is the outer edge data:

[0043] The "normal temperature" referred to in the present invention is a temperature range between 20°C and 30°C.

[0044] 1.1 On a weighing sensor with a range of 3kg, a weighing aging experiment was carried out with a constant load of 500g and a duration of 6 hours to obtain weight data and temperature data.

[0045] 1.2 Preprocess the acquired valid data: Preprocessing includes data cleaning, data alignment, filtering, division and normalization.

[0046] Data cleaning: Check whether there are missing values, outliers or noise in the two sets of data, and perform cleaning processing.

[0047] Data alignment: Ensure that weight data and temperature data are aligned in time for subsequent analysis.

[0048] Filtering: Use the local weighted regression method to smooth the two sets of data. The width of the smoothing window is 1% of the total length of the original data. The specific formulas are shown in (1) and (2):

[0049] Calculate the weight w t :

[0050]

[0051] Among them, x t represents the horizontal coordinate of the original data, x represents the horizontal coordinate of the current point, τ is the bandwidth parameter of the smoothing window, which controls the decay speed of the weight, and w t Represents weight.

[0052] Compute the weighted linear regression:

[0053]

[0054] in, represents the smoothed result at x, y t Represents the ordinate of the original data, w t Represents weight. Division: The present invention divides data in chronological order to avoid future information leakage, and divides weight data and temperature data into a weight training set, a weight test set, a temperature training set, and a temperature test set.

[0055] Normalization: Normalize the smoothed data. The specific formula is shown in (3):

[0056]

[0057] in, represents the smoothed result at x, is the minimum value in the smoothed data, is the maximum value in the smoothed data, and Y is the normalized data.

[0058] Step 2: Input the temperature data into the LSTM model network, and use the weight data training set and the temperature data training set to train the model to obtain multiple different LSTM prediction models;

[0059] 2.1 Build the LSTM model network structure and set network parameters:

[0060] The LSTM model network structure of the present invention consists of an input layer, an LSTM layer, a fully connected layer and a regression layer. Figure 2 , Figure 2 This is the internal structure diagram of the repeating unit in the LSTM network, which uses temperature data and weight data as inputs to the repeating unit of the LSTM network at the same time.

[0061] The relationship formulas between the previous moment unit, the current moment unit and the next moment unit in the LSTM layer are shown in (4)-(9).

[0062] The formulas of the LSTM model network repetition unit are shown in (4)-(9):

[0063] f t =σ[W f ·(h t-1 , X t , Y t )+b f ] (4)

[0064] i t =σ[W i ·(h t-1 ,X t , Y t )+b i ] (5)

[0065]

[0066] o t =σ[W o ·(h t-1 , X t , Y t )+b o ] (8)

[0067]

[0068] Among them, W f , W i , W c and W o is the weight matrix, b f 、b i 、b c and b o is the bias term, (h t-1 ,X t , Y t ) represents the output vector (h t-1 ) and (X t , Y t ) are connected to form a longer vector, σ is the sigmoid function, f t is the output of the forget gate at time step t, X t , Y t is the input of the current time step, i t is the output of the input gate at time step t, is the candidate memory vector, c tis the cell state update at time step t, represents element-wise multiplication, o t is the output of the output gate at time step t, h t is the hidden state at the next time step.

[0069] In addition to the weight data of the weighing sensor, the present invention adds temperature data as peripheral data to predict the future weight information of the weighing sensor. The model is called an improved LSTM model.

[0070] 2.2 Train the model to get different multiple LSTM predictions:

[0071] In this step, the weight data training set and the temperature data training set are used, and the mean square error loss function and the classic Adam optimizer are used, combined with the piecewise linear learning rate decay strategy, and the reshuffled data is used to train the improved LSTM model to obtain different multiple LSTM prediction models. The piecewise linear learning rate decay strategy significantly improves the training speed and accuracy of the model, and the use of reshuffled data increases the randomness of training, which helps to improve the generalization ability of the improved LSTM model, thereby effectively reducing the bit error rate and improving the overall performance of the system.

[0072] Step 3: Use the temperature test set and predicted weight data as the input of multiple trained LSTM prediction models to perform multi-step prediction, and average the multi-step prediction results of multiple LSTMs to obtain the predicted degradation trend of the weighing sensor;

[0073] 3.1 Use the temperature test set and predicted weight data as input to the trained model for multi-step prediction:

[0074] In the prediction stage, the edge data temperature and predicted weight data are used as input and input into the trained model to achieve multi-step prediction of weight; Figure 3 The figure shows the results of single-step prediction of the classic LSTM model and the improved LSTM model when the split point is 7920s. It can be seen from the figure that the prediction results of the improved LSTM model have not been significantly improved when performing single-step prediction. After precise calculation, the RMSE of the prediction accuracy of the classic LSTM model is 0.2865, and the RMSE of the prediction accuracy of the improved LSTM model is 0.3415. The prediction accuracy of the improved LSTM model is 0.055 lower than that of the classic LSTM model. In summary, when the edge data temperature is used as the enhanced driving data for single-step prediction, the prediction results are not significantly improved.

[0075] See also Figure 4 , Figure 4The degradation trend of the multi-step weight prediction of the weighing sensor, the intuitive comparison of the confidence interval and the mean are shown. The figure shows the effect of the multi-step prediction of the classic LSTM model and the improved LSTM model when the split point is 7920s. It can be seen intuitively from the figure that when performing multi-step prediction, the prediction results of the improved LSTM model solve the limitations of the classic LSTM model and can better track the trend of data changes. After precise calculation, the RMSE of the prediction accuracy of the classic LSTM model is 0.6539, and the RMSE of the prediction accuracy of the improved LSTM model is 0.4312. In summary, when the outer edge data temperature is used as the enhanced driving data for multi-step prediction, the prediction results are significantly improved, which solves the limitations of the classic LSTM model for multi-step prediction.

[0076] 3.2 Calculate the mean of the multi-step prediction results of multiple LSTMs to obtain the predicted degradation trend of the weighing sensor:

[0077] Calculate the mean and confidence interval. By calculating the confidence interval, the mean is finally used as the final degradation trend.

[0078] In order to reflect the uncertainty of the estimated degradation trend, the mean and confidence interval method is used. The calculation formulas are shown in (10) and (11):

[0079]

[0080] in, is the sample mean, x i is the output result of the i-th prediction, n is the number of predictions, (t α / 2,n-1 ) is the critical value on the t distribution with (n-1) degrees of freedom, satisfying (P(T≤t α / 2,n-1 )=1-α / 2), usually obtained using statistical software or table lookup; s is the sample standard deviation; n is the sample size; μ is the population mean.

[0081] Step 4: Select a test point i (i=1, ..., n), obtain the degradation trend of the weighing sensor at different test points with the help of step 3, and calculate and predict the remaining life RUL.

[0082] Step 4.1: Select test point i (i = 1, ..., n), and use the degradation trend of the weighing sensor obtained in step 3 to calculate the remaining service life RUL. The formula is shown in formula (13):

[0083] The specific formula for calculating the real RUL is shown in (12):

[0084] RUL 真实 =EOL 真实 -T t (12)

[0085] RUL预测 =EOL 预测 -T t (13)

[0086] Among them, EOL 真实 Is the true end of life, EOL 预测 is the predicted end-of-life, T t is the time used.

[0087] Step 4.2: Calculate the error between the actual remaining life and the predicted remaining life.

[0088] The present invention adopts mean square error (MSE) as the loss function, and the mean square error is used to measure the difference between the model prediction value and the true value.

[0089] The evaluation indicators include root mean square error (RMSE) and relative error (RE). The specific formulas are shown in (14)-(15):

[0090]

[0091] Where n is the number of samples; y i is the true value; is the predicted value.

[0092] The method of the present invention is verified by selecting different test points, and the results show that the accuracy is significantly improved. The "different test points" referred to in this embodiment refer to "5940 seconds, 7920 seconds, 9900 seconds, 11880 seconds, 13860 seconds, and 15840 seconds". See the table below:

[0093]

[0094] It is very important to estimate the RUL of the weighing sensor. The remaining service life of the weighing sensor can be obtained by calculating the degradation trend of different test points.

[0095] The predicted RUL is compared with the actual RUL, and the minimum relative error is calculated to be 4.3874%, and the RUL prediction accuracy is 95.6126%, which meets the requirements of the remaining life prediction accuracy. Therefore, the predicted remaining life of the weighing sensor can be obtained through this model.

[0096] Figure 5 Based on the degradation trends at 5610s, 6600s, 7590s, 8580s, 9570s, and 10560s, a comparison chart of the predicted RUL and the actual RUL of the weighing sensor with a total of 6 test points is calculated.

[0097] Figure 6This is a graph that calculates the corresponding RE based on the RUL values ​​at different test points. The results show that RE has a relevant fluctuation trend. In the 6 experiments conducted at 6 different test points, the RE of all items is less than 50%. Among them, the minimum RE of the experimental results is 4.3874%, that is, the RUL prediction accuracy at this time is 95.6126%.

[0098] The above is an explanation of the specific implementation of the present invention, rather than a limitation of the present invention. Those skilled in the relevant technical field can also make various equivalent technical solutions without departing from the scope of the present invention, so all equivalent technical solutions should be included in the patent protection scope of the present invention.

Claims

1. A method for predicting the remaining life of a weighing sensor based on peripheral data enhancement, characterized in that: The following steps are involved: Step 1, obtaining weight data and temperature data of a constant load of a weighing sensor at room temperature, and preprocessing them respectively, wherein the preprocessing includes filtering, dividing and normalizing, and dividing the weight data and temperature data into a weight training set, a weight test set, a temperature training set and a temperature test set; Step 2: Input the temperature data into the LSTM model network, and use the weight training set and temperature training set to train the model to obtain multiple different LSTM prediction models; Step 3: Use the temperature test set and the predicted weight data as the input of multiple trained LSTM prediction models, perform multi-step prediction to obtain the predicted weight data, average the multi-step prediction results of multiple LSTMs, and obtain the predicted degradation trend of the weighing sensor; Step 4: Select a test point i (i=1, ..., n), obtain the degradation trend of the weighing sensor at different test points with the help of step 3, and calculate and predict the remaining life RUL.

2. The method for predicting the remaining life of a weighing sensor based on peripheral data enhancement according to claim 1 is characterized in that: In step 1, the normalized formula is shown in (3): in, represents the smoothed result at x, is the minimum value in the smoothed data, is the maximum value in the smoothed data, and Y is the normalized data.

3. The method for predicting the remaining life of a weighing sensor based on peripheral data enhancement according to claim 2 is characterized in that: In step 2, the formula of the LSTM model network structure unit is shown in (4)-(9): f t =σ[W f ·(h t-1 ,X t ,Y t )+b f ] (4) i t =σ[W i ·(h t-1 ,X t ,Y t )+b i ] (5) the t =σ[W o ·(h t-1 ,X t ,Y t )+b o ] (8) Among them, W f , W i , W c and W o is the weight matrix, b f 、b i 、b c and b o is the bias term, (h t-1 ,X t ,Y t ) represents the output vector (h t-1 ) and (X t ,Y t ) are connected to form a longer vector, σ is the sigmoid function, f t is the output of the forget gate at time step t, X t ,Y t is the input of the current time step, i t is the output of the input gate at time step t, is the candidate memory vector, c t is the cell state update at time step t, represents element-wise multiplication, o t is the output of the output gate at time step t, h t is the hidden state at the next time step.

4. The method for predicting the remaining life of a weighing sensor based on peripheral data enhancement according to claim 3 is characterized in that: In step 4, the specific formulas for calculating RUL are shown in (12) and (13): RUL 真实 =EOL 真实 -T t (12) RUL 预测 =EOL 预测 -T t (13) Among them, EOL 真实 Is the true end of life, EOL 预测 is the predicted end-of-life, T t is the time used.

Images