Traceability and analysis method for measurement laboratory testing data based on intelligent Internet of Things and deep belief network

4.1 Constructing a sample library for measuring laboratory testing ledger

Before conducting an abnormal analysis of measurement laboratory testing data, it is necessary to analyze the influencing factors of the normal operation of measurement devices and determine the dimensions of data mining. Mainly considering five aspects: electricity consumption environment, meteorological factors, different types of metering devices, operating hours, engineering quality, and historical electricity consumption curve data. The first four aspects belong to ledger data, and the historical electricity consumption curve data belong to real-time data of the electricity consumption information collection system. In order to have sufficient samples for the model to learn, the curve length is taken as 1 month, with 96 points per day.

The electrical environment and meteorological factors can have a significant impact on the reliability of metering devices and even cause meter failures, such as abnormal energy meters caused by atmospheric corrosion of chloride salts in the operating environment, external force damage during operation, lightning strikes, etc. Due to the large number of suppliers of measuring devices and the differences in internal design and components among different measuring devices, the failures of measuring devices often exhibit familial defects. There is a high possibility of the same type of failure occurring in the same batch of measuring devices from the same manufacturer. An increase in operating time will lead to a decrease in the operational reliability of the metering device, which has a certain expected lifespan. As the operating time increases, the aging of its components and battery loss will both make the meter more prone to failure. For example, hardware damage include error tolerance, capacitor damage, battery damage, etc. Engineering quality issues can also affect the operation status of electricity meters. Engineering quality issues can be divided into meter quality and installation quality issues, including false soldering, foreign objects in the meter, short circuits of solder joints, incorrect parameter settings, reverse connection of incoming and outgoing lines, or phase errors.

Based on the analysis of the influencing factors of metering device failures mentioned above, multiple ledger attributes were selected as feature vectors, and the data consist of two categories: (1) Components of electricity consumption environment and meteorological factors. Dimensions such as usage region and industry category can be selected for electricity consumption environment and meteorological factors. (2) Different types of measuring device components. Select three dimensions for the components of the measuring device, namely category, manufacturer, and rated current. Select 8,000 pieces of measurement device data from the marketing business system and electricity information collection system as samples, and after screening, the remaining 6,498 pieces of data are used as samples to construct a sample library of measurement laboratory testing accounts. The sample data include various attribute values of normal and faulty devices, and data collection and organization of account attributes such as usage area, category, manufacturer, industry category, and calibration current are carried out.

4.2 GBRBM

The restricted Boltzmann machine (RBM) is an undirected probability graph model with a bipartite graph structure, including a visible layer v composed of m neurons and a hidden layer h composed of n neurons. Each node in the layer is not connected to each other and is independent of each other [22]. Due to the fact that both the visible and hidden layers of traditional RBM are binary units, the values are both 0 and 1. The vibration signal is a set of continuous time series data, and the binary unit model is not conducive to the operation of continuous data. Therefore, the proposed method introduces continuous values of independent Gaussian distribution in traditional RBM models to process continuous data, limiting the output value of RBM’s structural units to continuous values between 0 and 1.

GBRBM is jointly proposed by Krizhevsky and Hinton and is used to model the Boltzmann machine model of nonbinomial distribution data [23]. GBRBM assumes that the variable of the visible layer is the Gaussian distribution and the variable of the hidden layer is the Bernoulli distribution. Its energy function is defined as

The conditional probability of visible and hidden layers are

4.3 Improved DBN model

In order to solve the problem of the poor fitting effect of traditional RBM for data reconstruction of non-binomial distribution, a GBRBM-DBN model is proposed, which consists of a GBRBM, several RBMs, and a Softmax classifier, as shown in Figure 2.

Figure 2

Structure of an improved DBN.

The input training data are used as the input v 0 for the visible layer of the underlying GBRBM, and the output of its hidden layer h 0 is used as the input v 1 for the visible layer of the second layer RBM. Similarly, the contrast divergence (CD) algorithm is used to achieve layer-by-layer unsupervised pre-training of GBRBM and RBM, ensuring optimal feature vector mapping within the layer and maximizing the preservation of original feature information. Finally, the characteristic signals obtained from the layer-by-layer pre-training are input to the Softmax layer through a fully connected network. With cross entropy as the loss function, the Adam optimization algorithm is used to carry out a top-down supervised reverse fine-tuning of the network parameters to complete the training of the entire GBRBM-DBN network. The training time of the GBRBM-DBN network model is significantly reduced, and it is closer to the true value, which has a significant advantage in the fast and accurate processing of massive data.

Among them, GBRBM, like traditional RBM, uses the CD algorithm for pre-training. The CD algorithm uses the K-L distance between the estimated probability distribution and the true probability distribution as a metric criterion and performs g-step Gibbs sampling on each batch of training samples to generate samples with maximum probability.

The maximum value of the likelihood function is obtained by using the random gradient rise method, and the network parameter update rules are as follows:

4.4 Analysis process for abnormal detection data in metrology laboratory

Due to its ability to capture high-level abstract features in data, strong representation ability for complex data structures, and more stable model training, the DBN model is used to analyze complex metrology device data. Meanwhile, the DBN model utilizes RBM to generate distribution factors for training data, resulting in higher learning efficiency. Additionally, the improved DBN model based on GBRBM can quickly and accurately process massive amounts of data. Therefore, the GBRBM-DBN model is used for the analysis of massive detection data in metrology laboratories. Specifically, the detection ledger sample library data are input into the improved DBN model for learning to obtain abnormal detection data. The overall process is shown in Figure 3.

Figure 3

Data anomaly analysis process based on improved DBN.

First, divide 6,498 samples into a training set and a testing set in a 5:1 ratio. In the process of using the training set for model training, the training data are input to the underlying nodes of GBRBM-DBN, and the weights ω , visible node biases, and hidden layer node biases (a,b) between each layer are trained. The obtained feature data are then input to the Softmax classifier for analysis. According to the network update rules and iteration termination conditions, the optimal GBRBM-DBN model is obtained.

Then, during the model testing process, the test data are input into the trained GBRBM-DBN model, and the Softmax classifier is used to output the abnormal analysis results of the detection data to determine whether the measuring device is abnormal. Finally, trace and analyze the abnormal measuring devices to identify the cause of the abnormality and improve maintenance efficiency.

5.1 SDAE

The autoencoder includes encoding and decoding processes, which can convert input data X to hidden layer data Y and then decode it to Z [24]. The automatic encoder directly analyzes the complete input data, and once the input dataset changes, the adaptability of the model will also be significantly reduced. Therefore, in order to improve the adaptability of the model, a denoising automatic encoder (DAE) is proposed, which can set the values of certain positions in the input data to 0 with a certain probability distribution, equivalent to involving “bad” data X ˜ in the iterative process of the network, in order to learn robust features [25,26]. The mathematical expression of the DAE learning process is as follows:

Due to the limited ability of a single DAE to mine hidden data information, it is usually necessary to stack multiple DAEs, resulting in an SDAE model composed of multiple DAE models. SDAE can learn deeper features in input data than a single DAE, and its network structure is shown in Figure 4.

Figure 4

Structure of SDAE network.

The training process of SDAE includes unsupervised layer-by-layer pre-training and supervised fine-tuning. After randomly mapping the hidden layers of each individual DAE, set the data at certain positions to 0 and then use it as input for the next DAE, and so on, to form a complete SDAE network.

5.2 Proposed traceability method for measurement laboratory testing data

The historical operation data of the metering device reflect the operating health of the electricity meter. For the smart meter in normal operation, the metering and detection organization can normally collect its current, voltage, active power, reactive power, meter reading electricity, and other data, which reflects the operating status of the device. Devices with a high potential for failure often operate in a hidden and abnormal state, and their operating data differ from the original normal operating data, but the difference is small and cannot meet the fault diagnosis standards of measuring and testing institutions. Therefore, the extraction of abnormal indicators from real-time operational data can be used as an evaluation of the health status of measuring devices.

In order to provide sufficient samples for the model to learn, the curve length is taken as 1 month, with 96 points per day, and the curve categories are current, voltage, and power. Due to the large amount of current and power data collected by measuring devices, directly using curve data as model attributes can lead to excessive dimensionality, directly affecting the calculation time and traceability accuracy of the model [27,28,29,30,31]. Meanwhile, short-term curve data cannot reflect the operating status of the device over a longer period of time and commonly used neural networks such as LSTM are difficult to handle long time series, and the model is more complex. SDAE has good deep-level feature extraction and adaptive capabilities. In order to obtain the overall situation of real-time electricity consumption curve data as much as possible and facilitate deep feature extraction of the model, SDAE is used to extract features of current and power, and a fault feature sample library of the electricity consumption curve is constructed.

The input curve fault feature sample library of the measuring device that is determined to have abnormalities through the ledger sample library will be used to calculate the similarity between the operating curve characteristics of the testing measuring device and the fault curve characteristics. Based on the results, it will be determined whether the measuring device has malfunctioned, achieving abnormal traceability of the measuring device. The abnormal traceability process of measuring devices based on SDAE is shown in Figure 5.

1. Set up 96 input nodes in the SDAE input layer and receive a current or power sequence of 1 96 point input. Among them, the GBRBM-DBN model determines the existence of abnormal metering devices, and the input current or power is the sequence from the day before the abnormality occurs.

2. The input layer maps abnormal data to the hidden layer through random mapping and learns key feature information from the abnormal data X ˜ m . The features of X ˜ m are represented by f

   (8) f = φ ( X ˜ m ) = δ ( ϖ X m + d ) ,

   where ϖ is the weight value.

3. Reconstruct abnormal data X Θ output without noise, i.e., a 96 point time series. In order to ensure that the output time series is as consistent as possible with the initial time series while reducing noise and retaining all key information of the initial series, it is necessary to minimize the reconstruction error function to obtain optimal parameters. The reconstruction error function L is shown as follows:

   (9) L = ζ 1 2 n ∥ X m − X Θ ∥ 2 + 1 2 n ∥ φ ( X ˜ m ) − X Θ ∥ 2 .

   After the training of the SDAE network is completed, the reconstruction error is small, and it is considered that the extracted features at this time are typical. The features extracted from the hidden layer, namely the historical current curve and power curve, are condensed into two 24 point feature sequences. Classify the current characteristics and power characteristics of faulty meters into two categories and construct a fault feature sample library for electricity consumption curves.

4. Due to the time correlation of the data collected by the measuring device, its distribution will change over time. Therefore, a sliding window model is used to place the newly collected data and the latest time data in the same window, and the output is compared and analyzed with the fault curve characteristics. In the sliding window model, k represents the size of the sliding window, with windows 1 and 2 at time t − 1 and t, respectively. When new data are collected at the next moment, the window slides to the right, sliding the samples to be analyzed into the window, and sliding some of the analyzed samples out of the window. Finally, output the sample data in the new window.

5. When comparing the power consumption curve features of the target metering device output through the sliding window model with the fault features, cosine similarity is used to evaluate the similarity between the two features. The cosine similarity D cos i reflects the similarity between the fault characteristic curve x of the electricity consumption curve and the change trend of the electricity consumption curve characteristic y of the target metering device. The closer the value is to 1, the more similar it is. D cos i is calculated as follows:

where Q is the length of the characteristic vector of the electricity consumption curve, with a value of 24; x q , y q represents the characteristic components of the q-th point on the power fault characteristic curve and the target metering device’s power consumption characteristic curve, respectively. When D cos i approaches 1, it indicates that the reason for the abnormality of the target metering device is the existence of this type of electrical fault, and the tracing is completed. On the contrary, the target metering device is normal.

Figure 5

The abnormal traceability process of measuring devices based on SDAE.

6.1 Evaluation index

The experiment measures the performance of the algorithm from two aspects: accuracy and computational time. The algorithm accuracy includes the accuracy, false positive rate (FPR), and recall rate (Recall) of anomaly detection analysis, which are calculated as follows:

where Y represents the true abnormal state in the original data, Y ˆ is the abnormal state obtained through abnormal analysis method analysis, and Y ¯ is the average abnormal state.

In addition, the time indicator is chosen as the average execution time of different methods executing the same number of times.

6.2 Determination of model structure

The number of hidden layers of the model has a certain impact on the experimental results. If the number of layers is too large, it will not only increase the operation time, but also cause overfitting, affecting the accuracy of analysis. The selection of layers is too small to deeply learn feature phasors and cannot extract abnormal features well, resulting in insufficient accuracy. Therefore, 200 samples were selected for training and 50 samples for testing. The relationship between the number of hidden layers and analysis accuracy is shown in Figure 6.

Figure 6

The relationship between accuracy and the number of hidden layers: (a) improved DBN model and (b) SDAE model.

From Figure 6, it can be seen that when the hidden layer of the improved DBN model is three layers, the analysis accuracy reaches the highest 95%. Increasing the number of hidden layers does not increase the accuracy but may actually decrease it. When the number of hidden layers reaches the seventh layer, the accuracy is even lower than 93%. Therefore, the hidden layer of the improved DBN model is set to three layers. Similarly, when the hidden layer of the SDAE model is four layers, the analysis accuracy reaches the highest 94.6%. The increase in the number of hidden layers does not improve the analysis accuracy, but rather decreases. When the hidden layer is eight layers, the accuracy of the SDAE model decreases to 92.5%, and an increase in the number of hidden layers greatly increases the learning time of the model. Therefore, considering all factors, the hidden layer of the SDAE model is set to four layers.

6.3 Comparison of abnormal analysis results between different models

The improved DBN model processes continuous data by introducing an independent Gaussian distribution. In order to demonstrate the performance difference between it and traditional DBN models, the two were compared and trained, and the results are shown in Figure 7.

Figure 7

Performance comparison of the DBN model before and after improvement.

From Figure 7, it can be seen that after 350 iterations, the improved GBRBM-DBN model tends to stabilize with an accuracy rate of over 95%. The DBN model tends to converge after 420 iterations, and the accuracy decreases by 2% compared to the improved DBN model. This indicates that introducing the GBRBM layer into the DBN model can improve the model’s anomaly analysis ability and accelerate convergence.

In order to verify the data anomaly analysis performance of the improved DBN model, it was compared and analyzed by Wu et al. [10] and Xu et al. [12]. Among them, Wu et al. [10] use traditional detection models, while Xu et al. [12] use deep learning models, which are comparable to the proposed improved DBN model. Input the sample database of the ledger into four models, and the evaluation indicators (accuracy, FPR, and recall rate) of the analysis results are shown in Table 3.

From Table 3, the proposed GBRBM-DBN model has the best performance, with accuracy, FPR, and recall rate reaching 95.51, 4.49, and 93.08%, respectively. This is because the improved DBN model adopts a layer of GBRBM, which can better handle continuous sample library data. At the same time, compared to other models, the model has stronger deep learning ability, so the anomaly analysis effect is the most ideal. Wu et al. [10] use the improved kernel density estimation method to detect data anomalies, but this method has greater limitations, and the detection accuracy is 8.32% lower than the proposed method. The combination of LAL and AdaBoost algorithms in Xu et al. [12] for data anomaly detection can ensure a good detection accuracy of 92.25%. However, the uncertainty of the LAL model leads to unstable analysis results of this method. Therefore, overall, the proposed method has significant advantages in analyzing abnormal metering devices.

6.4 Analysis of traceability results

The accuracy of traceability analysis results is crucial. Compare the proposed SDAE + GBRBM-DBN method with SDAE, SDAE + DBN, and select 200 electricity consumption curves as samples for training. The FPR under different iteration times is shown in Figure 8.

Figure 8

Analysis error under different iterations.

From Figure 8, it can be seen that when the number of iterations reaches 150, the error detection rate of the SDAE + GBRBM-DBN model reaches a stable state, approaching 8%. However, the SDAE + DBN model has some fluctuations in the entire training process due to not considering the time continuity of the data, and the error detection rate has improved. The learning and analysis ability of a single SDAE model is poor, resulting in a false detection rate of nearly 20%. From this, it can be argued that the proposed SDAE + GBRBM-DBN method has the best effect on the traceability analysis of anomalies in measuring devices.

Select 20 non-abnormal electricity meters and 20 abnormal electricity meter ledger data as the test set and use the proposed method to test the electricity meters in this test set. Extract the electricity consumption curve data judged as abnormal electricity meters and input it into the electricity consumption curve fault feature sample library for anomaly tracing. The electricity consumption curve attributes are the current and power sequences extracted through feature extraction. Using cosine similarity for similarity calculation, determine a faulty meter as one that has a similarity of more than 0.8 between the electricity consumption characteristic curve and the fault characteristic curve. The ID and fault tracing results of some test meters are shown in Table 4.

According to the test results in Table 4, for the smart meter with abnormal points, the accuracy of the fault traceability test is 86.7%. However, for the smart meter without abnormality, there is also some judgment deviation in fault traceability. For example, for the normal meter with ID 1006581, the traceability result is fault. Overall, the proposed method has a traceability accuracy of over 85% and has high detection accuracy.

6.5 Comparative analysis of different traceability methods

To demonstrate the performance of the proposed method, it was compared with the studies by Fu et al. [15] and Wang et al. [17]. Select 120 measuring devices to be analyzed and divide them into six groups. Use four methods to perform traceability analysis on the six groups of devices. The accuracy of the actual status and judgment results calculated is shown in Figure 9.

Figure 9

Traceability analysis results of different methods.

From Figure 9, it can be seen that the proposed method has the highest overall traceability analysis accuracy, and the data in each group has little fluctuation and good stability. The Byzantine fault-tolerant algorithm for the alliance chain proposed by Fu et al. [15] has certain requirements for system settings, resulting in poor analysis performance, low accuracy, and significant fluctuations. Both Wang et al. [17] and Li et al. [18] involve reasonable traceability strategies, ensuring the reliability of traceability results with an accuracy rate of approximately 80%. However, compared to the proposed method, its analysis results are unstable, and accuracy still needs to be improved.

Based on the above analysis, the analysis accuracy and traceability time of different traceability methods are shown in Table 5.

From Table 5, it can be seen that the accuracy and time consumption of the proposed method’s traceability results are 88.72% and 3.949 s, respectively. Compared with other methods, the traceability accuracy has a significant advantage. This is because it combines SDAE and GBRBM-DBN models, ensures the accuracy of the measurement device’s detection of data anomalies through the GBRBM-DBN model, and extracts high-precision features in combination with SDAE to ensure the reliability of the traceability results at multiple levels. However, this method combines two deep learning algorithms, which take a long learning time and therefore increase the time consumption compared to some methods. Fu et al. [15] use the Byzantine fault-tolerant algorithm based on the alliance chain for traceability, which is simple and easy to implement. Therefore, the traceability time is short, only 2.196 s, but the accuracy is less than 70%. Wang et al. [17] determined the defect cause tag through text similarity calculation, extracted defect features using L1 regularization and XGBoost integration methods, and determined the fault source based on D-S evidence theory. Although this method improves the traceability accuracy compared to Fu et al. [15], the lack of defect identification methods directly affects the traceability effect, reducing the accuracy by 12.77% compared to the proposed method. Li et al. [18] proposed the LSTM-based quality early warning method and the quality data backtracking query optimization algorithm and identified defects through the LSTM model. However, compared with the SDAE model of the proposed method, the query optimization algorithm has poor learning performance, so the traceability accuracy is reduced by 6.05%. Due to the short time consumption of the query optimization algorithm, the traceability time of this method is reduced to 2.571 s.