An elevator abnormal energy consumption identification and alarm method

Abstract

The application discloses a kind of identification and alarm method of elevator abnormal energy consumption, comprising the following steps: training data set is input into LSTM neural network and is trained and tested, and the trained prediction model and prediction error are output;Through the preliminary screening of prediction error to elevator energy consumption abnormal data, on this basis, 0 data is identified, k-means clustering and CART algorithm are used to perceive energy abnormal data, and the decision-making process in early warning strategy is proposed by using the way of typical scene analysis to detect mutation data, and alarm decision is made to abnormal data by combining artificial rule.The present application can effectively identify 0 data, energy abnormal data and mutation data, and set alarm decision to corresponding abnormal data, so that false positive rate and false negative rate are lower than traditional alarm strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of elevator detection technology, and particularly relates to a method for identifying and alarming abnormal energy consumption of an elevator. BACKGROUND

[0002] Due to the complexity of the elevator operation system, the actual energy consumption has the characteristics of segment fluctuation and unknown. Due to this problem, the implementation of energy consumption alarm becomes extremely difficult. In view of this problem, the current strategies for solving the identification and alarm of abnormal energy consumption mainly include statistical methods and data mining methods:

[0003] (1) Statistical method

[0004] The basic principle of the statistical method is to approximate fit the preprocessed samples by a mathematical distribution method, and then use statistical reasoning to judge whether the target data deviates from the distribution interval, and identify the abnormality according to the inconsistency of the distribution region.

[0005] (2) Data mining method

[0006] The data mining method includes clustering analysis, classification identification and nearest neighbor analysis. The essence of the clustering analysis method is to approximate separate the samples by an unsupervised analysis method, and to distinguish into several classification items, and the identification of abnormal data is through a method different from the known classification items, and the principle is to determine whether the distance value of the data to the nearest cluster center point exceeds the clustering range. The common method is fuzzy C-means clustering.

[0007] The above methods are essentially an analysis of the data fluctuation characteristics, and the research on energy consumption alarm is mainly based on the judgment and identification of energy consumption abnormalities. In practice, it is known that the energy consumption alarm of the elevator mainly relies on the artificial experience to set the energy consumption threshold, and the subsequent judgment of energy consumption abnormalities is mainly based on whether the energy consumption fluctuation is within the threshold. Although this method is simple in calculation process and can quickly determine the energy consumption abnormal data, the disadvantage of this method is that the artificial subjective deviation is large, and it is difficult to determine the threshold value. Improper value will have a very bad impact on the decision. If the threshold is small, the false alarm rate will increase, and if the threshold is large, the missed alarm rate will increase. On the one hand, it increases the probability of false alarm, and on the other hand, it cannot immediately discover the failure of the device. An inaccurate threshold value cannot perceive the energy consumption change. Therefore, it is urgent to provide a method for identifying and alarming abnormal energy consumption of an elevator. SUMMARY

[0008] In order to solve the above technical problems, the present application provides a method for identifying and alarming abnormal energy consumption of an elevator, which can effectively identify normal data, abnormal energy consumption data and mutation data, and set corresponding alarm decisions for the corresponding abnormal data, so that the false alarm rate and the missed alarm rate are lower than those of the traditional alarm strategy.

[0009] In order to achieve the above object, the technical scheme of the present application is as follows:

[0010] An elevator abnormal energy consumption identification and alarm method, comprising the following steps:

[0011] Step S1, obtaining a training data set, wherein the training data set comprises elevator historical energy consumption data;

[0012] Step S2, inputting the training data set into an LSTM neural network for training and testing until a preset requirement is met, and outputting a trained prediction model and a prediction error;

[0013] Step S3, inputting current actual collected data into the trained prediction model for prediction, and obtaining a prediction curve and a prediction data set;

[0014] Step S4, judging whether the next time actual collected elevator energy consumption data is 0 data or null data, if yes, determining that the next time actual collected elevator energy consumption data is 0 data abnormal data and recording; if not, turning to step S5;

[0015] Step S5, judging whether the absolute value of the difference between the next time actual collected elevator energy consumption data and the corresponding prediction data in the prediction data set is greater than the prediction error, if yes, turning to step S6; if not, turning to step S7;

[0016] Step S6, determining the energy consumption mode in which the next time actual collected elevator energy consumption data is located by using k-means clustering and CART algorithm, and judging whether the next time actual collected elevator energy consumption data is energy consumption abnormal data under the current energy consumption mode, if yes, determining that the next time actual collected elevator energy consumption data is energy consumption abnormal data and recording, if not, turning to step S7;

[0017] Step S7, judging whether the next time actual collected elevator energy consumption data meets a preset mutation rule, if yes, determining that the next time actual collected elevator energy consumption data is mutation data and recording, if not, determining that the next time actual collected elevator energy consumption data is normal data.

[0018] Preferably, the elevator historical energy consumption data is obtained and preprocessed, and the processed elevator historical energy consumption data is divided into a training data set by a sliding window method.

[0019] Preferably, the prediction error is half of the MAE value of the training data set.

[0020] Preferably, the k-means clustering and CART algorithm are used to determine the energy consumption mode in which the next time actual collected elevator energy consumption data is located, and the method comprises the following steps:

[0021] Based on k-means clustering, the historical elevator energy consumption data set is clustered and analyzed to identify different energy consumption modes;

[0022] Based on the CART algorithm, a classification model is constructed, and the collected elevator energy consumption data is determined by the classification model.

[0023] Preferably, it is judged whether the actual collected elevator energy consumption data at the next moment is abnormal data under the current energy consumption mode, which specifically includes the following steps:

[0024] During the clustering analysis process, the clustering center and data point distribution of the energy consumption mode in which the actual collected elevator energy consumption data at the next moment is located are obtained;

[0025] The difference between the farthest data point in the energy consumption mode cluster in which the actual collected elevator energy consumption data at the next moment is located and the cluster center point is calculated, and the absolute value of the calculated difference is taken as the maximum clustering range under the corresponding energy consumption mode. It is judged whether the absolute value of the difference between the actual collected elevator energy consumption data Di at the next moment and the cluster center point Cm exceeds the maximum clustering range Rm.

[0026] Preferably, the preset mutation rule includes:

[0027] Rule 1: P i-1 , P i , P i+1 are three consecutive values in the prediction curve, T i-1 , T i , T i+1 are consecutive actual collected data, when there are P i-1 <P i <P i+1 or P i-1 >P i >P i+1 , if:

[0028] T i >P i or T i <P i

[0029] T i is considered as energy consumption mutation data;

[0030] Rule 2: P i-1 , P i , P i+1 are three consecutive values in the prediction curve, T i-1 , T i , T i+1 are consecutive actual collected data, when there is a monotonically increasing prediction curve, if:

[0031] T i-1 <P i <T i and T i <P i+1 <T i+1

[0032] or for the case of monotone decreasing of the prediction curve, the same reasoning follows as:

[0033] T i-1 >P i >T i and T i >P i+1 >T i+1

[0034] then T i-1 , T i are energy consumption mutation data;

[0035] Set rule 3: P i-1 , P i , P i+1 are three consecutive values in the prediction curve, T i-1 , T i , T i+1 are consecutive actual collected data, if there are:

[0036] T i-1 >P i -1and T i >P i and T i+1 >P i+1

[0037] Continue to judge the next time, until a time k is found, so that the actual value restores the prediction trend, and the difference between them is within the prediction error range, at time (i-1, k), it is considered that the actual value has a growth mutation in the time period, wherein T i-1 ,T i ,…,T k are energy consumption mutation data;

[0038] Set rule 4: P i-1 , P i , P i+1 are three consecutive values in the prediction curve, T i -1, T i , T i+1 are consecutive actual collected data, if there are:

[0039] T i-1 <P i-1 and T i <P iand T i+1 <P i+1

[0040] The next moment is continuously determined until a moment k is found, such that the actual value restores the predicted trend, and the difference value is within the predicted error range, at the moment (i-1, k), it is considered that the actual value has a reducing mutation in the time period, wherein T i-1 ,T i ,…,T k is the energy consumption mutation value.

[0041] Preferably, the method further comprises the following steps:

[0042] The number of continuously detected 0 data abnormal data is counted, and when the number of continuously detected abnormal data exceeds M0, a 0 data alarm signal is triggered;

[0043] The number of continuously detected energy consumption abnormal data is counted, and when the number of continuously detected energy consumption abnormal data exceeds M 用 , an energy consumption data abnormal alarm signal is triggered;

[0044] The number of continuously detected mutation data according to the set rules 2, 3 or 4 is counted, and when the number of continuously detected mutation data exceeds M 突 , a mutation data alarm signal is triggered.

[0045] Preferably, the M0 is an integer not less than 6; M 用 and M 突 are both integers not less than 2.

[0046] Preferably, the method further comprises the following steps: sending the 0 data alarm signal, the energy consumption data abnormal alarm signal or the mutation data alarm signal to the relevant elevator maintenance personnel.

[0047] Based on the above technical scheme, the beneficial effects of the present application are: the present application preliminarily screens the elevator energy consumption abnormal data by setting the prediction error value, and further analyzes the suspected abnormal energy consumption data on this basis. The abnormal data is divided into three types of 0 data, energy consumption abnormal data and mutation data, the data mining method based on k-means clustering and decision tree is used to deeply perceive the energy consumption abnormal data, the mutation data is detected by using the typical scene analysis method, and the decision-making process in the early warning strategy is proposed by combining the artificial rules to make alarm decision for the abnormal data. Finally, the above work is applied to the actual scene and the results are analyzed, and more accurate alarm results are obtained, and the false alarm rate and the missed alarm rate are lower than those of the traditional alarm strategy. The energy consumption alarm strategy can be embedded into the elevator control system to help the maintenance personnel to accurately judge the energy consumption abnormality of the elevator, reduce the probability of abnormal false alarm, and improve the safety of the elevator and increase the life cycle of the elevator. From data acquisition to data analysis, first "monitoring" and then "analysis", from "energy consumption management" to "abnormal judgment", the management and technology are combined, which provides a scheme and theoretical research with reference significance for the in-depth application of elevator energy consumption supervision, and promotes the establishment and perfection of the elevator management system. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 It is a flow chart of an elevator abnormal energy consumption identification method in an embodiment;

[0049] Figure 2 It is a flow chart of an energy consumption abnormal data determination method in an embodiment;

[0050] Figure 3 It is a single-point mutation curve in an embodiment, wherein (a) is a scenario higher than the prediction value, and (b) is a scenario lower than the prediction value;

[0051] Figure 4 It is a multi-point cross mutation curve in an embodiment;

[0052] Figure 5 It is a growth mutation curve in an embodiment;

[0053] Figure 6 It is a reduction mutation curve in an embodiment;

[0054] Figure 7 It is a flow chart of an alarm decision method based on artificial decision in an embodiment;

[0055] Figure 8 It is a structure diagram of an EEFG-LSTM-Attention neural network prediction model in an embodiment;

[0056] Figure 9 It is an error tracking flow chart in an embodiment;

[0057] Figure 10 is a structural schematic diagram of an EEFG-LSTM unit structure in an embodiment. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application.

[0059] As Figure 1 shown, the embodiment provides an elevator abnormal energy consumption identification and alarm method, including the following steps:

[0060] Step S1, obtaining a training data set, the training data set including elevator historical energy consumption data.

[0061] In the embodiment, the elevator historical energy consumption data is collected, including normal data and abnormal data, wherein the abnormal data is mainly divided into three categories: 0 data, mutation data and energy consumption abnormal data.

[0062] The elevator historical energy consumption data is preprocessed, and the processed elevator historical energy consumption data is divided by a sliding window method to form a training data set.

[0063] Step S2, inputting the training data set into an LSTM neural network for training and testing until a preset requirement is met, outputting a trained prediction model and a prediction error.

[0064] In the embodiment, an EEFG-LSTM-Attention neural network is specifically used for training and testing, and an error forgetting long short-term memory unit (EEFG-LSTM) and an attention mechanism (Attention-mechanism) are used to save the trained prediction model and the prediction error, wherein the prediction error is half of the MAE value of the training data set.

[0065] Specifically, the prediction model based on the EEFG-LSTM-Attention neural network is constructed, and the constitution thereof is respectively: a data input layer, three EEFG-LSTM unit layers, a Dropout layer, an Attention layer, two fully connected layers and an output layer. The specific structure of the prediction model is shown in Figure 8 . The prediction model is explained as follows:

[0066] (1) Input layer

[0067] The input layer is a three-dimensional array, and its form is: (samples, timestamp, features), representing the number of samples, time steps, and input dimensions, respectively. The lookback parameter can be understood as a time step here, and the input data is (x1, x2, x3,..., x n) is input data, and the prediction output value. The input dimension is represented as several features of energy consumption at a single time step. In this embodiment, the data only has average power as a sample feature, so features defaults to 1.

[0068] (2) Hidden layer

[0069] The first three layers of EEFG-LSTM unit layers in the hidden layer automatically learn the energy consumption data input, extract the features and related rule information of energy consumption changes, and fit the prediction curve through a large number of parameters. In this paper, the first layer of EEFG-LSTM unit layer is composed of 64 units; the second layer is composed of 128 EEFG-LSTM units; and the third layer is composed of 256 EEFG-LSTM units. After learning through three layers and passing information to the next layer, each EEFG-LSTM unit includes an input gate, a forgetting gate, and an output gate. The error between the predicted value at the historical time and the actual value at the time is introduced into the forgetting gate of the next hidden layer unit. The information size of the forgetting gate is updated through the error size, so as to further determine the forgetting of the historical information, and correspondingly improve the information size of the current input gate. The prediction accuracy of the peak value of the elevator energy consumption data can be effectively improved, that is, the historical information is dynamically adjusted through the error transmission to improve the accuracy of the current prediction value.

[0070] Specifically, the error forgetting long short-term memory unit LSTM (EEFG-LSTM) is introduced. In the LSTM unit, reducing the effect of historical time information on the current prediction is equivalent to reducing the LSTM unit state c(t-1) at t-1 time. This step can be realized by changing the input amount of the forgetting gate. Therefore, the forgetting gate adjusts the input amount according to the deviation between the predicted value and the actual value at t-1 time, and performs error forgetting. For this purpose, the EEFG-LSTM changes the absolute error between the predicted value and the actual value at t-1 time and the memory c(t-1) at the previous time as the input amount of the forgetting gate update. Such a change considers the adjustment effect of the deviation between the predicted value and the actual value at t-1 time on the forgetting gate at t time, as shown in Figure 9 .

[0071] In the EEFG-LSTM unit structure, as shown in Figure 10 , the activation function of the forgetting gate is adjusted, and the sigmoid function in the standard LSTM is replaced by the ReLU function. The reason for this is that the derivative value of the sigmoid and tanh activation functions is less than 1, which is prone to gradient disappearance phenomenon.

[0072] In the formula, K = |h t-1 -h t-1 |, h t-1 is the actual value at t-1 time; h t-1is the predicted value at time t-1; C t-1 is the memory state at the previous time; ReLU is a new activation function instead of sigmoid.

[0073] f t = ReLU(W f · [K, C t-1 ]+b f )

[0074] i t = ReLU(W i · [h t-1 ,x t ]+b i )

[0075] C t '= tanh(W c [h t-1 ,x t ]+b c )

[0076] C t = f t · C t-1 +i t · C t '

[0077] O t = ReLU(W o · [h t -1,x t ]+b o )

[0078] h t = O t · tanh(C t )

[0079] The formula of ReLU activation function:

[0080] ReLU(x) = max(0, x)

[0081] In the above formula, f t forget gate, according to the state K at the previous time and the input x t at the current time, the memory state at the current time C is calculated and added with the weight number b c , and finally the result value is calculated by ReLU function.

[0082] i t input gate, according to the state h t-1 at the previous time and the input x t at the current time, the memory state at the current time W i is calculated and added with the weight number b i, and the result value is calculated by ReLU function.

[0083] C` t The current memory output is calculated according to the state h t-1 and the input x t at the previous moment, and the memory state W c at the current moment, and the weight b c is added, and finally the result value is calculated by tanh function.

[0084] C t The memory output is calculated by multiplying the memory output at the previous moment with the forgetting gate, and adding the product of the input gate and the current memory output.

[0085] O t The output gate is calculated according to the state h t-1 and the input x t at the previous moment, and the memory state W o at the current moment, and the weight b o is added, and finally the result value is calculated by ReLU function.

[0086] h t is the product of the output gate O t and the C t calculated by tanh function.

[0087] h t = O t · tanh(C t )

[0088] The fourth layer of the hidden layer is the Dropout layer, which mainly functions to randomly omit the connection between the output of the previous LSTM layer and the input of the next layer, preventing overfitting phenomenon in the training process. In this paper, the parameter value of this layer is 0.1. Dropout is a common method to prevent overfitting in machine learning.

[0089] The fifth layer of the hidden layer is the Attention layer. The influence of multi-step input on output is not the same, which affects the prediction result. The Attention mechanism is used to assign weights to different time step inputs to improve the energy consumption prediction effect. The main principle is to save the intermediate results generated by the LSTM network for time series input, and link this result with the output value, so that the model learns how to selectively focus on data and assign more reasonable weights to these data. The number of units is 50.

[0090] The sixth and seventh layers of the hidden layer are two fully connected layers (Dense Layer) with 50 and 1 units, respectively. The number of units in the sixth layer depends on the training effect and model convergence speed, and its main function is to extract features from the information transmitted by the previous layers. After the non-linear transformation of the Dense layer, the correlation between the energy consumption data can be extracted, and finally mapped on the seventh layer. The seventh layer has only one unit, because the problem solved by the prediction model is a regression problem, and the output value is the result, not a classification result, so a Dense unit is used to output the predicted energy consumption value.

[0091] (3) Output layer

[0092] The output of the output layer is the output of the seventh layer of the hidden layer after ReLU activation function calculation and output.

[0093] Parameter summary table of the prediction model based on EEFG-LSTM-Attention neural network

[0094]

[0095] Step S3: input the current actual collected data into the trained prediction model for prediction to obtain the prediction curve and the prediction data set.

[0096] Step S4: determine whether the actual collected elevator energy consumption data at the next time is 0 data or null data. If yes, it is determined that the actual collected elevator energy consumption data at the next time is 0 data abnormal data and is recorded. If not, go to step S5.

[0097] In this embodiment, the elevator has the phenomenon of stopping maintenance or communication equipment damage, so that the data transmission is 0 data or null data, and such data needs to be attributed to 0 data abnormality.

[0098] Step S5: determine whether the absolute value of the difference between the actual collected elevator energy consumption data at the next time and the corresponding predicted data in the prediction data set is greater than the prediction error. If yes, go to step S6; if not, go to step S7.

[0099] In this embodiment, since MAE represents the average value of the absolute error between the actual value and the predicted value, the setting of this value represents that the actual curve has a certain fluctuation interval. If the absolute value of the difference between the actual collected elevator energy consumption data Di at the next time and the corresponding predicted data Pi in the prediction data set is greater than e / 2, e is the MAE value, it means that this point may have abnormal conditions and needs to be further determined.

[0100] Step S6, the k-means clustering and CART algorithm are used to determine the energy consumption mode in which the actual energy consumption data of the elevator collected at the next time point is located, and whether the actual energy consumption data of the elevator collected at the next time point is abnormal energy consumption data in the current energy consumption mode is judged, if yes, the actual energy consumption data of the elevator collected at the next time point is determined as abnormal energy consumption data and recorded, if not, step S7 is entered.

[0101] In this embodiment, considering that the energy consumption sequence of the elevator is a single variable time sequence, its number increases linearly with time. Because of its large data volume and fast growth rate, a method capable of processing large data and finding data features needs to be found. Therefore, the k-means clustering and CART algorithm are used to solve the problem.

[0102] Abnormal energy consumption data is a radical data of the energy consumption of the elevator, which is different from the small range of data change value. The difference between the data and the normal predicted value is large. The threshold estimation method in the traditional alarm strategy will miss such phenomena to some extent. After all, the total energy consumption threshold or the weighted summation method cannot reflect obvious abnormalities in the overall energy consumption. If the energy consumption values are not checked one by one by artificial, the abnormal point is easy to be ignored. As shown in FIG. 1, the abnormal energy consumption data determination specifically includes the following steps: Figure 2

[0103] 1) Cluster analysis is performed on the historical elevator energy consumption data set, and the clustering method based on k-means is used to divide the elevator energy consumption into four energy consumption modes (low, medium, medium-high, and high). The number of classifications is derived from the running mode of the elevator. In the elevator dispatching strategy, the running mode of the elevator is classified according to the characteristics of the passenger flow. Generally, it is divided into up peak, down peak, inter-floor traffic, and idle mode. The energy consumption of the elevator is different in different running modes.

[0104] 2) The collected data is subjected to energy consumption mode determination. This method is to define the range between energy consumption data. Energy consumption data is not comparable in different characteristic environments. In this paper, the CART algorithm is used to determine the energy consumption mode. The historical elevator energy consumption data with labels is classified and trained to find the related factors that determine the energy consumption mode in the data, construct an energy consumption mode determination model, and generate a determination rule.

[0105] 3) The cluster center obtained in the cluster analysis is used, and the difference between the farthest data point in the cluster and the cluster center point is calculated. The calculated difference is used as the maximum cluster range Rm in the energy consumption mode. Then, it is determined whether the absolute value of the difference between the actual energy consumption data Di of the elevator collected at the next time point and the cluster center point Cm exceeds the maximum cluster range Rm, if yes, it is abnormal energy consumption data.

[0106] ​K-means clustering is a clustering analysis and a classification method. Its essence is to divide target objects into several categories, and ensure that the difference between objects in the same category is minimized, while the difference between objects in different categories should be maximized. The clustering analysis has an advantage that it starts from the data itself without considering the relevant prior knowledge before division. The algorithm is fast, accurate and simple to use in large sample data sets. The main steps of the k-means clustering algorithm are as follows:

[0107] (a) Input n data {i1, i2, i3, …, in} and give the number of categories k; n

[0108] (b) Optionally, select k cluster center points {c1, c2, c3, …, ck}; k

[0109] (c) For each data, calculate the distance between it and the k center points, and take the nearest distance as the cluster center point of the current data;

[0110] (d) After completing a round of clustering analysis, recalculate the cluster center points c j , which is exactly equal to the average value of all objects in the category; j

[0111] (e) Repeat the above two steps until the clustering termination condition E is met, as shown below.

[0112]

[0113] In the formula, i represents the input data point, c represents the different cluster centers, each input value is subtracted from the cluster center, the absolute distance square of each combination is calculated, and the termination result E is obtained by summation.

[0114] The essence of CART algorithm is the processing mode of decision tree, which is a discipline of pattern classification. Its essence is to learn from the labeled data to form a classifier with internal rule system. The general classifier training process is to train the labeled data with selected classification algorithm, and the minimum loss value is the training termination condition. After the energy consumption data is labeled, the new energy consumption data can be classified by the trained classifier, which is convenient for subsequent energy consumption data analysis and detection.

[0115] CART algorithm details:

[0116] Assume that the sample set D contains K classes of data, P m ​​​Let Gini represent the probability that data belongs to the k-th class. The Gini index for a node A is defined as follows: When all samples belong to the same class, Gini(A) = C / (C-1) / 2, and the Gini index is at its maximum when all classes appear with equal probability in the node.

[0117]

[0118] In the formula, the Gini index of node A is calculated by subtracting the total misclassification probability from the total probability, and the probability of class k from class m-1 to class M.

[0119] If the sample set D is divided into D1 and D2 according to the value of node X, the Gini value of this partition is:

[0120]

[0121] Where D, D1, and D2 represent the number of samples in sample sets D, D1, and D2, respectively.

[0122] For each feature in the sample set, CART calculates the Gini coefficient of each feature, finds the smallest Gini coefficient as the best split point for the corresponding feature, and then compares the Gini coefficients of the best split points of all features. The feature with the smallest split Gini coefficient is taken as the optimal feature, forming the final classification rule.

[0123] Table 4.1 CART Algorithm Steps

[0124]

[0125] Step S7: Determine whether the elevator energy consumption data actually collected at the next moment meets the preset mutation rule. If yes, determine that the elevator energy consumption data actually collected at the next moment is mutation data and record it. If no, determine that the elevator energy consumption data actually collected at the next moment is normal data.

[0126] In this embodiment, since the mutation data is mainly determined by the monotonicity of the predicted curve and the fluctuation of the actual curve, it is analyzed through typical abnormal scenarios, which mainly fall into four categories: single-point mutation, multi-point crossover mutation, growth mutation, and reduction mutation. Corresponding mutation rules are set for different scenario types:

[0127] 1) Single point mutation

[0128] Single-point mutation data refers to a situation where the actual curve changes under the monotonic trend of the predicted curve, but there is an abrupt point in the process where energy consumption exceeds the prediction error range, and there is only one such point. Afterwards, the actual curve returns to the same trend as the predicted curve. This is called a single-point mutation. (See attached image) Figure 3 As shown. Based on this scenario, rule 1 is established: Assume P i-1 Pi , P i+1 are three consecutive values in the predicted curve, T i-1 , T i , T i+1 are consecutive actual collected data. When there is a monotone increasing of the predicted curve, if there is: i-1 < P i < P i+1 or P i-1 > P i > P i+1 , then T

[0129] T i > P i or T i < P i

[0130] T i are energy consumption mutation values.

[0131] 2) Multi-point cross mutation

[0132] Multi-point cross mutation is very similar to single-point mutation, the biggest difference is that multi-point mutation, and is continuous. From the attached Figure 4 , it can be seen that at 3 o'clock and 4 o'clock, there are mutation phenomena. This phenomenon is a more complex mutation phenomenon. Set rule 2: assuming P i-1 , P i , P i+1 are three consecutive values in the predicted curve, T i-1 , T i , T i+1 are consecutive actual collected data. When there is a monotone increasing of the predicted curve, if there is:

[0133] T i-1 < P i < T i and T i < P i+1 < T i+1

[0134] or for the monotone decreasing of the predicted curve, the same.

[0135] T i-1 > P i > T i and T i > P i+1 > T i+1

[0136] T i-1 , T i are energy consumption mutation values.

[0137] 3) Growth mutation

[0138] A growth mutation occurs when the actual curve is consistently higher than the predicted curve over a certain period, with the difference exceeding the prediction error, and then, after a certain time, reverts to the same trend as the predicted curve. This type of mutation is defined as a growth mutation. (From the appendix...) Figure 5 In the data, between time points 2 and 5, there is a significant data discrepancy between the actual and predicted values, and the predicted trend begins to recover at time point 6. Rule 3 is set: if there is...

[0139] T i-1 >P i-1 and T i >P i and T i+1 >P i+1

[0140] Continue judging the next time step until a time step k is found where the actual value recovers the predicted trend and the difference between the two is within the prediction error range. At time (i-1, k), it is considered that the actual value has undergone a sudden increase during this time period, where T i-1 ,T i ,…,T k It is the energy consumption mutation value.

[0141] 4) Reduced sex mutations

[0142] In contrast to growth mutations, rule 4 is set: if it exists

[0143] T i-1 <P i-1 and T i <P i and T i+1 <P i+1

[0144] Continue judging the next time step until a time step k is found where the actual value recovers the predicted trend and the difference between the two is within the prediction error range. At time (i-1, k), it is considered that the actual value underwent a decrease abruptly during this time period, where T i-1 ,T i ,…,T k This is a sudden change in energy consumption value; see [link / reference]. Figure 6 .

[0145] like Figure 7 As shown, in one embodiment of an elevator abnormal energy consumption identification and alarm method, an alarm method is also provided. In real-world scenarios, a small number of abnormal points are insufficient to require manpower for inspection; an alarm signal is only triggered after a certain number of occurrences. Therefore, based on practical engineering experience, different decision thresholds are set for different abnormal data to reduce the false alarm rate. Specifically, the steps include the following:

[0146] 0 data, generally there are stop maintenance, communication or metering equipment failure. Suppose after continuous detection of 0 data more than 8 times, trigger 0 data alarm signal. According to the time as the time unit, the time of small repair will not exceed this period, unless there is a large repair or equipment failure.

[0147] With energy anomaly data, if it appears a few times in a short period of time, it means that the elevator running state is still healthy, and occasional fluctuations can be ignored. This consideration reduces the false positive rate to some extent, and reduces the disturbance problem of maintenance personnel. However, if it appears continuously for many times in a short period of time, it means that the elevator running state becomes unstable and needs to be repaired immediately. According to engineering experience, set the continuous detection of energy anomaly data more than 2 times to trigger the energy data anomaly alarm signal.

[0148] Mutation data, unlike the change range of energy anomaly data, is characterized by small fluctuations and high frequency. The premise of judging abnormal data is that the collected data does not exceed the prediction range of the corresponding predicted data, that is, the prediction error. Under this condition, the number of mutation data will be more. Single-point mutation data does not need to trigger an alarm signal. For multi-point cross, continuous increase and decrease mutation data, threshold needs to be set for alarm prevention. According to engineering experience, set the continuous mutation data more than 2 times to trigger the mutation data alarm signal.

[0149] The above only describes a preferred embodiment of the elevator abnormal energy consumption recognition and alarm method disclosed by the present application, and is not used to limit the protection scope of the embodiments of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. A method for identifying and alarming abnormal energy consumption in elevators, characterized in that, Includes the following steps: Step S1: Obtain the training dataset, which includes historical energy consumption data of elevators; Step S2: Input the training dataset into the LSTM neural network for training and testing until the preset requirements are met, and output the trained prediction model and prediction error. Step S3: Input the currently collected data into the trained prediction model to make a prediction and obtain the prediction curve and prediction dataset. Step S4: Determine whether the actual elevator energy consumption data collected in the next moment is 0 data or null data. If so, determine that the actual elevator energy consumption data collected in the next moment is 0 data, which is abnormal data, and record it. If not, proceed to step S5. Step S5: Determine whether the absolute value of the difference between the actual elevator energy consumption data collected at the next moment and the corresponding predicted data in the predicted dataset is greater than the prediction error. If yes, proceed to step S6; otherwise, proceed to step S7. Step S6: Use k-means clustering and CART algorithm to determine the energy consumption mode of the elevator energy consumption data collected in the next moment, and determine whether the elevator energy consumption data collected in the next moment is abnormal energy consumption data under the current energy consumption mode. If so, determine the elevator energy consumption data collected in the next moment as abnormal energy consumption data and record it. If not, proceed to step S7. Step S7: Determine whether the elevator energy consumption data actually collected at the next moment meets the preset mutation rule. If yes, determine that the elevator energy consumption data actually collected at the next moment is mutation data and record it; otherwise, determine that the elevator energy consumption data actually collected at the next moment is normal data. The preset mutation rule includes: Rule 1: P i-1 , P i , P i+1 To predict three consecutive values ​​in the curve, T i-1 , T i , T i+1 For continuous, actual collected data, when there exists P i-1 < P i < P i+1 or P i-1 > P i > P i+1 If the following situations apply: T i > P i or T i < P i Then it is considered T i This is data on sudden changes in energy consumption; Rule 2: P i-1 , P i , P i+1 To predict three consecutive values ​​in the curve, T i-1 , T i , T i+1 For continuously collected actual data, if the prediction curve shows a monotonically increasing trend, then: T i-1 <P i <T i and T i <P i+1 <T i+1 Or, for the case where the prediction curve is monotonically decreasing, the same logic applies: T i-1 >P i >T i and T i >P i+1 >T i+1 Then see T i-1 , T i This is data on sudden changes in energy consumption; Rule 3: P i-1 , P i , P i+1 To predict three consecutive values ​​in the curve, T i-1 , T i , T i+1 For continuous, actual collected data, if: T i-1 >P i-1 and T i >P i and T i+1 >P i+1 Continue judging the next moment until a moment is found. k This ensures that the actual value recovers to the predicted trend, and the difference between the two is within the prediction error range, over a time period ( i-1, k ), suggesting that the actual value experienced a sudden increase during this period, among which T i-1 , T i ,…,T k These are data on sudden changes in energy consumption; Rule 4: P i-1 , P i , P i+1 To predict three consecutive values ​​in the curve, T i-1 , T i , T i+1 For continuous, actual collected data, if: T i-1 <P i-1 and T i <P i and T i+1 <P i+1 Continue judging the next moment until a moment is found. k This ensures that the actual value recovers to the predicted trend, and the difference between the two is within the prediction error range, over a time period ( i-1 , k ), suggesting that the actual value underwent a decrease during this period, among which T i-1 , T i ,…,T k It is the energy consumption mutation value.

2. The method for identifying and alarming abnormal energy consumption in elevators according to claim 1, characterized in that, Historical energy consumption data of elevators is acquired and preprocessed. The processed historical energy consumption data of elevators is then segmented using the sliding window method to form a training dataset.

3. The method for identifying and alarming abnormal energy consumption in elevators according to claim 1, characterized in that, The prediction error is half of the MAE value of the training dataset.

4. The method for identifying and alarming abnormal energy consumption in an elevator according to claim 1, characterized in that, The energy consumption pattern of the elevator energy consumption data collected at the next time step is determined using k-means clustering and the CART algorithm. The specific steps include the following: Cluster analysis was performed on the historical elevator energy consumption dataset based on k-means clustering to identify different energy consumption patterns; A classification model is built based on the CART algorithm, and the energy consumption mode is determined by the collected elevator energy consumption data.

5. The method for identifying and alarming abnormal energy consumption in an elevator according to claim 4, characterized in that, Determining whether the elevator energy consumption data collected at the next moment is abnormal energy consumption data under the current energy consumption mode involves the following steps: During the cluster analysis process, the cluster centers and data point distribution of the energy consumption pattern of the elevator energy consumption data collected at the next moment are obtained; The difference between the furthest data point in the energy consumption pattern cluster of the elevator energy consumption data collected at the next moment and the cluster center point is used. The absolute value of the calculated difference is taken as the maximum cluster range in the corresponding energy consumption pattern. Rm Determine the actual elevator energy consumption data collected at the next moment. Di with cluster center Cm Does the absolute value of the difference exceed the maximum clustering range? Rm .

6. The method for identifying and alarming abnormal energy consumption in an elevator according to claim 1, characterized in that, It also includes the following steps: The number of consecutive zero-data abnormal data detections is counted. When the number of consecutive zero-data abnormal data detections exceeds M0, a zero-data alarm signal is triggered. Count the number of times abnormal energy consumption data is continuously detected. When the number of times abnormal energy consumption data is continuously detected exceeds M... 用 This triggers an alarm signal for abnormal energy consumption data; Count the number of consecutive mutation detections that meet the criteria of rules 2, 3, or 4. When the number of consecutive mutation detections exceeds M... 突 This triggers a mutation data alarm signal.

7. The method for identifying and alarming abnormal energy consumption in an elevator according to claim 6, characterized in that, M0 is an integer not less than 6; M 用 and M 突 All are integers not less than 2.

8. The method for identifying and alarming abnormal energy consumption in an elevator according to claim 6, characterized in that, It also includes the following steps: sending 0 data alarm signals, abnormal energy consumption data alarm signals, or sudden data alarm signals to the relevant elevator maintenance personnel.

Images