A fault diagnosis and residual life prediction method for a construction site intelligent distribution box

Abstract

The application discloses a kind of fault diagnosis and residual life prediction method for construction site intelligent distribution box, comprising the following steps: data collected from distribution box is transmitted to cloud platform by wireless network;Build pre-processing model based on LSTM, fault diagnosis model of unsupervised HI based on DBN and residual life prediction model based on DBN and IPF;The data received by cloud platform is put into neural network for training;The components in distribution box are detected by the network model after training, whether a single part will fail in the future a period of time and the remaining use time are predicted, finally the training result is returned to cloud platform end display and whether to be artificially intervened.The application can better maintain the use of distribution box in construction site scene compared with prior art, predict the time of distribution box failure in advance, improve maintenance efficiency, and greatly reduce labor cost.

Description

Technical Field

[0001] This invention relates to the fields of construction site power distribution boxes and deep learning technology, and in particular to a method for fault diagnosis and remaining life prediction of intelligent power distribution boxes on construction sites. Background Technology

[0002] With rapid economic development, my country's electricity consumption has reached an unprecedented peak, bringing with it a large number of electrical safety accidents. Statistics show that from 2011 to 2016, my country experienced 524,000 electrical fires, resulting in 3,261 deaths, 2,063 injuries, and direct economic losses exceeding 9.2 billion yuan. These figures account for more than 30% of the total number of fires and related casualties nationwide. Therefore, ensuring safe electricity use is extremely important.

[0003] In recent years, the number of construction projects in China has been steadily increasing, leading to increasingly stringent requirements for safe electricity use on construction sites. Compared to electricity use in ordinary homes and factories, construction sites are more prone to electric shock accidents due to the variety of electrical equipment, large power capacity, unpredictable working environments, outdoor operations, and temporary use. Furthermore, construction site environments often present challenges such as strong electromagnetic interference, weak wireless communication, and extreme rainy weather. Traditional construction site distribution boxes only offer basic functions like power supply, leakage protection, overload protection, and short-circuit protection, and cannot monitor, control, protect, or analyze the electrical equipment within them. Additionally, the fixed installation locations on construction sites make it very difficult to relocate traditional distribution boxes, lacking portability. Therefore, to improve the safety of electricity use in power systems, intelligent distribution boxes have become a key focus of current research and development in safe electricity use technology.

[0004] With the rapid development of computer technology, computer vision has been integrated with numerous industries, achieving significant breakthroughs in fields such as facial recognition and vehicle recognition. However, its application in the construction industry is still in its early stages. Simultaneously, as the construction industry becomes increasingly sophisticated, higher demands are being placed on the safety of construction sites. Therefore, real-time identification and monitoring of various data from construction site electrical distribution boxes is of significant and practical importance. Currently, most construction sites still rely on manual labor to monitor electrical facilities such as distribution boxes, which inevitably leads to oversights and omissions. Furthermore, repairs are only carried out after a fault occurs, resulting in low efficiency and difficulty in guaranteeing accuracy. Therefore, a fault diagnosis and remaining life prediction method for intelligent distribution boxes on construction sites has significant practical value for strengthening the management of distribution boxes, ensuring the safety of workers and materials, predicting potential faults in advance, and reducing manual maintenance costs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for fault diagnosis and remaining life prediction of intelligent power distribution boxes on construction sites. This method enables the prediction of faults and remaining life of various components in the power distribution box, allowing for the timely detection of potential problems and the development of maintenance plans in advance. Furthermore, it allows for real-time monitoring of the remaining usage time of the power distribution box, effectively reducing the labor required for manual monitoring, improving maintenance efficiency, and promoting the construction of smart and safe construction sites.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for fault diagnosis and remaining life prediction of intelligent power distribution boxes on construction sites includes the following steps:

[0008] Step 1: Collect relevant physical quantities using the universal data acquisition device in the distribution box to obtain the collected data;

[0009] Step 2: Transmit the collected data to the cloud platform, and then feed the collected data into the constructed neural network through the cloud platform;

[0010] Step 3: Build a preprocessing model based on LSTM and an unsupervised HI fault diagnosis model based on deep neural network DBN, and improve the existing PF method and integrate it into DBN to predict the remaining usage time.

[0011] Step 4: Train all neural network models using the preprocessed data;

[0012] Step 5: Use the trained neural network to diagnose faults and predict the remaining lifespan of the intelligent power distribution box on the construction site, and finally return the results to the cloud platform.

[0013] Based on the above technical solution, the LSTM hidden layer contains one memory unit and three "gate" structures: an input gate, a forget gate, and an output gate; the loss function is chosen to be the mean-square error (MSE).

[0014]

[0015] Where h is the true value of the data, and h′ is the predicted value of the data;

[0016] Historical time series data is divided into training sets, standardized, and split into data segments to generate a one-dimensional vector {x}. k The k generated one-dimensional vectors are then input into the L hidden layers of an LSTM, and finally, the predicted parameters for a future time period are output. k , k = 1, 2, ..., L}.

[0017] Based on the above technical solution, the Deep Neural Network (DBN) is used to classify data and extract features of that data type. The DBN is composed of multiple stacked Restricted Block Models (RBMs). The network structure mainly consists of one explicit layer and one hidden layer. There are connections between layers, but units within a layer are independent. The explicit layer receives data, the hidden layer extracts features, the top two layers are undirected, and the bottom layer represents a data vector. The input to the DBN network is the output data of the LSTM neural network. k The output is a feature vector set constructed as the health index HI.

[0018] Based on the above technical solution, the training process of the deep neural network DBN is carried out layer by layer. Each layer uses data vectors to infer the hidden layer, and then uses the hidden layer as the input of the data vector of the next layer. After the deep neural network DBN is trained upward, the state of each layer is obtained and optimized through backpropagation.

[0019] Based on the above technical solution, the health index HI is obtained by feature quantification through unsupervised learning of the neural network DBN. The HI degradation process is the process in which a fault is about to occur. The features learned by the neural network DBN can be used to distinguish between fault states and degradation states in the data. That is, when the HI degradation is 0 under the fault state, the fault time FT is automatically obtained.

[0020] Based on the above technical solution, the remaining life prediction model is obtained by integrating a neural network DBN and an improved particle filter PF. The improved particle filter PF effectively improves particle diversity and avoids particle degeneracy by adding a fuzzy inference system to the basic particle filter, thereby improving resampling accuracy. The input of the improved particle filter PF is the constructed health index HI, which is used to predict the downward trend of the health index HI to obtain the remaining life time.

[0021] Based on the above technical solution, the improved particle filter PF is as follows:

[0022] Step A: Normalize the feature vector set output by the DBN neural network, then select an appropriate feature layer size and project the input variables to obtain a feature set z. i At the same time, a fault feature set z can also be obtained. fault Therefore, h can be represented i h i The value ranges from 0 to 1, where 0 indicates a fault has occurred and 1 indicates that everything is working normally.

[0023] h i =min(∥z) i -z fault ∥2)

[0024] Step B, construct h i As the input to the particle filter, i.e., the measurement vectors at different times, particle initialization is first performed. N represents the number of particles, and p(x0) is generated using prior probabilities; the dynamic equation used for particle updates is x k =f(x) k-1 )+w k-1 , z k =h(x k )+v k , where x k Let z represent the target state vector at time k. k This represents the measurement vector at time k; then the particle weights are updated. in Represents the importance density function, This represents the observation probability density corresponding to the system observation model. This indicates whether the state transition probability density corresponds to a system state transition; subsequently, the weights are smoothed using the MF equation (Membership Functions), and the processed weights are normalized to obtain... New particles are obtained by resampling at 1 / N. Ultimately, the new state can be estimated using the new particles and weights, i.e.:

[0025] The remaining life prediction model uses the currently estimated state The parameters are used to predict the future health index HI, and then the remaining usage time is obtained. Specifically, based on the health index HI having t measurements, the state is determined using the SMC method. The posterior probability density function at time t is δ(·) is the Dirac function, N is the number of particles, and x t This represents the device operation record input at time t. z represents the weight at time t. 0：t Representing the range from 0 to t, the probability density function predicted after k steps is: The current state prediction value is Finally, the particle state is extrapolated through the state model until the fault time FT is reached. The k time periods that are elapsed are the remaining lifetime time.

[0026] Preferably, the performance verification is performed by comparing the built network DBN-IPF (IPF: Improved PF) with the network DBN-PF. Firstly, performance metrics AI and PI are proposed, with corresponding formulas Al = |RUL|. t -RULe |, PI = H_CI - L_CI, where RUL t For the true remaining lifespan, RUL e To estimate the remaining lifetime, H_CI and L_CI represent the upper and lower bounds of the remaining lifetime estimation confidence interval at a 95% confidence level, respectively. Smaller values ​​for both AI and PI are preferred. A HI descent curve was constructed using 5000 particles as a reference. Experiments showed that the HI curve constructed by DBN-IPF was smoother. At the same initial prediction time (green line), the failure time probability density function estimated by DBN-IPF was higher and narrower than that of DBN-PF. The actual remaining lifetime in this experiment was calculated to be 108 periods, while the predicted periods by DBN-IPF and DBN-PF were 99 and 95, respectively. DBN-IPF performed better in terms of AI value. The calculated PI values ​​for DBN-IPF and DBN-PF were 47 and 42, respectively, with DBN-IPF performing better in terms of PI value. In conclusion, the DBN-IPF model constructed in this paper is innovative and performs better than traditional models.

[0027] Preferably, the fault prediction of the intelligent power distribution box at the construction site is performed through a neural network. If a fault is likely to occur, a warning is issued in advance through the cloud platform and corresponding maintenance measures are provided. The remaining life prediction will be displayed in real time on the terminal device through the cloud platform.

[0028] Because the present invention adopts the above-described technical solution, it has the following beneficial effects:

[0029] This invention transmits data collected from distribution boxes to a cloud platform via a wireless network; it constructs a preprocessing model based on LSTM (Long Short-Term Memory), an unsupervised HI (Health Indicator) fault diagnosis model based on DBN (Deep Belief Network), and a remaining life prediction model based on DBN and IPF (Improved Particle Filter); it feeds the data received from the cloud platform into the neural network for training; and it uses the trained network model to detect components in the distribution box, predicting whether a single component will fail and its remaining usage time in the future. Finally, the training results are returned to the cloud platform for display, indicating whether human intervention is needed. This invention combines construction site distribution boxes and deep learning network models, using corresponding neural networks to determine whether the distribution box will fail and to obtain its remaining usage time. Compared with existing technologies, it can better maintain the use of distribution boxes in construction site scenarios, predict the time of distribution box failure in advance, improve maintenance efficiency, and significantly reduce labor costs. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process of the present invention.

[0031] Figure 2(a) is a diagram of the LSTM neural network model of the present invention.

[0032] Figure 2(b) is a diagram of the LSTM hidden structure of the present invention.

[0033] Figure 3(a) is a diagram of the hidden layer framework of the neural network model DBN of the present invention.

[0034] Figure 3(b) is a diagram of the DBN neural network model of the present invention. Detailed Implementation

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, a method for fault diagnosis and remaining life prediction of intelligent distribution boxes on construction sites includes the following steps:

[0037] 1. Collect relevant physical quantities using a general-purpose data acquisition device in the distribution box to obtain the collected data;

[0038] 2. Transmit the collected data to the cloud platform, and then feed the collected data into the constructed neural network through the cloud platform;

[0039] 3. Build a preprocessing model based on LSTM and an unsupervised HI fault diagnosis model based on deep neural network DBN, and improve the existing PF method and integrate it into DBN to predict the remaining usage time.

[0040] 4. Train all neural network models using the preprocessed data;

[0041] 5. The trained neural network is used to diagnose faults and predict the remaining lifespan of the intelligent power distribution box on the construction site, and the results are finally returned to the cloud platform for manual processing.

[0042] The physical quantities collected by the general-purpose data acquisition device include voltage, current, line temperature, module temperature, energy consumption, residual current, power, and ambient temperature.

[0043] The proportional division of the dataset includes dividing it into a training set and a validation set, with the training set used for validation after the overall model is built.

[0044] The cloud platform is used to analyze and process the data of each component in the intelligent power distribution box at the construction site, and displays it in real time on the terminal.

[0045] The aforementioned neural network LSTM (Long-Short Term Memory) predicts the changing trends of various physical quantities over a future period using historical time series data, and uses this as the basis for fault diagnosis. Its model is shown in Figure 2(a). The historical time series data is divided into a training set, standardized, and segmented to generate a one-dimensional vector {x}. k Let k = 1, 2, ..., L. Then, these k historical data points are input into the L hidden layers of the LSTM. The hidden structure of the LSTM is shown in Figure 2(b), which includes an input gate, a forget gate, and an output gate. The forget gate is used to select and forget previously unused information; the input gate determines which relevant information from the previous memory unit needs to be stored in the neuron state; the output gate determines which part of the neuron state should be output. t h represents the device operation record input at time t. t-1 s represents the output at the previous time t-1. t This represents the current state value of the neuron at time t; the historical information matrix before time t serves as the neuron's state. Finally, the predicted parameter values ​​for a future time period are output. k , k = 1, 2, ..., L}.

[0046] The neural network DBN (Deep Belief Network) consists of multiple stacked Restricted Boltzmann Machines (RBMs). The network structure primarily comprises one explicit layer and one hidden layer, with connections between layers and no inter-layer connections between units. The explicit layer receives data, and the hidden layer extracts features. The top two layers are undirected, and the bottom layer represents the data vector. In the DBN, the output of the previous RBM serves as the input to the next RBM. Each layer uses the data vector to infer the hidden layer, and then uses the hidden layer as the input to the next layer. Figure 3(a) shows the hidden layer framework of the DBN model of this invention. The DBN is optimized using a greedy layered training method. The general training process is as follows: the DBN transmits the input data from top to bottom and extracts features, using the Contrastive Divergence (CD) algorithm for layered pre-training and iterative weight updates from top to bottom. Then, the DBN internally reconstructs the input data by transmitting features from top to bottom, using the mean squared error (MSE) between the input data and the reconstructed data as an evaluation metric, and fine-tuning the weights of the entire network is achieved through backpropagation. The overall DBN architecture is shown in Figure 3(b). The input to the DBN network is the output data of the neural network LSTM. k The output is a feature vector set constructed as the health index HI.

[0047] The Health Indicator (HI) is obtained by quantifying features learned through unsupervised learning of a deep neural network (DBN). The HI degradation process represents the imminent occurrence of a fault. The features learned by the DBN can be used to distinguish between fault states and degradation states in the data. Specifically, when the HI degrades to 0 under a fault state, the fault time (FT) is automatically obtained.

[0048] The feature vector set output by the deep neural network (DBN) is normalized, and then an appropriate feature layer size is selected. The input variables are then projected to obtain a feature set z. i At the same time, a fault feature set z can also be obtained. fault Therefore, h can be represented i h i The value ranges from 0 to 1, where 0 indicates a fault has occurred and 1 indicates that everything is working normally.

[0049] h i =min(∥z) i -z fault ∥2)

[0050] The remaining life expectancy prediction model is obtained by integrating a neural network DBN and an improved particle filter PF, which is used to predict the downward trend of the health index HI.

[0051] The particle filter PF is a Bayesian estimation-based algorithm that uses a large number of random samples and performs the Bayesian recursive filtering process through Monte Carlo simulation. Here, the particle filter PF is used to predict the descent process of HI. The particle filter PF suffers from severe particle degeneracy; the improved particle filter introduces a fuzzy inference system to increase particle diversity before resampling.

[0052] The constructed h i As the input to the particle filter, i.e., the measurement vectors at different times, particle initialization is first performed. N represents the number of particles, and p(x0) is generated using prior probabilities; the dynamic equation used for particle updates is x k =f(x) k-1 )+w k-1 , z k =h(x k )+v k , where x k Let z represent the target state vector at time k. k This represents the measurement vector at time k; then the particle weights are updated. in Represents the importance density function, This represents the observation probability density corresponding to the system observation model. This indicates whether the state transition probability density corresponds to a system state transition; subsequently, the weights are smoothed using the MF equation (Membership Functions), and the processed weights are normalized to obtain... New particles are obtained by resampling at 1 / N. Ultimately, the new state can be estimated using the new particles and weights, i.e.:

[0053] The remaining lifetime prediction model predicts the future HI based on the current estimated state and parameters, thereby obtaining the remaining usage time.

[0054] First, assume that HI consists of t measurements. Based on the Sequential Monte Carlo (SMC) method, the posterior probability density function of the state variable at time t is: δ(·) is the Dirac function, N is the number of particles, and x t This represents the device operation record input at time t. z represents the weight at time t. 0：t Representing the range from 0 to t, the probability density function predicted after k steps is: The predicted value of the time state is Finally, the particle state is extrapolated through the state model until the fault time FT is reached. The k time intervals that are elapsed are the remaining lifetime time.

[0055] The system uses neural networks to diagnose faults in intelligent power distribution boxes at construction sites, issues warnings via a cloud platform, and provides corresponding human intervention. The remaining lifespan prediction will display the available time in real time on terminal devices via the cloud platform.

[0056] This invention combines a construction site electrical distribution box with a deep learning network model, using a corresponding neural network to determine whether a fault will occur and to obtain the remaining usage time of the distribution box. Compared with existing technologies, it can better maintain the use of distribution boxes in construction site scenarios, provide early warnings of possible faults, improve maintenance efficiency, and greatly reduce labor costs.

[0057] The above description represents preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technical or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for fault diagnosis and residual life prediction of a job site intelligent distribution box, characterized in that, The method comprises the following steps: Step 1, collecting relevant physical quantities through a general collector in the power distribution box to obtain collected data; Step 2, transmitting the collected data to a cloud platform, and placing the collected data into a built neural network model through the cloud platform; Step 3, building a pretreatment model based on a long short-term memory network (LSTM), a health index (HI) fault diagnosis model based on a deep neural network (DBN), and an improved particle filter (PF) method to predict the remaining useful time; Step 4, training all neural network models by using the pretreated data; Step 5, diagnosing faults and predicting the remaining life of the intelligent power distribution box through the trained neural network model, and finally returning the results to the cloud platform; The improved particle filter (PF) is as follows: Step A, the feature vector set output by the neural network DBN is normalized, and then a suitable feature layer size is selected to project the input variable to obtain a feature set , and a fault feature set is obtained , so as to represent , the value of which is between 0 and 1, 0 indicating reaching failure and 1 indicating complete normality: Step B, the constructed As the input of particle filter, the measurement vector of different time is firstly initialized , , represents the number of particles, and is generated using the prior probability ; the dynamic equation used in particle update is , , where represents the target state vector at time k, represents the measurement vector at time k; then the particle weight is updated , where represents the importance density function, represents the observation probability density corresponding to the system observation model, represents whether the state transition probability density corresponds to the system state transition; subsequently, the weight is smoothed through the MF equation, and the processed weight is normalized to obtain ; Obtain new particles according to 1 / N resampling , Finally, estimate the new state using the new particles and weights: ; The remaining life prediction model predicts the future health index HI through the current estimated state and the posterior probability, and then obtains the remaining use time, specifically, according to the health index HI with t measurement values, based on the SMC method, the state The posterior probability density function at time t is , is the Dirac function, is the particle number, represents the current device operation record input at time t, represents the weight at time t, represents from 0 to t, thereby predicting the probability density function after t steps , at this time the state prediction value is , and finally the future particle state is extrapolated through the formula to obtain the HI value at the future time until the failure time FT is reached. The k time periods experienced are the remaining life time.

2. The method for fault diagnosis and residual life prediction of a site intelligent distribution box according to claim 1, wherein, The hidden layer of the LSTM contains a memory unit and three "gate" structures, namely an input gate, a forgetting gate and an output gate; and the loss function is selected as a mean square error: where h is the data true value, h is the data predicted value; The historical time series data is divided into a training set, standardized, and segmented to generate a one-dimensional vector The k generated one-dimensional vectors are then input into an L hidden layer structure of the LSTM, and finally, a prediction result of the parameters in the future period of time is output .

3. The method for fault diagnosis and remaining life prediction of intelligent distribution boxes on construction sites as described in claim 1, characterized in that: The deep neural network DBN is used for classifying data and extracting features of the data, the deep neural network DBN is composed of a plurality of RBM stacks, the network structure is one visible layer and one hidden layer, there is a link between the layers, and the units in the layer are not associated with each other, the visible layer is used for receiving data, and the hidden layer is used for feature extraction, the top two layers are not directly linked, and the bottom layer represents a data vector, input of the deep neural network DBN is output data of the neural network LSTM , and the output is a feature vector set constructed as a health index HI.

4. The method for fault diagnosis and residual life prediction of a job site intelligent distribution box according to claim 3, characterized in that, The training process of the deep neural network (DBN) is performed layer by layer, each layer uses a data vector to infer a hidden layer, and then the hidden layer is input as a data vector of the next layer; after the deep neural network (DBN) is trained upward, each layer state is obtained and optimized through back propagation.

5. The method for fault diagnosis and residual life prediction of a job site intelligent distribution box according to claim 4, characterized in that, The health index (HI) is obtained by quantizing the features learned by the neural network (DBN), the HI degradation process is a process in which a fault is about to occur, the features learned by the neural network (DBN) are used to distinguish the fault state and the degradation state in the data, when the HI degradation is 0 in the fault state, the fault time (FT) is automatically obtained.

6. The method for fault diagnosis and residual life prediction of a job site intelligent distribution box according to claim 1, characterized in that, The remaining life prediction model is obtained by integrating the neural network (DBN) and the improved particle filter (PF), the improved particle filter (PF) adds a fuzzy reasoning system to the basic particle filter, effectively improves the particle diversity, avoids the particle degeneracy problem, improves the resampling accuracy, and the input of the improved particle filter (PF) is the health index (HI) constructed, which is used to predict the health index (HI) decline trend to obtain the remaining life time.

Images