A task timing allocation method for on-site detection system based on hybrid density deep learning

Through the task timing allocation method and genetic algorithm based on hybrid density deep learning, the working order of the detection equipment is automatically arranged, and the problem of low efficiency of on-site detection experiments of power distribution is solved, achieving more efficient detection efficiency and experimental module utilization.

CN114358599BActive Publication Date: 2024-12-03STATE GRID JIANGSU ELECTRIC POWER CO LTD MARKETING SERVICE CENT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210006558.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-05
Publication Date
2024-12-03
Estimated Expiration
2042-01-05

AI Technical Summary

Technical Problem

The on-site inspection and testing of existing household distribution equipment is low, the utilization efficiency of the test module is low and often idle, so the inspection efficiency needs to be improved.

Method used

Using a task timing allocation method based on hybrid density deep learning, the central control platform uses historical data to allocate optimal timing, and uses genetic algorithms to automatically arrange the working order of the equipment, switch the working state to complete the detection test. .

Benefits of technology

The efficiency of the on-site inspection experiment of household power distribution is improved, the utilization rate of the test module is optimized, and the occurrence of idle situations is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114358599B_ABST
    Figure CN114358599B_ABST
Patent Text Reader

Abstract

The present invention discloses a task timing allocation method for a residential distribution on-site detection system based on hybrid density deep learning, the method comprising: ① inputting detection equipment and detection project requirements through a handheld terminal and uploading them to a central control platform; ② using hybrid density deep learning based on historical data to give the optimal initial solution for test task timing allocation; ③ using a genetic algorithm to solve the optimal timing allocation, and the central control module automatically arranges the working order of each test module, switches its working state to complete the detection test. The present invention proposes a task timing allocation method for a residential distribution on-site detection system based on hybrid density deep learning, which provides a theoretical basis for on-site detection of residential distribution equipment and improvement of detection efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of on-site detection of residential power distribution equipment, and specifically to a task timing allocation method for a residential power distribution on-site detection system based on mixed density deep learning. Background Art

[0002] Economic development starts with electricity, and the economic development of a region cannot be separated from the supply of electric energy. With the expansion of urban scale and more and more people flowing to cities, the scale of residential power distribution projects has exploded, and higher requirements have been put forward for the efficiency of residential power distribution on-site detection experiments. At present, the residential power distribution project site mainly tests 8 types of equipment and materials, such as distribution transformers (dry type), high-voltage (switch / ring network) cabinets, and low-voltage switch cabinets. Field tests usually involve multiple devices, and the same device also contains multiple test items. Considering economy and portability, the number of each test module is limited. At present, each test module is planned and used according to the experience of the test personnel. The utilization efficiency of the test module is low, and it is often idle. Therefore, there is an urgent need for a method that can realize the timing allocation of residential power distribution on-site detection test tasks to improve the efficiency of residential power distribution on-site detection experiments. Summary of the invention

[0003] In view of the low efficiency of existing on-site detection tests of residential power distribution equipment, the present invention provides a task timing allocation method for a residential power distribution on-site detection system based on mixed density deep learning, so as to improve the efficiency of on-site detection experiments of residential power distribution equipment.

[0004] In order to achieve the above object, a method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning of the present invention comprises the following steps:

[0005] S1: Enter the testing equipment and testing project requirements through the handheld terminal, and upload the entered testing equipment and testing project requirements to the central control platform;

[0006] S2: The central control platform uses mixed density deep learning based on historical detection data to provide the optimal initial solution for the test task timing allocation;

[0007] S3: Genetic algorithm is used to solve the optimal timing allocation. The central control module automatically arranges the working order of each detection equipment and switches its working state to complete the detection test.

[0008] In a preferred embodiment of the present invention, the inputting of the testing equipment and testing project requirements through the handheld terminal and uploading the inputted testing equipment and testing project requirements to the central control platform specifically includes:

[0009] According to the test requirements, the test personnel select the model of the equipment to be tested and the test items on the handheld terminal, enter the work location and time, and upload it to the central control platform via the wireless LAN.

[0010] In a preferred embodiment of the present invention, the central control platform runs on an on-site mobile server, and the entire on-site detection task timing distribution and automatic switching of the test module working status are realized by configuring the on-site detection system software in the server.

[0011] In a preferred embodiment of the present invention, the central control platform uses mixed density deep learning based on historical data to provide an optimal initial solution for the timing allocation of test tasks, specifically:

[0012] The central control platform extracts the historical test data of the same model of testing equipment from the server database, or when there is less data, it can simultaneously extract the historical test data of testing equipment of similar models and use these historical test data as mixed density deep learning training data sets. These historical test data include the model of the testing equipment, test items, and test task timing distribution;

[0013] The test task timing allocation is a time series prediction problem. The probability density of the detection equipment performing a certain test task at a certain time point after the start of work is expressed as the following formula (1):

[0014]

[0015] In formula (1), p(y|x) is the probability density distribution expression, πn(x,t) is the weight of the nth Gaussian kernel function, is the nth Gaussian kernel function, μ n (x, t) is the mean of the nth Gaussian kernel function, is the variance of the nth Gaussian kernel function;

[0016] Using the maximum likelihood method, the mean and variance in formula (1) are calculated by maximizing the log-likelihood function and used as the loss function.

[0017]

[0018] Since the maximum likelihood method may lead to overfitting, in order to solve this problem, in a preferred embodiment of the present invention, a regularization term based on adversarial training is introduced into the loss function.

[0019] In a preferred embodiment of the present invention, the adversarial training is to add interference to the main training output of the neural network. As an example of adversarial training, a fast gradient sign method is used to quickly generate an example of adversarial training. The added disturbance is proportional to the sign of the gradient back-propagated from the output layer to the input layer. At the same time, in order to maintain the consistency of the hybrid network training, L-2 regularization is added to the loss function. The loss function in formula (2) is rewritten as:

[0020]

[0021] Where: λ is the weight of the adversarial example, which represents the impact of adversarial training on the loss function; ε is a constant that limits the maximum norm of the perturbation; γ is a regularization hyperparameter; K is the total number of learning weights; θ k is the kth learning weight.

[0022] Since the main goal of the training process is to minimize the loss function, in order to prevent the appearance of non-numbers due to low values ​​of the logarithm and denominator terms and high values ​​of the exponential terms during the optimization of the loss function, in a preferred embodiment of the present invention, gradient clipping is used to prevent the appearance of excessively large gradient points by cutting off the gradient.

[0023] In a preferred embodiment of the present invention, the mixed density deep learning is divided into the following three step modules:

[0024] Step 1 module: pre-process the training data set and input it into the convolutional neural network module;

[0025] Step 2 Module: The output of the convolutional neural network module is flattened into a one-dimensional data feature through the Flatten layer and input into the gated recurrent unit module; the gated recurrent unit module learns the short-term and long-term dependencies of the data and captures the temporal features therein;

[0026] Step three module: fully connected neural network module.

[0027] In a preferred embodiment of the present invention, the convolutional neural network module selects a rectified linear unit (ReLU) as an activation function, convolves the feature map of the previous layer with a filter set in the convolution layer to output a feature map, and the pooling layer takes the maximum value of the input value to create its own feature map and applies it to each feature map, thereby forming a time series allocation feature vector.

[0028] To prevent overfitting during the training process, in a preferred embodiment of the present invention, the convolutional neural network module randomly discards 20% of the data when outputting data.

[0029] In a preferred embodiment of the present invention, the gated recurrent unit module selects a linear rectification unit as an activation function, and includes two gated recurrent unit layers to determine the amount of information used by the mixture density deep learning in each time interval; each gated recurrent unit layer is composed of an update gate u(t) and a reset gate r(t), and their functional expressions are respectively formula (4), formula (5), formula (6) and formula (7):

[0030] u(t)=f[ω u x(t)+R u h(t-1)+b u ] (4)

[0031] r(t)=f[ω r x(t)+R r h(t-1)+b r ] (5)

[0032] h(t)=(1-u(t))eh(t-1)+u(t)eh%(t) (6)

[0033] h%(t)=f[ω u x(t)+R u (r(t)eh(t-1))] (7)

[0034] In equations (4), (5), (6) and (7), x(t) is the input vector at the current moment, ω u , ω r They are update gate and reset gate learning weights, R u , R r are the loop parameters of the update gate and the reset gate, respectively, u 、b r are the update gate and reset gate deviations respectively, h(t) and h(t-1) are the state variables at the current moment and the previous moment respectively, and h%(t) is the candidate state variable at the current moment.

[0035] In a preferred embodiment of the present invention, the update gate controls how many hidden state variables must be updated, and the reset gate controls the amount of information written to the current candidate state at the last moment.

[0036] To prevent overfitting during training and reduce the information required for mixed density deep learning, in a preferred embodiment of the present invention, 50% of the data is randomly discarded when the two gated recurrent unit layers output data.

[0037] In a preferred embodiment of the present invention, the fully connected neural network module includes four fully connected neural network layers and an output layer; the four fully connected neural network layers all select linear correction units as activation functions and perform end-to-end training on them to improve performance; the output layer includes three parallel fully connected neural networks, and selects linear exponential units (ELUs) as activation functions to fit the output probability density function weight π n (x, t) and mean μ n (x, t), using the softmax activation function to fit the output variance Substitute the fitting results into (1) to obtain the probability density of the test module executing a certain test task at a certain time point after the start of work.

[0038] At the same time, in order to prevent overfitting, in a preferred embodiment of the present invention, the first two fully connected neural network layers randomly discard 25% of the data when outputting data. The output layer includes three parallel fully connected neural networks, and the linear exponential unit (ELU) is selected as the activation function to fit the output probability density function weight π n (x, t) and mean μ n (x, t), using the softmax activation function to fit the output variance Substitute the fitting results into (1) to obtain the probability density of the test module executing a certain test task at a certain time point after the start of work.

[0039] In a preferred embodiment of the present invention, the optimal initial solution for the timing allocation of test tasks is specifically provided as follows: according to the working time required for each test task, the probability density function is integrated within the time region, which is the probability that the test module executes the test task within the time region; on the time axis, starting from the starting point of the test task, each test module selects the test task with the highest probability in the same time region, thereby obtaining the optimal initial solution for the timing allocation of the test tasks.

[0040] In a preferred embodiment of the present invention, the genetic algorithm is used to solve the optimal timing distribution, and the central control module automatically arranges the working order of each detection device and switches its working state to complete the detection test. Specifically,

[0041] Taking the minimum total working time from the start of the test task to the completion of the last test as the objective function, and the constraint that the detection equipment can only perform one test task at the same time, the genetic algorithm is used to obtain the optimal initial solution as the starting point for optimization, and the optimal timing allocation of the test tasks of each detection equipment is solved; the central control module generates a test wiring plan based on the optimal timing allocation result and transmits it back to the handheld terminal; after the test personnel complete the wiring, the central control module automatically completes the detection test by switching the working status of each detection equipment.

[0042] Due to the adoption of the above technical solution, the present invention can implement a method for timing distribution of residential power distribution on-site detection test tasks to improve the efficiency of residential power distribution on-site detection experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the process of the present invention.

[0044] Figure 2 This is a diagram of the hybrid deep learning steps of the present invention.

[0045] Figure 3 The test system connection diagram of the present invention DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with specific implementation modes.

[0047] See also Figure 1 The present invention provides a method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning, comprising the following steps:

[0048] S1: Enter the testing equipment and testing project requirements through the handheld terminal, and upload the entered testing equipment and testing project requirements to the central control platform;

[0049] S2: The central control platform uses mixed density deep learning based on historical detection data to provide the optimal initial solution for the test task timing allocation;

[0050] S3: Genetic algorithm is used to solve the optimal timing allocation. The central control module automatically arranges the working order of each detection equipment and switches its working state to complete the detection test.

[0051] The inputting of the testing equipment and testing project requirements through the handheld terminal and uploading the inputted testing equipment and testing project requirements to the central control platform specifically includes:

[0052] According to the test requirements, the test personnel select the equipment model and test items to be tested on the mobile handheld terminal, enter the work location and time, and upload it to the central control platform through the wireless WIFI local area network. The central control platform runs on an on-site mobile server computer. Through the system software, the entire on-site detection task timing allocation and the automatic switching of the test module working status can be realized.

[0053] The central control platform uses mixed density deep learning based on historical data to provide the optimal initial solution for the timing allocation of test tasks, specifically:

[0054] The central control platform extracts historical inspection data of the same model of equipment (when there is less data, similar models of equipment can be extracted at the same time) from the database, including equipment model, test items, test task timing distribution, etc., and uses it as a mixed density deep learning training data set.

[0055] The timing allocation of test tasks is a time series prediction problem. The probability density of a test module executing a test task at a certain time point after the work starts can be expressed as:

[0056]

[0057] Where: p(y|x) is the probability density distribution expression, πn(x,t) is the weight of the nth Gaussian kernel function, is the nth Gaussian kernel function, μ n (x, t) is the mean of the nth Gaussian kernel function, is the variance of the nth Gaussian kernel function.

[0058] Using the maximum likelihood method, the mean and variance in formula (1) are calculated by maximizing the log-likelihood function and used as the loss function.

[0059]

[0060] The maximum likelihood method can lead to overfitting. To solve this problem, a regularization term based on adversarial training is introduced into the loss function. Adversarial training adds interference to the main training output of the neural network as an adversarial example. The fast gradient sign method is used to quickly generate adversarial training examples. The added perturbation is proportional to the sign of the gradient backpropagated from the output layer to the input layer. At the same time, in order to maintain the consistency of the hybrid network training, L-2 regularization is added to the loss function. The loss function in formula (2) is rewritten as,

[0061]

[0062] Where: λ is the weight of the adversarial example, which represents the impact of adversarial training on the loss function; ε is a constant that limits the maximum norm of the perturbation; γ is a regularization hyperparameter; K is the total number of learning weights; θ k is the kth learning weight.

[0063] The main goal of the training process is to minimize the loss function. In order to prevent the appearance of non-numbers due to low values ​​of the logarithm and denominator and high values ​​of the exponential term during the optimization of the loss function, gradient clipping is used to prevent the appearance of excessively large gradient points by cutting off the gradient.

[0064] The steps of mixed density deep learning are as follows Figure 2 shown.

[0065] Mixed density deep learning is divided into three steps. First, the training data set is preprocessed and input into the convolutional neural network module. The convolutional neural network module selects the rectified linear unit (ReLU) as the activation function. In the convolution layer, the feature map of the previous layer is convolved with the filter set to output the feature map. The pooling layer takes the maximum value of the input value to create its own feature map and applies it to each feature map to form a time series distribution feature vector. To prevent overfitting during training, the module randomly discards 20% of the data when outputting data.

[0066] The output of the convolutional neural network module is flattened into a one-dimensional data feature by the Flatten layer and input into the gated recurrent unit module. The gated recurrent unit module learns the short-term and long-term dependencies of the data and captures the temporal features therein. The gated recurrent unit module selects the linear rectification unit as the activation function and includes two gated recurrent unit layers to determine the amount of information used by the mixture density deep learning in each time interval. Each gated recurrent unit layer consists of an update gate u(t) and a reset gate r(t), and their function expressions are respectively,

[0067] u(t)=f[ω u x(t)+R u h(t-1)+b u ] (4)

[0068] r(t)=f[ω r x(t)+R r h(t-1)+b r ] (5)

[0069] h(t)=(1-u(t))eh(t-1)+u(t)eh%(t) (6)

[0070] h%(t)=f[ω u x(t)+R u (r(t)eh(t-1))] (7)

[0071] Where: x(t) is the input vector at the current moment, ω u ,ω r They are update gate and reset gate learning weights, R u , R r are the loop parameters of the update gate and the reset gate, respectively, u 、b rare update gate and reset gate biases, h(t) and h(t-1) are the state variables at the current and previous moments, respectively, and h%(t) is the candidate state variable at the current moment. The update gate controls how many hidden state variables must be updated, and the reset gate controls the amount of information written to the current candidate state at the previous moment. To prevent overfitting during training and reduce the information required for deep learning of mixed density, 50% of the data is randomly discarded when the two gated recurrent unit layers output data.

[0072] The fully connected neural network module includes 4 fully connected neural network layers and an output layer. The first 4 fully connected neural network layers all select the linear rectified unit as the activation function and perform end-to-end training on them to improve performance. At the same time, in order to prevent overfitting, the first two fully connected neural network layers randomly discard 25% of the data when outputting data. The output layer includes three parallel fully connected neural networks, and the linear exponential unit (ELU) is selected as the activation function to fit the output probability density function weight π n (x, t) and mean μ n (x, t), using the softmax activation function to fit the output variance Substitute the fitting results into (1) to obtain the probability density of the test module executing a certain test task at a certain time point after the start of work.

[0073] According to the working time required for each test task, the probability density function is integrated in the time region, which is the probability of the test module executing the test task in the time region. On the time axis, starting from the starting point of the test task, each test module selects the test task with the highest probability in the same time region, and then obtains the optimal initial solution for the test task timing allocation.

[0074] Genetic algorithm is used to solve the optimal timing allocation. The central control module automatically arranges the working order of each test module and switches its working state to complete the detection test:

[0075] Taking the minimum total working time from the start of the test task to the completion of the last test as the objective function, and the constraint that the test module can only execute one test task at the same time, the genetic algorithm is used to find the optimal timing allocation of the test tasks of each test module with the optimal initial solution obtained in step 2 as the starting point. The central control module generates a test wiring plan based on the optimal timing allocation result and transmits it back to the handheld terminal. After the test personnel complete the wiring, the central control module automatically completes the detection test by switching the working status of each test module. The connection between the central control platform, the handheld terminal, and the test module is as follows: Figure 3 as shown in .

Claims

1. A task timing allocation method for a home delivery on-site detection system based on mixed density deep learning, characterized in that: The following steps are involved: S1: Enter the testing equipment and testing project requirements through the handheld terminal, and upload the entered testing equipment and testing project requirements to the central control platform; S2: The central control platform uses mixed density deep learning based on historical detection data to provide the optimal initial solution for the timing allocation of test tasks; S3: Genetic algorithm is used to solve the optimal timing allocation, and the central control module automatically arranges the working order of each detection equipment and switches its working state to complete the detection test; The central control platform uses mixed density deep learning based on historical data to provide the optimal initial solution for the timing allocation of test tasks, specifically: The central control platform extracts the historical test data of the same model of testing equipment from the server database, or extracts the historical test data of testing equipment of similar models when there is little data, and uses these historical test data as mixed density deep learning training data sets. These historical test data include the model of the testing equipment, test items, and test task timing distribution; The test task timing allocation is a time series prediction problem. The probability density of the detection equipment performing a certain test task at a certain time point after the start of work is expressed as formula (1): In formula (1), p(y|x) is the probability density distribution expression, π n (x, t) is the weight of the nth Gaussian kernel function, is the nth Gaussian kernel function, μ n (x, t) is the mean of the nth Gaussian kernel function, is the variance of the nth Gaussian kernel function; Using the maximum likelihood method, the mean and variance in formula (1) are calculated by maximizing the log-likelihood function and used as the loss function. Introducing a regularization term based on adversarial training into the loss function; The adversarial training is to add interference to the main training output of the neural network. As an example of adversarial training, a fast gradient sign method is used to quickly generate an adversarial training example. The added disturbance is proportional to the sign of the gradient back-propagated from the output layer to the input layer. At the same time, in order to maintain the consistency of the hybrid network training, L-2 regularization is added to the loss function. The loss function in formula (2) is rewritten as: Where: λ is the weight of the adversarial example, which represents the impact of adversarial training on the loss function; ε is a constant that limits the maximum norm of the perturbation; γ is a regularization hyperparameter; K is the total number of learning weights; θ k is the kth learning weight; Gradient shearing is used to prevent the appearance of excessively large gradient points by cutting off the gradient; The mixed density deep learning is divided into the following three step modules: Step 1 module: pre-process the training data set and input it into the convolutional neural network module; Step 2 Module: The output of the convolutional neural network module is flattened into a one-dimensional data feature through the Flatten layer and input into the gated recurrent unit module; the gated recurrent unit module learns the short-term and long-term dependencies of the data and captures the temporal features therein; Step 3 module: fully connected neural network module; The gated recurrent unit module selects a linear rectification unit as an activation function and includes two gated recurrent unit layers to determine the amount of information used by the mixture density deep learning in each time interval; each gated recurrent unit layer consists of an update gate u(t) and a reset gate r(t), and their function expressions are respectively: u(t)=f『ω u x(t)+R u h(t-1)+b u ] (4) r(t)=f[ω r x(t)+R r h(t-1)+b r ] (5) h(t)=(1-u(t))eh(t-1)+u(t)eh%(t) (6) h=(t)=f[ω u xt)+R r (r(t)e(t-1))] (7) In equations (4), (5), (6) and (7), x(t) is the input vector at the current moment, ω u ,ω r They are update gate and reset gate learning weights, R u , R r are the loop parameters of the update gate and the reset gate, respectively, u , b r are the update gate and reset gate deviations, h(t) and h(t-1) are the state variables at the current and previous moments, respectively, and h%(t) is the candidate state variable at the current moment; The optimal initial solution for the test task timing allocation is specifically provided as follows: according to the working time required for each test task, the probability density function is integrated in the time region, that is, the probability of the test module executing the test task in the time region; on the time axis, starting from the starting point of the test task, each test module selects the test task with the highest probability in the same time region, thereby obtaining the optimal initial solution for the test task timing allocation; The genetic algorithm is used to solve the optimal timing distribution, and the central control module automatically arranges the working order of each detection device and switches its working state to complete the detection test, specifically: Taking the minimum total working time from the start of the test task to the completion of the last test as the objective function, and the constraint that the detection equipment can only perform one test task at the same time, the genetic algorithm is used to obtain the optimal initial solution as the starting point for optimization, and the optimal timing allocation of the test tasks of each detection equipment is solved; the central control module generates a test wiring plan based on the optimal timing allocation result and transmits it back to the handheld terminal; after the test personnel complete the wiring, the central control module automatically completes the detection test by switching the working status of each detection equipment.

2. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 1, characterized in that: The inputting of the testing equipment and testing project requirements through the handheld terminal and uploading the inputted testing equipment and testing project requirements to the central control platform specifically includes: According to the test requirements, the test personnel select the model of the equipment to be tested and the test items on the handheld terminal, enter the work location and time, and upload it to the central control platform via the wireless LAN.

3. The task timing allocation method of the on-site detection system based on mixed density deep learning as described in claim 2 is characterized by: The central control platform runs on a mobile server on site, and realizes the timing distribution of the entire operation site detection task and the automatic switching of the working status of the test module through the on-site detection system software configured in the server.

4. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 1, characterized in that: The convolutional neural network module selects the rectified linear unit (ReLU) as the activation function, convolves the feature map of the previous layer with the filter set in the convolution layer to output the feature map, and the pooling layer takes the maximum value of the input value to create its own feature map and applies it to each feature map, thereby forming a time series allocation feature vector.

5. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 4, characterized in that: The convolutional neural network module randomly discards 20% of the data when outputting data.

6. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 1, characterized in that: The update gate controls how much hidden state variables must be updated, and the reset gate controls how much information was written to the current candidate state at the previous moment.

7. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 6, characterized in that: The two GRU layers randomly drop 50% of the output data.

8. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 1, characterized in that: The fully connected neural network module includes four fully connected neural network layers and one output layer; the four fully connected neural network layers all select linear rectified units as activation functions and perform end-to-end training on them to improve performance; the output layer includes three parallel fully connected neural networks, and selects linear exponential units (ELUs) as activation functions to fit the output probability density function weight π n (x, t) and mean μ n (x, t), using the softmax activation function to fit the output variance Substitute the fitting results into (1) to obtain the probability density of the test module executing a certain test task at a certain time point after the start of work.

9. The method for allocating task timing of a home delivery on-site detection system based on mixed density deep learning as claimed in claim 8, characterized in that: The first two fully connected neural network layers randomly discard 25% of the data when outputting data; the output layer includes three parallel fully connected neural networks, and the linear exponential unit (ELU) is selected as the activation function to fit the output probability density function weight π n (x, t) and mean μ n (x, t), using the softmax activation function to fit the output variance Substitute the fitting results into (1) to obtain the probability density of the test module executing a certain test task at a certain time point after the start of work.