A method and system for upgrading the algorithm of a high-resistance grounding identification device in a distribution network

Abstract

The invention discloses an algorithm upgrade method and system for a high-resistance grounding identification device in a distribution network. A preset decay coefficient is used as a reference. When a judgment value calculated by a built-in algorithm model according to input signal characteristics is less than or equal to the decay coefficient, it indicates that the algorithm model is in an activated decay period, and there is a risk of being unable to effectively identify characteristics, and the weight coefficient and offset need to be re-updated. At this time, the weight coefficient and offset of the neural network model are updated online according to a first weight coefficient update formula and a first offset update formula. There is no need to change the neural network structure, and there is no need to pre-train the built-in algorithm model through a training platform before each upgrade and update, and then burn it offline into the high-resistance grounding identification device controller. The invention solves the technical problem that the existing algorithm upgrade method for the high-resistance grounding identification device in the distribution network adopts an offline burning method, which is complicated to operate, inefficient, and cannot meet the requirements of high efficiency and intelligence.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault identification, and in particular to an algorithm upgrading method and system based on a distribution network high-resistance grounding identification device. Background Art

[0002] A high-resistance ground fault (HIF) refers to a ground fault that occurs when a power line passes through conductive media such as roads, soil, branches or concrete buildings. A high-resistance ground fault in a distribution network can easily cause fires or electric shocks, affect the reliability of the distribution network, and cause a significant social impact.

[0003] The existing built-in algorithm of high-resistance grounding identification device does not support online upgrade and update. It is often necessary to complete the algorithm research, experiment and finalization in the background offline, and then burn it into the controller of the high-resistance grounding identification device through USB, SWD, JTAG, UART and other interfaces. The operation is complicated and inefficient, and cannot meet the needs of efficient intelligence. Therefore, it is urgent to study a method that can support the online upgrade of the built-in algorithm of the high-resistance grounding identification device to meet the needs. Summary of the invention

[0004] The embodiment of the present invention provides an algorithm upgrade method and system based on a distribution network high-resistance grounding identification device, which is used to solve the technical problem that the existing algorithm upgrade method of the distribution network high-resistance grounding identification device adopts an offline burning method, which is complex to operate, inefficient, and cannot meet the requirements of high efficiency and intelligence.

[0005] In view of this, a first aspect of the present invention provides an algorithm upgrade method for a distribution network high-resistance grounding identification device, the method comprising:

[0006] Obtaining a determination value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the determination value is a calculated value of a neural network activation function;

[0007] When the judgment value is less than or equal to the decay coefficient, the weight coefficient and offset of the neural network model are updated according to the first weight coefficient update formula and the first offset update formula, wherein the first weight coefficient update formula is: ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the decision value, and δ and χ are the update coefficients.

[0008] Optionally, the parameter setting module is specifically used to set the number of tests, the high and low thresholds of the real waveform (l1, h1), the high and low thresholds of the simulated waveform (l2, h2) and the interference waveform threshold l3.

[0009] Optionally, it also includes:

[0010] Calculate the activation value of the neuron according to the neural network activation function;

[0011] When the activation value is less than or equal to the information coefficient, and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted;

[0012] When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, where the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k.

[0013] Optionally, it also includes:

[0014] Calculate the activation value of the neuron according to the neural network activation function;

[0015] The output deviation of the neuron is calculated according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, the two neurons are merged, and the weight coefficient and offset of the neural network model are updated according to the third weight coefficient update formula and the third offset update formula, where the calculation formula for the output deviation of the neuron is: The third weight coefficient update formula is w′ ij =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k).

[0016] Optionally, the calculation formula of the judgment value is: g = f(z), f(z) is the neural network activation function, n ij is the node element, x ij is the node value, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, a ij is the offset of the jth node in the i+1th layer, c ij is the connection relationship of the jth node in the i+1th layer, L is the number of neural network layers, and N is the maximum number of nodes in a single-layer neural network.

[0017] Optionally, the built-in algorithm model is a convolutional neural network model.

[0018] Optionally, before obtaining the determination value calculated by the built-in algorithm model of the distribution network high-resistance grounding identification device according to the input signal characteristics, it also includes:

[0019] Read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device and convert them into a public and open neural network exchange format;

[0020] According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol;

[0021] Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model;

[0022] The network parameters of the trained built-in algorithm model are reverse parsed into a public and open neural network exchange format and imported into the distribution network high-resistance grounding identification device.

[0023] A second aspect of the present invention provides an algorithm upgrade system for a distribution network high-resistance grounding identification device, the system comprising:

[0024] An acquisition module is used to obtain a determination value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the determination value is a calculated value of a neural network activation function;

[0025] An updating module is used to update the weight coefficient and offset of the neural network model according to the first weight coefficient updating formula and the first offset updating formula when the judgment value is less than or equal to the decay coefficient, wherein the first weight coefficient updating formula is: w′ ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the decision value, and δ and χ are the update coefficients.

[0026] Optionally, the update module is also used to:

[0027] Calculate the activation value of the neuron according to the neural network activation function;

[0028] When the activation value is less than or equal to the information coefficient, and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted;

[0029] When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, where the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k.

[0030] Optionally, the update module is also used to:

[0031] Calculate the activation value of the neuron according to the neural network activation function;

[0032] The output deviation of the neuron is calculated according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, the two neurons are merged, and the weight coefficient and offset of the neural network model are updated according to the third weight coefficient update formula and the third offset update formula, where the calculation formula for the output deviation of the neuron is: The third weight coefficient update formula is w′ ij =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k).

[0033] Also includes offline preprocessing modules for:

[0034] Read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device and convert them into a public and open neural network exchange format;

[0035] According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol;

[0036] Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model;

[0037] The network parameters of the trained built-in algorithm model are reverse parsed into a public and open neural network exchange format and imported into the distribution network high-resistance grounding identification device.

[0038] It can be seen from the above technical solutions that the embodiments of the present invention have the following advantages:

[0039] The algorithm upgrade method of the distribution network high-resistance grounding identification device provided in the present invention uses a preset decay coefficient as a reference. When the judgment value calculated by the built-in algorithm model according to the input signal characteristics is less than or equal to the decay coefficient, it indicates that the algorithm model is in an activated decay period and there is a risk of being unable to effectively identify the characteristics. The weight coefficient and offset need to be updated again. At this time, the weight coefficient and offset of the neural network model are updated online according to the first weight coefficient update formula and the first offset update formula. There is no need to change the neural network structure, nor is there a need to pre-train the built-in algorithm model on a training platform for each upgrade and update before offline burning it into the high-resistance grounding identification device controller. This solves the technical problem that the existing algorithm upgrade method of the distribution network high-resistance grounding identification device adopts an offline burning method, which is complicated to operate, inefficient, and cannot meet the requirements of high efficiency and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a first flow chart of an algorithm upgrade method for a distribution network high-resistance grounding identification device provided in an embodiment of the present invention;

[0041] Figure 2 A second flow chart of an algorithm upgrade method for a distribution network high-resistance grounding identification device provided in an embodiment of the present invention;

[0042] Figure 3 A third flow chart of the algorithm upgrade method of the distribution network high-resistance grounding identification device provided in an embodiment of the present invention;

[0043] Figure 4 A schematic diagram of format conversion of an offline processing process of an algorithm of a device for identifying high-resistance grounding in a distribution network provided in an embodiment of the present invention;

[0044] Figure 5 It is a structural schematic diagram of an algorithm upgrading system for a distribution network high-resistance grounding identification device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] definition:

[0047] Decay coefficient: A coefficient used to characterize the neural network model entering the activation decay period.

[0048] Information coefficient: A coefficient used to characterize the amount of information carried by a neuron in a neural network.

[0049] Deviation coefficient: used to characterize the information overlap between two neurons.

[0050] For easier understanding, see Figure 1 , Figure 1 The algorithm upgrade method of the distribution network high-resistance grounding identification device provided in the embodiment of the present invention is as follows: Figure 1 As shown, the algorithm upgrade method of the distribution network high-resistance grounding identification device includes:

[0051] Step 101, obtaining a judgment value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the judgment value is a calculated value of a neural network activation function.

[0052] In the present invention, the distribution network high-resistance grounding identification device adopts a built-in algorithm model based on a neural network to identify high-resistance faults, uses a signal acquisition module to collect zero-sequence voltage and current signals at a frequency of 10kHz according to set requirements, converts the signal into an energy spectrum through wavelet transform, uses a trained neural network to extract features, and uses n ij To represent the output nodes, the number of neural network layers, and the number of nodes in each layer. ij Indicates the jth node of the i-th layer neural network, which is a node element. When the value is 1, it means that the node exists, and when the value is 0, it means that the node does not exist. i=1,…L, L is the number of neural network layers, j=1,…N, N is the maximum number of neural network nodes in a single-layer neural network. ,x ij is the node value, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, a ij is the offset of the jth node in the i+1th layer, c ij is the connection relationship of the jth node in the i+1th layer.

[0053] Define the judgment value g = f(z), f(z) is the neural network activation function, which can be sigmoid function, tanh function, ReLu function or softmax function, and z is the activation value.

[0054] Step 102: When the judgment value is less than or equal to the decay coefficient, the weight coefficient and the offset of the neural network model are updated according to the first weight coefficient update formula and the first offset update formula.

[0055] According to the built-in algorithm model of the high-resistance grounding identification device of the distribution network, the voltage signal or current signal feature is input into the built-in algorithm model, and the judgment value g of the built-in algorithm model is calculated. When g≤ε, ε is the decay coefficient, indicating that the built-in algorithm model is in the activated decay period, and there is a risk of being unable to effectively identify the feature. The weight coefficient and offset of the model need to be updated again, but the neural network structure does not need to be changed. In the embodiment of the present invention, the first weight coefficient update formula is: w′ ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the decision value, and δ and χ are the update coefficients.

[0056] The algorithm upgrade method of the distribution network high-resistance grounding identification device provided in the embodiment of the present invention uses a preset decay coefficient as a reference. When the judgment value calculated by the built-in algorithm model according to the input signal characteristics is less than or equal to the decay coefficient, it indicates that the algorithm model is in an activated decay period and there is a risk of being unable to effectively identify the characteristics. The weight coefficient and offset need to be updated again. At this time, the weight coefficient and offset of the neural network model are updated online according to the first weight coefficient update formula and the first offset update formula. There is no need to change the neural network structure, nor is there a need to pre-train the built-in algorithm model on the training platform for each upgrade and update before offline burning it into the high-resistance grounding identification device controller. This solves the technical problem that the existing algorithm upgrade method of the distribution network high-resistance grounding identification device adopts an offline burning method, which is complicated to operate, inefficient, and cannot meet the requirements of efficient intelligence.

[0057] See also Figure 2 In one embodiment, the following steps may also be included:

[0058] Step 103: Calculate the activation value of the neuron according to the neural network activation function.

[0059] Step 104: When the activation value is less than or equal to the information coefficient and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted.

[0060] Step 105: When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, wherein the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k.

[0061] In the embodiment of the present invention, in addition to using the decay coefficient ε as the basis for upgrading and updating, the information coefficient β can also be used as the criterion for upgrading and updating the algorithm. Specifically, the activation value of the neuron is calculated according to the neural network activation function, and the calculation formula is:

[0062]

[0063] Among them, g(k) is the activation value of neuron k.

[0064] When g(k)≤β, and g>ε, it indicates that neuron k has less information in the neural network, so neuron k is deleted. When g(k)≤β, and g≤ε, it indicates that the current neuron k can no longer handle more fault features, so 1 neuron is added to this layer as the k+1 neuron, and then the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula. This method is an upgrade method that changes the structure of the neural network.

[0065] See also Figure 3 In one embodiment, the following steps may also be included:

[0066] Step 103: Calculate the activation value of the neuron according to the neural network activation function.

[0067] Step 106: Calculate the output deviation l(k) of the neuron according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, update the weight coefficient and offset of the neural network model according to the third weight coefficient update formula and the third offset update formula.

[0068] In addition to using the decay coefficient ε and information coefficient β as the criteria for algorithm upgrades and updates, the deviation coefficient λ can also be used as the criterion for algorithm upgrades and updates. The calculation formula for the output deviation of the neuron is: When the output deviation of two neurons is less than or equal to the deviation coefficient λ, that is, |l(k)-l(k-1)|≤λ, it indicates that the information overlap of the two neurons is high, and the neurons can be merged, and then the weight coefficient and offset can be updated. The third weight coefficient update formula is w′ ij =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k). This method is also an upgrade method that changes the structure of the neural network.

[0069] See also Figure 4 and Figure 5 In one embodiment, before step 101, the following steps may also be included:

[0070] Step 001, read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device, and convert them into a public and open neural network exchange format;

[0071] Step 002: According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol;

[0072] Step 003: Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model;

[0073] Step 004: reverse-parse the network parameters of the trained built-in algorithm model into a public and open neural network exchange format, and import it into the distribution network high-resistance grounding identification device.

[0074] According to the requirements of the offline training platform, an open neural network exchange model is used as a medium to realize the conversion of the device neural network algorithm and the mainstream artificial intelligence algorithm training platform files. For example, it can support the *.mdb format of the Caffe2 platform, the *.pt format of the Pytorch platform, and the *.ckpt format of the TensorFlow platform, realize the unbinding of the output results from the environment and the platform, and support the offline upgrade of the algorithm. Therefore, the present invention not only supports the online upgrade of the algorithm, but also supports the offline upgrade, has strong applicability, and is easy to promote.

[0075] For easier understanding, see Figure 5 The present invention provides an embodiment of an algorithm upgrade system for a distribution network high-resistance grounding identification device, including:

[0076] An acquisition module 501 is used to obtain a determination value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the determination value is a calculated value of a neural network activation function;

[0077] The updating module 502 is used to update the weight coefficient and the offset of the neural network model according to the first weight coefficient updating formula and the first offset updating formula when the judgment value is less than or equal to the decay coefficient, wherein the first weight coefficient updating formula is: ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the decision value, and δ and χ are the update coefficients.

[0078] The calculation formula of the judgment value is: g = f (z), f(z) is the neural network activation function, n ij is the node element, x ij is the node value, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, a ij is the offset of the jth node in the i+1th layer, c ij is the connection relationship of the jth node in the i+1th layer, L is the number of neural network layers, and N is the maximum number of nodes in a single-layer neural network.

[0079] The update module 502 is also used for:

[0080] Calculate the activation value of the neuron according to the neural network activation function;

[0081] When the activation value is less than or equal to the information coefficient, and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted;

[0082] When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, where the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k.

[0083] The update module 502 is also used for:

[0084] Calculate the activation value of the neuron according to the neural network activation function;

[0085] The output deviation of the neuron is calculated according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, the two neurons are merged, and the weight coefficient and offset of the neural network model are updated according to the third weight coefficient update formula and the third offset update formula, where the calculation formula for the output deviation of the neuron is: The third weight coefficient update formula is w′ ij =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k).

[0086] It also includes an offline pre-processing module 500, which is used to:

[0087] Read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device and convert them into a public and open neural network exchange format;

[0088] According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol;

[0089] Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model;

[0090] The network parameters of the trained built-in algorithm model are reverse parsed into a public and open neural network exchange format and imported into the distribution network high-resistance grounding identification device.

[0091] The distribution network high-resistance grounding identification device algorithm upgrade system provided in the present invention is used to execute the distribution network high-resistance grounding identification device algorithm upgrade method in the aforementioned distribution network high-resistance grounding identification device algorithm upgrade method embodiment. Its working principle is the same as that of the distribution network high-resistance grounding identification device algorithm upgrade method embodiment, and will not be repeated here.

[0092] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for upgrading the algorithm of a high-resistance grounding identification device in a distribution network, characterized in that: include: Obtaining a determination value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the determination value is a calculated value of a neural network activation function; When the judgment value is less than or equal to the decay coefficient, the weight coefficient and offset of the neural network model are updated according to the first weight coefficient update formula and the first offset update formula, wherein the first weight coefficient update formula is: ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the judgment value, δ and χ are the update coefficients, ε is the decay coefficient, and the decay coefficient is the coefficient used to characterize the neural network model entering the activation decay period.

2. The algorithm upgrade method of the distribution network high-resistance grounding identification device according to claim 1 is characterized in that: Also includes: Calculate the activation value of the neuron according to the neural network activation function; When the activation value is less than or equal to the information coefficient, and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted; When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, where the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k, and β is the information coefficient.

3. The algorithm upgrade method of the distribution network high-resistance grounding identification device according to claim 1 is characterized in that: Also includes: Calculate the activation value of the neuron according to the neural network activation function; The output deviation of the neuron is calculated according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, the two neurons are merged, and the weight coefficient and offset of the neural network model are updated according to the third weight coefficient update formula and the third offset update formula, where the calculation formula for the output deviation of the neuron is: The third weight coefficient update formula is w′ i ' j =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k), g(k) is the activation value of neuron k, x ij is the node value of the jth node in the i+1th layer.

4. The algorithm upgrade method of the distribution network high-resistance grounding identification device according to claim 1 is characterized in that: The calculation formula of the judgment value is: g = f (z), f(z) is the neural network activation function, n ij is the node element, x ij is the node value, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, a ij is the offset of the jth node in the i+1th layer, c ij is the connection relationship of the jth node in the i+1th layer, L is the number of neural network layers, and N is the maximum number of nodes in a single-layer neural network.

5. The algorithm upgrade method of the distribution network high-resistance grounding identification device according to claim 1 is characterized in that: The built-in algorithm model is a convolutional neural network model.

6. The algorithm upgrade method of the distribution network high-resistance grounding identification device according to claim 1 is characterized in that: Before obtaining the judgment value calculated by the built-in algorithm model of the distribution network high-resistance grounding identification device according to the input signal characteristics, it also includes: Read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device and convert them into a public and open neural network exchange format; According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol; Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model; The network parameters of the trained built-in algorithm model are reverse parsed into a public and open neural network exchange format and imported into the distribution network high-resistance grounding identification device.

7. An algorithm upgrade system for a high-resistance grounding identification device in a distribution network, characterized in that: include: An acquisition module is used to obtain a determination value calculated by a built-in algorithm model of a distribution network high-resistance grounding identification device according to input signal characteristics, wherein the built-in algorithm model is a neural network model, the input signal characteristics are voltage signal characteristics or current signal characteristics, and the determination value is a calculated value of a neural network activation function; An updating module is used to update the weight coefficient and offset of the neural network model according to the first weight coefficient updating formula and the first offset updating formula when the judgment value is less than or equal to the decay coefficient, wherein the first weight coefficient updating formula is: w′ ij =w ij +δ(g-ε), the first offset update formula is: a′ ij =a ij +χ(g-ε),a ij is the offset of the jth node in the i+1th layer, w ij is the weight between the jth node in the i+1th layer and the node in the ith layer, g is the judgment value, δ and χ are the update coefficients, ε is the decay coefficient, and the decay coefficient is the coefficient used to characterize the neural network model entering the activation decay period.

8. The algorithm upgrade system for identifying a high-resistance grounding device in a distribution network according to claim 7 is characterized in that: The update module is also used to: Calculate the activation value of the neuron according to the neural network activation function; When the activation value is less than or equal to the information coefficient, and the judgment value is greater than the decay coefficient, the corresponding neuron is deleted; When the activation value is less than or equal to the information coefficient, and the judgment value is less than or equal to the decay coefficient, one neuron is added to the corresponding layer, and the weight coefficient and offset of the neural network model are updated according to the second weight coefficient update formula and the second offset update formula, where the second weight coefficient update formula is: The second offset update formula is: g(k) is the activation value of neuron k, and β is the information coefficient.

9. The algorithm upgrade system for identifying a high-resistance grounding device in a distribution network according to claim 7, characterized in that: The update module is also used to: Calculate the activation value of the neuron according to the neural network activation function; The output deviation of the neuron is calculated according to the activation value of the neuron. When the output deviation of two neurons is less than or equal to the deviation coefficient, the two neurons are merged, and the weight coefficient and offset of the neural network model are updated according to the third weight coefficient update formula and the third offset update formula, where the calculation formula for the output deviation of the neuron is: The third weight coefficient update formula is w′ ij =w ij +l(k), the first offset update formula is: a′ ij =a ij +l(k), g(k) is the activation value of neuron k, x ij is the node value of the jth node in the i+1th layer.

10. The algorithm upgrade system for identifying a high-resistance grounding device in a distribution network according to claim 7, characterized in that: Also includes offline preprocessing modules for: Read the network parameters of the algorithm model built into the distribution network high-resistance grounding identification device and convert them into a public and open neural network exchange format; According to the private protocol of the neural network training platform, the built-in algorithm model of the distribution network high-resistance grounding identification device is converted into a data packet of the neural network training platform protocol; Train the built-in algorithm model in the neural network training platform and save the trained built-in algorithm model; The network parameters of the trained built-in algorithm model are reverse parsed into a public and open neural network exchange format and imported into the distribution network high-resistance grounding identification device.

Images