Amplitude processing network training method, amplitude processing method, equipment and storage medium

Abstract

The invention discloses an amplitude processing network training method and device, an amplitude processing method and device and a storage medium, and the method comprises the steps: loading an amplitude processing network when original amplitude data is collected for a pumped storage unit at a plurality of moments; respectively inputting the original amplitude data at each moment into a first encoder to encode a first original amplitude feature; inputting the original amplitude data at multiple moments into a second encoder at the same time to encode a second original amplitude feature; inputting the plurality of first original amplitude features and the second original amplitude features into a feature interaction module to interact into fused amplitude features; inputting the fused amplitude characteristics into a decoder, and decoding the fused amplitude characteristics into target amplitude characteristics; inputting the target amplitude characteristics into a multi-task processing module to generate multiple pieces of amplitude task information; generating a total loss value according to the multiple pieces of amplitude task information; and updating the amplitude processing network according to the total loss value. According to the embodiment of the invention, the robustness of the amplitude processing network is improved, and the accuracy of multi-task processing is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning, and in particular to a training of an amplitude processing network and an amplitude processing method, device and storage medium thereof. Background Art

[0002] A pumped-storage power plant is a complex automated intelligent operation system, which has several core devices including a turbine pumped-storage machine and a pumped-storage unit. The pumped-storage unit has several important equipment including an upper frame, a lower frame, a water-guided top cover, a thrust bearing, etc. These devices will produce a vibration amplitude with a certain changing pattern during normal operation.

[0003] During the operation of pumped-storage power plants, there are different complex working conditions at different times. In order to ensure that the pumped-storage units maintain a stable and safe operating state, the threshold method, Seq2Seq (sequence to sequence) and other methods are currently used to monitor the amplitude of the upper frame, lower frame, water guide top cover, thrust bearing and other equipment in many aspects, formulate and quickly initiate corresponding safety measures to ensure the safe operation of the pumped-storage units and avoid sudden accidents.

[0004] Among them, the threshold method means that when the vibration amplitude exceeds a certain threshold, an alarm signal or a fault signal is triggered. However, the threshold is fixed. If the range of the threshold is too narrow, it is easy to produce false alarms. If the range of the threshold is too wide, the sensitivity is low, resulting in low efficiency of the threshold method.

[0005] Seq2Seq is usually a time series sequence model based on RNN (Recurrent Neural Network), which connects the encoder and decoder in series, reduces the amount of extracted feature information, and results in lower robustness of the time series sequence model. Summary of the invention

[0006] In view of this, the present invention provides a training of an amplitude processing network and an amplitude processing method, device and storage medium thereof, so as to improve the efficiency and robustness of monitoring the amplitude of a pumped storage unit.

[0007] A first aspect of the present invention provides a training method for an amplitude processing network, comprising:

[0008] When the original amplitude data of the pumped storage unit is collected at multiple times, an amplitude processing network is loaded; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0009] Inputting the original amplitude data at each moment into the first encoder to encode the first original amplitude feature;

[0010] Inputting the original amplitude data at multiple moments into the second encoder simultaneously to encode a second original amplitude feature;

[0011] Inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact with each other to form a fused amplitude feature;

[0012] Inputting the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0013] Inputting the target amplitude feature into the multi-task processing module to generate multiple amplitude task information;

[0014] generating a total loss value according to the plurality of amplitude task information;

[0015] The amplitude processing network is updated according to the total loss value.

[0016] A second aspect of the present invention provides an amplitude processing method, comprising:

[0017] When the original amplitude data of the pumped storage unit is collected at multiple moments, the amplitude processing network trained by the method described in the first aspect is loaded; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0018] Inputting the original amplitude data at each moment into the first encoder to encode the first original amplitude feature;

[0019] Inputting the original amplitude data at multiple moments into the second encoder simultaneously to encode a second original amplitude feature;

[0020] Inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact with each other to form a fused amplitude feature;

[0021] Inputting the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0022] The target amplitude feature is input into the multi-task processing module to generate a plurality of amplitude task information.

[0023] A third aspect of the present invention provides a training device for an amplitude processing network, comprising:

[0024] An amplitude processing network loading module is used to load the amplitude processing network when the original amplitude data of the pumped storage unit is collected at multiple times; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0025] A first original amplitude feature encoding module, used for inputting the original amplitude data at each moment into the first encoder to encode a first original amplitude feature;

[0026] A second original amplitude feature encoding module, used for simultaneously inputting the original amplitude data at multiple moments into the second encoder to encode a second original amplitude feature;

[0027] A fused amplitude feature interaction module, used for inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact as a fused amplitude feature;

[0028] A target amplitude feature decoding module, used for inputting the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0029] An amplitude task information generating module, used for inputting the target amplitude feature into the multi-task processing module to generate a plurality of amplitude task information;

[0030] A total loss value generating module, used for generating a total loss value according to the plurality of amplitude task information;

[0031] An amplitude processing network updating module is used to update the amplitude processing network according to the total loss value.

[0032] A fourth aspect of the present invention provides an amplitude processing device, comprising:

[0033] An amplitude processing network loading module is used to load the amplitude processing network trained by the method described in the first aspect above when the original amplitude data is collected from the pumped storage unit at multiple moments; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0034] A first original amplitude feature encoding module, used for inputting the original amplitude data at each moment into the first encoder to encode a first original amplitude feature;

[0035] A second original amplitude feature encoding module, used for simultaneously inputting the original amplitude data at multiple moments into the second encoder to encode a second original amplitude feature;

[0036] A fused amplitude feature interaction module, used for inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact as a fused amplitude feature;

[0037] A target amplitude feature decoding module, used for inputting the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0038] The amplitude task information generating module is used to input the target amplitude feature into the multi-task processing module to generate a plurality of amplitude task information.

[0039] A fifth aspect of the present invention provides an electronic device, the electronic device comprising:

[0040] at least one processor; and

[0041] a memory communicatively connected to the at least one processor; wherein,

[0042] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the training method of the amplitude processing network as described in the first aspect or the amplitude processing method as described in the second aspect.

[0043] A sixth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the training method of the amplitude processing network as described in the first aspect or the amplitude processing method as described in the second aspect.

[0044] A seventh aspect of the present invention provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the training method of the amplitude processing network as described in the first aspect or the amplitude processing method as described in the second aspect.

[0045] In this embodiment, when the original amplitude data of the pumped storage unit is collected at multiple moments, the amplitude processing network is loaded; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module; the original amplitude data at each moment is respectively input into the first encoder to encode the first original amplitude feature; the original amplitude data at multiple moments is simultaneously input into the second encoder to encode the second original amplitude feature; the multiple first original amplitude features and the second original amplitude features are input into the feature interaction module to interact as a fused amplitude feature; the fused amplitude feature is input into the decoder to be decoded into a target amplitude feature; the target amplitude feature is input into the multi-task processing module to generate multiple amplitude task information; a total loss value is generated based on the multiple amplitude task information; and the amplitude processing network is updated based on the total loss value. In this embodiment, the features are interactively fused between the encoder and the decoder, which can increase the amount of information of the features, thereby improving the robustness of the amplitude processing network, and improving the accuracy of multi-task processing. In addition, the amplitude features are adaptively used to perform multi-task processing, which has high flexibility, thereby improving processing efficiency.

[0046] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 It is a flow chart of a training method for an amplitude processing network provided in Embodiment 1 of the present invention.

[0049] Figure 2 It is a structural diagram of an amplitude processing network provided in Embodiment 1 of the present invention.

[0050] Figure 3 This is a flow chart of an amplitude processing method provided in Embodiment 2 of the present invention.

[0051] Figure 4 It is a structural schematic diagram of a training device for an amplitude processing network provided in Embodiment 3 of the present invention.

[0052] Figure 5 It is a structural schematic diagram of an amplitude processing device provided in Embodiment 4 of the present invention.

[0053] Figure 6 It is a structural schematic diagram of an electronic device provided in Embodiment 5 of the present invention. DETAILED DESCRIPTION

[0054] 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 should fall within the scope of protection of the present invention.

[0055] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can cover sequential implementations other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0056] Embodiment 1

[0057] See also Figure 1 , shows a flow chart of a method for training an amplitude processing network provided by a first embodiment of the present invention, the method can be executed by a training device for an amplitude processing network, the training of the amplitude processing network and its amplitude processing device can be implemented in the form of hardware and / or software, and the training device for the amplitude processing network can be configured in an electronic device. Figure 1 As shown, the method includes:

[0058] Step 101: When raw amplitude data of the pumped storage unit is collected at multiple time points, an amplitude processing network is loaded.

[0059] In this embodiment, vibration sensors can be installed on various equipment such as the upper frame, lower frame, water guide top cover, thrust bearing, etc. on the pumped-storage unit. During the operation of the pumped-storage unit, the vibration amplitude collected by the vibration sensor at each moment can be read and recorded as original amplitude data.

[0060] like Figure 2 As shown, the original vibration data at multiple moments can be used to construct the training data of each Batch [{x t-T+1 ,x t-T+2 ,…,x t},x t+1 ,x” t+1 ], where {x t-T+1 ,x t-T+2 ,…,x t} is the original amplitude data from time t-T+1 to time t, x t+1 is the original amplitude data at time t+1, x” t+1 Based on x t+1 Constructed abnormal amplitude data.

[0061] In this embodiment, an amplitude processing network can be constructed based on deep learning, wherein the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module.

[0062] When training the amplitude processing network, load the amplitude processing network into memory and run it.

[0063] Step 102: Input the original amplitude data at each moment into a first encoder to encode a first original amplitude feature.

[0064] In this embodiment, if Figure 2 As shown, the original amplitude data at each moment can be input into the first encoder respectively, and the first encoder encodes the original amplitude data at each moment respectively to obtain the first original amplitude feature C1.

[0065] Exemplarily, the first encoder is a backbone network Backbone that applies a self-attention mechanism. Optionally, the backbone network that applies the self-attention mechanism includes a neural network Transformer, and the backbone network is constructed using the Transformer as a module.

[0066] Step 103: Input the original amplitude data at multiple moments into the second encoder simultaneously to encode the second original amplitude feature.

[0067] In this embodiment, if Figure 2 As shown, the original amplitude data at multiple moments can be respectively input into the second encoder, and the second encoder encodes the original amplitude data at multiple moments respectively to obtain the second original amplitude feature C2.

[0068] Exemplarily, the second encoder is a backbone network Backbone using a recurrent neural network (RNN). Optionally, the backbone network using the recurrent neural network includes a long short-term memory network (LSTM).

[0069] Step 104: Input the plurality of first original amplitude features and the second original amplitude features into a feature interaction module to interact with each other to form a fused amplitude feature.

[0070] In this embodiment, if Figure 2 As shown, a plurality of first original amplitude features C1 and a second original amplitude feature C2 may be input into a feature interaction module, and the feature interaction module interactively fuses the plurality of first original amplitude features C1 and the second original amplitude features C2 to obtain a fused amplitude feature C3.

[0071] In one embodiment of the present invention, step 104 may include the following steps:

[0072] Step 1041: Perform a normalization operation on each first original amplitude feature to obtain a first candidate amplitude feature.

[0073] In this embodiment, if Figure 2 As shown, a normalization operation may be performed on each first original amplitude feature C1 to obtain a first candidate amplitude feature S1.

[0074] For example, Figure 2 As shown, the normalization operation is Instance Norm (instance normalization), and the first candidate amplitude feature is expressed as:

[0075]

[0076] Among them, i is the moment, i=1, 2,..., T, C1 is the first original amplitude feature, U1 is the average value of the first original amplitude feature, σ1 is the variance of the first original amplitude feature, ε is a hyperparameter (such as 0.00001), and S1 is the first candidate amplitude feature.

[0077] Step 1042: Perform a normalization operation on the second original amplitude feature to obtain a second candidate amplitude feature.

[0078] In this embodiment, if Figure 2 As shown, a normalization operation may be performed on each second original amplitude feature C2 to obtain a second candidate amplitude feature S2.

[0079] Generally, the normalization operation performed on each first original amplitude feature is the same as the normalization operation performed on the second original amplitude feature.

[0080] For example, the normalization operation is Instance Norm (instance normalization), and the second candidate amplitude feature is expressed as:

[0081]

[0082] Among them, C2 is the second original amplitude feature, U2 is the average value of the second original amplitude feature, σ2 is the variance of the second original amplitude feature, ε is a hyperparameter (such as 0.00001), and S2 is the second candidate amplitude feature.

[0083] Step 1043: Calculate the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature.

[0084] In this embodiment, for the same normalization operation, an attention mechanism may be used to calculate the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature.

[0085] In the specific implementation, Figure 2 As shown, in the attention mechanism, the first candidate amplitude feature S1 is used as the key value vector K, the second candidate amplitude feature S2 is used as the query vector Q, and multiple reference weights w are calculated.

[0086] At this time, the key value vector K (ie, the first candidate amplitude feature S1) is multiplied with the query vector Q (ie, the second candidate amplitude feature S2), and the result of the dot multiplication is normalized by the softmax function to obtain each of the multiple reference weights w.

[0087] Then, multiple benchmark weights w are expressed as:

[0088]

[0089] Among them, K∈R T*D , Q∈R 1*D , w∈R T*1 , T is the time period, and D is the characteristic dimension of each time series.

[0090] The alignment operation is performed on the average value. At this time, the product between the average value of the first original amplitude feature and the reference weight is added to obtain the average value after alignment. Then, the average value u after alignment can be expressed as

[0091] The average value of the second original amplitude feature and the aligned average value are linearly fused into a feature interaction average value. For example, the feature interaction average value U can be expressed as U=(u+U2)*0.5.

[0092] The variance is aligned. At this time, the product of the variance of the first original amplitude feature and the reference weight is added to obtain the aligned variance. Then, the aligned variance σ can be expressed as

[0093] The variance of the second original amplitude feature and the aligned variance are linearly fused into a feature interaction variance. For example, the feature interaction variance U can be expressed as Σ=(σ+σ2)*0.5.

[0094] Step 1044: align the second candidate vibration features into fused amplitude features according to the attention weights.

[0095] In this embodiment, if Figure 2 As shown, the second candidate vibration feature S2 is aligned (Align) to the target amplitude feature C3 according to the attention weight.

[0096] In the specific implementation, Figure 2 As shown, the fused amplitude feature is expressed as:

[0097] C3=S2×Σ+U;

[0098] Among them, C3 is the fused amplitude feature, S2 is the second candidate vibration feature, U is the feature interaction mean, and Σ is the feature interaction variance.

[0099] In this embodiment, according to the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature under the same normalization operation, the second candidate vibration feature is aligned to a fused amplitude feature, so that a more robust feature can be obtained.

[0100] Step 105: Input the fused amplitude feature into a decoder and decode it into a target amplitude feature.

[0101] In this embodiment, if Figure 2 As shown, the fused amplitude feature C3 can be input into the decoder Decoder, and the decoder Decoder decodes the fused amplitude feature C3 into the target amplitude feature.

[0102] For example, Figure 2 As shown, the decoder is a long short-term memory network LSTM.

[0103] Step 106: Input the target amplitude feature into the multi-task processing module to generate multiple amplitude task information.

[0104] In this embodiment, the target amplitude feature may be input into the multi-task processing module, and the multi-task processing module generates a plurality of amplitude task information according to the target amplitude feature in accordance with the business requirements.

[0105] In a specific implementation, the multi-task processing module includes a sequence prediction module, a working condition classification module and an abnormality discriminator. Accordingly, the multiple amplitude task information includes target amplitude data, target working condition status and abnormality discrimination labels.

[0106] Among them, the sequence prediction module is a plurality of fully connected layers FC, the working condition classification module is a plurality of fully connected layers FC, and the abnormal discriminator is a plurality of fully connected layers FC.

[0107] Exemplarily, the sequence prediction module is three fully connected layers FC, whose feature sizes are 256, 512, and 2 respectively.

[0108] The working condition classification module consists of three fully connected layers FC, whose feature sizes are 256, 512, and k respectively, where k is the number of target working condition states (i.e., categories).

[0109] The anomaly discriminator consists of three fully connected layers FC, whose feature sizes are 256, 512, and 2 respectively.

[0110] So, if Figure 2 As shown, the target amplitude feature is input into the sequence prediction module to generate the target amplitude data x' at the next moment. t+1 .

[0111] The target amplitude characteristics are input into the operating condition classification module to classify the target operating condition state of the pumped storage unit.

[0112] The target amplitude data is input into the abnormality discriminator to generate an abnormality discrimination label; wherein the abnormality discrimination label indicates whether the target amplitude data is abnormal.

[0113] Furthermore, for the abnormality discriminator, the real original amplitude data x t+1 Construct a batch of abnormal amplitude data x" t+1 .

[0114] Among them, the construction method includes: swapping the column position of the original amplitude data, changing the value of a certain measuring point data in the original amplitude data to make it exceed the normal data range, changing the data of multiple column position points in the original amplitude data to make it fluctuate more, and so on.

[0115] The true original amplitude data x t+1 , abnormal amplitude data x" t+1 With the target amplitude data x' t+1 , calculate the difference between the two, and input the difference into the abnormal discriminator for training until the difference (x' t+1 -x t+1 ) is used as the abnormality discrimination label for input, and all other differences are abnormal.

[0116] Step 107: Generate a total loss value according to multiple amplitude task information.

[0117] In this embodiment, multiple amplitude task information may be substituted into a preset loss function for calculation, thereby generating a total loss value.

[0118] In the specific implementation, Figure 2 As shown, on the one hand, the loss loss is calculated based on the difference (such as mean square error, etc.) between the target amplitude data and the original amplitude data at the next moment, thereby generating the first sub-loss value L 预测loss .

[0119] On the other hand, a true and false discriminator is constructed and trained using the training mode of the generative adversarial network (GAN). The target amplitude data x' predicted at time t+1 is t+1 The actual original amplitude data x t+1 Make a True / False judgment, that is, judge the current data (that is, the target amplitude data x' t+1 Or the actual raw amplitude data x t+1 ) is true data or false data.

[0120] The target amplitude data is input into a preset true and false discriminator to generate a true and false discrimination label; wherein the true and false discrimination label indicates whether the target amplitude data is true or false.

[0121] Generate the second sub-loss value L based on the true and false discrimination label gan .

[0122] At this time, the second sub-loss value L gan It is expressed as:

[0123]

[0124] Among them, G represents the generation network (i.e., the first encoder, the second encoder, the feature interaction module, the decoder and the sequence prediction module), D represents the true and false discriminator, P data (x) represents the distribution of the original amplitude data, P z (z) represents the distribution of target amplitude data, z represents the original amplitude data, z = {x t-T+1 ,x t-T+2 ,…,x t}, x represents the real original amplitude data x at the next moment t+1 .

[0125] On the other hand, the third sub-loss value L is generated according to the difference (such as cross entropy, etc.) between the target operating state and the sample operating state marked for the pumped storage unit. 工况Discriminator .

[0126] On the other hand, the fourth sub-loss value L is generated based on the abnormal discrimination label using binary cross entropy or other methods 异常 .

[0127] The first sub-loss value, the second sub-loss value, the third sub-loss value and the fourth sub-loss value are fused into a total loss value in a linear or nonlinear manner.

[0128] For example, the total loss value can be expressed as:

[0129] L total =r1*L 预测loss +r2* Lgan +r3*L 工况Discriminator +r4*L 异常

[0130] Among them, L total is the total loss value, r1, r2, r3, and r4 are all hyperparameters.

[0131] Step 108: Update the amplitude processing network according to the total loss value.

[0132] In this embodiment, the total loss value can be substituted into an optimization algorithm such as SGD (stochastic gradient descent) and Adam (Adaptive momentum) to calculate the update amplitude of the parameters in the amplitude processing network, and the parameters in the amplitude processing network are updated according to the update amplitude.

[0133] In addition, training conditions can be set in advance as conditions for stopping training the amplitude processing network, for example, the number of iterations reaches a certain threshold, the loss value is less than a certain threshold, the change amplitude of the loss value in multiple iterative training is less than a certain threshold, etc. In each round of iterative training, it is determined whether the current corresponding parameters meet the training conditions.

[0134] If the training conditions are met, the amplitude processing network can be considered to have completed training.

[0135] If the training conditions are not met, the next round of iterative training can be entered, and steps 102 to 108 are executed again, and the iterative training is repeated in this way until the amplitude processing network is trained.

[0136] In this embodiment, when the original amplitude data of the pumped storage unit is collected at multiple moments, the amplitude processing network is loaded; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module; the original amplitude data at each moment is respectively input into the first encoder to encode the first original amplitude feature; the original amplitude data at multiple moments is simultaneously input into the second encoder to encode the second original amplitude feature; the multiple first original amplitude features and the second original amplitude features are input into the feature interaction module to interact as a fused amplitude feature; the fused amplitude feature is input into the decoder to be decoded into a target amplitude feature; the target amplitude feature is input into the multi-task processing module to generate multiple amplitude task information; a total loss value is generated based on the multiple amplitude task information; and the amplitude processing network is updated based on the total loss value. In this embodiment, the features are interactively fused between the encoder and the decoder, which can increase the amount of information of the features, thereby improving the robustness of the amplitude processing network, and improving the accuracy of multi-task processing. In addition, the amplitude features are adaptively used to perform multi-task processing, which has high flexibility, thereby improving processing efficiency.

[0137] Embodiment 2

[0138] See also Figure 3 , shows a flow chart of an amplitude processing method provided by the first embodiment of the present invention, the method can be executed by an amplitude processing device, the training of the amplitude processing network and the amplitude processing device can be implemented in the form of hardware and / or software, and the amplitude processing device can be configured in an electronic device. Figure 3 As shown, the method includes:

[0139] Step 301: when raw amplitude data of the pumped storage unit is collected at multiple time points, an amplitude processing network is loaded.

[0140] In actual application, if the method of the first embodiment is applied to complete the training of the amplitude processing network, and the amplitude processing network is verified, tested and other operations are completed, the amplitude processing network can be deployed and run online.

[0141] The amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module.

[0142] Step 302: Input the original amplitude data at each moment into a first encoder to encode a first original amplitude feature.

[0143] Exemplarily, the first encoder is a backbone network that applies a self-attention mechanism, and the backbone network that applies the self-attention mechanism includes a neural network Transformer.

[0144] Step 303: Input the original amplitude data at multiple moments into the second encoder simultaneously to encode the second original amplitude feature.

[0145] Exemplarily, the second encoder is a long short-term memory network.

[0146] Step 304: Input the plurality of first original amplitude features and the second original amplitude features into a feature interaction module to interact with each other to form a fused amplitude feature.

[0147] In one embodiment of the present invention, step 304 may include the following steps:

[0148] Step 3041: Perform a normalization operation on each first original amplitude feature to obtain a first candidate amplitude feature.

[0149] Exemplarily, the first candidate amplitude feature is expressed as:

[0150]

[0151] Among them, i is the moment, C1 is the first original amplitude feature, U1 is the average value of the first original amplitude feature, σ1 is the variance of the first original amplitude feature, ε is the hyperparameter, and S1 is the first candidate amplitude feature.

[0152] Step 3042: Perform a normalization operation on the second original amplitude feature to obtain a second candidate amplitude feature.

[0153] Exemplarily, the second candidate amplitude feature is expressed as:

[0154]

[0155] Among them, C2 is the second original amplitude feature, U2 is the average value of the second original amplitude feature, σ2 is the variance of the second original amplitude feature, ε is a hyperparameter, and S2 is the second candidate amplitude feature.

[0156] Step 3043: Calculate the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature.

[0157] In the attention mechanism, the first candidate amplitude feature is used as the key value vector and the second candidate amplitude feature is used as the query vector to calculate multiple benchmark weights; the product of the average value of the first original amplitude feature and the benchmark weight is added to obtain the aligned average value; the average value of the second original amplitude feature and the aligned average value are linearly fused into the feature interaction average value; the product of the variance of the first original amplitude feature and the benchmark weight is added to obtain the aligned variance; the variance of the second original amplitude feature and the aligned variance are linearly fused into the feature interaction variance.

[0158] Step 3044: align the second candidate vibration feature into a fused amplitude feature based on the attention weight.

[0159] Exemplarily, the fused amplitude feature is expressed as:

[0160] C3=S2×Σ+U;

[0161] Among them, C3 is the fused amplitude feature, S2 is the second candidate vibration feature, U is the feature interaction average, and Σ is the feature interaction variance.

[0162] Step 305: Input the fused amplitude feature into a decoder and decode it into a target amplitude feature.

[0163] Exemplarily, the decoder is a long short-term memory network.

[0164] Step 306: Input the target amplitude feature into the multi-task processing module to generate multiple amplitude task information.

[0165] In a specific implementation, the multi-task processing module includes a sequence prediction module, a working condition classification module and an abnormality discriminator, and the multiple amplitude task information includes target amplitude data, target working condition status and abnormality discrimination labels.

[0166] Exemplarily, the sequence prediction module is a plurality of fully connected layers, the operating condition classification module is a plurality of fully connected layers, and the abnormality discriminator is a plurality of fully connected layers.

[0167] On the one hand, the target amplitude features are input into the sequence prediction module to generate the target amplitude data at the next moment.

[0168] On the other hand, the target amplitude characteristics are input into the operating condition classification module to classify the target operating condition state of the pumped storage unit.

[0169] On the other hand, the target amplitude data is input into an abnormality discriminator to generate an abnormality discrimination label; the abnormality discrimination label indicates whether the target amplitude data is abnormal.

[0170] In this embodiment, since steps 301 to 306 are basically similar to the application of embodiment 1, the description is relatively simple, and the relevant parts can be referred to the partial description of embodiment 1, and the embodiment of the present invention is not described in detail here.

[0171] In this embodiment, when the original amplitude data of the pumped storage unit is collected at multiple moments, the amplitude processing network is loaded; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module; the original amplitude data at each moment is respectively input into the first encoder to encode the first original amplitude feature; the original amplitude data at multiple moments is simultaneously input into the second encoder to encode the second original amplitude feature; the multiple first original amplitude features and the second original amplitude features are input into the feature interaction module to interact as a fused amplitude feature; the fused amplitude feature is input into the decoder to be decoded into a target amplitude feature; the target amplitude feature is input into the multi-task processing module to generate multiple amplitude task information. In this embodiment, the features are interactively fused between the encoder and the decoder, which can increase the information content of the features, thereby improving the robustness of the amplitude processing network, improving the accuracy of multi-task processing, and adaptively using the amplitude features to perform multi-task processing, which has high flexibility, thereby improving processing efficiency.

[0172] Embodiment 3

[0173] See also Figure 4 , shows a schematic diagram of the structure of a training device for an amplitude processing network provided by Embodiment 3 of the present invention. Figure 4 As shown, the device comprises:

[0174] The amplitude processing network loading module 401 is used to load the amplitude processing network when the original amplitude data of the pumped storage unit is collected at multiple times; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0175] A first original amplitude feature encoding module 402, configured to input the original amplitude data at each moment into the first encoder to encode a first original amplitude feature;

[0176] A second original amplitude feature encoding module 403, configured to simultaneously input the original amplitude data at multiple moments into the second encoder to encode a second original amplitude feature;

[0177] A fused amplitude feature interaction module 404 is used to input a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact as a fused amplitude feature;

[0178] A target amplitude feature decoding module 405 is used to input the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0179] An amplitude task information generating module 406 is used to input the target amplitude feature into the multi-task processing module to generate a plurality of amplitude task information;

[0180] A total loss value generating module 407, used to generate a total loss value according to the plurality of amplitude task information;

[0181] The amplitude processing network updating module 408 is used to update the amplitude processing network according to the total loss value.

[0182] In one embodiment of the present invention, the fused amplitude feature interaction module 404 includes:

[0183] A first candidate amplitude feature generating module, configured to perform a normalization operation on each of the first original amplitude features to obtain a first candidate amplitude feature;

[0184] A second candidate amplitude feature generating module, configured to perform a normalization operation on the second original amplitude feature to obtain a second candidate amplitude feature;

[0185] an attention weight calculation module, used to calculate the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature;

[0186] A fused amplitude feature alignment module is used to align the second candidate vibration feature into a fused amplitude feature according to the attention weight.

[0187] In one embodiment of the present invention, the first candidate amplitude feature is expressed as:

[0188]

[0189] Wherein, i is the moment, C1 is the first original amplitude feature, U1 is the average value of the first original amplitude feature, σ1 is the variance of the first original amplitude feature, ε is a hyperparameter, and S1 is the first candidate amplitude feature;

[0190] The second candidate amplitude feature is expressed as:

[0191]

[0192] Among them, C2 is the second original amplitude feature, U2 is the average value of the second original amplitude feature, σ2 is the variance of the second original amplitude feature, ε is a hyperparameter, and S2 is the second candidate amplitude feature.

[0193] In one embodiment of the present invention, the attention weight calculation module includes:

[0194] A reference weight calculation module, used to calculate a plurality of reference weights in an attention mechanism, using the first candidate amplitude feature as a key value vector and the second candidate amplitude feature as a query vector;

[0195] an average value alignment module, configured to add the product of the average value of the first original amplitude feature and the reference weight to obtain an aligned average value;

[0196] an average weight calculation module, used for linearly fusing the average of the second original amplitude feature and the aligned average into a feature interaction average;

[0197] A variance alignment module, configured to add the product of the variance of the first original amplitude feature and the reference weight to obtain an aligned variance;

[0198] A variance weight calculation module, used for linearly fusing the variance of the second original amplitude feature and the aligned variance into a feature interaction variance;

[0199] The fusion amplitude feature is expressed as:

[0200] C3=S2×Σ+U;

[0201] Among them, C3 is the fused amplitude feature, S2 is the second candidate vibration feature, U is the feature interaction average, and Σ is the feature interaction variance.

[0202] In one embodiment of the present invention, the multi-task processing module includes a sequence prediction module, a working condition classification module and an abnormality discriminator, and the plurality of amplitude task information includes target amplitude data, target working condition status and abnormality discrimination labels;

[0203] The amplitude task information generating module 406 includes:

[0204] A target amplitude data generating module, used for inputting the target amplitude feature into the sequence prediction module to generate target amplitude data at the next moment;

[0205] A target operating state classification module, used for inputting the target amplitude feature into the operating state classification module to classify the target operating state of the pumped storage unit;

[0206] The abnormality discrimination label generation module is used to input the target amplitude data into the abnormality discriminator to generate an abnormality discrimination label; the abnormality discrimination label indicates whether the target amplitude data is abnormal.

[0207] In one embodiment of the present invention, the total loss value generating module 407 includes:

[0208] A first sub-loss value generating module, configured to generate a first sub-loss value according to a difference between the target amplitude data and the original amplitude data at a next moment;

[0209] A true-false discrimination label generation module, used for inputting the target amplitude data into a preset true-false discriminator to generate a true-false discrimination label; the true-false discrimination label indicates whether the target amplitude data is true or false;

[0210] A second sub-loss value generating module, used to generate a second sub-loss value according to the true and false discrimination label;

[0211] A third sub-loss value generating module, configured to generate a third sub-loss value according to a difference between the target operating state and a sample operating state marked for the pumped-storage unit;

[0212] A fourth sub-loss value generating module, used to generate a fourth sub-loss value according to the abnormality discrimination label;

[0213] A total loss value fusion module is used to fuse the first sub-loss value, the second sub-loss value, the third sub-loss value and the fourth sub-loss value into a total loss value.

[0214] In one embodiment of the present invention, the first encoder applies a backbone network of a self-attention mechanism, the backbone network applying the self-attention mechanism includes a neural network Transformer, the second encoder is a long short-term memory network, and the decoder is a long short-term memory network;

[0215] The sequence prediction module is a plurality of fully connected layers, the operating condition classification module is a plurality of fully connected layers, and the abnormality discriminator is a plurality of fully connected layers.

[0216] The training device for an amplitude processing network provided in an embodiment of the present invention can execute the training method for an amplitude processing network provided in any embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the training method for an amplitude processing network.

[0217] Embodiment 4

[0218] See also Figure 5 , shows a schematic diagram of the structure of an amplitude processing device provided by Embodiment 4 of the present invention. Figure 5 As shown, the device comprises:

[0219] An amplitude processing network loading module 501 is used to load the amplitude processing network trained according to the method described in Embodiment 1 when the original amplitude data is collected from the pumped storage unit at multiple moments; the amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module;

[0220] A first original amplitude feature encoding module 502, used to input the original amplitude data at each moment into the first encoder to encode a first original amplitude feature;

[0221] A second original amplitude feature encoding module 503, configured to simultaneously input the original amplitude data at multiple moments into the second encoder to encode a second original amplitude feature;

[0222] A fused amplitude feature interaction module 504 is used to input a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact as a fused amplitude feature;

[0223] A target amplitude feature decoding module 505 is used to input the fused amplitude feature into the decoder to decode it into a target amplitude feature;

[0224] The amplitude task information generating module 506 is used to input the target amplitude feature into the multi-task processing module to generate a plurality of amplitude task information.

[0225] In one embodiment of the present invention, the fused amplitude feature interaction module 504 includes:

[0226] A first candidate amplitude feature generating module, configured to perform a normalization operation on each of the first original amplitude features to obtain a first candidate amplitude feature;

[0227] A second candidate amplitude feature generating module, configured to perform a normalization operation on the second original amplitude feature to obtain a second candidate amplitude feature;

[0228] an attention weight calculation module, used to calculate the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature;

[0229] A fused amplitude feature alignment module is used to align the second candidate vibration feature into a fused amplitude feature according to the attention weight.

[0230] In one embodiment of the present invention, the first candidate amplitude feature is expressed as:

[0231]

[0232] Wherein, i is the moment, C1 is the first original amplitude feature, U1 is the average value of the first original amplitude feature, σ1 is the variance of the first original amplitude feature, ε is a hyperparameter, and S1 is the first candidate amplitude feature;

[0233] The second candidate amplitude feature is expressed as:

[0234]

[0235] Among them, C2 is the second original amplitude feature, U2 is the average value of the second original amplitude feature, σ2 is the variance of the second original amplitude feature, ε is a hyperparameter, and S2 is the second candidate amplitude feature.

[0236] In one embodiment of the present invention, the attention weight calculation module includes:

[0237] A reference weight calculation module, used to calculate a plurality of reference weights in an attention mechanism, using the first candidate amplitude feature as a key value vector and the second candidate amplitude feature as a query vector;

[0238] an average value alignment module, configured to add the product of the average value of the first original amplitude feature and the reference weight to obtain an aligned average value;

[0239] an average weight calculation module, used for linearly fusing the average of the second original amplitude feature and the aligned average into a feature interaction average;

[0240] A variance alignment module, configured to add the product of the variance of the first original amplitude feature and the reference weight to obtain an aligned variance;

[0241] A variance weight calculation module, used for linearly fusing the variance of the second original amplitude feature and the aligned variance into a feature interaction variance;

[0242] The fusion amplitude feature is expressed as:

[0243] C3=S2×Σ+U;

[0244] Among them, C3 is the fused amplitude feature, S2 is the second candidate vibration feature, U is the feature interaction average, and Σ is the feature interaction variance.

[0245] In one embodiment of the present invention, the multi-task processing module includes a sequence prediction module, a working condition classification module and an abnormality discriminator, and the plurality of amplitude task information includes target amplitude data, target working condition status and abnormality discrimination labels;

[0246] The amplitude task information generation module 506 includes:

[0247] A target amplitude data generating module, used for inputting the target amplitude feature into the sequence prediction module to generate target amplitude data at the next moment;

[0248] A target operating state classification module, used for inputting the target amplitude feature into the operating state classification module to classify the target operating state of the pumped storage unit;

[0249] The abnormality discrimination label generation module is used to input the target amplitude data into the abnormality discriminator to generate an abnormality discrimination label; the abnormality discrimination label indicates whether the target amplitude data is abnormal.

[0250] In one embodiment of the present invention, the first encoder applies a backbone network of a self-attention mechanism, the backbone network applying the self-attention mechanism includes a neural network Transformer, the second encoder is a long short-term memory network, and the decoder is a long short-term memory network;

[0251] The sequence prediction module is a plurality of fully connected layers, the operating condition classification module is a plurality of fully connected layers, and the abnormality discriminator is a plurality of fully connected layers.

[0252] The amplitude processing device provided in the embodiment of the present invention can execute the amplitude processing method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects for executing the amplitude processing method.

[0253] Embodiment 5

[0254] See also Figure 6 , shows a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, blade servers, mainframe computers, and other suitable computers. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0255] like Figure 6As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0256] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0257] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 performs the various methods and processes described above, such as the training method of the amplitude processing network or the amplitude processing method.

[0258] In some embodiments, the training method of the amplitude processing network or the amplitude processing method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the training method of the amplitude processing network or the amplitude processing method described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to execute the training method of the amplitude processing network or the amplitude processing method by any other appropriate means (e.g., by means of firmware).

[0259] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0260] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0261] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0262] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0263] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0264] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0265] Embodiment 6

[0266] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the training method of the amplitude processing network or the amplitude processing method provided by any embodiment of the present invention is implemented.

[0267] In the process of implementation, the computer program product can be written in one or more programming languages ​​or a combination thereof to perform the computer program code of the present invention, including object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

[0268] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0269] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A training method for an amplitude processing network, characterized in that: include: When the original amplitude data of the pumped storage unit is collected at multiple times, the amplitude processing network is loaded; The amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module; Inputting the original amplitude data at each moment into the first encoder to encode the first original amplitude feature; Inputting the original amplitude data at multiple moments into the second encoder simultaneously to encode a second original amplitude feature; Inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact with each other to form a fused amplitude feature; Inputting the fused amplitude feature into the decoder to be decoded into a target amplitude feature; Inputting the target amplitude feature into the multi-task processing module to generate multiple amplitude task information; generating a total loss value according to the plurality of amplitude task information; The amplitude processing network is updated according to the total loss value.

2. The method according to claim 1, characterized in that The step of inputting the plurality of first original amplitude features and the second original amplitude features into the feature interaction module to interact with each other to form a fused amplitude feature comprises: Performing a normalization operation on each of the first original amplitude features to obtain a first candidate amplitude feature; Performing a normalization operation on the second original amplitude feature to obtain a second candidate amplitude feature; Calculating an attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature; The second candidate vibration feature is aligned as a fused amplitude feature according to the attention weight.

3. The method according to claim 2, characterized in that The first candidate amplitude feature is expressed as: Wherein, i is the moment, C1 is the first original amplitude feature, U1 is the average value of the first original amplitude feature, σ1 is the variance of the first original amplitude feature, ε is a hyperparameter, and S1 is the first candidate amplitude feature; The second candidate amplitude feature is expressed as: Among them, C2 is the second original amplitude feature, U2 is the average value of the second original amplitude feature, σ2 is the variance of the second original amplitude feature, ε is a hyperparameter, and S2 is the second candidate amplitude feature.

4. The method according to claim 3, characterized in that The calculating the attention weight of the first candidate amplitude feature relative to the second candidate amplitude feature comprises: In the attention mechanism, a plurality of reference weights are calculated using the first candidate amplitude feature as a key value vector and the second candidate amplitude feature as a query vector; Adding the product of the average value of the first original amplitude feature and the reference weight to obtain an aligned average value; Linearly fuse the average value of the second original amplitude feature and the average value after alignment into a feature interaction average value; Adding the product of the variance of the first original amplitude feature and the reference weight to obtain an aligned variance; Linearly fusing the variance of the second original amplitude feature and the aligned variance into a feature interaction variance; The fusion amplitude feature is expressed as: C3=S2×Σ+U; Among them, C3 is the fused amplitude feature, S2 is the second candidate vibration feature, U is the feature interaction average, and Σ is the feature interaction variance.

5. The method according to any one of claims 1 to 4, characterized in that The multi-task processing module includes a sequence prediction module, a working condition classification module and an abnormality discriminator, and the plurality of amplitude task information includes target amplitude data, target working condition status and abnormality discrimination labels; The step of inputting the target amplitude feature into the multi-task processing module to generate a plurality of amplitude task information includes: Inputting the target amplitude feature into the sequence prediction module to generate target amplitude data at the next moment; Inputting the target amplitude feature into the operating condition classification module to classify the target operating condition state of the pumped storage unit; The target amplitude data is input into the abnormality discriminator to generate an abnormality discrimination label; the abnormality discrimination label indicates whether the target amplitude data is abnormal.

6. The method according to claim 5, characterized in that The generating of the total loss value according to the plurality of amplitude task information comprises: generating a first sub-loss value according to a difference between the target amplitude data and the original amplitude data at a next moment; Inputting the target amplitude data into a preset true-false discriminator to generate a true-false discrimination label; the true-false discrimination label indicates whether the target amplitude data is true or false; Generating a second sub-loss value according to the true and false discrimination label; generating a third sub-loss value according to a difference between the target operating state and a sample operating state marked for the pumped-storage unit; Generating a fourth sub-loss value according to the abnormal discrimination label; The first sub-loss value, the second sub-loss value, the third sub-loss value and the fourth sub-loss value are merged into a total loss value.

7. The method according to claim 5, characterized in that The first encoder is a backbone network applying a self-attention mechanism, the backbone network applying the self-attention mechanism includes a neural network Transformer, the second encoder is a long short-term memory network, and the decoder is a long short-term memory network; The sequence prediction module is a plurality of fully connected layers, the operating condition classification module is a plurality of fully connected layers, and the abnormality discriminator is a plurality of fully connected layers.

8. An amplitude processing method, characterized in that: include: When raw amplitude data is collected from the pumped storage unit at multiple moments, an amplitude processing network trained according to the method described in any one of claims 1 to 7 is loaded; The amplitude processing network includes a first encoder, a second encoder, a feature interaction module, a decoder and a multi-task processing module; Inputting the original amplitude data at each moment into the first encoder to encode the first original amplitude feature; Inputting the original amplitude data at multiple moments into the second encoder simultaneously to encode a second original amplitude feature; Inputting a plurality of the first original amplitude features and the second original amplitude features into the feature interaction module to interact with each other to form a fused amplitude feature; Inputting the fused amplitude feature into the decoder to be decoded into a target amplitude feature; The target amplitude feature is input into the multi-task processing module to generate a plurality of amplitude task information.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can perform the training method of the amplitude processing network according to any one of claims 1 to 7 or the amplitude processing method according to claim 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the training method of the amplitude processing network according to any one of claims 1 to 7 or the amplitude processing method according to claim 8 is implemented.

Images