State detection method and device, detection equipment and storage medium

[0081] In order to make the object, technical solutions, and advantages of the present application, the embodiments will be further described in detail below with reference to the accompanying drawings.

[0082] It will be appreciated that the terms "first", "second", and the like as used herein can describe various concepts herein, unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish a concept with another concept. By way of example, without departing from the scope of the present application, the first node may be referred to as a second node, the second node is called a first node.

[0083] As used herein, the term "at least one", "plurality", "every", "any one" and the like, comprising a at least one, two or more, comprising a plurality of two or more, each a plurality of means corresponding to each one, according to any one refers to any of a plurality. For example, a plurality of training samples comprise three training samples, each training sample refers to a sample of each of the three training samples in the training, the training according to any one refers to any of these three samples a, it may be the first a, it may be a second, a third may be.

[0084] figure 1 It is a schematic diagram of an implementation environment provided by the embodiment of the present application. See figure 1 The environmental embodiment comprises: the detection device 101 and the electronic apparatus 102. The electronic device 101 and the detection device 102 are in the same space, the detection device 102 is configured with an audio capture device, such as a microphone or the like, audio data can be acquired during operation of the electronic device 101 is generated. Detection apparatus 102 through the audio collection device, audio data sample collected during the normal operation of the electronic device 101 is generated, so that the audio data based on the sample training state detection model, this model is used to determine the state detecting operating state of the electronic device 101. Then, the detection device 102 through the audio collection device to capture audio data process 101 to run the electronic device generated, the state detection model trained based on the audio data process 101 to run the electronic device generated processed to determine that the electronic device 101 operating condition.

[0085] figure 2 It is a schematic diagram of another embodiment of the present application provides environmental embodiment, see figure 2 The embodiment environment comprising: an electronic device 201, detecting device 202 and an audio capture device 203. The electronic device 201 and an audio capture device 203 is in the same space, between the detection apparatus 202 and the audio capture device 203 via wireless or wired network, audio capture device 203 can collect audio data sample 201 during operation of the electronic device generated in the detection the device 202 sends the collected audio data samples, sample based detection device 202 detects the audio data training state model, this model is used to determine the operating state detecting state of the electronic device 201. The case of audio capture device 201 203 produced during the operation of the electronic device to capture audio data sent to the collection device 202 to detect the audio data, the audio data after the detection of the state detection device 202 based on the trained model, received processed to determine the operating state of the electronic device 201.

[0086] In one possible implementation, the detecting apparatus 202 is a terminal, the terminal and the audio capture device connected via a wireless or wired network 203.

[0087] In another possible implementation, the detecting apparatus 202 is a server, and audio capture device connected via a wireless or wired network 203.

[0088] In another possible implementation, the detecting apparatus 202 includes a terminal and a server, between the terminal 203 and the audio capture device connected via a wireless or wired network, and between a terminal and a server connected via a wireless or wired network. Audio capture device 203 can collect samples of the audio data during operation of the electronic device 201 generates and transmits the collected sample audio data to the terminal, the terminal is forwarded to the server, the server training state based on the sample data of the audio detection model, this model is used for detecting the state of determine the operating state of the electronic device 201. Audio capture device 203 of the case 201 generated during operation of an electronic device to capture audio data, the terminal transmits the collected audio data, forwarded by the terminal to the server, then the server status detection model based on the training, the received the audio data are processed 201 to determine the operating state of the electronic device, back to the terminal, the terminal displays the operating state, so that the person operating the electronic device 201 can view the operating state on the terminal.

[0089] Alternatively, the server after the training state detection model to the terminal, in the case of an audio capture device 203 produced during 201 running collected into an electronic device audio data, transmitted the collected audio data to the terminal, after the training terminal based on the state detection model, the received audio data are processed 201 to determine the operating state of the electronic device, the terminal displays the operating state, so that the person operating the electronic device 201 can view the operating state on the terminal.

[0090]Alternatively, the terminal is a terminal for any type of smart mobile phone, a desktop computer, a smart wear device, or a smart home device. The server is a server, or a server cluster consisting of several servers, or a cloud computing service center. .

[0091] The methods provided in the present application embodiment can be applied to a variety of scenes.

[0092] For example, the method provided by the present application embodiment is applied to a scenario that performs state detection against any high voltage shift power station device. The detection device is generated by the sample audio data generated in the case of normal operation of the high-voltage substation device, and the state detection model is trained using the method provided by the present application embodiment, and thereafter, the operation process of any high-voltage variable power station device is operated based on the status detection model based on the training. The audio data generated in the middle is detected to determine the operating state of the high voltage substation device.

[0093] image 3 It is a flow chart of a state detecting method provided by the embodiment of the present application. The execution body of the present application embodiment is a detection device. See image 3 The method includes the following steps:

[0094] 301. Detecting equipment acquires multiple training samples.

[0095] In order to train the status detection model, the detection device first acquires multiple training samples. Wherein, each training sample includes a first number of sample audio features, the first number of sample audio features correspond to a first number of audio frames in the same sample audio data, and the sample audio data is operating normally in the electronic device. Under the case of collected.

[0096] 302. The detection device inputs each training sample to the state detection model to obtain the prediction probability corresponding to each training sample.

[0097] The detecting device inputs each training sample to the state detection model, that is, the first number of sample audio features in each training sample to the state detection model, thereby obtaining a prediction probability corresponding to each training sample. Among them, a training sample is based on a sample audio data, which is used to represent the characteristics of the sample audio data. Therefore, the prediction probability of the training sample corresponds to the normal operation of the electronic device, the first number of the training sample The probability of simultaneous appearance of sample audio features, that is, the probability that the sample audio data corresponding to the training sample is.

[0098] 303. The detection device conducts training of the status detection model based on the prediction probability corresponding to the plurality of training samples.

[0099] Since each training sample corresponds to the probability of the sample audio data appears in the case where the electronic device is operating normally, the probability of the electronic device is operating normally based on the prediction probability of the electronic device. Distribution, therefore, based on the prediction probability corresponding to the number of training samples, there is no supervised training, thereby training the state detection model to enable the state detection model to detect the capacity of the operating state of the electronic device, and can predict it in electronics When a device issues an audio data, whether the electronic device is operating normally.

[0100] The method provided herein provides a plurality of training samples to acquire multiple training samples by collecting sample audio data generated during normal operation during normal operation. No need to mark the sample audio data, nor does it need to collect audio data generated during exception operation, avoiding less accurate model training due to less audio data generated during exception operation, ensuring state detection The accuracy of the model, in turn, when the operating state of any electronic device is detected based on a state detection model, the accuracy of the status detection is also ensured.

[0101] Figure 4 It is a flow chart of another state detection method provided by the embodiment of the present application. The execution body of the present application embodiment is a detection device. See Figure 4 The method includes the following steps:

[0102] 401. Detecting equipment acquires multiple training samples.

[0103] In order to train the status detection model, the detection device first acquires multiple training samples. Wherein, each training sample includes a first number of sample audio features, the first number of sample audio features correspond to a first number of audio frames in the same sample audio data, and the sample audio data is operating normally in the electronic device. Under the case of collected.

[0104] Alternatively, in the case where either electronic device is operating, the audio frame generated by the electronic device is collected in real time; each of the first number of audio frames is acquired, the first number of audio frames constitute a sample audio data, and Extracting the sample audio feature of each audio frame in the sample audio data, the extracted first number of sample audio features constitutes a training sample until a second number of training samples are obtained.

[0105] Among them, the process of collecting audio frames with constructive training samples is performed in real time, and detecting a first number of audio frames per acquirement, the first number of audio frames constitutes a sample audio data, thereby based on the sample audio data. Construct a training sample. The procedure of the above acquisition audio frames and construct training samples will be repeated, resulting in more training samples. And also statistically the number of training samples obtained, stop collection after acquiring the second quantity training sample. Wherein, the first quantity and second quantity are pre-set by the detection device.

[0106] Alternatively, the audio data generated in the case where either electronic device is generated is acquired, and the audio data is divided into a plurality of audio data segments, wherein each audio data segment includes a first number of audio frames, extracts each audio. The sample audio features of each audio frame in the data segment constitute a training sample with a first number of sample audio features extracted in the same audio data segment to obtain multiple training samples.

[0107] Among them, the process of collecting audio data and construction training samples are asynchronously, and the detection device collects audio data of the electronic device, and stores the audio data. After the training status detection model is required, read the stored audio data, segment the audio data, constitute a training sample based on each audio data segment, and multiple training samples can be obtained.

[0108] It is assumed that the number of multiple training samples is the second number, the length of the audio data (i.e., the number included in the audio frame) is not less than the first number and the second number of products, to ensure that the audio data can be constructed from the audio data Two number of training samples. Wherein, the first quantity and second quantity are pre-set by the detection device.

[0109] In the above alternative, only one electronic device is an example, the detecting device acquires a plurality of training samples by collecting audio data during normal operation of this electronic device, and can acquire a plurality of training samples for subsequent state detection models. . In another embodiment, it is also possible to assemble a plurality of the same type of electronic device, collecting audio data during normal operation of each electronic device, and can obtain at least one training sample, and can be obtained by multiple electronic devices. To multiple training samples, the training process for subsequent state detection models. Wherein, the same type refers to the same electronic device, or a series of electronic devices that are the same or similar in the running process, such as electronic devices that are the same type of electronic devices, or, refrigerators and television are home equipment. Belong to the same type.

[0110] Alternatively, considering the possible audio data generated during the operation of different types of electronic devices, if the sample audio data training status detection model generated by some type of electronic device is subsequently based on the status detection model based on the status detection model. The electronic device performs status detection, and the detection accuracy cannot be guaranteed, and the normal operation of the electronic device may be misjected as an abnormal operation, and it is also possible to generate an exceptionally running electronic device. Therefore, the type of electronic device used in the process of training status detection model, the type of electronic device detected by this state detection model should be the same.

[0111] 402. For each training sample, the detection device inputs the first number of sample audio features in the training sample to the status detection model to obtain the conditional probability of each sample audio feature.

[0112] Since each training sample includes a first number of sample audio features, the detection device inputs the first number of sample audio features to the state detection model, thereby obtaining the conditional probability corresponding to each sample audio feature. Wherein, the probability is indicated in the case where the electronic device is operating normally, and in the case where the sample audio feature prior to the sample audio feature is present, the probability of the sample audio feature occurs.

[0113] For example, the detecting device acquires a plurality of training samples, each training sample includes N sample audio features, the detecting device needs to input N sample audio features in a training sample to the state detection model, thereby obtaining n samples The condition probability corresponding to the audio feature, where N is an integer greater than 1.

[0114] Alternatively, the state detection model includes a first number of input layer nodes, at least two hidden layer nodes, and a first number of output layer nodes, where at least two hidden layer nodes can be distributed in a hidden layer, or distributed Multiple hidden layers. In this state detection model, based on the features in the front layer node, the features in the latter layer node can be obtained, and the last layer node, that is, the characteristics of the output layer node can be obtained in this type.

[0115] The state detection model also includes a weight matrix and a mask matrix including the weight parameters between the two nodes of the adjacent two layers, including the state detection model including the state detection model. Mask parameters between the two layers of the two nodes. Among them, the two nodes are located in adjacent two layers, that is, the layer where one of the nodes is the layer of the layer where the other node is located, then one of the nodes can be referred to as the front layer node of another node. Further, the weight parameters and the mask parameters are used to adjust the features of the previous layer nodes in the two nodes to process based on the adjusted feature to obtain the characteristics of the rear layer nodes in the two nodes. .

[0116] Then, after obtaining a training sample, it is necessary to obtain the condition probability corresponding to each sample audio feature in the training sample based on the state detection model, and the first number of sample audio features in the training sample is input to the state detection model, respectively. The input layer node in each input layer node is a sample audio feature. Then follow the positional relationship between the nodes in the state detection model, sequentially.

[0117] Taking the first node and the second node in the state detection model as an example, the first node and the second node are two nodes located adjacent two layers, the first node being the first layer node of the second node, First determine the feature in the first node, such as the first node as an input layer node, the feature in the first node is the sample audio feature input to the first node, or the first audio is not an input layer node, then The first node is characterized by the features obtained based on the feature processing in the front layer node of the first node. And determine the weight parameters and mask parameters between the first node and the second node, determine the features in the first node, the weight parameters between the first node and the second node, and the product of the mask parameter, and the activation function, The determined product is processed, and the resulting feature is determined as features in the second node. For each node in the state detection model, the above-described similar steps are repeated, that is, the features in each output layer node are obtained, that is, the conditional probability of each output layer node output.

[0118]The first point of the description is that if there is at least two first nodes in the previous layer of the second node, when determining the feature in the second node, the characteristics of each first node are determined, the first node The product of the weight parameters and mask parameters between the second node, using the activation function, the determined product is processed, obtain the feature component corresponding to each of the first nodes, the feature component corresponding to at least two first nodes Fusion is carried out, and the fused feature is determined as features in the second node. Among them, the fusion method can be summed, or the average is given, or weighted or weighted.

[0119] The second point in which it is to be explained, assuming that the state detection model includes a M-layer node, m is an integer greater than 1, from the first layer node, the second layer node, until the m-layer node, can be connected to constitute a node path. In this node path, the feature in the previous node affects the features in the later nodes. The purpose of introducing a mask matrix in the state detection model is that by setting the mask parameter 1 or 0, it is possible to ensure that the mask parameter corresponds to the two nodes, the previous node. The features in the middle can maintain the impact of the features in the latter node. In the position where the mask parameter is 0, the mask parameters correspond to the characteristics in the previous node no longer affect the characteristics of the latter node, namely the node between the two nodes. The path is "disconnected".

[0120] In the present application embodiment, the mask matrix of the state detecting model is set to: in the node path of each output layer node and the preamble input layer node, the mask parameters between each two nodes are not 0, In each of the node paths composed of each output layer node and other input layer nodes, there is at least a mask parameter between two nodes of adjacent two layers. The preferences of the output layer node are: the input layer node corresponding to the other output layer nodes located before the output layer node.

[0121] This ensures that the feature in the output layer node is affected by the sample audio features in the pre-sequence input layer node, and is not affected by the sample audio features in other input layer nodes, in line with the "output layer node" In the case where the sample audio feature in the pre-sequence input layer node occurs, the probability of the sample audio feature in the input layer node corresponding to the output layer node. "

[0122] Alternatively, the activation function is non-linear, and the characteristic of the activation function is used to process the characteristics of the first node, the product of the weight parameters between the first node and the second node, and the product of the mask parameter, actually in the first node The characteristics, weight parameters, and mask parameters have been combined, and ensure that the combined result is a non-linear relationship with the features, weight parameters, and mask parameters in the first node. For example, the activation function can be a TANH (dual-dicing) function.

[0123] for example, Figure 5 It is a structural diagram of a state detection model provided by the embodiment of the present application, such as Figure 5 As shown in this state detection model, the three-layer node includes a three-layer node, which includes an input layer node 501, a hidden layer node 502, and an output layer node 503. Wherein, the input layer node 501 corresponds to one of the output layer node 502, the input layer node is the same, for example, the n input layer node in the figure corresponds to the N output layer nodes. The feature in the input layer node 501 is input to the sample audio feature in the input layer node 501, with x 1 , X 2... x n Express. The feature in the output layer node 502 represents the condition probability corresponding to the sample audio feature value of the corresponding input layer node, with P (X 1 ), P (x 2 | x 1 ) ... p (x n | x n-1... x 1 )Express. This state detection model also includes a weight matrix and a mask matrix.

[0124] With the first input layer node x 1 For example, determine the sample audio feature in the node, the weight parameter between the node and the hidden layer node and the mask parameter, using the above-described activation function h (x) = tanh (W 1 · M 1 · X 1 ), The features in the corresponding hidden layer nodes can be obtained. Similarly, the other nodes of the input layer are processed to obtain features in each node of the hidden layer. It should be noted that in the case where the mask parameter is 0, the feature in the hidden layer node is 0, indicating that the feature in the input layer node does not affect the hidden layer node. The processing procedure between the hidden layer node and the output layer node is similar to the above process, and details are not described herein.

[0125] For the first number of sample audio features in any training sample, since the mask matrix is ​​added, the influencing factors of the conditional probability corresponding to each sample audio feature are different. Taking N sample audio features in either training sample, for the first sample audio feature, the output layer node corresponding to the input layer node is 0, and then the output is indicated by the output layer node. The features in the layer node are independent of the characteristics in other nodes, that is, for the calculation of the conditional probability of the first sample audio feature, no reference other sample audio features; for the second sample audio feature, the conditional probability corresponds to The mask parameter between the input layer nodes corresponding to the first sample audio feature is not 0, ie the mask parameter between the pre-sequence input layer node is not 0, in addition to the other input layer node. The mask parameters are all 0, and the characteristics in the output layer node are only related to the first sample audio feature, and the other sample audio features are independent, that is, calculations for the critical probability of the second sample audio feature. It is only necessary to refer to the first sample audio feature, indicating that in the case where the first sample audio feature is present, the probability of the second sample audio feature; the same, for the Nth sample audio feature, its conditional probability The corresponding output layer node is not 0 between the mask parameters between the pre-sequence input layer nodes, and the features in the output layer node are related to the features in the forward input layer node, that is, for the nth sample audio feature. The calculation of the condition probability needs to consider the effects of other sample audio features except the sample audio feature, indicating the nth sample audio feature when the first sample audio feature to the nth-1 audio feature The probability of appearing.

[0126] 403. The detection device determines the prediction probability of the training sample corresponding to the conditional probability of each sample audio feature in the training sample.

[0127] Since each training sample includes a first number of sample audio features, each sample audio feature has a corresponding conditional probability, and for any training sample, the product probability corresponding to the first number of sample audio features Determine the prediction probability corresponding to the training sample. Since the obtained sample audio data is generated in the case of normal operation, the training sample is positive, and the prediction probability corresponding to the training sample represents the training sample in the case where the electronic device is operating normally. The first number of sample audio features simultaneously appeared, that is, the probability indicating the appearance of the sample audio data.

[0128] It should be noted that since the detection device acquires a plurality of training samples, the above processes 402-403 are required to obtain the above processes 402-403, respectively, thereby obtaining the prediction probability corresponding to each training sample in multiple training samples.

[0129] 404. The detection device uses a loss function to process the prediction probability of a plurality of training samples to obtain a loss value.

[0130] Alternatively, the loss function is a negative log, or other functions.

[0131] Since the plurality of training samples correspond to the probability of the sample audio data, the probability distribution of the sample audio data that occurs in the case where the electronic device is operating, and therefore the loss function can be processed to process the prediction probability corresponding to the plurality of training samples. Loss value.

[0132] 405. The detection device is based on the loss value, and the status detection model is trained.

[0133] Among them, when training the status detection model, the detection device is updated based on the loss value, which is updated to each weight parameter in the weight matrix to make the update weight parameters. The post-state detection model is more accurate.

[0134] It should be noted that steps 403-405 described above only illustrates the one-time training process, and during the training status detection model, there is a need to carry out multiple training in the state detection model. In one possible implementation, the detection device stops training for the status detection model in response to the number of training times; or, the loss value obtained in response to the number of current training times is not greater than the second threshold, stopping the status detection Training of the model. Wherein, both the first threshold and the second threshold are arbitrary, for example, the first threshold is 10 or 15, etc., the second threshold is 0.4 or 0.3, and the like.

[0135] The method provided herein provides a plurality of training samples to acquire multiple training samples by collecting sample audio data generated during normal operation during normal operation. No need to mark the sample audio data, nor does it need to collect audio data generated during exception operation, avoiding less accurate model training due to less audio data generated during exception operation, ensuring state detection The accuracy of the model, in turn, when the operating state of any electronic device is detected based on a state detection model, the accuracy of the status detection is also ensured.

[0136] Further, the product of the condition probability corresponding to each sample audio feature in the training sample is used as the prediction probability corresponding to the training sample, so that the prediction probability corresponding to the training sample can represent the first number of sample audio features in the training sample. The probability of appearing to ensure that the probability is more in line with the actual probability distribution in normal operation of electronic equipment, and improves accuracy.

[0137] Further, by setting the weight matrix and the mask matrix in the state detection model, the feature based on the weight parameter and the mask parameter are adjusted to the two nodes in order to process the adjusted feature. , The characteristics of the rear layer node in these two nodes are obtained, ensuring the effects of the characteristics in the output layer node are affected by the sample audio features in the pre-sequence input layer node, and are not subject to sample audio features in other input layer nodes. The impact, in line with the actual conditional probability of each audio frame in the audio frames generated in the audio data generated during normal operation, and the objective situation is more consistent, and the accuracy is improved.

[0138] Further, the acquisition process of the audio frame generated during normal operation of the electronic device and the construction process of the training sample can be real-time, saving the time spent on the acquisition of multiple training samples, thereby improving the training status detection model. effectiveness.

[0139] Further, the acquisition process of the sample audio data generated during normal operation of the electronic device and the construction process of the training sample can also be asynchronously, first collect and store the sample audio data generated during normal operation of the electronic device. The sample audio data is then processed, thereby obtaining a plurality of training samples, which ensures that the acquisition process of sample audio data and the construction of the training sample does not affect each other, and improves the reliability of training samples.

[0140] Based on the above embodiment, after the state detection model is trained, the operating state of either electronic device can be detected based on the state detection model. The specific process is detailed with the following examples.

[0141] Image 6 It is a flow chart of still another state detection method provided by the embodiment of the present application. The execution body of the present application embodiment is a detection device. See Image 6 The method includes the following steps:

[0142]601. Detecting the device acquired the target audio data generated during the operation of the target electronic device, which includes a first number of audio frames.

[0143] Alternatively, the audio frame generated during the operation of the target electronic device is acquired, and each of the first number of audio frames constitutes a sample audio data for each of the first number of audio frames. This execution process is similar to the execution process in step 401, and details are not described herein again.

[0144] Alternatively, the audio data generated during the operation of the target electronic device is obtained, and the audio data is divided into a plurality of audio data segments, each of which includes a first number of audio frames. This execution process is similar to the execution process in step 401, and details are not described herein again.

[0145] Alternatively, the target electronic device and the electronic device at the time of collecting the training sample may be the same device, or may be the same type of electronic device.

[0146] 602. The detection device extracts the audio feature of each audio frame in the target audio data, and the extracted first number of audio features are input to the status detection model after the training to obtain the prediction probability corresponding to the target audio data.

[0147] The detection device extracts the acquired target audio data, that is, the characteristic extraction of each audio frame in the target audio data is extracted to obtain a first number of audio features. The first number of audio features are input to the status detection model after the training to obtain the conditional probability of each audio feature, thereby obtaining the prediction probability corresponding to the target audio data.

[0148] 603. Detecting equipment determines the score of the target electronic device based on the prediction probability corresponding to the target audio data, which is negatively related to the probability.

[0149] Alternatively, the detection device is based on the predicted probability corresponding to the target audio data, and the prediction probability corresponding to the target audio data is used to determine the score of the target electronic device.

[0150] Among them, the number of loads used is monotonous decreasing function, and the score of the target electronic device is negatively correlated with the prediction probability of the target audio data.

[0151] 604. In the case where the detection device is greater than the target score, it is determined that the target electronic device is in an abnormal operating state.

[0152] The target score is used to indicate the probability of generating the sample audio data during normal operation. In the case where the above score is greater than the target score, the predicted probability corresponding to the target audio data is smaller than the probability of generating the sample audio data during normal operation, and the target audio data is determined as an electronic device in an abnormal operation. The audio data generated during the process is that it is determined that the target electronic device is in an abnormal operating state.

[0153] In another embodiment, in the case where the above score is not greater than the target score, the predicted probability corresponding to the target audio data is not smaller than the probability of generating the sample audio data during normal operation, and the target will be The audio data is determined as audio data generated during normal operation, that is, it is determined that the target electronic device is in normal operation.

[0154] The method provided by the embodiment of the present application is based on the training state detection model, and the operational state of the target electronic device is an abnormal operation state by collecting the audio data generated during the operation of the target electronic device. Further, since the status detection model after training is more accurate, the accuracy of the status detection is guaranteed.

[0155] The embodiment of this application also provides a state detection method, and provides a state detection method. Figure 7 It is a flow chart of still another state detection method provided by the embodiment of the present application. Such as Figure 7 As shown, the state detection model according to the present application embodiment is the self-encoder model. Further, in order to realize the node path in the state detection model, the mask matrix is ​​introduced into the state detection model, thereby constructing the self-wodulator model of the plus mask.

[0156] First, sample audio data generated in the case of normal operation of high-voltage variable power station equipment is acquired, and audio feature extraction of each audio frame in sample audio data; the first number of audio frames in the acquired audio data will then be collected The corresponding sample audio feature is determined as a training sample.

[0157] Thereafter, the acquired plurality of training samples are input to the self-encoder model, give each training sample corresponding to the prediction probability, resulting in a plurality of training samples corresponding to the prediction probability distribution; based on a plurality of training samples corresponding to prediction The probability, using an unwailed training method, training the self-encoder model; the above steps are performed multiple times until the training is more accurate self-encoder model; finally, based on the training self-proder model, high pressure to detect detection The audio data generated during the operation of the substation device is collected, and the acquired target audio data is processed. The target audio data is scored by the target audio data; when the score exceeds the threshold, it is determined that high voltage substation equipment In an abnormal operating state.

[0158] Figure 8 It is a schematic structural diagram of a state detecting apparatus according to the embodiment of the present application. See Figure 8 The device includes:

[0159] The module 801 is acquired to acquire a plurality of training samples, each training sample includes a first number of sample audio features, and the first number of sample audio features correspond to a first number of audio frames in the same sample audio data, and the sample Audio data is collected in the normal operation of electronic equipment;

[0160] The first detecting module 802 is configured to input each training sample to the state detection model, and obtain the probability of the prediction probability corresponding to each training sample, the probability indicates the probability of sample audio data in the case of normal operation of the electronic device;

[0161] The training module 803 is used to train the status detection model based on the prediction probability corresponding to the plurality of training samples, and the state detection model after training is used to detect the operating state of any electronic device.

[0162] The apparatus provided herein provides a plurality of training samples to acquire multiple training samples by collecting sample audio data generated in the process of normal operation during normal operation. No need to mark the sample audio data, nor does it need to collect audio data generated during exception operation, avoiding less accurate model training due to less audio data generated during exception operation, ensuring state detection The accuracy of the model, in turn, when the operating state of any electronic device is detected based on a state detection model, the accuracy of the status detection is also ensured.

[0163] In one possible implementation, the first detection module 802 includes:

[0164] Detection unit, for each training sample, input the first number of sample audio features in the training sample to the status detection model to obtain the conditional probability of each sample audio feature, and the conditional probability is indicated in the normal operation of the electronic device, and training In the case where the sample audio feature appeared in the sample audio feature, the probability of sample audio feature occurs;

[0165] The first determination unit is used to determine the product probability of the condition probability of each sample audio feature in the training sample, determined as the prediction probability corresponding to the training sample.

[0166] In one possible implementation, the status detection model includes a first number of input layer nodes, at least two hidden layer nodes, and the first number output layer node, the status detection model also includes a weight matrix and a mask matrix, and the weight matrix includes In the state detection model, the weight parameters located between adjacent two layers, the mask matrix includes a mask parameter between the two nodes of the adjacent two layers in the state detection model; wherein at each In the node path composed of the output layer node and the front input layer node, the mask parameters between each two nodes are not 0; at least in each output layer node and other input layer nodes, at least The mask parameter between the two nodes of the adjacent two layers is 0; the preamble input layer node of the output layer node is: the input layer node corresponding to the other output layer nodes before the output layer node.

[0167] In a possible implementation, the detection unit, including:

[0168] The input subunit is used to input the first number of sample audio features in the training sample to the input layer node in the state detection model, respectively;

[0169] The subunit is determined for the first node and the second node in the state detection model, determines the characteristics in the first node, the weight parameters of the first node and the second node, and the product of the mask parameter, using the activation function For the determined product, the obtained feature is determined as the feature in the second node, wherein the first node and the second node are two nodes located adjacent two layers until each output layer node is output. Conditional Probability.

[0170] In one possible implementation, the training module 803, including:

[0171] The processing unit is used to use the loss function to process the prediction probability corresponding to the plurality of training samples to obtain a loss value;

[0172] Training units are used to train status detection models based on loss values.

[0173] In one possible implementation, the acquisition module 801, including:

[0174] The acquisition unit is used to collect the audio frame generated by the electronic device when the electronic device is operating normally;

[0175] The first extracting unit is used for each of the first number of audio frames, and the first number of audio frames constitutes a sample audio data, and extracts the sample audio characteristics of each audio frame in the sample audio data. The number of sample audio features constitute a training sample until a second number of training samples are obtained.

[0176] In one possible implementation, the acquisition module 801, including:

[0177] The acquisition unit is used to obtain audio data generated in the case of normal operation;

[0178] The division unit is used to divide the audio data into a plurality of audio data segments, where each audio data segment includes a first number of audio frames;

[0179] The second extracting unit is used to extract the sample audio feature of each audio frame in each audio data segment, and the first number of sample audio features extracted in the same audio data segment constitute a training sample to obtain multiple training samples.

[0180] In one possible implementation, the device also includes:

[0181] The acquisition module is used to acquire the target audio data generated during the operation of the target electronic device, and the target audio data includes a first number of audio frames;

[0182] The second detection module is used to extract audio features of each audio frame in the target audio data, and the extracted first number of audio features are input to the status detection model after training to obtain the prediction probability corresponding to the target audio data;

[0183] The module is determined to determine the operational state of the target electronic device based on the prediction probability corresponding to the target audio data.

[0184] In one possible implementation, the module, including:

[0185]The second determination unit is used to determine the score of the target electronic device, the score and probability of the target electronic device is negatively correlated;

[0186] The third determination unit is used to determine the target electronic device in an abnormal operating state in the case where the score is greater than the target score.

[0187] All of the above alternative techniques can be used to form an alternative embodiment of the present application, which is not repeated herein.

[0188] It should be noted that when the state detecting device provided by the above embodiment is detected in the state detection of any device, only the partition of each functional module is described. In the actual application, the above function allocation can be different from different. The function module is complete, so that the internal structure of the detection device is divided into different functional modules to complete all or part of the above described above. Further, the state detecting means and state detecting method according to the above embodiment belongs to the same concept, and the process is described in the method embodiment, and details are not described herein again.

[0189] The present application embodiment also provides a detection device including a processor and a memory, and at least one program code is stored in the memory. The at least one program code is loaded and executed by the processor to achieve the state detection of the above embodiment. The operation performed in the method.

[0190] Optionally, the detection device is provided as a terminal. Figure 9 It is a structural diagram of a terminal 900 provided by the embodiment of the present application. The terminal 900 can be a portable mobile terminal, such as a smartphone, a tablet, MP3 player (MovingPicture Experts Group Audio Layer III, Dynamic Image Expert Compressed Standard Audio Level 3), MP4 (Moving Picture Experts Group Audio Layer IV, Dynamic Image Expert Compression Standard Audio League 4) Player, laptop or desktop computer. The terminal 900 may also be referred to as other names such as user equipment, portable terminals, laptop terminals, and desktop terminals.

[0191] The terminal 900 includes a processor 901 and a memory 902.

[0192] Processor 901 can include one or more processing cores, such as 4 core processors, 8 core processors, and the like. The processor 901 can use the DSP (Digital Signal Processing, Digital Signal Processing), FPGA (Field-Programmable Gate Array, Field Programmable Gate Array), PLA (Programmable Logic Array, Programmable Logic Array) in at least one of hardware forms. accomplish. Processor 901 may also include a main processor and a coprocessor, a processor for processing data in a wake-up state, also known as the CPU (Central ProcessingUnit, central processor); coprocessor is used A low power processor processed on the data in the standby state. In some embodiments, processor 901 can integrate GPU (Graphics Processing Unit, Image Processor), and GPUs are used to be rendered and drawn by the contents of the display you want to display. In some embodiments, processor 901 may also include AI (Artificial Intelligence, artificial intelligence) processor, which is used to process computing operations related to machine learning.

[0193] Memory 902 can include one or more computer readable storage media, which can be readable in non-transient. Memory 902 can also include high-speed random access memory, and non-volatile memory, such as one or more disk storage devices, flash storage devices. In some embodiments, the non-transitory computer readable storage medium in the memory 902 is used to store at least one program code, which is used to implement the processor 901 to implement the embodiments of the present application. State detection method.

[0194] In some embodiments, the terminal 900 is also optionally: peripheral device interface 903 and at least one peripheral device. Processor 901, memory 902, and peripheral interface 903 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 903 via a bus, a signal line or a circuit board. Specifically, the peripherals include: radio frequency circuit 904, at least one of the audio circuit 905, and power source 906.

[0195] Peripheral interface 903 can be used to connect at least one peripheral device associated with I / O (input / OUTPUT, input / output) to processor 901 and memory 902. In some embodiments, processor 901, memory 902, and peripheral interface 903 are integrated on the same chip or circuit board; in some other embodiments, the processor 901, memory 902, and peripheral device interface 903 or Two can be implemented on a separate chip or circuit board, which is not limited thereto.

[0196] RF circuit 904 is used to receive and transmit RF (Radio Frequency, RF) signal, also known as electromagnetic signals. RF circuit 904 communicates with the communication network and other communication devices by electromagnetic signals. The radio frequency circuit 904 converts the electrical signal into an electromagnetic signal to transmit, or convert the received electromagnetic signal into an electrical signal. Alternatively, RF circuit 904 includes: antenna system, RF transceiver, one or more amplifiers, tuner, oscillator, digital signal processor, codec chip group, user identity module card, etc. Radio frequency circuit 904 can communicate with other terminals by at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to, web, metropolitan networks, intranets, all-generation mobile communication networks (2G, 3G, 4G, and 5G), wireless LAN and / or WiFi (WiReless Fidelity, Wireless Fire Fire "network. In some embodiments, the radio frequency circuit 904 may also include a circuit related to NFC (Near Field Communication, near-distance wireless communication), which is not limited thereto.

[0197] The audio circuit 905 can include a microphone and a speaker. The microphone is used to collect acoustic waves of the user and the environment and convert the acoustic wave into the electrical signal input to processor 901 for processing, or input to the radio frequency circuit 904 to achieve voice communication. For the purpose of stereo collection or noise reduction, the microphone can be plural, respectively, which are disposed at different parts of the terminal 900. The microphone can also be an array microphone or a full-to-pending microphone. The speaker is used to convert the electrical signal from the processor 901 or the radio frequency circuit 904 into sound waves. The speaker can be a conventional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, the electrical signal can be converted to the sound waves that the human audible sound can also be converted into the sound wave of human being invisible to ranging. In some embodiments, the audio circuit 905 can also include a headphone jack.

[0198] The power source 906 is used to power the various components in the terminal 900. Power supply 906 can be an alternating current, direct current, disposable battery, or rechargeable battery. When the power source 906 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. The wired charging battery is a battery that is charged by a wired line, and a wireless rechargeable battery is charged through a wireless coil. This rechargeable battery can also be used to support fast charge technology.

[0199] Those skilled in the art will appreciate that Figure 9 The structure shown is not constituied to the terminal 900, which may include more or fewer components, or in combination, or different component arrangements.

[0200] Alternatively, the detection device is provided as a server. Figure 10 It is a structural diagram of a server provided by the embodiment of the present application. The server 1000 can produce a relatively large difference in configuration or performance, and may include one or more processors (CPUs) 1001 and one or one. The above memory 1002, wherein at least one program code is stored in the memory 1002, which is loaded and executed by the processor 1001 to achieve the method provided by each method embodiment. Of course, the server can also have a wired or wireless network interface, a keyboard, and a component such as an input / output interface for input and output, which may also include other components for implementing device functions, not described here.

[0201] In an exemplary embodiment, a computer readable storage medium is also provided, and at least one program code is stored in the computer readable storage medium, which is loaded and executed by the processor in the server to achieve the above. The state detecting method in the examples. The computer readable storage medium can be a memory. For example, the computer readable storage medium can be a ROM (Read-Only Memory, read-only memory), RAM (Random Access Memory, Random Access Memory), CD-ROM (Compact Disc Read-Only Memory, compact disc read only) Storage device, tape, floppy disk, and optical data storage device, and the like.

[0202] In an exemplary embodiment, a computer program product is also provided, the computer program product comprising at least one program code, and at least one program code can be performed by a processor of the detection device to achieve the state detection method shown in the above embodiment. .

[0203] One of ordinary skill in the art will appreciate that all or some of the steps that implement the above embodiment can be accomplished by hardware, or the hardware can be done by a program to perform, which can be stored in a computer readable storage medium. The storage medium can be read-only memory, disk, or disc.

[0204] The above is only an alternative embodiment of the present application embodiment, and is not intended to limit the embodiments of the present application, and any modification, equivalent replacement, improvement, etc. according to the spirit and principles of the present application embodiments should be included. Within the protection of this application.