Root cause location model training method, root cause location method, device and equipment

Abstract

The present disclosure provides a root cause localization model training method, a root cause localization method, an apparatus and a device, wherein the method comprises: obtaining a multimodal training sample; wherein the training sample comprises at least one of the following: network element indicator data, a network element label and a network element analysis text; extracting features of the network element label and the network element indicator data through a feature extraction network in the root cause localization model, and obtaining classification label features and network element features in sequence; generating a prediction analysis text of the training sample based on the network element features and the network element analysis text through a text generation network in the root cause localization model; training the root cause localization model based on the prediction analysis text, the classification label features and the network element features, and obtaining a target root cause localization model after training.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a root cause localization model training method, a root cause localization method, a device, and an apparatus. Background Art

[0002] Root cause location is a very necessary function in the field of intelligent operation and maintenance. This function aims to use algorithms to increase the explainability of service level agreements (SLAs) to facilitate operation and maintenance personnel to understand the root causes of problems in a timely manner when anomalies or failures occur, thereby reducing the impact of the failure.

[0003] In the related art, data can be classified by comparative learning in root cause location, so as to obtain the root cause type of the network element. In order to further analyze the root cause type, the indicator anomaly detection model can be called to perform anomaly detection on various indicators of the network element, so as to locate the degradation indicator when the anomaly occurs. By locating the degradation indicator, the abnormal indicators in various indicators of the network element can be analyzed, but not all abnormal indicators are associated with the root cause type. Therefore, the related art cannot give the classification reason of the root cause type for the abnormal indicator. Summary of the invention

[0004] The embodiments of the present disclosure at least provide a root cause localization model training method, a root cause localization method, an apparatus, and a device.

[0005] In a first aspect, an embodiment of the present disclosure provides a root cause localization model training method, comprising:

[0006] Acquire multimodal training samples; wherein the training samples include at least one of the following: network element indicator data, network element labels, and network element analysis texts;

[0007] Extracting features of the network element label and the network element indicator data to obtain classification label features and network element features;

[0008] Generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text;

[0009] Based on the prediction analysis text, the classification label features and the network element features, a root cause location model is trained to obtain a target root cause location model.

[0010] In an optional implementation manner, the generating the prediction analysis text of the training sample based on the network element features and the network element analysis text by the text generation network in the root cause localization model includes:

[0011] Processing the network element feature and the network element analysis text through an attention module to obtain a first processing result;

[0012] The first processing result is processed by a first feedforward neural network to obtain the predictive analysis text.

[0013] In an optional embodiment, each of the attention modules includes: a causal self-attention model, a cross attention module, and a second feedforward neural network connected in sequence;

[0014] The processing of the network element feature and the network element analysis text by the attention module to obtain a first processing result includes:

[0015] Processing the network element analysis text through the causal self-attention model to obtain text features;

[0016] Processing the text feature and the network element feature by the cross attention module to obtain a second processing result;

[0017] The second processing result is processed by the second feedforward neural network, and the first processing result is determined according to the output result.

[0018] In an optional implementation manner, the processing of the text feature and the network element feature by the cross attention module to obtain a second processing result includes:

[0019] Performing linear mapping processing on the text features through the cross attention module to obtain a query matrix;

[0020] Performing linear mapping processing on the network element features through the cross attention module to obtain a key value matrix and a value matrix;

[0021] The second processing result is calculated by using the query matrix, the key value matrix and the value matrix.

[0022] In an optional implementation, there are multiple attention modules, the input of each attention module is the output result of the previous attention module, the input of the cross-attention module in each attention module is the network element feature, and the output result of the last attention module is the input of the first feedforward neural network.

[0023] In an optional implementation manner, extracting the features of the network element label and the network element indicator data through a feature extraction network in the root cause location model to sequentially obtain classification label features and network element features includes:

[0024] Extracting features of the network element labels by a single-mode encoder to obtain the classification label features;

[0025] The network element indicator data is subjected to feature extraction through a network pre-training large model to obtain the network element features.

[0026] In an optional implementation, the root cause location model is trained based on the prediction analysis text, the classification label features, and the network element features, and a target root cause location model is obtained after the training, including:

[0027] Processing the classification label features and the network element features respectively through a third feedforward neural network to obtain a third processing result, and calculating a first loss function value according to the third processing result;

[0028] Calculating a second loss function value according to the network element analysis text and the prediction analysis text;

[0029] A weighted sum is performed on the first loss function value and the second loss function value to obtain a target loss function value, and a root cause localization model is trained according to the target loss function value to obtain the target root cause localization model.

[0030] In an optional implementation, performing weighted summation on the first loss function value and the second loss function value to obtain a target loss function value includes:

[0031] Determine a target training round of the root cause localization model at a current moment;

[0032] Based on the target training round and the weight update rate, determine the weights of the first loss function value and the second loss function value at the current moment to obtain a first weight and a second weight;

[0033] Based on the first weight and the second weight, the first loss function value and the second loss function value are weightedly summed to obtain a target loss function value.

[0034] In an optional implementation manner, training the root cause localization model according to the target loss function value to obtain the target root cause localization model includes:

[0035] The network parameters of the attention module, the first feedforward neural network and the third feedforward neural network in the root cause localization model are adjusted according to the target loss function value, and the target root cause localization model is determined according to the root cause localization model after adjusting the network parameters; wherein the attention module and the first feedforward neural network are networks that generate predictive analysis texts of the training samples based on the network element features and the network element analysis texts.

[0036] In a second aspect, an embodiment of the present disclosure provides a root cause location method, including:

[0037] Obtaining network element indicator data to be predicted;

[0038] Extracting characteristics of the network element indicator data to obtain target network element characteristics;

[0039] The classification result corresponding to the network element indicator data is determined according to the target network element characteristics, and the target network element characteristics are processed through a target root cause positioning model to obtain an analysis text of the classification result.

[0040] In an optional implementation manner, in an optional implementation manner, the processing of the target network element features by the target root cause location model to obtain the analysis text of the classification result includes:

[0041] Processing the target network element feature through the attention module in the target root cause positioning model to obtain a first target processing result;

[0042] The first target processing result is linearly mapped by the first feedforward neural network in the target root cause location model to obtain an analysis text of the classification result.

[0043] In an optional embodiment, the attention module includes: a causal self-attention model, a cross attention module, and a second feedforward neural network connected in sequence;

[0044] The processing of the target network element feature by the attention module in the target root cause positioning model to obtain a first target processing result includes:

[0045] Processing the target input data through the causal self-attention model to obtain target text features;

[0046] Processing the target text feature and the target network element feature through the cross attention module to obtain a second target processing result;

[0047] The second target processing result is processed by the second feedforward neural network, and the first target processing result is determined according to the output result.

[0048] In an optional embodiment, the target input data of the causal self-attention model in the first attention module is the initial text vector, the target input data of the causal self-attention model in the non-first attention module is the text vector output by the previous attention module, and the output result of the second feedforward neural network in the last attention module is the first target processing result.

[0049] In an optional implementation manner, extracting the characteristics of the network element indicator data to obtain the target network element characteristics includes:

[0050] The network element indicator data is subjected to feature extraction through a network element pre-training large model to obtain the target network element features.

[0051] In a third aspect, an embodiment of the present disclosure provides a root cause location model training device, including:

[0052] A first acquisition unit is used to acquire a multimodal training sample; wherein the training sample includes at least one of the following: network element indicator data, network element label and network element analysis text;

[0053] A first feature extraction unit, used to extract the features of the network element label and the network element indicator data, and obtain the classification label feature and the network element feature in sequence;

[0054] A first text generation unit, configured to generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text;

[0055] A training unit is used to train the root cause location model based on the prediction analysis text, the classification label features and the network element features to obtain a target root cause location model.

[0056] In a fourth aspect, an embodiment of the present disclosure provides a root cause locating device, including:

[0057] A second acquisition unit, used to acquire network element indicator data to be predicted;

[0058] A second feature extraction unit is used to extract the features of the network element indicator data to obtain target network element features;

[0059] The second text generation unit is used to determine the classification result corresponding to the network element indicator data according to the target network element characteristics, and process the target network element characteristics through the target root cause positioning model to obtain the analysis text of the classification result.

[0060] In a fifth aspect, an embodiment of the present disclosure further provides an electronic device, comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps in any possible implementation of the first aspect or the second aspect are performed.

[0061] In a sixth aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any possible implementation of the first aspect or the second aspect are executed.

[0062] In the seventh aspect, the embodiments of the present disclosure further provide a computer program product, which is stored in a storage medium and is executed by at least one processor to implement the steps in any possible implementation of the first aspect or the second aspect mentioned above.

[0063] In an embodiment of the present disclosure, first, a multimodal training sample is obtained; wherein the training sample includes at least one of the following: network element indicator data, network element labels, and network element analysis texts; then, features of the network element labels and the network element indicator data are extracted to obtain classification label features and network element features; next, a predictive analysis text of the training sample is generated based on the network element features and the network element analysis text; finally, a root cause location model is trained based on the predictive analysis text, the classification label features, and the network element features to obtain a target root cause location model.

[0064] In the above implementation mode, by guiding the root cause location model to generate network element analysis text based on the network element characteristics of the network element indicator data, a detailed analysis description of the classification result can be obtained while achieving the classification result through comparative learning, and the correlation between the abnormal indicator and the classification result can be established, thereby improving the interpretability and trustworthiness of the classification decision of the root cause location model.

[0065] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0067] Figure 1 A flow chart of a root cause location model training method provided by an embodiment of the present disclosure is shown;

[0068] Figure 2 A schematic diagram of the model structure of a root cause location model provided by an embodiment of the present disclosure is shown;

[0069] Figure 3 A flow chart of a root cause location method provided by an embodiment of the present disclosure is shown;

[0070] Figure 4A result schematic diagram showing an output result of a root cause location model provided by an embodiment of the present disclosure;

[0071] Figure 5 A schematic diagram of a root cause location model training device provided by an embodiment of the present disclosure is shown;

[0072] Figure 6 A schematic diagram of a root cause locating device provided by an embodiment of the present disclosure is shown;

[0073] Figure 7 A schematic diagram of an electronic device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0075] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0076] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0077] Root cause location is a very necessary function in the field of intelligent operation or maintenance. This function aims to use algorithms to increase the explainability of SLA (service level agreement) to facilitate operation and maintenance personnel to understand the root cause of the problem in a timely manner when anomalies or failures occur, thereby reducing the impact scope of the failure.

[0078] In the related art, data can be classified by comparative learning in root cause location, so as to obtain the root cause type of the network element. In order to further analyze the root cause type, the indicator anomaly detection model can be called to perform anomaly detection on various indicators of the network element, so as to locate the degradation indicator when the anomaly occurs. By locating the degradation indicator, the abnormal indicators in various indicators of the network element can be analyzed, but not all abnormal indicators are associated with the root cause type. Therefore, the related art cannot give the classification reason of the root cause type for the abnormal indicator.

[0079] Based on the above research, the present disclosure provides a root cause location model training method, root cause location method, device and equipment. In the embodiment of the present disclosure, first, a multimodal training sample is obtained; wherein the training sample includes at least one of the following: network element indicator data, network element label and network element analysis text; then, the features of the network element label and the network element indicator data are extracted to obtain the classification label features and the network element features in turn; next, a prediction analysis text of the training sample is generated based on the network element features and the network element analysis text; finally, based on the prediction analysis text, the classification label features and the network element features, the root cause location model is trained to obtain the target root cause location model.

[0080] In the above implementation, by using the network element characteristics of the network element indicator data as a condition to guide the text generation network in the root cause location model to generate network element analysis text, it is possible to obtain a detailed analysis description of the classification result while achieving the classification result through comparative learning, establish the correlation between the abnormal indicator and the classification result, and improve the interpretability and trustworthiness of the classification decision of the root cause location model.

[0081] To facilitate understanding of this embodiment, a root cause location model training method disclosed in the embodiment of the present disclosure is first introduced in detail. The execution subject of the root cause location model training method provided in the embodiment of the present disclosure is generally an electronic device with certain computing capabilities. In some possible implementations, the root cause location model training method can be implemented by a processor calling a computer-readable instruction stored in a memory.

[0082] See also Figure 1 FIG. 1 is a flowchart of a root cause localization model training method provided by an embodiment of the present disclosure, wherein the method includes steps S101 to S104, wherein:

[0083] S101: Acquire multimodal training samples; wherein the training samples include at least one of the following: network element indicator data, network element labels, and network element analysis texts.

[0084] In the disclosed embodiment, each training sample can be described as follows: network element indicator-network element label-network element analysis text. The multimodal training sample contains structured data and unstructured data, wherein the structured data is the network element indicator data, and the unstructured data includes the network element label and the network element analysis text. For example, the network element indicator data includes but is not limited to the following data: network element performance indicators, MR performance indicators, industrial parameters and other time series type data; wherein the time granularity of the network element indicator data can be N minutes (or seconds), and the data dimension can be M dimensions, for example, the time granularity is 15 minutes, and the data dimension is 159 dimensions. Here, the specific values ​​of the time granularity and the data dimension are not specifically limited, and are subject to implementation.

[0085] In the disclosed embodiment, the network element label is a true value label pre-set for the network element indicator data, through which the network element label can be used to determine the network element fault corresponding to the network element indicator data, that is, the classification result. For example, the network element label can be a classification result of "coverage", "interference", or "neighboring area". Here, the user can set the network element label of the training sample according to actual needs. The network element analysis text is an analysis text pre-set based on the network element indicator data and the network element label, through which the cause of the classification result can be explained. Here, the network element analysis text can be set to a sentence of no more than P words, where P can be 128 or other achievable values. For example, the network element analysis text can be: "The proportion of cell RSRP higher than -110 dbm is 85.89%, the proportion of uplink RSRP less than -130 is greater than 40% for multiple time periods, and the cell is over-covered."

[0086] S102: Extracting the network to extract the features of the network element label and the network element indicator data to obtain classification label features and network element features.

[0087] In the embodiment of the present disclosure, the features of the network element label and the network element indicator data can be extracted through the feature extraction network in the root cause location model to obtain the classification label feature and the network element feature respectively. In addition, the classification label feature and the network element feature can also be extracted by other feature extraction methods, and the present disclosure does not make specific restrictions on this, which is subject to what can be implemented.

[0088] Here, the root cause localization model includes a feature extraction network and a text generation network; wherein the feature extraction network includes a unimodal encoder, a network pre-trained large model and two feedforward neural networks (i.e., the third feedforward neural network in the following implementation); and the text generation network includes a multimodal encoder and a first feedforward neural network. wherein the unimodal encoder and the network pre-trained large model constitute a classification module based on contrastive learning.

[0089] Here, the input end of one of the two feedforward neural networks is connected to the output end of the unimodal encoder, and the input end of the other of the two feedforward neural networks is connected to the output end of the network pre-training large model. The output end of the network pre-training large model is connected to the text generation network.

[0090] After the training samples are obtained, the network element labels and network element indicator data in the training samples can be input into the feature extraction network for feature extraction processing, thereby obtaining classification label features and network element features in turn.

[0091] S103: Generate a prediction analysis text of the training sample based on the network element features and the network element analysis text through a text generation network in a root cause localization model.

[0092] In the disclosed embodiment, the network element features and the network element analysis text can be input into the multimodal encoder, so that the multimodal encoder generates the prediction analysis text based on the network element features. Through this processing method, a detailed analysis description of the classification result can be obtained, thereby improving the interpretability and trustworthiness of the classification decision.

[0093] Here, the network element analysis text can be understood as the real network element analysis text, and the prediction analysis text is the analysis text generated by the model. For example, the real detailed description can be: query the historical work order, there is no such problem work order in the past 30 days, observe the indicators in the past 24 hours, the current wireless drop rate is 99.00%, the indicator is restored, it is an occasional indicator fluctuation, and it is directly evaluated. The prediction detailed description can be: query the historical work order, there is no such problem work order in the past 7 days, observe the indicators in the past 24 hours, the current switching success rate is 99.12%, the indicator is restored, it is an occasional indicator fluctuation, and it is directly evaluated.

[0094] S104: Based on the prediction analysis text, the classification label features and the network element features, the root cause localization model is trained to obtain a target root cause localization model after the training.

[0095] After generating the prediction analysis text of the training sample, the target loss function value can be calculated based on the prediction analysis text, the classification label features and the network element features, and then the network parameters of the network to be adjusted in the root cause location model can be adjusted according to the target loss function value. Here, the above steps can be repeatedly performed until a root cause location model that meets the training accuracy requirements is obtained, and the root cause location model that meets the training accuracy requirements is determined as the target root cause location model.

[0096] In the above implementation, by using the network element characteristics of the network element indicator data as a condition to guide the text generation network in the root cause location model to generate network element analysis text, it is possible to obtain a detailed analysis description of the classification result while achieving the classification result through comparative learning, establish the correlation between the abnormal indicator and the classification result, and improve the interpretability and trustworthiness of the classification decision of the root cause location model.

[0097] In an optional implementation, the step S102 extracts the network to extract the features of the network element label and the network element indicator data to obtain the classification label features and the network element features, which specifically includes the following steps:

[0098] Step S11, extracting features of the network element label by a single-mode encoder to obtain the classification label features;

[0099] Step S12, extracting features from the network element indicator data through a network pre-trained large model to obtain the network element features.

[0100] In the disclosed embodiment, the input data of the network pre-trained large model is the network element indicator, and the dimension of the input data of the network pre-trained large model is [batch, s, k], wherein batch represents the batch size, s represents the time series length, and k represents the number of indicators in the network element indicator data. The output dimension of the output data of the network pre-trained large model is [batch, s, 1024], wherein the output data can be recorded as C network element features.

[0101] The input data of the unimodal encoder is the network element label, where the dimension of the input data is [batch, label_max_length], batch represents the batch size, and label_max_length represents the maximum text length of the network element label; the output dimension of the output data of the unimodal encoder is [batch, 768], where the output data is recorded as W classification label features.

[0102] Here, the network element label can be input into the single-mode encoder for feature extraction, thereby extracting the classification label feature; at the same time, the network element indicator data is input into the network pre-trained large model for feature extraction, thereby obtaining the network element feature.

[0103] Next, two feedforward neural networks (i.e., the third feedforward neural network described below) can be used to map the C network element features and the W classification label features to the same feature space, respectively, thereby obtaining the following result (i.e., the third processing result in the following embodiment): and .in, It represents the processing result of the Bth training sample in the batch after the unimodal encoder and the feedforward neural network; The Bth training sample in the batch is processed by the network pre-training large model and the feedforward neural network.

[0104] In the embodiment of the present disclosure, a unimodal encoder is selected as a pre-trained model. By selecting a feature extraction network as a pre-trained model, the amount of training tasks for the neural network can be reduced, so that the root cause localization model can converge quickly during the training process, thereby accelerating the training of the root cause localization model.

[0105] In an optional implementation, the step S103 generates the prediction analysis text of the training sample based on the network element feature and the network element analysis text, and specifically includes the following steps:

[0106] Step S21: Processing the network element feature and the network element analysis text through an attention module to obtain a first processing result;

[0107] Step S22: Processing the first processing result through a first feedforward neural network to obtain the predictive analysis text.

[0108] In the embodiment of the present disclosure, the attention module is the multimodal encoder described above. Here, the network element features and the network element analysis text can be input into the multimodal encoder for processing, so as to guide the multimodal encoder to generate the prediction analysis text based on the network element features.

[0109] In an embodiment of the present disclosure, network element features and network element analysis text may be input into an attention module for processing to obtain a first processing result; next, the first processing result may be input into a first feedforward neural network for processing to obtain a predictive analysis text.

[0110] In the disclosed embodiment, by adding a multimodal encoder and generating network element analysis text guided by network element features, it is possible to obtain a detailed analysis description of the classification result while achieving the classification result through comparative learning, so as to explain the classification result through the detailed analysis description, thereby improving the interpretability and trustworthiness of the classification decision, and gradually establishing the user's trust in the root cause location model.

[0111] Here, the number of attention modules can be multiple, and multiple attention modules are connected in series; wherein the output result of the previous attention module is the input of the next attention module; and the output result of the last attention module is the first processing result. wherein the input of each attention module includes network element features, and the input of the first attention module also includes network analysis text.

[0112] For example, the multiple attention modules include: attention module 1, attention module 2 and attention module 3. At this time, first, the network element features and the network analysis text are input into the attention module 1 for processing to obtain output result 1; the output result 1 and the network element features are input into the attention module 2 for processing to obtain output result 2; the output result 2 and the network element features are input into the attention module 3 for processing to obtain output result 3, that is, the first processing result. Afterwards, the first processing result can be processed by the first feedforward neural network to obtain the predictive analysis text.

[0113] In the embodiment of the present disclosure, each of the attention modules includes: a causal self-attention model, a cross attention module, and a second feedforward neural network, which are connected in sequence.

[0114] From the above description, it can be seen that there are multiple attention modules, the input of each attention module is the output result of the second feedforward neural network of the previous attention module, the input of the cross attention module in each attention module is the network element feature, and the output result of the second feedforward neural network of the last attention module is the first processing result, that is, the input of the first feedforward neural network.

[0115] In this case, the above steps process the network element features and the network element analysis text through the attention module to obtain a first processing result, which specifically includes the following steps:

[0116] First, the network element analysis text is processed by the causal self-attention model to obtain text features;

[0117] Secondly, the text feature and the network element feature are processed by the cross attention module to obtain a second processing result;

[0118] Next, the second processing result is processed by the second feedforward neural network, and the first processing result is determined according to the output result.

[0119] In the disclosed embodiment, the dimension of the network element analysis text T is [batch, text_max_length], where batch represents the batch size, and text_max_length represents the maximum length of the network element analysis text. For example, text_max_length can be set to 128. The network element analysis text and the network element analysis text T dimension are concatenated, and after the tokenization processing by the tokenizer, they are input into the causal self-attention model causal self-attention of the multimodal encoder. The output dimension of the causal self-attention model causal self-attention is [batch, token_number, 768]. Where batch represents the batch size, token_number represents the number of segmented words obtained after the segmentation processing, and 768 represents the output dimension of the causal self-attention model.

[0120] After the network element analysis text is processed through causal self-attention to obtain text features, the text features and network element features can enter cross attention for processing to obtain a second processing result.

[0121] Specifically, the text feature and the network element feature are processed by the cross attention module to obtain a second processing result, which specifically includes the following steps:

[0122] First, the text features are linearly mapped through the cross attention module to obtain a query matrix;

[0123] Secondly, the network element features are linearly mapped through the cross attention module to obtain a key value matrix and a value matrix;

[0124] Next, the second processing result is calculated through the query matrix, the key value matrix and the value matrix.

[0125] In the disclosed embodiment, first, the text feature can be input into the cross attention module for linear mapping processing, wherein the text feature can be used as the query matrix Q after the linear mapping processing. Here, the network element feature C can also be input into the cross attention module for linear mapping processing, wherein the network element feature can obtain the key value matrix K and the value matrix V after the linear mapping processing.

[0126] After obtaining the query matrix, the key value matrix and the value matrix, the query matrix Q determined based on the text features and the key value matrix K determined based on the network element features can be multiplied to obtain the target product; wherein the target product is used to indicate the correlation between the network element analysis text and the network element indicator data; the target product is multiplied by After scaling and softmax operation, the attention score matrix is ​​obtained. The attention score matrix is ​​multiplied by the value matrix V determined based on the network element features to obtain the cross attention layer output, that is, the second processing result, where the dimension of the cross attention layer output is [batch, token_number, 768].

[0127] The process of calculating the second processing result by using the query matrix, the key value matrix and the value matrix can also be expressed by the following attention module calculation formula:

[0128] ;in, That is the output of the crossattention layer, that is, the second processing result.

[0129] In the above implementation, the way of processing network element features and network element analysis text through the attention model can reduce the dependence on external information, and can more accurately capture the internal correlation between network element features and network element analysis text, thereby providing a basis for guiding the generation of network element analysis text, and can generate more accurate and reasonable network element analysis text; and then achieve the generation of decision analysis text while obtaining the classification result in the root cause location module, and realize the interpretation and analysis of the classification result.

[0130] In an optional implementation, the step S104 trains the root cause location model based on the prediction analysis text, the classification label features, and the network element features to obtain a target root cause location model, which specifically includes the following steps:

[0131] Step S31: respectively processing the classification label feature and the network element feature through a third feedforward neural network to obtain a third processing result, and calculating a first loss function value according to the third processing result;

[0132] Step S32: Calculating a second loss function value according to the network element analysis text and the prediction analysis text;

[0133] Step S33: performing weighted summation on the first loss function value and the second loss function value to obtain a target loss function value, and training the root cause location model according to the target loss function value to obtain a target root cause location model after training.

[0134] It can be seen from the above description that in the embodiment of the present disclosure, after the network element label is processed by the single-mode encoder, the classification label feature is obtained; after the network element indicator data is processed by the network pre-training large model, the network element feature is obtained. Here, the third feedforward neural network includes two feedforward neural networks, which are recorded as feedforward neural network 1 and feedforward neural network 2. The classification label feature can be processed by feedforward neural network 1 to obtain processing result 1 ( ); The network element features can be processed through the feedforward neural network 2 to obtain the processing result 2 ( ); wherein processing result 1 and processing result 2 are the third processing result described above.

[0135] Here, the network structure of the feedforward neural network 1 may be the same as the network structure of the feedforward neural network 2, and the network parameters of the feedforward neural network 1 may not be the same as the network parameters of the feedforward neural network 2.

[0136] After obtaining processing result 1 and processing result 2, processing result 1 and processing result 2 can be normalized respectively to obtain normalized vectors; then, the similarity between the two normalized vectors is calculated, so as to calculate the first loss function value according to the similarity, wherein the first loss function value can be the function value of the cross entropy loss function, or can also be other types of loss functions that can replace the cross entropy loss function.

[0137] against and , the corresponding network element analysis text and network element feature pairs in the sample order are positive samples, that is, is a positive sample, is a negative sample. The goal of training the root cause localization model is to maximize the similarity of the network element features and text feature pairs on the diagonal. Here, the first loss function can be recorded as .

[0138] In the embodiment of the present disclosure, after obtaining the prediction analysis text, a second loss function value may be calculated based on the prediction analysis text and the network element analysis text, wherein the loss function corresponding to the second loss function value may be a cross entropy loss function, which is recorded as , or other loss functions that can replace the cross entropy loss function, or other loss functions that can achieve the same function.

[0139] From the above description, we can know that the text generation network includes multiple attention modules, each of which is composed of a causal self-attention layer, a cross attention layer and a feed forward layer. The input of causal self-attention comes from the previous attention module, and the input of cross attention is the network element feature C output by the network pre-trained large model, K and V as attention operations, and the output of the causal self-attention layer, Q as attention operations. The output of the last attention module is linearly mapped to [batch, token_number, vocab_size] through the first feedforward neural network feed forward layer, where vocab_size represents the size of the word segmentation dictionary. At this point, the output of the first feedforward neural network feed forward layer is the predictive analysis text. At this point, the cross entropy loss can be calculated based on the predictive analysis text and the network element analysis text, recorded as .

[0140] After obtaining the first loss function value and the second loss function value, the weights of the first loss function value and the second loss function value can be determined, and then the first loss function value and the second loss function value are weightedly summed according to the weights to obtain the target loss function value.

[0141] In the above implementation, the root cause localization model is trained by using the first loss function value and the second loss function value, so as to accelerate the convergence speed of the model and obtain a model that can simultaneously determine the classification result and the detailed analysis description of the classification result.

[0142] Here, in the process of training the root cause localization model, the weights of the first loss function value and the second loss function value are not fixed, but dynamically change, and the change of the weights of the first loss function value and the second loss function value is associated with the training rounds and hyperparameter update rate (i.e., the weight update rate described below) of the root cause localization model at the current moment.

[0143] Specifically, performing weighted summation on the first loss function value and the second loss function value to obtain a target loss function value includes the following steps:

[0144] First, determine the target training round of the root cause localization model at the current moment;

[0145] Secondly, based on the target training round and the weight update rate, determine the weights of the first loss function value and the second loss function value at the current moment to obtain a first weight and a second weight;

[0146] Next, based on the first weight and the second weight, a weighted sum is performed on the first loss function value and the second loss function value to obtain a target loss function value.

[0147] In the embodiment of the present disclosure, the weights of the first loss function value and the second loss function value are set as hyperparameters, and the hyperparameters are dynamically updated during the root cause location model training process, wherein the update rules of the hyperparameters are described as follows:

[0148] Determine the training round of the root cause location model at the current moment, and obtain the target training round, recorded as epoch. For example, the training round at the current moment is A. Next, determine the weight update rate of the first loss function value and the second loss function value, recorded as γ, and determine the weight update base, recorded as ; Then, perform exponential operation on the weight update rate γ and the target training round epoch to obtain the exponential operation result, that is, ; Then, the product between the weight update rate γ and the result of the exponential operation is calculated, so that the weights of the first loss function value and the second loss function value at the current moment can be determined according to the product. For example, the product is recorded as , at this time, the weight of the first loss function value (i.e., the first weight) can be determined as ; and determine the weight of the second loss function value as 1- (ie, the second weight).

[0149] The above process can be expressed by the following formula: ; Among them, the hyperparameter Dynamic update, where epoch is the target training round, γ is the weight update rate, usually, the weight update rate can be set to 0.995, 0.994, or 0.996. Here, the specific value of the weight update rate can be determined according to actual needs, and the weight update base Generally, it can be set to values ​​such as 0.9, 0.89, 0.91, etc. Here, the specific value of the weight update base can be determined according to actual needs.

[0150] After determining the first weight and the second weight, the first loss function value and the second loss function value may be weighted and summed based on the first weight and the second weight to obtain a target loss function value. The specific formula is as follows: .

[0151] In the disclosed embodiment, the single-modal encoder adopts a pre-trained model. During the training of the root cause localization model, the parameters of the network pre-trained large model and the single-modal encoder are fixed and are not updated.

[0152] In the disclosed embodiment, since the unimodal encoder and the network pre-trained large model under the classification task are both pre-trained models, at this time, classification only needs to update the network parameters of the two feedforward neural networks; at the same time, the classification task is simpler than the text generation task. Therefore, setting the weights of the first loss function value and the second loss function value to be dynamically adjusted with the training process can improve the convergence speed of the classification module in the root cause localization model, so that the classification module can converge faster, thereby shortening the training time of the root cause localization model.

[0153] In the embodiment of the present disclosure, the root cause localization model is trained according to the target loss function value to obtain the target root cause localization model, which specifically includes the following steps:

[0154] The network parameters of the attention module, the first feedforward neural network and the third feedforward neural network in the root cause localization model are adjusted according to the target loss function value, and the target root cause localization model is determined according to the root cause localization model after adjusting the network parameters; wherein the attention module and the first feedforward neural network are networks that generate predictive analysis texts of the training samples based on the network element features and the network element analysis texts.

[0155] In the disclosed embodiment, the unimodal encoder is a pre-trained model. At this time, it is necessary to adjust the network parameters of the text generation network (i.e., the attention module, the first feedforward neural network) and the third feedforward neural network according to the target loss function value, that is, adjust the causal self-attention layer, the cross attention layer, the feed forward layer, and the network parameters of the third feedforward neural network. Here, the above steps can be repeatedly performed until a root cause localization model that meets the training accuracy requirements is obtained, and the root cause localization model that meets the training accuracy requirements is determined as the target root cause localization model.

[0156] From the above description, it can be seen that the technical solution disclosed in the present invention, based on the existing use of contrastive learning to complete the root cause location of network elements, introduces a multimodal encoder to realize the generation of network element analysis text. For the input network element indicator data, an analysis text is generated while obtaining the root cause decision; the technical solution disclosed in the present invention can realize end-to-end training, and analyze and describe the reasons for making decisions without the need for additional training of interpretable models (such as linear models) or LIME models, thereby improving the interpretability of the classification module.

[0157] Combine the following Figure 2 The root cause localization model training method is introduced.

[0158] like Figure 2As shown in the figure, the root cause localization model mainly includes the following two parts: feature extraction network and text generation network; wherein, the feature extraction network includes a unimodal encoder, a network pre-trained large model and two feedforward neural networks (i.e., feedforward neural network 1 and feedforward neural network 2); the text generation network includes a multimodal encoder and a feedforward neural network 3. Among them, the unimodal encoder and the network pre-trained large model constitute a classification module based on contrastive learning.

[0159] The text generation network includes multiple attention modules, each of which consists of a causal self-attention layer, a cross attention layer, and a feed forward layer. The input of causal self-attention comes from the previous attention module, and the input of cross attention is the network element feature C output by the network pre-trained large model, K and V as attention operations, and the output of the causal self-attention layer, Q as attention operation.

[0160] Here, the input of feedforward neural network 1 is connected to the output of the unimodal encoder, and the input of feedforward neural network 2 is connected to the output of the network pre-trained large model. The output of the network pre-trained large model is connected to the cross attention layer in the text generation network.

[0161] like Figure 2 As shown, the input data of the network pre-trained large model is the network element index, and the dimension of the input data of the network pre-trained large model is [batch, s, k]. The output dimension of the output data of the network pre-trained large model is [batch, s, 1024], where the output data can be recorded as C network element features. The input data of the unimodal encoder is the network element label, where the dimension of the input data is [batch, label_max_length], and the output dimension of the output data of the unimodal encoder is [batch, 768], where the output data is recorded as W classification label features.

[0162] Here, the network element label can be input into the single-mode encoder for feature extraction, thereby extracting the classification label feature; at the same time, the network element indicator data can be input into the network pre-trained large model for feature extraction, thereby obtaining the network element feature. Next, the feedforward neural network 1 and the feedforward neural network 2 can be used to map the C network element features and the W classification label features to the same feature space, respectively, to obtain the result and .in, It can be understood as a classification label feature with a dimension of [batch, feature_size], It is understood as a network element feature with a dimension of [batch, feature_size]. The loss function value of the contrastive learning loss (i.e., the first loss function value) can be calculated through the network element features and the classification label features.

[0163] like Figure 2 As shown, the network element analysis text and the network element analysis text T dimension are spliced, and after the tokenize word segmentation processing, they are input into the causal self-attention model of the multimodal encoder. The output dimension of the causal self-attention model is [batch, token_number, 768]. After the network element analysis text is processed by causal self-attention to obtain the text features, the text features can be input into the cross attention module cross attention for linear mapping processing, wherein the text features can be used as the query matrix Q after the linear mapping processing. Here, the network element features C can also be input into the cross attention module crossattention for linear mapping processing, wherein the network element features can obtain the key value matrix K and the value matrix V after the linear mapping processing. Afterwards, the query matrix Q determined based on the text features and the key value matrix K determined based on the network element features can be multiplied to obtain the target product; wherein, the target product is used to indicate the correlation between the network element analysis text and the network element indicator data; the target product is After scaling and softmax operations, the attention score matrix is ​​obtained. The attention score matrix is ​​multiplied by the value matrix V determined based on the network element features to obtain the cross attention layer output. Finally, the cross attention layer output is processed by the feedforward neural network 4 in the multimodal encoder to obtain the output of the attention module. Among them, the output of the last attention module is processed by the feedforward neural network 3 to obtain the predictive analysis text. Then, the loss function value (i.e., the second loss function value) of the language model loss can be calculated through the predictive analysis text and the network element analysis text; and then the first loss function value and the second loss function value are used to calculate the loss function value of the language model loss. Figure 2 The network parameters of feedforward neural network 1, feedforward neural network 2 and text generation network are adjusted.

[0164] See also Figure 3 FIG. 1 is a flowchart of a root cause location method provided by an embodiment of the present disclosure, wherein the method includes steps S301 to S303, wherein:

[0165] S301, obtaining network element indicator data to be predicted.

[0166] The network element indicator data includes but is not limited to the following data: network element performance indicators, MR performance indicators, industrial parameters and other time series data; wherein the time granularity of the network element indicator data may be N minutes (or seconds), and the data dimension may be M dimensions, for example, the time granularity is 15 minutes, and the data dimension is 159 dimensions. Here, the specific values ​​of the time granularity and the data dimension are not specifically limited, and are subject to implementation.

[0167] S302, extracting the characteristics of the network element indicator data to obtain target network element characteristics.

[0168] Here, we can Figure 2 The network pre-trained large model shown extracts the features of network element indicator data to obtain network element features.

[0169] S303: Determine the classification result corresponding to the network element indicator data according to the target network element feature, and process the target network element feature through a text generation network in a target root cause location model to obtain an analysis text of the classification result.

[0170] In the disclosed embodiment, the network element indicator data may be subjected to feature extraction through a network element pre-training large model to obtain target network element features.

[0171] In specific implementation, the network element indicator data can be input into the network pre-training large model; then, the network pre-training large model is used to extract the features to obtain the target network element features; the target network element features are fed into the feedforward neural network (i.e., Figure 2 After the feedforward neural network 2) is processed, the processing result is output. At this time, the classification result of the network element indicator data can be obtained based on the processing result. Here, the specific structure of the network element pre-training large model is as follows Figure 2 shown.

[0172] After obtaining the target network element features, the target network element features can also be input into the text generation network of the target root cause location model for processing, thereby obtaining the analysis text of the classification result. The specific processing process of the text generation network is the same as the processing process of the text generation network in the above embodiment.

[0173] In the embodiment of the present disclosure, the target root cause localization model is obtained by training in the embodiment described by the root cause localization model training method, and the specific training process will not be described in detail here.

[0174] The accuracy of the classification task in the disclosed technical solution is relatively high, such as Figure 4As shown in the figure, it can reach 88.22%. At the same time, by taking the network element characteristics of the network element indicator data as a condition to guide the text generation network in the root cause location model to generate network element analysis text, it is possible to obtain a detailed analysis description of the classification results while achieving the classification results through comparative learning, establish the correlation between the abnormal indicators and the classification results, and improve the interpretability and trustworthiness of the classification decisions of the root cause location model.

[0175] In an optional implementation, the above step processes the target network element features through the target root cause location model to obtain an analysis text of the classification result, specifically including:

[0176] First, the target network element features are processed by the attention module in the target root cause positioning model to obtain a first target processing result;

[0177] Secondly, the first target processing result is linearly mapped by the first feedforward neural network in the target root cause location model to obtain an analysis text of the classification result.

[0178] In the embodiment of the present disclosure, the text generation network of the target root cause localization model includes an attention model and a first feedforward neural network, namely Figure 2 Here, the target network element feature can be input into the attention module for processing, thereby obtaining a first target processing result; then, the first target processing result can be input into the first feedforward neural network for linear mapping processing, thereby obtaining an analysis text of the classification result.

[0179] Here, there can be multiple attention modules, and multiple attention modules are connected in series; wherein the output result of the previous attention module is the input of the next attention module; and the output result of the last attention module is the first target processing result. wherein the target network element feature is the input of each attention module.

[0180] In the above implementation, by using the network element characteristics of the network element indicator data as a condition to guide the text generation network in the root cause location model to generate network element analysis text, it is possible to obtain a detailed analysis description of the classification result while achieving the classification result through comparative learning, establish the correlation between the abnormal indicator and the classification result, and improve the interpretability and trustworthiness of the classification decision of the root cause location model.

[0181] In an optional embodiment, the attention module includes: a causal self-attention model, a cross-attention module and a second feedforward neural network connected in sequence.

[0182] In this case, the above steps process the target network element features through the attention module in the target root cause positioning model to obtain a first target processing result, specifically including:

[0183] First, the target input data is processed by the causal self-attention model to obtain the target text features;

[0184] Secondly, the target text feature and the target network element feature are processed by the cross attention module to obtain a second target processing result;

[0185] Next, the second target processing result is processed by the second feedforward neural network, and the first target processing result is determined according to the output result.

[0186] Here, the target input data of the causal self-attention model in the first attention module is the initial text vector, and the target input data of the causal self-attention model in the non-first attention module is the text vector output by the previous attention module; the output result of the second feedforward neural network in the last attention module is the first target processing result.

[0187] In the embodiment of the present disclosure, for the first attention module, the initial text vector can be input into the causal self-attention model for processing to obtain the target text feature, and then the target network element feature and the target text feature are input into the cross attention module for processing to obtain the second target processing result; wherein the processing principle of the cross attention module is the same as Figure 2 The processing principle of the cross-attention module is the same and will not be described in detail here. After obtaining the second target processing result, the second target processing result can be processed by a second feedforward neural network to obtain an output result, that is, the text vector output by the attention module. Next, the output result can be used as the target input data of the next attention module, and the target input data and target network element features can be processed according to the process described above until the last attention module outputs the result, and the output result of the last attention module is determined as the first target processing result.

[0188] In the above implementation, the way of processing network element features and network element analysis text through the attention model can reduce the dependence on external information, and can more accurately capture the internal correlation between network element features and network element analysis text, thereby providing a basis for guiding the generation of network element analysis text, and can generate more accurate and reasonable network element analysis text; and then achieve the generation of decision analysis text while obtaining the classification result in the root cause location module, and realize the interpretation and analysis of the classification result.

[0189] The disclosed technical solution can be applied to the end-to-end service optimization self-closed loop solution of wireless network load scenarios. It has been successfully deployed in a certain place and is being tested on the live network. The model version is iteratively optimized every two weeks online. Figure 4 As shown, the accuracy f1-score of the root cause classification result of the disclosed technical solution reaches more than 85%, and the text generation Rouge-L index (Recall-Oriented Understudy for Gisting Evaluation) reaches more than 80%.

[0190] The technical solution disclosed in the present invention can realize the perception of multiple types of network element indicator data, for example, it supports the perception of 22 degradation indicators; it can also realize multi-level root cause positioning, for example, it can effectively identify the real problems of the network while outputting detailed text descriptions; it can also realize multi-dimensional optimization, for example, it can output capacity-type network element parameter adjustment plans, support optimization plans such as capacity expansion, idle demolition and new site construction; it can also perform all-round evaluation, for example, it can comprehensively consider the effects of multiple standards, multiple frequency bands, and multiple network elements for the best overall result.

[0191] The input and output parameters of the online application interface of the disclosed technical solution include the following:

[0192] Input parameters:

[0193] snap_shot: The type is pandas.Dataframe, which is a required input parameter. The meaning of this parameter is: 512 network element snapshots before the dispatch time [512, 1025], where the network element snapshot is obtained by calling the network element pre-trained large model, that is, the network element snapshot is the output result of the network element pre-trained large model, for example, the network element snapshot, (512, 1024), the first column is the time column, and the next 1024 columns are data columns.

[0194] classify_start_date and classify_end_date: type datetime, required input parameter, which means: the time period the user wants to classify the data; for example, datetime(2023, 5, 27, 1,15) and datetime(2023, 5, 27, 1,15).

[0195] network_type: The type is a string and is a required input parameter. The meaning of this parameter is: 4G and 5G are optional; the current default is 4G.

[0196] model_type: The type is a string, which is a required input parameter. The meaning of this parameter is: single network element and multiple network elements can be selected; the current default is single network element.

[0197] Labels: The type is list, which is a required input parameter. The meaning of this parameter is: a list of secondary root cause categories, including all types of secondary root causes.

[0198] tertiary_labels: The type is list, which is a required input parameter. The meaning of this parameter is: a list of tertiary root cause categories, including all types of tertiary root causes.

[0199] Different granularity definitions can be made for the same problem, so multiple levels of definition can be made for the same problem through a multi-level root cause category list.

[0200] order_id: The type is a string, not a required input parameter, and the meaning of this parameter is: work order number. rulename: The type is a string, not a required input parameter, and the meaning of this parameter is: order dispatching rule.

[0201] Output parameters:

[0202] classify_result: The output format is a string. The meaning of this parameter is: first-level root cause result, such as 'capacity'. second_result: The output format is a string. The meaning of this parameter is: second-level root cause result, such as 'high load'. tertiary_result: The output format is a string. The meaning of this parameter is: third-level root cause result, such as 'capacity expansion required'.

[0203] In the embodiment of the present disclosure, the content returned by the above interface includes multiple levels of root cause results (i.e., the classification results described above). For example, it includes three levels of root cause results, i.e., first-level root cause results, second-level root cause results, and third-level root cause results; wherein, from the first-level root cause results to the third-level root cause results, a fine-grained description of the problem can be made.

[0204] analysis_text: The output format is a string. The meaning of this parameter is: network element analysis text, such as "The coverage scenario is an industrial park. There are no faults or interference in the cell during the statistical period. The recent load is high and the PRB busy hour utilization rate is about 74.51%. It is determined to be caused by high business." order_id: The output format is a string. The meaning of this parameter is: if the work order number is entered, the work order number should be output. If the work order number is not entered, an empty string is output.

[0205] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0206] Based on the same inventive concept, the embodiment of the present disclosure also provides a root cause localization model training device corresponding to the root cause localization model training method. Since the principle of solving the problem by the device in the embodiment of the present disclosure is similar to the above-mentioned root cause localization model training method in the embodiment of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0207] Reference Figure 5 As shown, it is a schematic diagram of a root cause localization model training device provided by an embodiment of the present disclosure, the device comprises: a first acquisition unit 51, a first feature extraction unit 52, a first text generation unit 53 and a training unit 54; wherein,

[0208] A first acquisition unit is used to acquire a multimodal training sample; wherein the training sample includes at least one of the following: network element indicator data, network element label and network element analysis text;

[0209] A first feature extraction unit, used to extract features of the network element label and the network element indicator data, and obtain classification label features and network element features in sequence;

[0210] A first text generation unit, configured to generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text;

[0211] A training unit is used to train the root cause location model based on the prediction analysis text, the classification label features and the network element features to obtain a target root cause location model.

[0212] In a possible implementation manner, the first text generation unit is used to:

[0213] Processing the network element feature and the network element analysis text through an attention module to obtain a first processing result;

[0214] The first processing result is processed by a first feedforward neural network to obtain the predictive analysis text.

[0215] In a possible implementation, when each of the attention modules includes: a causal self-attention model, a cross attention module, and a second feedforward neural network connected in sequence, the first text generation unit is used to:

[0216] Processing the network element analysis text through the causal self-attention model to obtain text features;

[0217] Processing the text feature and the network element feature by the cross attention module to obtain a second processing result;

[0218] The second processing result is processed by the second feedforward neural network, and the first processing result is determined according to the output result.

[0219] In a possible implementation manner, the first text generation unit is used to:

[0220] Performing linear mapping processing on the text features through the cross attention module to obtain a query matrix;

[0221] Performing linear mapping processing on the network element features through the cross attention module to obtain a key value matrix and a value matrix;

[0222] The second processing result is calculated by using the query matrix, the key value matrix and the value matrix.

[0223] In one possible implementation, there are multiple attention modules, the input of each attention module is the output result of the previous attention module, the input of the cross-attention module in each attention module is the network element feature, and the output result of the last attention module is the input of the first feedforward neural network.

[0224] In a possible implementation manner, the first feature extraction unit is used to:

[0225] Extracting features of the network element labels by a single-mode encoder to obtain the classification label features;

[0226] The network element indicator data is subjected to feature extraction through a network pre-training large model to obtain the network element features.

[0227] In a possible implementation manner, the training unit is used to:

[0228] Processing the classification label features and the network element features respectively through a third feedforward neural network to obtain a third processing result, and calculating a first loss function value according to the third processing result;

[0229] Calculating a second loss function value according to the network element analysis text and the prediction analysis text;

[0230] A weighted sum is performed on the first loss function value and the second loss function value to obtain a target loss function value, and a root cause localization model is trained according to the target loss function value to obtain the target root cause localization model.

[0231] In a possible implementation manner, the training unit is used to:

[0232] Determine a target training round of the root cause localization model at a current moment;

[0233] Based on the target training round and the weight update rate, determine the weights of the first loss function value and the second loss function value at the current moment to obtain a first weight and a second weight;

[0234] Based on the first weight and the second weight, the first loss function value and the second loss function value are weightedly summed to obtain a target loss function value.

[0235] In a possible implementation manner, the training unit is used to:

[0236] The network parameters of the attention module, the first feedforward neural network and the third feedforward neural network in the root cause localization model are adjusted according to the target loss function value, and the target root cause localization model is determined according to the root cause localization model after adjusting the network parameters; wherein the attention module and the first feedforward neural network are networks that generate predictive analysis texts of the training samples based on the network element features and the network element analysis texts.

[0237] Reference Figure 6 , which is a schematic diagram of a root cause location device provided by an embodiment of the present disclosure, the device comprises: a second acquisition unit 61, a second feature extraction unit 62, and a second text generation unit 63; wherein,

[0238] A second acquisition unit 61 is used to acquire network element indicator data to be predicted;

[0239] A second feature extraction unit 62 is used to extract features of the network element indicator data to obtain target network element features; wherein the target root cause location model is trained using the root cause location model training method described in any one of the above embodiments;

[0240] The second text generation unit 63 is used to determine the classification result corresponding to the network element indicator data according to the target network element characteristics, and process the target network element characteristics through the text generation network in the target root cause positioning model to obtain the analysis text of the classification result.

[0241] In a possible implementation manner, the second text generation unit is used to:

[0242] Processing the target network element feature through the attention module in the target root cause positioning model to obtain a first target processing result;

[0243] The first target processing result is linearly mapped by the first feedforward neural network in the target root cause location model to obtain an analysis text of the classification result.

[0244] In a possible implementation, when the attention module includes: a causal self-attention model, a cross attention module, and a second feedforward neural network connected in sequence, the second text generation unit is used to:

[0245] Processing the target input data through the causal self-attention model to obtain target text features;

[0246] Processing the target text feature and the target network element feature through the cross attention module to obtain a second target processing result;

[0247] The second target processing result is processed by the second feedforward neural network, and the first target processing result is determined according to the output result.

[0248] In one possible implementation, the target input data of the causal self-attention model in the first attention module is the initial text vector, the target input data of the causal self-attention model in the non-first attention module is the text vector output by the previous attention module, and the output result of the second feedforward neural network in the last attention module is the first target processing result.

[0249] In a possible implementation manner, the second feature extraction unit is used to:

[0250] The network element indicator data is subjected to feature extraction through a network element pre-training large model to obtain the target network element features.

[0251] For descriptions of the processing flow of each module in the device and the interaction flow between each module, reference may be made to the relevant descriptions in the above method embodiment, which will not be described in detail here.

[0252] Corresponds to Figure 1 The root cause positioning model training method and the root cause positioning method in the present disclosure embodiment also provide an electronic device 700, such as Figure 7 FIG. 7 is a schematic diagram of the structure of an electronic device 700 provided in an embodiment of the present disclosure, including:

[0253] Processor 71, memory 72, and bus 73; memory 72 is used to store execution instructions, including internal memory 721 and external memory 722; the internal memory 721 is also called internal memory, which is used to temporarily store the operation data in the processor 71 and the data exchanged with the external memory 722 such as a hard disk. The processor 71 exchanges data with the external memory 722 through the internal memory 721. When the electronic device 700 is running, the processor 71 communicates with the memory 72 through the bus 73, so that the processor 71 executes the following instructions:

[0254] Acquire multimodal training samples; wherein the training samples include at least one of the following: network element indicator data, network element labels, and network element analysis texts;

[0255] Extracting the network to extract the features of the network element label and the network element indicator data, and obtaining the classification label features and the network element features in turn;

[0256] Generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text;

[0257] Based on the prediction analysis text, the classification label features and the network element features, the root cause location model is trained to obtain a target root cause location model.

[0258] or

[0259] Obtaining network element indicator data to be predicted;

[0260] Extracting the network to extract the characteristics of the network element indicator data to obtain the target network element characteristics;

[0261] The classification result corresponding to the network element indicator data is determined according to the target network element feature, and the target network element feature is processed by the text generation network in the target root cause positioning model to obtain an analysis text of the classification result.

[0262] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the root cause location model training method or the root cause location method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0263] The embodiments of the present disclosure also provide a computer program product, which carries a program code. The instructions included in the program code can be used to execute the root cause localization model training method or the steps of the root cause localization method described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0264] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is implemented as a computer storage medium. In another optional embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0265] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0266] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0267] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0268] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0269] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A root cause location model training method, characterized in that: include: Acquire multimodal training samples; wherein the training samples include at least one of the following: network element indicator data, network element labels, and network element analysis texts; Extracting features of the network element label and the network element indicator data to obtain classification label features and network element features; Generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text; Based on the prediction analysis text, the classification label features and the network element features, a root cause location model is trained to obtain a target root cause location model; The step of generating the prediction analysis text of the training sample based on the network element feature and the network element analysis text includes: Processing the network element analysis text through a causal self-attention model to obtain text features; Processing the text features and the network element features through a cross attention module to obtain a second processing result; processing the second processing result through a second feedforward neural network, and determining the first processing result according to the output result; wherein the causal self-attention model, the cross attention module and the second feedforward neural network are connected in sequence; The first processing result is processed by a first feedforward neural network to obtain the predictive analysis text.

2. The method according to claim 1, characterized in that: The processing of the text feature and the network element feature by the cross attention module to obtain a second processing result includes: Performing linear mapping processing on the text features through the cross attention module to obtain a query matrix; Performing linear mapping processing on the network element features through the cross attention module to obtain a key value matrix and a value matrix; The second processing result is calculated by using the query matrix, the key value matrix and the value matrix.

3. The method according to claim 1, characterized in that There are multiple attention modules, the input of each attention module is the output result of the previous attention module, the input of the cross-attention module in each attention module is the network element feature, and the output result of the last attention module is the input of the first feedforward neural network.

4. The method according to claim 1, characterized in that: The extracting the features of the network element label and the network element indicator data to obtain the classification label features and the network element features in sequence includes: Extracting features of the network element labels by a single-mode encoder to obtain the classification label features; The network element indicator data is subjected to feature extraction through a network element pre-training large model to obtain the network element features.

5. The method according to claim 1, characterized in that The step of training the root cause location model based on the prediction analysis text, the classification label features, and the network element features to obtain a target root cause location model includes: Processing the classification label features and the network element features respectively through a third feedforward neural network to obtain a third processing result, and calculating a first loss function value according to the third processing result; Calculating a second loss function value according to the network element analysis text and the prediction analysis text; A weighted sum is performed on the first loss function value and the second loss function value to obtain a target loss function value, and a root cause localization model is trained according to the target loss function value to obtain the target root cause localization model.

6. The method according to claim 5, characterized in that The step of performing a weighted summation on the first loss function value and the second loss function value to obtain a target loss function value includes: Determine a target training round of the root cause localization model at a current moment; Based on the target training round and the weight update rate, determine the weights of the first loss function value and the second loss function value at the current moment to obtain a first weight and a second weight; Based on the first weight and the second weight, the first loss function value and the second loss function value are weightedly summed to obtain a target loss function value.

7. The method according to claim 5, characterized in that The step of training the root cause location model according to the target loss function value to obtain the target root cause location model includes: The network parameters of the attention module, the first feedforward neural network and the third feedforward neural network in the root cause localization model are adjusted according to the target loss function value, and the target root cause localization model is determined according to the root cause localization model after adjusting the network parameters; wherein the attention module and the first feedforward neural network are networks that generate predictive analysis texts of the training samples based on the network element features and the network element analysis texts.

8. A root cause location method, characterized in that: include: Obtaining network element indicator data to be predicted; Extracting characteristics of the network element indicator data to obtain target network element characteristics; Determine a classification result corresponding to the network element indicator data according to the target network element characteristics, and process the target network element characteristics through a target root cause location model to obtain an analysis text of the classification result; The processing of the target network element features by the target root cause location model to obtain the analysis text of the classification result includes: The target input data is processed through the causal self-attention model to obtain the target text features; The target text feature and the target network element feature are processed by a cross attention module to obtain a second target processing result; the second target processing result is processed by a second feedforward neural network, and the first target processing result is determined according to the output result; wherein the causal self-attention model, the cross attention module and the second feedforward neural network are connected in sequence; The first target processing result is linearly mapped by the first feedforward neural network in the target root cause location model to obtain an analysis text of the classification result.

9. The method according to claim 8, characterized in that The target input data of the causal self-attention model in the first attention module is the initial text vector, the target input data of the causal self-attention model in the non-first attention module is the text vector output by the previous attention module, and the output result of the second feedforward neural network in the last attention module is the first target processing result.

10. The method according to claim 8, characterized in that The extracting the characteristics of the network element indicator data to obtain the target network element characteristics includes: The network element indicator data is subjected to feature extraction through a network element pre-training large model to obtain the target network element features.

11. A root cause location model training device, characterized in that: include: A first acquisition unit is used to acquire a multimodal training sample; wherein the training sample includes at least one of the following: network element indicator data, network element label and network element analysis text; A first feature extraction unit, used to extract the features of the network element label and the network element indicator data, and obtain the classification label feature and the network element feature in sequence; A first text generation unit, configured to generate a prediction analysis text of the training sample based on the network element feature and the network element analysis text; A training unit, configured to train the root cause location model based on the prediction analysis text, the classification label features and the network element features to obtain a target root cause location model; Wherein, the first text generation unit is further used for: Processing the network element analysis text through a causal self-attention model to obtain text features; Processing the text features and the network element features through a cross attention module to obtain a second processing result; processing the second processing result through a second feedforward neural network, and determining the first processing result according to the output result; wherein the causal self-attention model, the cross attention module and the second feedforward neural network are connected in sequence; The first processing result is processed by a first feedforward neural network to obtain the predictive analysis text.

12. A root cause locating device, characterized in that: include: A second acquisition unit, used to acquire network element indicator data to be predicted; A second feature extraction unit, used to extract features of the network element indicator data to obtain target network element features; A second text generating unit is used to determine the classification result corresponding to the network element indicator data according to the target network element feature, and process the target network element feature through a target root cause location model to obtain an analysis text of the classification result; Wherein, the second text generation unit is further used for: The target input data is processed through the causal self-attention model to obtain the target text features; The target text feature and the target network element feature are processed by a cross attention module to obtain a second target processing result; the second target processing result is processed by a second feedforward neural network, and the first target processing result is determined according to the output result; wherein the causal self-attention model, the cross attention module and the second feedforward neural network are connected in sequence; The first target processing result is linearly mapped by the first feedforward neural network in the target root cause location model to obtain an analysis text of the classification result.

13. An electronic device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of the root cause localization model training method according to any one of claims 1 to 7 are executed; or, the steps of the root cause localization method according to any one of claims 8 to 10 are executed.

14. 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 computer program executes the steps of the root cause localization model training method according to any one of claims 1 to 7; or executes the steps of the root cause localization method according to any one of claims 8 to 10.

15. A computer program product, characterized in that The computer program product is stored in a storage medium, and the program product is executed by at least one processor to implement the steps of the root cause localization model training method according to any one of claims 1 to 7; or, to execute the steps of the root cause localization method according to any one of claims 8 to 10.

Images