AI training exception diagnosis method, electronic device, and storage medium

Abstract

This invention discloses an AI training anomaly diagnosis method, electronic device, and storage medium. The method includes acquiring training data and the operation logs of the storage device; determining a data quality feature vector based on the training data; determining a storage performance feature vector based on the operation logs; calculating a first causal influence coefficient and a second causal influence coefficient of the storage performance feature vector on the data quality feature vector; and generating anomaly diagnosis results based on the first and second causal influence coefficients. This application, by calculating the first and second causal influence coefficients of the data quality feature vector on the storage performance feature vector and generating anomaly diagnosis results based on these coefficients, can perform bidirectional causal correlation analysis on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI training anomaly diagnosis method, electronic device, and storage medium. Background Technology

[0002] With the exponential growth in the scale of large-scale model training, the stability of AI training has become a key bottleneck restricting R&D efficiency. When anomalies occur in AI training tasks, it is necessary to investigate the cause.

[0003] In existing technical solutions, troubleshooting typically requires examining both the algorithm and infrastructure sides separately. Specifically, algorithm-side troubleshooting focuses on separately checking the training data, AI model structure optimization, or parameter tuning. For example, when a model's accuracy declines, algorithm engineers usually first suspect noisy data or labeling errors in the training data. Infrastructure-side troubleshooting focuses on separately checking for problems with the hardware itself, such as issues with the storage device.

[0004] However, the distribution characteristics of training data can affect the storage performance of storage devices, such as input / output (I / O) patterns. For example, when the training dataset contains a large number of small files and the data loader is not adequately prefetched, it can lead to extremely high random read loads on the storage device, causing I / O bottlenecks. In this case, the cause of the anomaly is easily misdiagnosed as a problem with the storage device itself. Therefore, current AI training anomaly diagnosis methods are not accurate enough. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides an AI training anomaly diagnosis method, electronic device, and storage medium. By determining the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector, and generating anomaly diagnosis results based on the first and second causal influence coefficients, it is possible to perform bidirectional causal correlation analysis on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results.

[0006] To address the above problems, the present invention provides the following technical solution: In a first aspect, embodiments of this application provide an AI training anomaly diagnosis method, including: acquiring training data of an AI training task and operation logs of a storage device; Determine the data quality feature vector based on the training data; Determine the storage performance feature vector based on the aforementioned runtime logs; Determine the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector; Anomaly diagnosis results are generated based on the first causal influence coefficient and the second causal influence coefficient.

[0007] Optionally, determining the data quality feature vector based on the training data includes: During the execution of the AI ​​training task, a data quality evaluation value is determined based on a portion of the training data currently read by the AI ​​model at every first preset time interval, resulting in multiple data quality evaluation values ​​at different times. The data quality feature vector is determined based on the data quality evaluation values ​​at the multiple time points; The process of determining the storage performance feature vector based on the runtime logs includes: During the execution of the AI ​​training task, a storage performance evaluation value is determined based on the running log every first preset time interval, resulting in storage feature vectors at multiple times. The storage performance feature vector is determined based on the storage performance evaluation values ​​at the multiple time points; wherein the time point corresponding to the storage performance evaluation value is the same as the time point corresponding to the data quality evaluation value.

[0008] Optionally, determining the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector includes: Determine the first residual vector of the data quality feature vector and the second residual vector of the storage performance feature vector; Based on the first residual vector and the second residual vector, the conditional entropy of the first residual vector with respect to the second residual vector is determined and used as the first causal influence coefficient; Based on the first residual vector and the second residual vector, the conditional entropy of the second residual vector with respect to the first residual vector is determined and used as the second causal influence coefficient.

[0009] Optionally, generating the abnormal diagnosis result based on the first causal influence coefficient and the second causal influence coefficient includes: The root cause of the anomaly is determined based on the first causal influence coefficient and the second causal influence coefficient; The accuracy of causal judgment is determined based on the difference and sum of the first causal influence coefficient and the second causal influence coefficient; The abnormal diagnosis result is generated based on the abnormal root cause and the accuracy of the causal judgment.

[0010] Optionally, determining the root cause of the anomaly based on the first causal influence coefficient and the second causal influence coefficient includes: When both the first causal influence coefficient and the second causal influence coefficient are greater than the corresponding first judgment threshold, the root cause of the anomaly is determined to be a storage device problem; When both the first causal influence coefficient and the second causal influence coefficient are less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data. When the first causal influence coefficient and the second causal influence coefficient are not both greater than the corresponding first judgment threshold and are not both less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data and a problem with the storage device.

[0011] Optionally, the method further includes: The training phase type, training data type, and data volume of the AI ​​training task are determined based on the training data. The first judgment threshold is determined based on the training stage type, the training data type, and the data volume.

[0012] Optionally, determining the accuracy of causal judgment based on the difference and sum of the first causal influence coefficient and the second causal influence coefficient includes: Obtain historical normal training data and normal operation logs of storage devices; The third residual vector of the data quality feature vector under normal conditions is determined based on the historical normal training data, and the fourth residual vector of the storage performance feature vector under normal conditions is determined based on the normal operation log. Background noise is determined based on the third residual vector and the fourth residual vector; The accuracy of the causal judgment is determined based on the background noise, the difference between the first causal influence coefficient and the second causal influence coefficient.

[0013] Optionally, generating the abnormal diagnosis result based on the first causal influence coefficient and the second causal influence coefficient includes: The expected load information of the storage device is determined based on the training data using a pre-trained data loading model. Based on the aforementioned operation logs, determine the historical load information of the storage device; Determine the similarity between the expected load information of the storage device and the historical load information of the storage device; Anomaly diagnosis results are generated based on the similarity, the first causal influence coefficient, and the second causal influence coefficient.

[0014] Secondly, embodiments of this application provide an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the AI ​​training anomaly diagnosis method as described in the first aspect.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing an executable program, which is executed by a processor to implement the AI ​​training anomaly diagnosis method as described in the first aspect.

[0016] This application provides an AI training anomaly diagnosis method, electronic device, and storage medium. This application calculates a first causal influence coefficient of the data quality feature vector on the storage performance feature vector and a second causal influence coefficient of the storage performance feature vector on the data quality feature vector. Based on the first causal influence coefficient and the second causal influence coefficient, anomaly diagnosis results are generated. This method can perform bidirectional causal correlation analysis on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the AI ​​training anomaly diagnosis method provided in the embodiments of this application.

[0018] Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0019] Figure 3 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "multiple" means two or more, unless otherwise explicitly specified.

[0022] This application provides an AI training anomaly diagnosis method, electronic device, and storage medium. By calculating the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector, anomaly diagnosis results are generated based on the first and second causal influence coefficients. This method can perform bidirectional causal correlation analysis on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results.

[0023] This application relates to the field of next-generation information network industry, and its classification number should be G06F11 / 07. This application also relates to the field of artificial intelligence.

[0024] The AI ​​training anomaly diagnosis method provided in this application will be described in detail below with reference to the accompanying drawings.

[0025] Please see Figure 1 , Figure 1 This is a flowchart illustrating the AI ​​training anomaly diagnosis method provided in an embodiment of this application. Figure 1 As shown, the AI ​​training anomaly diagnosis method includes steps S100 to S500.

[0026] Step S100: Obtain the training data for the AI ​​training task and the operation logs of the storage device.

[0027] Step S200: Determine the data quality feature vector based on the training data.

[0028] Optionally, the data quality feature vector includes multiple data quality evaluation values.

[0029] Optionally, during the execution of the AI ​​training task, a data quality evaluation value is determined based on the portion of training data currently read by the AI ​​model every first preset time interval, resulting in multiple data quality evaluation values ​​at different times. A data quality feature vector is then determined based on these multiple data quality evaluation values. Each data quality evaluation value is one of a first preset number of preset data quality evaluation values.

[0030] For example, the first preset quantity is 2, 3 or 4, etc.

[0031] Optionally, a pre-trained data quality evaluation value is used to determine the data quality evaluation value based on at least one of the following features of the training data: data size, data type, data file size distribution statistics, data sampling randomness parameters, and data distribution entropy.

[0032] Optionally, the data quality evaluation value can be determined using a neural network model, a deep Q-network model, or a support vector machine model, etc.

[0033] Alternatively, data size refers to the scale of the training data, measured by the number of files in the training data or the total amount of training data.

[0034] Optionally, the data file size distribution statistics include at least one of the mean, standard deviation, maximum, and minimum values ​​of the data file size distribution.

[0035] Optionally, the data sampling randomness parameter refers to the parameter used to control randomness when sampling training data.

[0036] Optionally, the data distribution entropy is a metric that measures the uncertainty of the training data or dataset. Based on the probability distribution of data file categories in the vector data statistics, the probability values ​​of each data file category are substituted into the information entropy formula to calculate the contribution of each category and sum them, ultimately obtaining the data distribution entropy.

[0037] Step S300: Determine the storage performance feature vector based on the running logs.

[0038] Optionally, during the execution of the AI ​​training task, a storage performance evaluation value is determined based on the running log every first preset time interval, resulting in storage performance evaluation values ​​at multiple times. A storage performance feature vector is then determined based on these multiple storage performance evaluation values. The time corresponding to each storage performance evaluation value is the same as the time corresponding to the data quality evaluation value. Each storage performance evaluation value is one of a second preset number of preset storage performance evaluation values.

[0039] Optionally, a pre-trained storage performance evaluation value determination model can be used to determine the storage performance evaluation value based on at least one of the following features: read frequency, write frequency, I / O latency distribution statistics, error count, and I / O queue statistics of the running log.

[0040] Optionally, the storage performance evaluation value can be determined by a neural network model, a deep Q-network model, or a support vector machine model, etc.

[0041] Read frequency refers to the number of read operations performed on a storage device per unit of time, used to measure the device's read capability. Write frequency refers to the number of write operations performed on a storage device per unit of time, used to measure its write capability. I / O latency distribution statistics refer to the latency distribution of I / O requests processed by the storage device, including average latency, maximum latency, and minimum latency. Error count is the number of errors that occur during storage device operation, used to assess system reliability. I / O queue statistics refer to the status of the storage device's I / O request queue, including queue length, used to reflect the storage device's load.

[0042] For example, the second preset quantity is 2, 3 or 4, etc.

[0043] Optionally, the second preset quantity is the same as the first preset quantity.

[0044] Optionally, the second preset quantity is different from the first preset quantity.

[0045] Preferably, the first preset quantity is 2, and the second preset quantity is 2. This reduces the amount of computation.

[0046] Step S400: Calculate the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector.

[0047] Optionally, step S400 includes steps S410 to S430.

[0048] Step S410: Calculate the first residual vector of the data quality feature vector and the second residual vector of the storage performance feature vector.

[0049] Optionally, a preset normal data quality evaluation value is subtracted from all data quality evaluation values ​​in the data quality feature vector to obtain a first residual vector. The first residual vector is used to reflect the degree to which the data quality evaluation values ​​deviate from the normal range.

[0050] Optionally, a second residual vector is obtained by subtracting a preset normal storage performance evaluation value from all storage performance evaluation values ​​in the storage performance feature vector. The second residual vector is used to reflect the degree to which the storage performance evaluation values ​​deviate from the normal range.

[0051] Step S420: Calculate the conditional entropy of the first residual vector with respect to the second residual vector based on the first residual vector and the second residual vector, and determine it as the first causal influence coefficient.

[0052] Optionally, the first causal influence coefficient is calculated multiple times based on multiple preset first lag time intervals, and the calculation result with the lowest value among the multiple first causal influence coefficients is determined as the final first causal influence coefficient.

[0053] The lower the value of the first causal influence coefficient, the greater the impact of the training data on storage quality.

[0054] For example, the formula for calculating the first causal influence coefficient is: , in, Indicates the first lag time interval. This represents the first causal influence coefficient under the first lag time interval. This represents the mutual information calculation function. Indicates time, Indicates the time corresponding to the second residual vector Quantitative state, Represents the time corresponding to the first residual vector The quantified state.

[0055] Specifically, to calculate the first causal influence coefficient, a joint state matrix containing time lag relationships is constructed based on the first residual vector and the second residual vector, and then the probability required to calculate the conditional entropy is determined based on the joint state matrix.

[0056] For example, the joint state matrix is: ,in, Indicates for The joint state matrix, This represents the value corresponding to the first time step in the first residual vector. The value corresponding to the second time step in the first residual vector. Represents the first residual vector in the first residual vector. The value corresponding to each moment, and so on. This represents the dimension of the first residual vector, which is the same as the dimension of the second residual vector. This represents the value corresponding to the first time step in the second residual vector. The second residual vector in the first The value corresponding to each time point Represents the first in the second residual vector The value corresponding to each moment, and so on.

[0057] Optionally, the type of each value in the first residual vector is determined based on multiple preset data quality deviation evaluation intervals. For example, the type of the value may include large deviation, medium deviation, and small deviation. The type of each value in the second residual vector is also determined based on multiple preset data quality deviation evaluation intervals. For example, the type of the value may include large deviation, medium deviation, and small deviation. Then, the conditional entropy is calculated based on the types of all values ​​in the first residual vector and the types of all values ​​in the second residual vector.

[0058] It's understandable that data quality itself isn't affected by storage performance, but data quality can influence storage performance. If the data quality is normal, and a storage device problem causes training failure, data quality cannot affect storage performance, and storage performance cannot affect data quality. In this case, both the second and first causal factors will be significant. Conversely, if training data issues cause training failure, the training data will also lead to a decrease in storage performance. In this case, data quality will affect storage performance, but both the first and second causal factors will be small.

[0059] The first causal influence coefficient measures the impact of the first residual vector on the future storage state, given the past storage state. If the first causal influence coefficient is very close to zero, it indicates that changes in the training data have led to changes in storage performance.

[0060] Step S430: Calculate the conditional entropy of the second residual vector with respect to the first residual vector based on the first residual vector and the second residual vector, and determine it as the second causal influence coefficient.

[0061] Optionally, based on the range of lag time intervals, the first residual vector, and the second residual vector, a second lag time interval is calculated that maximizes the conditional entropy of the second residual vector relative to the first residual vector, and the conditional entropy of the second residual vector relative to the first residual vector is determined as the second causal influence coefficient.

[0062] Optionally, the second causal influence coefficient is calculated multiple times based on multiple preset second lag time intervals, and the calculation result with the lowest value among the multiple second causal influence coefficients is determined as the final second causal influence coefficient.

[0063] The lower the value of the second causal influence coefficient, the greater the impact of storage quality on the training data.

[0064] For example, the formula for calculating the second causal influence coefficient is as follows: , in, Indicates the second lag time interval. This represents the second causal influence coefficient under the second lag time interval. Indicates the time corresponding to the second residual vector Quantitative state, Represents the time corresponding to the first residual vector The quantified state.

[0065] The method for calculating the second causal influence coefficient is the same as the method for calculating the first causal influence coefficient.

[0066] The second causal influence coefficient is used to measure whether "abnormal storage performance led to a decline in the quality of subsequent training data". If the second causal influence coefficient is very close to zero, it indicates that the change in storage performance led to the change in training data.

[0067] Step S500: Generate an abnormal diagnosis result based on the first causal influence coefficient and the second causal influence coefficient.

[0068] Optionally, step S500 includes steps S510 to S530.

[0069] Step S510: Determine the root cause of the anomaly based on the first causal influence coefficient and the second causal influence coefficient.

[0070] Optionally, step S510 includes steps S511 to S513.

[0071] Step S511: When both the first causal influence coefficient and the second causal influence coefficient are greater than the corresponding first judgment threshold, the root cause of the anomaly is determined to be a storage device problem.

[0072] Optionally, the first judgment threshold corresponding to the first causal influence coefficient may be the same as or different from the first judgment threshold corresponding to the second causal influence coefficient.

[0073] Step S512: When both the first causal influence coefficient and the second causal influence coefficient are less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data.

[0074] Optionally, the second preset judgment threshold corresponding to the first causal influence coefficient may be the same as or different from the second preset judgment threshold corresponding to the second causal influence coefficient.

[0075] Step S513: When the first causal influence coefficient and the second causal influence coefficient are not both greater than the corresponding first judgment threshold and are not both less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data and a problem with the storage device.

[0076] Optionally, the first judgment threshold is preset.

[0077] Optionally, the first judgment threshold is calculated in real time, and the method further includes steps S514 to S515.

[0078] Step S514: Determine the training stage type, training data type, and data volume of the AI ​​training task based on the training data.

[0079] Optionally, the training phase types include an initialization phase, a main training phase, a fine-tuning phase, and a validation phase.

[0080] Optionally, the training data types include image data, text data, audio data, tabular data, and video data, etc.

[0081] Step S515: Determine the first judgment threshold based on the training phase type, training data type, and data volume.

[0082] Optionally, based on historical screening data, a set of recommended first judgment thresholds are pre-calculated for each combination of training stage type, training data type, and data volume, and a threshold table is generated. Then, the first judgment threshold to be used is calculated using an interpolation method based on the current training stage type, training data type, data volume, and threshold table.

[0083] In this way, the first judgment threshold can be adjusted to adapt to the AI ​​training scenario, thereby improving the accuracy of the judgment.

[0084] Step S520: Determine the accuracy of causal judgment based on the difference and sum of the first causal influence coefficient and the second causal influence coefficient.

[0085] Optionally, the formula for calculating the accuracy of causal judgment is: ,in, Indicates the accuracy of causal judgments. This represents the difference between the first causal influence coefficient and the second causal influence coefficient. This represents the sum of the first causal influence coefficient and the second causal influence coefficient. This represents a preset denominator coefficient that is greater than 0 and less than 1.

[0086] Optionally, step S520 includes steps S521 to S524.

[0087] Step S521: Obtain historical normal training data and normal operation logs of storage devices.

[0088] Step S522: Calculate the third residual vector of the data quality feature vector under normal conditions based on historical normal training data, and calculate the fourth residual vector of the storage performance feature vector under normal conditions based on normal operation logs.

[0089] The method for calculating the third residual vector is the same as the method for calculating the first residual vector, and the method for calculating the fourth residual vector is the same as the method for calculating the second residual vector.

[0090] Step S523: Calculate the background noise based on the third residual vector and the fourth residual vector.

[0091] Optionally, a sliding window method is used to slide the first preset time window multiple times, and the first variance and the second variance of the values ​​of the third residual vector and the fourth residual vector within the first preset time window are calculated, and the background noise is calculated based on the first variance and the second variance.

[0092] For example, the formula for calculating background noise is: , in, Indicates background noise. This represents the first variance corresponding to the third residual vector. This represents the second variance of the fourth residual vector.

[0093] Step S524: Determine the accuracy of causal judgment based on the background noise, the difference and sum of the first causal influence coefficient and the second causal influence coefficient.

[0094] Optionally, the accuracy of causal judgment can be determined based on background noise, the difference between the first causal influence coefficient and the second causal influence coefficient.

[0095] Optionally, the formula for calculating the accuracy of causal judgment is: ,in, The base of the natural logarithm. This represents the preset noise calculation coefficient. This represents a preset denominator coefficient that is greater than 0 and less than 1. It is a positive number. The value of makes Greater than 0.

[0096] For example, The values ​​are 0.01, 0.02, or 0.03, etc.

[0097] By using the methods described above, we can reduce the interference of background noise and improve the accuracy of causal judgments.

[0098] Step S530: Generate abnormal diagnosis results based on the accuracy of abnormal root cause and causal judgment.

[0099] Optionally, the abnormal diagnostic results include the accuracy of the abnormal root cause and causal judgment.

[0100] To further improve the accuracy of abnormal diagnostic results, especially the accuracy of abnormal root causes, step S500 may optionally include steps S540 to S570.

[0101] Step S540: Use a pre-trained data loading model to determine the expected load information of the storage device based on the training data.

[0102] The expected load information of the storage device includes storage characteristics that the storage device may exhibit, determined based on training data. These storage characteristics include at least one of the following: read frequency, write frequency, I / O latency distribution statistics, error count, and I / O queue statistics.

[0103] Step S550: Determine the historical load information of the storage device based on the operation log.

[0104] Optionally, the historical load information of the storage device includes at least one feature selected from read frequency, write frequency, I / O latency distribution statistics, error count, and I / O queue statistics. The feature types of the storage features included in the historical load information of the storage device are the same as the feature types of the storage features included in the expected load information of the storage device.

[0105] Step S560: Calculate the similarity between the expected load information of the storage device and the historical load information of the storage device.

[0106] Optionally, a storage device expected load feature vector is determined based on the storage device expected load information, and a storage device historical load feature vector is determined based on the storage device historical load information. The similarity between the storage device expected load feature vector and the storage device historical load feature vector is then calculated. The similarity can be cosine similarity, Euclidean distance, or Manhattan distance, etc.

[0107] Step S570: Generate anomaly diagnosis results based on similarity, the first causal influence coefficient, and the second causal influence coefficient.

[0108] Optionally, when the similarity is greater than the corresponding preset similarity threshold, and the causal judgment accuracy is determined to be greater than the first preset accuracy threshold using the method described above, the abnormal root cause is determined using the method described above, and the causal judgment accuracy is determined to be high.

[0109] Optionally, when the similarity is not greater than the corresponding preset similarity threshold, the accuracy of the causal judgment is determined to be medium.

[0110] The above methods can further improve the accuracy of abnormal diagnostic results.

[0111] In summary, the AI ​​training anomaly diagnosis method provided in this application has the following advantages: 1. By calculating the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector, and generating anomaly diagnosis results based on the first and second causal influence coefficients, bidirectional causal correlation analysis can be performed on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results.

[0112] 2. By calculating the first judgment threshold in real time, the first judgment threshold can be adjusted to adapt to the AI ​​training scenario, thereby improving the accuracy of the judgment.

[0113] 3. By calculating background noise, the interference of background noise can be reduced, and the accuracy of the root cause of the anomaly can be improved.

[0114] 4. By calculating the similarity between the expected load information of the storage device and the historical load information of the storage device, the accuracy of anomaly diagnosis results can be further improved.

[0115] Optionally, the above method is used to eliminate anomalies during AI model training to train memory type recognition model, first EEG signal data recognition model, second EEG signal data recognition model and third EEG signal data recognition model, etc., and then these models are applied to the field of brain-computer interface.

[0116] Optionally, the method further includes steps S610 to S640.

[0117] Step S610: Acquire the user's first EEG signal data and cerebral blood oxygenation monitoring data.

[0118] Optionally, the first EEG signal data is a multi-channel time-domain signal, acquired by electrode channels located on multiple brain regions of the user. Each brain region is also an acquisition area.

[0119] Optionally, the user's initial brainwave signal data can be acquired via a brain-computer interface.

[0120] Optionally, the brain oxygenation monitoring data is a multi-channel time-domain signal, acquired by brain oxygenation monitoring located in multiple brain regions of the user.

[0121] Optionally, brain oxygenation monitoring data of the user can be obtained through a brain-computer interface.

[0122] Optionally, the user's cerebral oxygenation monitoring data can be obtained through a non-invasive cerebral oxygenation monitoring device.

[0123] For example, multiple brain regions include the frontal lobe, parietal lobe, temporal lobe, and occipital lobe of the left brain, and the frontal lobe, parietal lobe, temporal lobe, and occipital lobe of the right brain.

[0124] Step S620: Use a pre-trained first EEG signal data recognition model to identify the user's cognitive state based on the first EEG signal data.

[0125] Optionally, the first EEG signal data is preprocessed first, and then the preprocessed first EEG signal data is input into the first EEG signal data recognition model to obtain the user's cognitive state.

[0126] All parameters used in the calculations in this application have been normalized. For example, the Min-Max normalization algorithm is used to normalize all parameters used in the calculations.

[0127] For example, the independent component analysis method is used to separate the EEG signal and artifact components in the first EEG signal data to obtain the reconstructed EEG signal. Then, a bandpass filter is used to filter the EEG signal based on a preset EEG signal frequency band to obtain the EEG signal related to the cognitive state. Finally, the filtered signal is normalized to obtain the standardized EEG signal.

[0128] Optionally, the first EEG signal data recognition model is an artificial intelligence model such as a neural network model, a Transformer architecture model, or a generative adversarial network model.

[0129] Optionally, the cognitive state can be a focused state, a fatigued state, or a relaxed state.

[0130] Step S630: When the cognitive state is focused, identify the user's memory type based on the first EEG signal data and cerebral blood oxygenation monitoring data.

[0131] Among them, the memory types are episodic memory, semantic memory, or procedural memory.

[0132] Episodic memory refers to the memory of events that one has personally experienced, typically including specific information such as time, place, people, and plot. The hippocampus is an important brain region for the formation and storage of episodic memories.

[0133] Semantic memory refers to the memory of general knowledge and concepts, including facts, rules, and principles. This type of memory is independent of specific time and place and is an abstract form of knowledge-based memory. Semantic memory is related to the temporal and parietal lobes of the brain.

[0134] Procedural memory refers to the memory of how to perform skills or actions, often involving unconscious, habitual behaviors. This type of memory is associated with behaviors and motor skills, such as riding a bicycle or playing the piano. Procedural memory is related to the motor areas of the brain.

[0135] Optionally, step S630 includes steps S631 to S633.

[0136] Step S631: When the cognitive state is focused, extract the first feature vector of the first EEG signal data.

[0137] The first feature vector includes the variation characteristics and acquisition area of ​​the EEG signal in each channel of the first EEG signal data.

[0138] Optionally, the variation characteristics include time-domain characteristics, frequency-domain characteristics, and spatial-domain characteristics.

[0139] Optionally, the temporal, frequency, and spatial features of each channel of the first EEG signal data are extracted, and the region identifier is obtained and concatenated into a first feature vector.

[0140] Optionally, the temporal features include the mean, variance, skewness, and kurtosis of the EEG signal for each channel.

[0141] Optionally, frequency domain features include the spectrum and the relative power spectral density of theta, alpha, and beta waves. The relative power spectral density reflects the energy proportion of different brainwave rhythms.

[0142] Optionally, spatial features include the Pearson correlation coefficient of EEG signals from two channels in adjacent acquisition regions. The Pearson correlation coefficient is used to reflect the synergistic effect of neural activity in different brain regions.

[0143] Step S632: Extract the second feature vector of cerebral blood oxygenation monitoring data.

[0144] The second feature vector includes the variation characteristics and acquisition area of ​​the cerebral oxygenation signal in each channel of the cerebral oxygenation monitoring data.

[0145] Optionally, the variation characteristics include time-domain characteristics, frequency-domain characteristics, and spatial-domain characteristics.

[0146] The first eigenvector and the second eigenvector have the same dimension.

[0147] Step S633: Identify the user's memory type based on the first feature vector and the second feature vector.

[0148] Optionally, a pre-trained memory type recognition model can be used to identify the user's memory type based on a first feature vector and a second feature vector.

[0149] Optionally, the memory type recognition model can be an artificial intelligence model such as a neural network model, a Transformer architecture model, or a generative adversarial network model.

[0150] Optionally, the first feature vector and the second feature vector are fused across modalities to obtain a fused feature vector, and the user's memory type is identified based on the fused feature vector.

[0151] Optionally, the user's memory type can be obtained by fusing the memory type recognition model with the feature vector input.

[0152] Optionally, the first feature vector and the second feature vector are fused across modes to obtain a fused feature vector, including steps S6331 to S6333.

[0153] Step S6331: Calculate the information entropy of the first eigenvector, the information entropy of the second eigenvector, and the mutual information between the first eigenvector and the second eigenvector.

[0154] Mutual information is used to reflect the degree of correlation between the first eigenvector and the second eigenvector.

[0155] Optionally, the kernel density estimation (KDE) algorithm is used to calculate the information entropy of the first eigenvector, the information entropy of the second eigenvector, and the mutual information between the first and second eigenvectors.

[0156] For example, the formula for calculating the information entropy of the first feature vector is: , in, The information entropy of the first feature vector is represented by... Let represent the probability density function of the first eigenvector.

[0157] Optionally, a kernel density estimation algorithm is used to determine the probability density function of the first eigenvector. The formula for calculating the probability density function of the first eigenvector is as follows: , in, This indicates the preset number of samples. This represents the preset kernel function. Indicates the first Each sample value.

[0158] The method for calculating the information entropy of the first eigenvector can be similarly derived from the method for calculating the information entropy of the second eigenvector.

[0159] For example, the formula for calculating the mutual information between the first feature vector and the second feature vector is: , , in, This represents the mutual information between the first eigenvector and the second eigenvector. The information entropy of the second feature vector, This represents the joint information entropy of the first and second eigenvectors. This represents the joint probability density function of the first eigenvector and the second eigenvector.

[0160] Optionally, a two-dimensional kernel density estimation algorithm is used to calculate the kernel density. .

[0161] Step S6332: Calculate the modal weights of the first and second eigenvectors based on the information entropy of the first eigenvector, the information entropy of the second eigenvector, and the mutual information between the first and second eigenvectors.

[0162] For example, the formula for calculating the modal weights of the first feature vector is: , , , in, The effective information entropy of the first eigenvector is represented. The effective information entropy of the second eigenvector is represented. This represents the modal weights of the first eigenvector.

[0163] For example, the formula for calculating the modal weights of the second feature vector is: , in, This represents the modal weights of the second eigenvector.

[0164] Optionally, when or When the value is negative, it indicates that the corresponding mode has no unique information. Therefore, the value is set to 0, and only the feature vectors with positive effective information are retained to participate in the weight allocation.

[0165] Using the above methods, if the user's memory type is episodic memory, which is mainly reflected by brain electrical activity, then... High proportion Larger, and when fusion is performed, the focus is on EEG signal characteristics. If the user's memory type is procedural memory, which is mainly reflected by the brain's blood oxygen metabolism activity in the motor cortex, then... High proportion Larger, and when fused, it emphasizes cerebral blood oxygenation characteristics. If the two modalities have strong information complementarity, then... and All values ​​are kept high, with weights evenly distributed to fully integrate dual-modal information. If the signal quality of one modality is poor, its effective information entropy ratio is low, and the modality weight will automatically decrease to avoid low-quality signals interfering with the recognition results.

[0166] Step S6333: The first feature vector and the second feature vector are weighted according to the modal weights of the first feature vector and the second feature vector respectively. Then the weighted first feature vector and the second feature vector are concatenated to obtain the fused feature vector.

[0167] Step S640: Based on the memory type and the preset correspondence between memory type and storage space, store the prompt data of the current first prompt device in the corresponding storage space.

[0168] Optionally, the first prompting device includes at least one of a display screen, a sound system, and a vibration motor.

[0169] Optionally, the first notification device can be a mobile phone, computer, or smartwatch.

[0170] Optionally, the cue data may include at least one of voice, text, image, video, and haptic cues.

[0171] Optionally, the prompt data is data obtained by the first prompting device from recording audio, taking photos, and / or recording video of the user.

[0172] Optionally, step S640 includes steps S641 to S643.

[0173] Step S641: When the memory type is episodic memory, the prompt data will be stored in the first priority storage space.

[0174] Step S642: When the memory type is semantic memory, the prompt data will be stored in the second priority storage space.

[0175] Step S643: When the memory type is procedural memory, the prompt data will be stored in the third priority storage space.

[0176] Among them, the reliability of the first priority storage space is higher than that of the second priority storage space, and the reliability of the second priority storage space is higher than that of the third priority storage space.

[0177] For example, the first priority storage space is magnetic tape storage space.

[0178] For example, the second priority storage space is solid-state drive storage space.

[0179] For example, the third priority storage space is the mechanical hard disk storage space.

[0180] In this way, more complex and important informational data can be stored in more reliable storage space based on the type of memory, thus achieving adaptive scheduling of storage resources.

[0181] For example, memory training prompts for patients with cognitive impairment and video recordings of daily life memories for the elderly are often prompts corresponding to episodic memory and are stored in the first priority storage space. Prompts for students / working professionals often correspond to semantic memory and are stored in the second priority storage space.

[0182] Optionally, step S640 includes steps S644 to S647.

[0183] Step S644: Use a pre-trained second EEG signal data recognition model to identify the user's alertness level based on the first EEG signal data.

[0184] Optionally, the second EEG signal data recognition model is an artificial intelligence model such as a neural network model, a Transformer architecture model, or a generative adversarial network model.

[0185] Optionally, in the second EEG signal data recognition model, the relative power ratio of β waves to θ waves of each channel of the first EEG signal data is calculated, and a feature vector is determined based on the relative power ratio of β waves to θ waves corresponding to the EEG signals of all channels. The user's alertness level is then identified based on this feature vector.

[0186] The relative power ratio of beta waves to theta waves can reflect the level of alertness in the brain. The higher the proportion of beta waves, the higher the level of alertness, and the higher the proportion of theta waves, the lower the level of alertness.

[0187] Step S450: Select a target encoding method from a variety of preset encoding methods based on the level of alertness.

[0188] Optionally, a target encoding method can be selected from a variety of preset encoding methods based on the level of alertness and the preset correspondence between the level of alertness and the encoding method.

[0189] Optionally, based on the preset correspondence between alertness levels and encoding methods, the characteristics of the encoding methods are matched according to the alertness level. For example, when the alertness level is a preset high alertness type, the encoding method with the highest security is selected as the target encoding method. When the alertness level is a preset medium alertness type, the encoding method with the highest encoding efficiency is selected as the target encoding method. When the alertness level is a preset low alertness type, the encoding method with the highest encoding speed is selected as the target encoding method.

[0190] Step S646: Encode the prompt data using the target encoding method to obtain the encoded prompt data.

[0191] Optionally, the encoded prompt data can be stored using the method described in step S640.

[0192] The above method enables adaptive encoding of alert data based on alert level, thereby improving data security and encoding efficiency.

[0193] Step S647: Based on the memory type and the preset correspondence between memory type and storage space, store the encoded prompt data in the corresponding storage space.

[0194] Optionally, the method further includes steps S651 to S652.

[0195] Step S651: Use a pre-trained third EEG signal data recognition model to recognize the user's reading command based on the first EEG signal data.

[0196] Optionally, the third EEG signal data recognition model is an artificial intelligence model such as a neural network model, a Transformer architecture model, or a generative adversarial network model.

[0197] Step S652: When a read command is detected, read the corresponding prompt data according to the read command and send the prompt data to the second prompt device.

[0198] Optionally, the second prompting device includes at least one of a display screen, a sound system, and a vibration motor.

[0199] Optionally, the second notification device can be a mobile phone, computer, or smartwatch.

[0200] In this way, users can easily read the prompt data, thus enabling the storage device to act as a "second brain" for users and making it easier for people with cognitive impairments to live normal lives.

[0201] Optionally, the method further includes steps S661 to S664.

[0202] Step S661: Collect the user's second EEG signal data whenever the user reads the prompt data.

[0203] Step S662: Count the number of times the prompt data is read and the time interval between readings.

[0204] Step S663: Determine the user's interest rating in the prompt data based on the second EEG signal data, the number of times the prompt data was read, and the reading time interval.

[0205] Optionally, step S663 includes steps S6631 to S6635.

[0206] Step S6631: Obtain the time period during which the second prompting device will provide prompts based on the prompting data.

[0207] Step S6632: Extract the first type of feature signal of the second EEG signal data within the time period and the two second type of feature signals acquired from the prefrontal and parietal regions of the brain, respectively.

[0208] Among them, the first type of feature signal is different from the second type of feature signal.

[0209] Optionally, the first type of feature signal includes EEG signals from multiple channels in the acquisition region other than the prefrontal and parietal regions.

[0210] The synchronicity of neural activity between the prefrontal and parietal lobes of the brain can reflect a user's level of interest; the higher the synchronicity, the higher the level of interest.

[0211] Step S6633: Determine the EEG interest value based on the first type of feature signal and two second type feature signals.

[0212] Optionally, step S6633 includes steps S66331 to S66333.

[0213] Step S66331: Calculate the power of the first type of characteristic signal.

[0214] Optionally, the power of the first type of feature signal is the average power of the EEG signals from multiple channels in the acquisition region, excluding the prefrontal and parietal regions.

[0215] Step S66332: Calculate the phase lock value of the two second-type characteristic signals.

[0216] Optionally, the formula for calculating the phase-locked value of the two type-2 characteristic signals is as follows: , in, This represents the phase-locked value of the two type II characteristic signals. This represents the first type of characteristic signal. This represents the second type of characteristic signal. This indicates the total number of sampling time points within the time period during which the second prompting device provides prompts based on the prompting data. The variable representing the sampling time point of the count. express exist Phase of time, express exist The phase of time.

[0217] Step S66333: Determine the EEG interest value based on the power and phase-locked value of the first type of feature signal.

[0218] All parameters used in the calculations in this application have been normalized. For example, the Min-Max normalization algorithm is used to normalize all parameters used in the calculations.

[0219] For example, the power and phase-locked values ​​of the first type of feature signal are normalized to obtain the normalized power and phase-locked values ​​of the first type of feature signal.

[0220] Optionally, the formula for calculating the EEG interest value is: , in, Indicates EEG interest value, This represents the power of the first type of characteristic signal after normalization. Indicates the first calculated coefficient. This represents the second calculation coefficient. and All are preset values.

[0221] Step S6634: Determine the behavioral interest value based on the number of times the prompt data is read and the reading time interval.

[0222] Optionally, the formula for calculating behavioral interest values ​​is: , in, Indicates behavioral interest value. This indicates the number of times the prompt data has been read. The base of the natural logarithm. This represents the preset exponential coefficient. This represents the normalized average read time interval calculated based on all read time intervals of the prompt data.

[0223] The formula for calculating behavioral interest values ​​reflects the trend of decreasing behavioral interest values ​​over time intervals.

[0224] Step S6635: Determine the user's interest rating for the prompt data based on EEG interest values ​​and behavioral interest values.

[0225] Optionally, the EEG interest value and behavioral interest value can be input into a pre-trained interest rating model to obtain an interest rating.

[0226] Optionally, the interest scoring model can be an artificial intelligence model such as a neural network model, a Transformer architecture model, or a generative adversarial network model.

[0227] Optionally, the formula for calculating the interest score is: ,in, This indicates the level of interest.

[0228] Step S664: Compress and store the prompt data based on the interest score.

[0229] Optionally, when the cognitive state is fatigued, or the interest score is in a preset low interest score range, the highest compression ratio compression algorithm is selected from a preset number of compression algorithms to compress the prompt data, and the compressed prompt data is stored in a preset cold data storage space.

[0230] Optionally, when the interest score is within a preset medium interest score range, a compression algorithm with a compression ratio that is neither the highest nor the lowest is selected from a preset variety of compression algorithms to compress the prompt data, and the compressed prompt data is stored in a preset hot data storage space.

[0231] Optionally, when the interest score is within a preset high interest score range, the lowest compression ratio compression algorithm is selected from a preset set of compression algorithms to compress the prompt data, and the compressed prompt data is stored in a preset hot data storage space.

[0232] By compressing and storing prompt data based on the user's level of interest, the utilization rate of storage resources can be improved. In particular, by applying the brain region coordination mechanism from neuroscience, the "latent" interests of users can be quantified, compensating for the shortcomings of relying solely on explicit behavioral data to quantify interest. The resulting hierarchical compression and storage strategy dynamically allocates storage resources to high-value, high-frequency data, improving the overall efficiency and personalized service level of storage devices, making them intelligent storage managers that can "understand" and "adapt" to the cognitive needs of users.

[0233] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. For example... Figure 2 As shown, the electronic device 400 includes: one or more processors 410 and a memory 420. Figure 2 Take a processor 410 as an example.

[0234] Alternatively, the processor 410 and the memory 420 can be connected via a bus or other means. Figure 2 Taking the example of a connection between China and Israel via a bus.

[0235] Optionally, the processor 410 is used to acquire training data of the AI ​​training task and the operation log of the storage device; determine the data quality feature vector based on the training data; determine the storage performance feature vector based on the operation log; calculate the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector; and generate anomaly diagnosis results based on the first causal influence coefficient and the second causal influence coefficient.

[0236] Optionally, the memory 420, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules of the AI ​​training anomaly diagnosis method in the embodiments of this application. The processor 410 executes various functional applications and data processing of the electronic device 400 by running the non-volatile software programs, instructions, and modules stored in the memory 420, thereby implementing the AI ​​training anomaly diagnosis method of the above-described method embodiments.

[0237] Optionally, the memory 420 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device 400, etc. Furthermore, the memory 420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. Optionally, the memory 420 may include memory remotely located relative to the processor 410, and these remote memories may be connected to the controller via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0238] Optionally, one or more modules are stored in memory 420, and when executed by one or more processors 410, they execute the AI ​​training anomaly diagnosis method in any of the above method embodiments, for example, executing the method described above. Figure 1 The method steps S100 to S500.

[0239] Please refer to Figure 3 , Figure 3 This is a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. The computer-readable storage medium 500 stores program code 510, which can be called by a processor to execute the AI ​​training anomaly diagnosis method described in the above method embodiments.

[0240] The computer-readable storage medium 500 may be an electronic storage device such as flash memory, electrically erasable programmable read-only memory (EEPROM), hard disk, or read-only memory (ROM). Optionally, the computer-readable storage medium includes a non-volatile computer-readable medium. The computer-readable storage medium 500 has storage space for program code that performs any of the method steps of the AI ​​training anomaly diagnosis method described above. This program code can be read from or written to one or more computer program products. The program code may, for example, be compressed in an appropriate form.

[0241] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described AI training anomaly diagnosis method.

[0242] In summary, this application provides an AI training anomaly diagnosis method, electronic device, and storage medium. The AI ​​training anomaly diagnosis method includes: acquiring training data for an AI training task and the operation logs of a storage device; determining a data quality feature vector based on the training data; determining a storage performance feature vector based on the operation logs; calculating a first causal influence coefficient and a second causal influence coefficient of the storage performance feature vector on the data quality feature vector; and generating anomaly diagnosis results based on the first and second causal influence coefficients. This application, by calculating the first and second causal influence coefficients of the data quality feature vector on the storage performance feature vector and generating anomaly diagnosis results based on these coefficients, can perform bidirectional causal correlation analysis on training data problems and storage device problems, thereby accurately generating anomaly diagnosis results.

[0243] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for diagnosing AI training anomalies, characterized in that, include: Acquire training data and operation logs from storage devices for AI training tasks; Determine the data quality feature vector based on the training data; Determine the storage performance feature vector based on the aforementioned runtime logs; Determine the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector; Anomaly diagnosis results are generated based on the first causal influence coefficient and the second causal influence coefficient; The step of generating an abnormal diagnostic result based on the first causal influence coefficient and the second causal influence coefficient includes: The root cause of the anomaly is determined based on the first causal influence coefficient and the second causal influence coefficient; The accuracy of causal judgment is determined based on the difference and sum of the first causal influence coefficient and the second causal influence coefficient; The abnormal diagnosis result is generated based on the abnormal root cause and the accuracy of the causal judgment.

2. The AI ​​training anomaly diagnosis method according to claim 1, characterized in that, The step of determining the data quality feature vector based on the training data includes: During the execution of the AI ​​training task, a data quality evaluation value is determined based on a portion of the training data currently read by the AI ​​model at each first preset time interval, resulting in multiple data quality evaluation values ​​at different times. The data quality feature vector is determined based on the data quality evaluation values ​​at the multiple time points; The process of determining the storage performance feature vector based on the runtime logs includes: During the execution of the AI ​​training task, a storage performance evaluation value is determined based on the running log every first preset time interval, resulting in storage feature vectors at multiple times. The storage performance feature vector is determined based on the storage performance evaluation values ​​at the multiple time points; wherein the time point corresponding to the storage performance evaluation value is the same as the time point corresponding to the data quality evaluation value.

3. The AI ​​training anomaly diagnosis method according to claim 2, characterized in that, Determining the first causal influence coefficient of the data quality feature vector on the storage performance feature vector and the second causal influence coefficient of the storage performance feature vector on the data quality feature vector includes: Determine the first residual vector of the data quality feature vector and the second residual vector of the storage performance feature vector; Based on the first residual vector and the second residual vector, the conditional entropy of the first residual vector with respect to the second residual vector is determined and used as the first causal influence coefficient; Based on the first residual vector and the second residual vector, the conditional entropy of the second residual vector with respect to the first residual vector is determined and used as the second causal influence coefficient.

4. The AI ​​training anomaly diagnosis method according to claim 1, characterized in that, The determination of the root cause of the anomaly based on the first causal influence coefficient and the second causal influence coefficient includes: When both the first causal influence coefficient and the second causal influence coefficient are greater than the corresponding first judgment threshold, the root cause of the anomaly is determined to be a storage device problem; When both the first causal influence coefficient and the second causal influence coefficient are less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data. When the first causal influence coefficient and the second causal influence coefficient are not both greater than the corresponding first judgment threshold and are not both less than the corresponding second preset judgment threshold, the abnormal root cause is determined to be a problem with the training data and a problem with the storage device.

5. The AI ​​training anomaly diagnosis method according to claim 4, characterized in that, The method further includes: The training phase type, training data type, and data volume of the AI ​​training task are determined based on the training data. The first judgment threshold is determined based on the training stage type, the training data type, and the data volume.

6. The AI ​​training anomaly diagnosis method according to claim 1, characterized in that, The method of determining the accuracy of causal judgment based on the difference and sum of the first causal influence coefficient and the second causal influence coefficient includes: Obtain historical normal training data and normal operation logs of storage devices; The third residual vector of the data quality feature vector under normal conditions is determined based on the historical normal training data, and the fourth residual vector of the storage performance feature vector under normal conditions is determined based on the normal operation log. Background noise is determined based on the third residual vector and the fourth residual vector; The accuracy of the causal judgment is determined based on the background noise, the difference between the first causal influence coefficient and the second causal influence coefficient.

7. The AI ​​training anomaly diagnosis method according to claim 1, characterized in that, The step of generating an abnormal diagnostic result based on the first causal influence coefficient and the second causal influence coefficient includes: The expected load information of the storage device is determined based on the training data using a pre-trained data loading model. Based on the aforementioned operation logs, determine the historical load information of the storage device; Determine the similarity between the expected load information of the storage device and the historical load information of the storage device; Anomaly diagnosis results are generated based on the similarity, the first causal influence coefficient, and the second causal influence coefficient.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the AI ​​training anomaly diagnosis method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an executable program, which is executed by a processor to implement the AI ​​training anomaly diagnosis method as described in any one of claims 1 to 7.

Images