Abnormality monitoring and fault identification method and device

Abstract

The present invention provides a method for abnormal monitoring and fault identification, which relates to the field of digital signal processing technology and can solve the problems of low efficiency in abnormal data detection and inaccurate fault identification. The method includes: acquiring real-time data, preprocessing the real-time data, and obtaining real-time target data; using a preset monitoring data knowledge base to perform state evaluation on the real-time target data to obtain abnormal data; inputting the abnormal data into a SincNet filter for pooling operation, and performing convolution layer standardization; using a gradient descent method with momentum to perform incremental learning on the abnormal data after convolution layer standardization, and determining the fault type corresponding to the abnormal data. The present invention effectively improves the accuracy of abnormal detection, has efficient low-dimensional feature extraction capabilities and the ability to identify fault types, and can be continuously updated in combination with the input of real-time data to improve the accuracy of fault identification.

Description

Technical Field

[0001] The present invention relates to the technical field of digital signal processing, and in particular to a method and device for abnormality monitoring and fault identification. Background Art

[0002] With the rapid development of modern industrialization and informatization, the requirements for the reliability of engineering systems are gradually increasing. In many industrial and civil fields, how to detect anomalies based on system output and how to classify the fault types and levels of the system have become an urgent need. At the same time, the existing fault detection and fault identification methods are mostly at the algorithm and software level. Moreover, in the actual system environment, the processing resources used for anomaly detection and fault identification are often limited by many conditions such as space, integration, and application scenarios. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] In view of the above-mentioned deficiencies, the main purpose of the present invention is to provide a method and device for abnormality monitoring and fault identification.

[0005] (II) Technical solution

[0006] In order to achieve the above-mentioned purpose, the first aspect of the present invention provides a method for abnormal monitoring and fault identification, which includes: acquiring real-time data, preprocessing the real-time data, and obtaining real-time target data; using a preset monitoring data knowledge base to perform state evaluation on the real-time target data to obtain abnormal data; inputting the abnormal data into a SincNet filter for pooling operation, and performing convolution layer standardization; using a gradient descent method with momentum to perform incremental learning on the abnormal data after convolution layer standardization, and determining the fault type corresponding to the abnormal data.

[0007] In the above scheme, the monitoring data knowledge base is preset in the following manner: obtaining raw data, preprocessing the raw data to obtain target data; performing density clustering-based learning on the target data to obtain learning results; and creating a monitoring data knowledge base based on the learning results.

[0008] In the above scheme, the real-time data includes continuous data and discrete data, and the real-time data is preprocessed, including: using the Z-score normalization algorithm to perform normalization preprocessing on the continuous data; using the longest common subsequence model to make the discrete data continuous, and then using the Z-score normalization algorithm to perform normalization preprocessing.

[0009] In the above scheme, the target data is learned based on density clustering, including: calculating the local density of the target data i and the repulsion distance δ i, where the target data i corresponds to multiple data clusters; the local density of the target data i With the repulsion distance δ i The clustering moments are obtained by multiplication, and a decision graph of the target data is constructed based on the clustering moments.

[0010] In the above scheme, the method also includes: after acquiring the real-time target data, continuously and step by step iteratively training the real-time target data and the target data, and the monitoring data knowledge base is composed of all data clusters obtained through training, wherein the updating of the monitoring data knowledge base is completed based on adjusting the hyperparameters of each data cluster, and the hyperparameters include a cutoff distance and a clustering threshold, the cutoff distance is used to adjust the size range of each data cluster, and the clustering moment threshold is used to adjust the number of each data cluster.

[0011] In the above scheme, a preset monitoring data knowledge base is used to perform status evaluation on real-time target data, including: in fast mode, if the real-time target data meets the condition that it cannot correspond to any data cluster, the real-time target data is determined to be abnormal data; in distance mode, the exclusion distance of the real-time target data is calculated, and when the exclusion distance exceeds a preset threshold, the real-time target data is determined to be abnormal data; in distance mode, if the real-time target data meets the condition of corresponding to any data cluster, the exclusion distance is non-zero, and the real-time target data is determined to be abnormal data.

[0012] In the above scheme, the abnormal data is input into the SincNet filter for pooling operation and convolution layer standardization, including: performing SincNet filtering on the abnormal data, wherein the parameters of the SincNet filtering are SincNet network parameters that meet the incremental learning training accuracy standards; pooling the abnormal data processed by the SincNet filtering, and performing convolution layer standardization on the pooled abnormal data.

[0013] In the above scheme, the gradient descent method with momentum is used to perform incremental learning on the abnormal data after normalization of the convolution layer to determine the fault type corresponding to the abnormal data, including: activating the standardized abnormal data with leakage; randomly discarding the activation results, and performing standard convolutional neural network inference on the abnormal data after random discarding, wherein the standard convolutional neural network parameters used are SincNet network parameters that meet the incremental learning training accuracy standards; the abnormal data after standard convolutional neural network inference is processed by a normalized exponential function to obtain a Softmax result, and the fault type corresponding to the abnormal data is determined.

[0014] In a second aspect of the present invention, an integrated chip device for abnormality monitoring and fault identification is provided, comprising: an input-output interface module, a system bus module, a system microcontroller module, a storage control module, an on-chip storage module, and a visual processor module; the input-output interface module is used to obtain real-time data and related network parameters; the system bus module is used to allocate and control the timing logic of the input-output interface module, the system microcontroller module, the storage control module, and the visual processor module; the system microcontroller module includes a system microcontroller memory and a RISC-V instruction set processor, which is used to act as a control core and realize complex irregular serial operations such as calculating the target position; the storage control module is used to store data obtained from outside the integrated chip device; the on-chip storage module includes a plurality of memories arranged in parallel, which are used to store all data during the execution of the device; the visual processor module includes a plurality of processing unit arrays, and the visual processor module is used to execute the abnormality monitoring and fault identification method in the above scheme.

[0015] In the above scheme, each processing unit array includes: a local register; an arithmetic logic unit; and a processing unit, which is used to execute the abnormality monitoring and fault identification method in the above scheme.

[0016] (III) Beneficial effects

[0017] The technical solution of the embodiment of the present invention has at least the following beneficial effects:

[0018] (1) Anomaly detection based on density clustering learning can effectively improve the accuracy of anomaly detection in application scenarios where the labeled data is far less than the unlabeled data;

[0019] (2) The SincNet filter is highly suitable for processing one-dimensional signals collected in time series, such as system health management monitoring signals. The fault identification model based on SincNet deep incremental learning has efficient low-dimensional feature extraction capabilities, which effectively improves the model's ability to identify fault types. In combination with real-time data input, the model is continuously updated to improve the accuracy of fault identification.

[0020] (3) The anomaly detection and fault identification device can efficiently implement the anomaly detection and fault identification method, which can effectively improve the integration of the processing system and enable it to have a wider and more flexible application scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of an abnormality monitoring and fault identification method according to an embodiment of the present invention is schematically shown;

[0022] Figure 2 The flowchart of the abnormal monitoring process based on density clustering learning in an embodiment of the present invention is schematically shown;

[0023] Figure 3 The flowchart of the fault identification process based on SincNet deep incremental learning in an embodiment of the present invention is schematically shown;

[0024] Figure 4 The structure diagram of the abnormality monitoring and fault identification device according to the embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0026] Figure 1 The flowchart of the abnormality monitoring and fault identification method according to the embodiment of the present invention is schematically shown.

[0027] Please refer to Figure 1 The abnormality monitoring and fault identification method of the embodiment of the present invention specifically includes operations S110 to S140.

[0028] In operation S110, real-time data is acquired and preprocessed to obtain real-time target data.

[0029] In operation S120, a preset monitoring data knowledge base is used to perform a status assessment on the real-time target data to obtain abnormal data.

[0030] In operation S130, the abnormal data is input into the SincNet filter for pooling operation and convolution layer normalization.

[0031] In operation S140 , incremental learning is performed on the abnormal data after the convolution layer normalization using a gradient descent method with momentum to determine the fault type corresponding to the abnormal data.

[0032] Through the embodiments of the present invention, the accuracy of anomaly detection is effectively improved, and the ability to extract low-dimensional features and identify fault types is provided. In addition, the system can be continuously updated in combination with the input of real-time data, thereby improving the accuracy of fault identification.

[0033] According to the following Figure 2 and Figure 3 The abnormality monitoring and fault identification method of the embodiment of the present invention is further described in detail.

[0034] Figure 2 The flowchart of the abnormal monitoring process based on density clustering learning according to an embodiment of the present invention is schematically shown.

[0035] Please refer to Figure 2 The abnormal monitoring process based on density clustering learning in the embodiment of the present invention specifically includes S210~S250.

[0036] In operation S210, original data is acquired and preprocessed to obtain target data.

[0037] In operation S220, density clustering-based learning is performed on the target data to obtain a learning result.

[0038] In operation S230 , a monitoring data knowledge base is created based on the learning result.

[0039] In operation S240, real-time data is acquired and preprocessed to obtain real-time target data.

[0040] In operation S250, a preset monitoring data knowledge base is used to perform a status assessment on the real-time target data to obtain abnormal data.

[0041] The above operations S210 to S230 are the process of presetting the monitoring data knowledge base, which will be further described in detail below.

[0042] Obtain raw data from the system to be monitored. For example, the raw data may include life characteristic information, state characteristic information, and fault characteristic information of the system to be monitored. Combined with the difficulty of obtaining raw data in actual work, it can be considered that the amount of data in normal state is much larger than the amount of data in fault state, and the amount of unlabeled data is much larger than the amount of labeled data.

[0043] Next, the original data is obtained and preprocessed to obtain the target data.

[0044] In an embodiment of the present invention, data such as raw data needs to be normalized and preprocessed, wherein the raw data includes a plurality of continuous data and / or discrete data. For example, a Z-score normalization algorithm is used to perform normalization and preprocessing on continuous data; the longest common subsequence model is used to make discrete data continuous, and then the Z-score normalization algorithm is used to perform normalization and preprocessing. The Z-score normalization algorithm includes calculating the mean and standard deviation of a plurality of data; subtracting the mean from the value of each data and dividing it by the standard deviation to generate normalized data.

[0045] In an embodiment of the present invention, density-based clustering learning is performed on target data, including: calculating the local density of target data i and the repulsion distance δ i , where the target data i corresponds to multiple data clusters; the local density of the target data With the repulsion distance δ i The clustering moments are obtained by multiplication, and a decision graph of the target data is constructed based on the clustering moments.

[0046] Specifically, the target data includes multiple,density-based clustering learning. For each target data i, the local density is calculated and the repulsion distance δ i To determine the cluster center, the clustering moment is calculated, which is the product of the local density and the repulsion distance ×δ i The clustering moments of all target data constitute the decision graph of target data i, and the clustering moment is greater than the clustering moment threshold The data point is the cluster center.

[0047] Among them, the local density of the target data Defined as the distance from it is less than d c The total number of points, that is:

[0048]

[0049] Where: χ is the indicator function: when d ij -d c When <0, the indicator function takes the value of 1, otherwise it takes the value of 0; d c is the cutoff distance; d ij is the distance between target data i and target data j.

[0050] The repulsion distance δ of target data i i It is defined as the minimum distance from the sample point to the higher density point, that is:

[0051]

[0052] represents another target data j, for which the local density of the target data j is greater than the local density of the target data i;

[0053] For the target data i with the largest density, its repulsion distance can be defined as:

[0054]

[0055] In the embodiment of the present invention, a monitoring data knowledge base is created based on the learning results of density clustering. It should be noted that the monitoring data knowledge base is composed of all data clusters obtained by continuously and step-by-step iterative training of real-time target data and target data.

[0056] Specifically, the normalized target data is obtained, and the corresponding clustering model is generated through density-based clustering learning. All data clusters are obtained by continuous iterative training based on the acquired real-time target data and target data, so as to continuously improve the generated clustering model and obtain a monitoring data knowledge base.

[0057] Among them, the monitoring data knowledge base is composed of all clusters obtained through training, and its composition generally includes information such as maximum value, minimum value, 1% quantile value, 99% quantile value, mean value, variance, etc. that characterizes the distribution characteristics of cluster data.

[0058] It should be noted that the density-based clustering model has the advantages of strong adaptability, high discrimination accuracy, and the ability to clean dirty data and filter abnormal data.

[0059] In an embodiment of the present invention, the updating of the monitoring data knowledge base is completed by adjusting the hyperparameters of each data cluster, and the hyperparameters include a cutoff distance and a clustering threshold. The cutoff distance is used to adjust the size range of each data cluster, and the clustering threshold is used to adjust the number of each data cluster.

[0060] It should be noted that, for the determination of the cutoff distance parameter and the clustering moment threshold parameter, the importance index of the cutoff distance parameter and the importance index of the clustering moment threshold parameter are different according to the different monitoring contents. For example, the cutoff distance parameter and the clustering moment threshold parameter can generally be screened by domain expert knowledge, and a quantitative measurement can be given according to the importance index of the parameter. In the absence of domain experts, a machine learning method, such as ROC curve, can be used to determine the parameters.

[0061] The above operations S240 to S250 are the process of obtaining abnormal data, which will be further described in detail below.

[0062] In an embodiment of the present invention, real-time data is acquired and preprocessed to obtain real-time target data.

[0063] It should be noted that, in the embodiments of the present invention, data such as real-time data is subjected to normalization preprocessing, wherein the real-time data includes continuous data and discrete data. For example, a Z-score normalization algorithm is used to perform normalization preprocessing on continuous data; the longest common subsequence model is used to make discrete data continuous, and then the Z-score normalization algorithm is used to perform normalization preprocessing.

[0064] In an embodiment of the present invention, a preset monitoring data knowledge base is used to perform status evaluation on real-time target data to obtain abnormal data.

[0065] Specifically, a preset monitoring data knowledge base is used to perform a status assessment on real-time target data, including: in the fast mode, if the real-time target data satisfies the condition that it cannot correspond to any data cluster, the real-time target data is determined to be abnormal data; in the distance mode, the exclusion distance of the real-time target data is calculated, and when the exclusion distance exceeds a preset threshold, the real-time target data is determined to be abnormal data; in the distance mode, if the real-time target data satisfies the condition of corresponding to any data cluster, the exclusion distance is non-zero, and the real-time target data is determined to be abnormal data.

[0066] For example, in the fast mode, monitor whether the input implementation target data belongs to a certain cluster. As long as it is evaluated to belong to any cluster, it is considered to be in a normal state, otherwise it is an abnormal state. In the distance mode, the distance between the input implementation target data and the nearest cluster is calculated, and the system state is further judged based on the distance. For example, when the exclusion distance exceeds the specified threshold, the real-time target data is considered to be in an abnormal state. In the distance mode, if the real-time target data input by the monitoring belongs to a certain cluster, the distance between the parameter vector and the nearest cluster is zero; otherwise it is non-zero, and the farther it deviates from the normal state, the greater the distance value, and the real-time target data is determined to be abnormal data. Therefore, based on the corresponding evaluation results, it can be judged whether the input implementation target data is abnormal data.

[0067] Figure 3 The flowchart of the fault identification process based on SincNet deep incremental learning in an embodiment of the present invention is schematically shown;

[0068] Please refer to Figure 3 The fault identification process based on SincNet deep incremental learning in an embodiment of the present invention specifically includes S310~S330.

[0069] In operation S310 , SincNet filtering is performed on abnormal data.

[0070] In operation S320, the abnormal data processed by the SincNet filter is pooled, and the pooled abnormal data is normalized by a convolution layer.

[0071] In operation S330 , incremental learning is performed on the abnormal data after the convolution layer normalization using a gradient descent method with momentum to determine the fault type corresponding to the abnormal data.

[0072] It should be noted that the parameters of the SincNet filtering processing are the SincNet network parameters that meet the incremental learning training accuracy standards.

[0073] In an embodiment of the present invention, a gradient descent method with momentum is used to perform incremental learning on abnormal data after normalization of the convolution layer to determine the fault type corresponding to the abnormal data.

[0074] Specifically, the standardized abnormal data is activated with leakage, the activation results are randomly discarded, and the abnormal data after random discarding is subjected to standard convolutional neural network inference, wherein the standard convolutional neural network parameters used are SincNet network parameters that meet the incremental learning training accuracy standards; the abnormal data after standard convolutional neural network inference is processed by a normalized exponential function to obtain the Softmax result, and the fault type corresponding to the abnormal data is determined.

[0075] For example, the fault identification model process based on SincNet deep incremental learning is as follows: abnormal data is pooled through the SincNet filter, activated with a leaky activation function (Leaky ReLU) after normalization at the convolution layer, randomly discarded (Dropout), and sent to the standard convolutional neural network for processing. The fault type is identified and classified through the normalized exponential function (Softmax).

[0076] Among them, the fault recognition model based on SincNet deep incremental learning receives a one-dimensional signal x[n] acquired in time series, where x[n] indicates that the time series signal sequence x consists of n signals, and n is a positive integer.

[0077] In the SincNet structure, the convolution operation uses a predefined function g, where g contains only a few learnable variables q, which are defined as follows:

[0078]

[0079] Among them: y[n] is the convolution output; g[n,q] means that this is the function g acting on x[n], n is a positive integer, and the neural network can be updated through the learnable variable q.

[0080] Function g is a rectangular bandpass filter, and its frequency domain characteristics are as follows:

[0081]

[0082] The rect function is the frequency domain transformation form of the sinc function, and the expression is as follows:

[0083]

[0084] Where F represents Fourier transform.

[0085] Among them, f1 and f2 are the low cutoff frequency and the high cutoff frequency respectively, both of which are learnable. In order to smooth the truncation characteristics of function g, a window function w needs to be multiplied on the basis of function g. To be precise, the g function class is multiplied by the w function class, where the g function class consists of n g functions, denoted as g[n], and the w function class consists of n w functions, denoted as w[n]. Then for each function in g and w, it can be denoted as g[n] ×w[n] or gw[n], where n is a positive integer.

[0086] get

[0087] The window function may be a Hamming window or other window functions.

[0088] For the samples of newly added abnormal data, the gradient descent method with momentum is used for incremental learning. By setting the learning rate 𝜂 and momentum parameter 𝛼, combined with the network initial parameters 𝜃 and initial speed 𝜈, the network parameters are updated. The update formula is as follows:

[0089]

[0090] Where: L represents the loss function, m represents the newly added abnormal data;

[0091] g new Represents the original g function (i.e. g old ) The new g function obtained after updating according to the formula (i.e., gradient descent);

[0092] v new Represents the original v function (i.e. v old ) The new v function obtained after updating according to the formula (i.e. gradient descent), the v function is the initial velocity function

[0093] θ new Represents the original θ function (i.e. θ old ) The new θ function obtained after updating according to the formula (i.e. gradient descent), the θ function is the initial parameter function of the network

[0094] It represents the derivative of θ, which is a common expression in mathematical formulas.

[0095] Through the embodiments of the present invention, the SincNet filter is efficiently suitable for processing one-dimensional signals such as system health management monitoring signals collected in time series. The fault recognition model based on SincNet deep incremental learning has efficient low-dimensional feature extraction capabilities, effectively improving the model's ability to identify fault types, and combined with the input of real-time data, continuously updating the model to improve the accuracy of fault recognition.

[0096] Based on the above-mentioned method for abnormality monitoring and fault identification, the present invention also provides a device for abnormality monitoring and fault identification.

[0097] Figure 4 A schematic diagram of the structure of an abnormality monitoring and fault identification device according to an embodiment of the present invention is shown.

[0098] Please refer to Figure 4 ,The abnormal monitoring and fault identification device includes an input and output interface module, a system bus module, a system microcontroller module, a storage control module, an on-chip storage module, and a visual processor module.

[0099] Among them, the input and output interface module is used to obtain real-time data and related network parameters; the system bus module is used to allocate and control the timing logic of the input and output interface module, the system microcontroller module, the storage control module, and the visual processor module; the system microcontroller module includes the system microcontroller memory and the RISC-V instruction set processor, which is used to act as the control core and realize complex irregular serial operations such as calculating the target position; the storage control module is used to store data obtained from outside the integrated chip device; the on-chip storage module includes multiple memories arranged in parallel, which are used to store all data during the execution of the device; the visual processor module includes multiple processing unit arrays, and the visual processor module is used to execute the above-mentioned abnormality monitoring and fault identification method.

[0100] Specifically, in an embodiment of the present invention, the system microcontroller module includes a system microcontroller memory and a RISC-V instruction set processor, which is used to act as a control core and implement complex irregular serial operations such as calculating target positions. The processor is composed of a 32-bit ALU, a barrel shift register, a 32×8 multiplier, a register stack, etc. The 32-bit ALU can perform operations such as addition, AND or NOT; the multiplier can perform multiplication operations.

[0101] In an embodiment of the present invention, each processing unit array includes: a local register; an arithmetic logic unit; and a processing unit (PE) for executing the above-mentioned abnormality monitoring and fault identification method.

[0102] Specifically, in an embodiment of the present invention, the visual processor module may include an array of M×M processing units in RISC-V working mode, and each processing unit array includes one or more of the following registers, such as scalar registers, globe registers, local and ai registers. Among them, the scalar register can be accessed by all PEs; the global register can be accessed by 8 PEs; the local and ai registers can be accessed by themselves and adjacent PEs. Register resources at different levels enable each PE to cache image data or algorithm parameters, and can implement parameter reuse between PEs to reduce storage consumption.

[0103] The PE array can be reconstructed into PE chains of different lengths, and the PE in each chain accesses the registers in multiple PEs adjacent to its left and right, thereby realizing the time shift of the one-dimensional SincNet filter in the fault identification process based on SincNet deep incremental learning.

[0104] In the embodiment of the present invention, in the PE array, the distances between all target data are calculated according to the density clustering algorithm to calculate the local density , the distance between the target data x(i) and the target data x(j) is d ij indicates that, and d ii =0.d ij It can be calculated by performing l cycles of calculations by a PE unit with l parallelism; It can be calculated by performing l instruction function cycles by a PE unit with l parallelism, where l is the number of target data.

[0105] Among them, the repulsion distance δ is calculated i The way is as follows:

[0106] Set the initial repulsion distance d i =max{d 0i ,d 1i ,d 2i …d (l-1)i};

[0107] Load the local density of the target data x(i) ;

[0108] For all target data x(j), 0≤j≤l, when the condition is met: And d i >d ji When d i =d ji .

[0109] All the above operations can be completed in parallel by the PE array, and the data are loaded in the PE registers. For a PE array of size M×M, the maximum parallelism can reach O(M×M).

[0110] In an embodiment of the present invention, when performing system status evaluation, it is necessary to calculate the distance between real-time data and stored data centers, and obtain real-time target data after preprocessing the newly input real-time data. Multiple PEs can be used to simultaneously calculate the distance between the real-time target data and multiple data centers, and then perform threshold judgment to complete the system status evaluation.

[0111] In an embodiment of the present invention, the SincNet network needs to perform a one-dimensional filter convolution operation during inference, and the convolution formula of the one-dimensional filter with a length of L is as follows:

[0112]

[0113] in, and n represents the first The first and nth; L represents the total number of one-dimensional filter w function classes, L-1 means starting from 0, the last filter in the total number of L filters is numbered L-1;

[0114] x(n) represents an input abnormal signal sequence, which contains n input abnormal signals;

[0115] w( -n) represents the first -n w functions.

[0116] Where x(n) is the input abnormal data signal sequence, w(n) is the filter coefficient of SincNet; its vector representation is as follows:

[0117] Where w[k] represents the kth function in the w function class, w[n] represents that the w function class consists of n w functions in total, and the time shift of the signal sequence is:

[0118]

[0119]

[0120] When the PE array is processed, each PE loads all the filter coefficients and the abnormal data signal at the corresponding time point and then performs multiplication and accumulation operations. Since PE can obtain the data of adjacent units, the time shift of the input abnormal data signal sequence can be achieved. For an abnormal data signal sequence of (M×M) length and a filter of length, the M×M PE array only needs to perform multiplication and accumulation cycles to complete the one-dimensional filter convolution operation.

[0121] In an embodiment of the present invention, the SincNet network includes a traditional two-dimensional convolution operation, which is essentially to use the input feature map of I input channels and K two-dimensional filters to perform convolution calculations to obtain output feature maps of K output channels, where I and K are both positive integers.

[0122] Among them, the input feature maps of I input channels are loaded into the globe register; the filter parameters of K two-dimensional filters are loaded into the local register; the parameters are read between adjacent PEs to form a filter window; multiplication and accumulation operations are performed within the PE to complete the convolution; and the convolution result is stored in the ai register.

[0123] Through the embodiments of the present invention, an abnormality detection and fault identification device provided by the present invention can efficiently implement an abnormality detection and fault identification method, and can effectively improve the integration of the processing system, so that it has a wider and more flexible application scenario.

[0124] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for abnormality monitoring and fault identification, characterized in that: The method comprises: Acquire real-time data, and pre-process the real-time data to obtain real-time target data; Using a preset monitoring data knowledge base to perform a status assessment on the real-time target data to obtain abnormal data; The abnormal data is input into the SincNet filter for pooling operation and convolution layer standardization; Using a gradient descent method with momentum to perform incremental learning on the abnormal data after the convolution layer is standardized, to determine the fault type corresponding to the abnormal data; The using a preset monitoring data knowledge base to perform status evaluation on the real-time target data includes: In the fast mode, if the real-time target data meets the condition that it cannot correspond to any data cluster, the real-time target data is determined to be abnormal data; In the distance mode, the exclusion distance of the real-time target data is calculated, and when the exclusion distance exceeds a preset threshold, the real-time target data is determined to be abnormal data; In the distance mode, if the real-time target data meets the condition corresponding to any data cluster, the exclusion distance is non-zero, and the real-time target data is determined to be abnormal data; The step of performing incremental learning on the abnormal data after the convolution layer normalization by using the gradient descent method with momentum to determine the fault type corresponding to the abnormal data includes: Activate with leakage on normalized abnormal data; The activation results are randomly discarded, and the abnormal data after random discarding is subjected to standard convolutional neural network inference, wherein the standard convolutional neural network parameters used are SincNet network parameters that meet the incremental learning training accuracy standards; The abnormal data after the standard convolutional neural network inference is processed by a normalized exponential function to obtain a Softmax result, and the fault type corresponding to the abnormal data is determined; The monitoring data knowledge base is preset in the following manner: Acquire original data, and preprocess the original data to obtain target data; Performing density clustering-based learning on the target data to obtain a learning result; Based on the learning result, creating the monitoring data knowledge base; The monitoring data knowledge base is composed of all data clusters obtained by continuous and step-by-step iterative training of real-time target data and target data; The learning of the target data based on density clustering includes: Calculate the local density 𝜌 of the target data i i and the repulsion distance δ i , wherein the target data i corresponds to multiple data clusters; The local density 𝜌 of the target data i i With the repulsion distance δ i Multiplying the clustering moments to obtain a clustering moment, and constructing a decision graph of the target data based on the clustering moment; After acquiring the real-time target data, the real-time target data and the target data are continuously and step by step iteratively trained, and the monitoring data knowledge base is composed of all data clusters obtained by the training. The updating of the monitoring data knowledge base is completed by adjusting the hyperparameters of each data cluster, wherein the hyperparameters include a cutoff distance and a clustering threshold. The cutoff distance is used to adjust the size range of each of the data clusters, and the clustering moment threshold is used to adjust the number of each of the data clusters.

2. The method for abnormality monitoring and fault identification according to claim 1, characterized in that: The real-time data includes continuous data and discrete data, and the preprocessing of the real-time data includes: The continuous data is preprocessed by using a Z-score normalization algorithm; The discrete data is continuousized using the longest common subsequence model, and then normalized preprocessing is performed using the Z-score normalization algorithm.

3. The method for abnormality monitoring and fault identification according to claim 1, characterized in that: The abnormal data is input into the SincNet filter for pooling operation and convolution layer standardization, including: Performing SincNet filtering on the abnormal data, wherein the parameters of the SincNet filtering are SincNet network parameters that meet the incremental learning training accuracy standards; The abnormal data processed by the SincNet filtering is pooled, and the pooled abnormal data is subjected to convolutional layer normalization processing.

4. An integrated chip device for abnormality monitoring and fault identification, characterized in that: include: Input and output interface module, system bus module, system microcontroller module, storage control module, on-chip storage module, and visual processor module; Input and output interface module, used to obtain real-time data and related network parameters; System bus module, used to distribute and control the timing logic of input and output interface module, system microcontroller module, storage control module and visual processor module; System microcontroller module, including system microcontroller memory and RISC-V instruction set processor, used to act as a control core and implement complex irregular serial operations such as calculating target positions; A storage control module, used for storing data acquired from outside the integrated chip device; An on-chip storage module, comprising a plurality of memories arranged in parallel, for storing all data during the execution of the device; A visual processor module, comprising a plurality of processing unit arrays, wherein the visual processor module is used to execute the abnormality monitoring and fault identification method according to any one of claims 1 to 3; Among them, the visual processor module includes an array of M×M processing units in RISC-V working mode.

5. The integrated chip device according to claim 4, characterized in that: Each of the processing unit arrays comprises: local registers; Arithmetic logic unit; A processing unit, used to execute the abnormality monitoring and fault identification method as described in any one of claims 1 to 3.

Images