Equipment abnormity detection method and device

Abstract

The invention discloses an equipment abnormity detection method and device. The method comprises the following steps: acquiring first detection data and second detection data; identifying the first detection data by applying a plurality of neural networks to obtain a plurality of groups of first detection results; identifying the second detection data to obtain multiple groups of second detection results; determining weights corresponding to the plurality of neural networks according to the plurality of groups of first detection results; for each group of second detection results, determining evidence support degrees corresponding to the second detection results according to weights corresponding to the neural networks to which the second detection results belong; the evidence support degree represents the accuracy degree of the second detection result; performing evidence fusion according to the plurality of evidence support degrees by applying an evidence fusion rule to obtain evidence credibility corresponding to the second detection data; and according to the evidence credibility, determining whether the device to which the second detection data belongs is abnormal at the moment of generating the second detection data, and performing early warning when the device is abnormal. And the equipment anomaly detection accuracy is improved.

Description

Technical Field

[0001] The present application relates to fault detection technology, and in particular to a method and a device for detecting equipment abnormality. Background Art

[0002] With the rapid development of industry, modern mechanical equipment has gradually become larger, more diversified and more complex. The service environment is diverse, and harsh working conditions such as variable loads and large disturbances are common. Once a failure occurs, it will cause serious economic losses and even casualties. The composition and structure of each large-scale equipment are becoming increasingly complex. When it fails, the losses caused by equipment failure are becoming more and more serious, and may even cause catastrophic accidents.

[0003] Therefore, how to quickly and accurately detect abnormal conditions of equipment in order to issue early warnings in a timely manner and then predict and evaluate the performance of mechanical equipment is crucial to ensuring the safe service of mechanical equipment, improving production efficiency and increasing economic benefits. Summary of the invention

[0004] In an exemplary embodiment of the present application, a method and a device for detecting an equipment abnormality are provided to improve the accuracy of equipment abnormality detection.

[0005] According to a first aspect of an exemplary embodiment, a method for detecting an abnormality of a device is provided, comprising:

[0006] Acquire first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times;

[0007] Applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results; wherein the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is a detection probability of different states of the device;

[0008] Determining weights corresponding to each of the plurality of neural networks according to the plurality of groups of first detection results;

[0009] For each group of second detection results, determining the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results belong; wherein the evidence support indicates the degree of accuracy of the second detection results;

[0010] Applying the evidence fusion rule, fusing the evidence according to the support of multiple pieces of evidence, and obtaining the credibility of the evidence corresponding to the second detection data;

[0011] According to the credibility of the evidence, determine whether the target device is abnormal at the target time, and issue an early warning when it is abnormal; wherein the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0012] According to a second aspect of the exemplary implementation, there is provided a detection device, comprising a data receiving unit and a processor;

[0013] The data receiving unit is configured to perform:

[0014] Receive first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times;

[0015] The processor is configured to execute:

[0016] Applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results; wherein the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is a detection probability of different states of the device;

[0017] Determining weights corresponding to each of the plurality of neural networks according to the plurality of groups of first detection results;

[0018] For each group of second detection results, determining the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results belong; wherein the evidence support indicates the degree of accuracy of the second detection results;

[0019] Applying the evidence fusion rule, fusing the evidence according to the support of multiple pieces of evidence, and obtaining the credibility of the evidence corresponding to the second detection data;

[0020] According to the credibility of the evidence, determine whether the target device is abnormal at the target time, and issue an early warning when it is abnormal; wherein the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0021] According to a third aspect of an exemplary embodiment, there is provided a device for detecting an abnormality of an equipment, comprising:

[0022] A data acquisition unit, used to: acquire first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times;

[0023] an identification unit, used to: apply multiple neural networks to identify the first detection data respectively to obtain multiple groups of first detection results; and apply multiple neural networks to identify the second detection data respectively to obtain multiple groups of second detection results; wherein the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is a detection probability of different states of the device;

[0024] A weight determination unit, used to: determine the weights corresponding to the plurality of neural networks respectively according to the plurality of groups of first detection results;

[0025] An evidence support determination unit is used to: for each group of second detection results, determine the evidence support corresponding to the second detection result according to the weight corresponding to the neural network to which each second detection result belongs; wherein the evidence support indicates the degree of accuracy of the second detection result;

[0026] A fusion unit is used to: apply evidence fusion rules, perform evidence fusion according to multiple evidence support, and obtain the evidence credibility corresponding to the second detection data;

[0027] The early warning unit is used to determine whether the target device is abnormal at the target time according to the credibility of the evidence, and issue an early warning when it is abnormal; wherein the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0028] According to a fourth aspect of the exemplary implementation, a computer storage medium is provided, in which computer program instructions are stored. When the instructions are executed on a computer, the computer executes the method for detecting an abnormality of a device according to the first aspect.

[0029] In the embodiment of the present application, the first detection data is input into multiple neural networks, and the weights of the multiple neural networks are determined according to the multiple groups of first detection results obtained; the second detection data is input into multiple neural networks to obtain multiple groups of second detection results, and the evidence support corresponding to the second detection results is determined according to the weights corresponding to the neural networks to which the second detection results belong; the evidence support indicates the degree of accuracy of the second detection result; the evidence fusion rule is applied to fuse the evidence according to the multiple evidence supports to obtain the evidence credibility corresponding to the second detection data; according to the evidence credibility, it is determined whether the target device is abnormal at the target time, and an early warning is issued when it is abnormal. The target device here is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated. Therefore, the embodiment of the present application applies the theory of evidence fusion to anomaly detection, which improves the accuracy of detection compared with directly using the output result of a neural network as the detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0031] Figure 1 A schematic diagram exemplarily shows a state transition relationship of a device provided in an embodiment of the present application;

[0032] Figure 2 A flowchart of a method for detecting device abnormality provided by an embodiment of the present application is exemplarily shown;

[0033] Figure 3 A schematic diagram of a cold water system provided in an embodiment of the present application is shown as an example;

[0034] Figure 4 A standardized numerical distribution diagram provided by an embodiment of the present application is exemplarily shown;

[0035] Figure 5 The structure diagram of a convolutional neural network provided by an embodiment of the present application is exemplarily shown;

[0036] Figure 6 A schematic diagram of a traditional activation function provided in an embodiment of the present application is exemplarily shown;

[0037] Figure 7 An exemplary comparison diagram of a new activation function and a ReLU function provided in an embodiment of the present application is shown;

[0038] Figure 8 An exemplary comparison diagram of the convergence speed of a training process provided by an embodiment of the present application is shown;

[0039] Fig. 9 A schematic diagram of a maximum pooling result provided by an embodiment of the present application is exemplarily shown;

[0040] Fig.10 A schematic diagram of a convolutional neural network provided in an embodiment of the present application is exemplarily shown;

[0041] Fig.11 A schematic diagram exemplarily shows a knowledge-based diagnosis principle provided by an embodiment of the present application;

[0042] Fig.12 A flowchart of a fault diagnosis method based on a convolutional neural network and DS evidence theory provided in an embodiment of the present application is exemplarily shown;

[0043] Fig.13A schematic diagram of outputting a warning result provided by an embodiment of the present application is exemplified;

[0044] Fig.14 A schematic diagram of a real-time data warning process provided by an embodiment of the present application is exemplified;

[0045] Fig.15 A schematic diagram exemplarily shows a model algorithm execution process provided by an embodiment of the present application;

[0046] Fig.16 The schematic diagram of the structure of a device for detecting abnormality of equipment provided by an embodiment of the present application is exemplarily shown;

[0047] Fig.17 The structural diagram of another detection device provided in an embodiment of the present application is exemplarily shown. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0049] For ease of understanding, the terms involved in the embodiments of the present application are explained below:

[0050] (1) DS evidence theory, also known as Dempster / Shafer evidence theory, has the ability to handle uncertain information. As an uncertain reasoning method, the main characteristics of evidence theory are: it meets weaker conditions than Bayesian probability theory; it has the ability to directly express "uncertainty" and "not knowing".

[0051] (2) argmax function: function y = f(x), x 0 =argmax(f(x)) means the parameter x 0 Satisfy f(x 0 ) is the maximum value of f(x), that is, argmax(f(x)) is the variable x that makes f(x) reach its maximum value.

[0052] The composition and structure of each large-scale equipment are becoming increasingly complex. When a failure occurs, the losses caused by the equipment failure are becoming more and more serious, and may even cause catastrophic accidents. Therefore, how to quickly and accurately detect abnormal conditions of equipment in order to give early warnings in time, and then predict and evaluate the development of mechanical equipment, is crucial to ensuring the safe service of mechanical equipment, improving production efficiency and increasing economic benefits.

[0053] To this end, the embodiment of the present application provides a method for detecting device anomalies, in which, for a trained model structure (including multiple neural networks), multiple neural networks are applied to identify the first detection data, and each neural network can obtain a group of first detection results; multiple neural networks are applied to identify the second detection data, and each neural network can obtain a group of second detection results. The first detection data and the second detection data here are detection data of two devices, or detection data of the same device at different times. Since the purpose of the embodiment of the present application is to identify the state of the device to which the second detection data belongs at the time of generating the second data, the weights corresponding to each of the multiple neural networks can be determined by multiple first detection results, and the purpose of this process is to apply to the theory of evidence fusion. Then, for each group of second detection results, the evidence support corresponding to the second detection result is determined according to the weight corresponding to the neural network to which the second detection result belongs; and the evidence support indicates the degree of accuracy of the second detection result, and finally the evidence fusion rule is applied to fuse the evidence according to multiple evidence supports to obtain the evidence credibility corresponding to the second detection data; according to the evidence credibility, it is determined whether the target device (the device to which the second detection data belongs) is abnormal at the target time (the time when the second detection data is generated), and an early warning is issued when it is abnormal. This process improves detection accuracy compared to directly applying a neural network to predict the state of the device.

[0054] After introducing the design ideas of the embodiments of the present application, the following briefly introduces the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present application and are not limited. In specific implementation, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0055] Figure 1 A schematic diagram of the state transition relationship of a device provided in an embodiment of the present application, wherein a possible change process of the fault process may be from normal operation to fault operation with different driving conditions, or from fault operation with different forms to normal operation.

[0056] Exemplarily, Table 1 is a description information of various states provided in an embodiment of the present application, wherein Table 1 shows five different states, normal operation, degraded operation, hidden fault operation, faulty operation and shutdown.

[0057] Table 1 Description of each status

[0058]

[0059] pass Figure 1As shown in Table 1, the equipment is in the expected state of normal operation, but in the related technology, users cannot perceive the state of degraded operation and hidden fault operation, and can only intervene and repair when the equipment fails and shows obvious phenomena (such as abnormal noise, abnormal vibration, etc.) or stops. At this time, the equipment has been damaged and has affected the operation, seriously affecting the operation of the equipment.

[0060] To further illustrate the technical solution provided by the embodiment of the present application, this is described in detail below in conjunction with the accompanying drawings and specific implementation methods. Although the embodiment of the present application provides the method operation steps shown in the following embodiments or drawings, more or fewer operation steps may be included in the method based on routine or no creative labor. In the steps where there is no necessary causal relationship logically, the execution order of these steps is not limited to the execution order provided in the embodiment of the present application.

[0061] Combine the following Figure 1 The application scenario shown is shown in Figure 2 A flowchart of a method for detecting an abnormality of a device is shown to illustrate the technical solution provided in an embodiment of the present application.

[0062] S201: Acquire first detection data and second detection data.

[0063] S202: applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results.

[0064] Among them, the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are represented by a vector, and each element in a vector is the detection probability of different states of the device.

[0065] S203: Determine weights corresponding to the plurality of neural networks according to the plurality of groups of first detection results.

[0066] S204: For each group of second detection results, determine the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results belong.

[0067] The evidence support indicates the degree of accuracy of the second test result.

[0068] S205: Apply evidence fusion rules to fuse the evidence according to the multiple evidence supports, and obtain the evidence credibility corresponding to the second detection data.

[0069] S206: Determine whether the target device is abnormal at the target time based on the credibility of the evidence, and issue an early warning if it is abnormal.

[0070] The target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0071] In the embodiment of the present application, the first detection data is input into multiple neural networks, and the weights of the multiple neural networks are determined according to the multiple groups of first detection results obtained; the second detection data is then input into multiple neural networks to obtain multiple groups of second detection results, and the evidence support corresponding to the second detection results is determined according to the weights corresponding to the neural networks to which the second detection results belong; the evidence support indicates the degree of accuracy of the second detection result; the evidence fusion rule is applied to fuse the evidence according to the multiple evidence supports to obtain the evidence credibility corresponding to the second detection data; according to the evidence credibility, it is determined whether the target device is abnormal at the target time, and an early warning is issued when it is abnormal. The target device here is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated. Therefore, the embodiment of the present application applies the theory of evidence fusion to anomaly detection, which improves the accuracy of detection compared to directly using the output result of a neural network as the detection result.

[0072] In relation to S201, the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times.

[0073] For example, the device to which the first detection data belongs is the first device, and the device to which the second detection data belongs is the second device. Alternatively, the devices to which the first detection data and the second detection data belong both belong to the same device (referred to as the target device), the collection time of the first detection data is the first moment, and the collection time of the second detection data is the target moment. In the embodiment of the present application, what is finally determined is the state or abnormality of the device to which the second detection data belongs at the collection time of the second detection data.

[0074] Exemplarily, the collection process of the first detection data or the second detection data is described.

[0075] The detection data may be realized by a sensor, for example, it may be a status signal of a device.

[0076] Taking the large system used to control the environment of public areas such as station halls and platforms in the environmental control system of the subway system as an example, the chiller subsystem under the ventilation and air-conditioning system is mainly used to illustrate the mobile phone process of equipment detection data. The subway ventilation system is complex in design and the pipelines occupy a large space. It mainly achieves the control goal of the internal environment. For example, a total of 14 parameters of the chiller subsystem are collected.

[0077] Figure 3 A schematic diagram of a chilled water system is provided for an embodiment of the present application, wherein different equipment or systems such as a cooling tower, a cooling water pump, and a chiller are shown.

[0078] In S202, the first detection data and the second detection data are respectively used as input data of the neural network, that is, the first detection data and the second detection data are identified by the neural network.

[0079] In this example, the input layer has 14 parameters. Here, the parameters of the chiller system are extracted from the subway system database. Table 2 shows the parameters of a chiller system.

[0080] Table 2 Parameters of chiller system

[0081]

[0082]

[0083] Among them, each parameter is selected with a certain length. Since the units of the parameters are inconsistent and the numerical ranges are quite different, they are converted into dimensionless pure values, which can improve the accuracy. Standardization processing can map the original detection data to a range with a mean of 0 and a standard deviation of 1. In a specific example, it can be achieved by the following formula:

[0084]

[0085] Among them, μ is the mean of the data and σ is the standard deviation of the data.

[0086] Ten samples are randomly selected as examples and compared before and after standardization. Figure 4 A standardized numerical distribution diagram provided in the embodiment of the present application is provided by Figure 4 It can be seen that most of the values ​​are distributed between -2 and 2.

[0087] Table 3 Examples of values ​​before normalization

[0088]

[0089]

[0090] Table 4 Examples of normalized values

[0091] x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14 -1.01 -1.18 -1.06 -1.14 1.02 1.23 -1.07 -1.27 -2.42 1.69 -0.93 2.07 -0.96 2.11 -1.00 -0.85 -1.04 -1.04 1.06 1.14 -1.22 -1.17 -0.89 1.29 -1.32 0.55 -1.27 0.40 -0.99 -0.88 -0.99 -0.92 0.98 0.93 -1.07 -0.95 -0.46 0.61 -1.13 0.05 -1.06 -0.07 -0.31 -0.98 -0.94 -0.92 1.29 1.09 -0.56 -0.63 -0.24 0.61 -0.74 0.55 -0.81 0.62 -0.20 1.02 0.96 0.96 0.05 -0.42 0.36 0.68 0.85 -0.74 0.78 -0.96 0.71 -0.89 -0.12 1.02 1.13 1.11 -0.64 -0.92 1.02 1.16 0.85 -0.88 1.16 -0.71 1.23 -0.72 -0.05 1.02 1.10 1.08 -1.42 -1.30 1.43 1.29 0.85 -1.15 1.16 -0.96 1.20 -1.01 -0.03 1.02 1.08 1.08 -1.53 -1.37 1.54 1.27 0.85 -1.15 1.16 -1.21 1.17 -1.06 1.81 -1.08 -0.94 -0.95 -0.22 0.30 -0.56 -0.72 0.20 0.47 -0.74 1.06 -0.75 1.13 1.89 0.92 0.70 0.74 -0.60 -0.67 0.14 0.34 0.41 -0.74 0.59 -0.45 0.54 -0.52

[0092] Next, the structure of the neural network is explained.

[0093] In an embodiment of the present application, the network model includes multiple neural networks, the number of neural networks is N, the N neural networks are connected in parallel, the structure and parameters of each neural network are the same, and each neural network can output a set of detection results.

[0094] In a specific example, the structure of each neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a first fully connected layer, a third activation layer, a second fully connected layer and a classifier. The neural network can be a convolutional neural network.

[0095] Figure 5 A structural diagram of a convolutional neural network provided in an embodiment of the present application, wherein: Figure 5 Only the input layer, hidden layer and output layer are shown in the figure. Multiple convolutional neural network models are established to create a classifier pool.

[0096] For example, in the first convolution layer, there are 20 convolution kernels, each of which is 5×5. Multiple convolutions are calculated in parallel at this layer to generate a set of linear activation responses. Convolution obtains local information of the input by acting on the local area with a convolution kernel of a certain size. Each element of the convolution kernel acts on each position of the input. The same parameters can be used in multiple functions of a model to achieve parameter sharing, ensuring that the convolution layer only needs to learn one set of parameters, rather than learning a separate set of parameters for each position, which can significantly reduce the number of parameters.

[0097] In the first activation function layer, each linear activation response passes through a nonlinear activation function, such as a sigmoid function. In the related art, a fixed activation function is used in the neural network, and the same nonlinear transformation is applied to all samples. The weight relationship between the sample and the feature is ignored. For different samples, each feature has different importance to it. Therefore, in the embodiment of the present application, a matching activation function is automatically set for each sample according to the different characteristics of the sample, which can not only improve the training convergence speed, but also improve the overall performance of the neural network model.

[0098] Therefore, in an embodiment of the present application, an activation function with an added adaptive coefficient is proposed. The adaptive coefficient can automatically adjust the parameters of the activation function during the training process, so that the activation function is no longer fixed, and different nonlinear transformations can be assigned to different sample data, with a certain degree of adaptability.

[0099] In the embodiment of the present application, the ReLU activation function, which is most commonly used in the convolutional neural network in the related art, is improved. The formula of the activation function in the related art is as follows:

[0100]

[0101] Figure 6 It is a schematic diagram of the traditional activation function. Figure 6It can be seen that when the input is negative, the activation values ​​are all zero, which may cause the gradient to disappear during model training, thereby affecting the accuracy of the neural network model.

[0102] In the embodiment of the present application, in order to avoid the problem of negative input values, the attenuation activation problem is introduced, and an adaptive coefficient is added to adjust the attenuation value. The new activation function is optimized when x < 0. Instead of letting the input activation value be directly 0, an adaptive decay is added. The formula of the new activation function is:

[0103]

[0104] Among them, α is a learnable adaptive decay parameter.

[0105] Comparing the new activation function with the ReLU curve, when α < 1 and tends to 0, the new function approaches the ReLU function, and when α > 1, the activation value will have a smooth decay when x < 0. The new activation function is differentiable and tends to saturate for large negative inputs, while reducing deviation offset and increasing robustness.

[0106] Figure 7 A comparison diagram of a new activation function and a ReLU function provided in an embodiment of the present application. Figure 8 A comparison diagram of adaptive activation function training curves provided in an embodiment of the present application. Figure 8 This is a comparison chart of the convergence speed of the training process after using the new activation function.

[0107] In the maximum pooling layer, the stride is 2×2, and the output of this layer is further adjusted to reduce the spatial dimension of the input, further perform feature extraction, and make the convolutional features more robust

[0108] The maximum pooling layer is implemented by the following formula:

[0109]

[0110] Among them, w(a,b) represents the pooling filter, a and b correspond to the dimensions of the pooling filter, where a and b are both equal to 2. Represents the jth channel output in convolution layer 1-1. A 2×2 pooling operation is performed on a 5×5 input.

[0111] Fig. 9 A schematic diagram of a maximum pooling result provided in an embodiment of the present application.

[0112] In addition, in the second convolutional layer, the number of convolution kernels is 50 and the size is still 5×5. The specific structure can be found in the first convolutional layer and will not be repeated here.

[0113] The second activation function layer still uses the new activation function. For details, please refer to the first activation function layer, which will not be repeated here.

[0114] The second maximum pooling layer can refer to the first maximum pooling layer and will not be described here.

[0115] Exemplarily, a third activation function layer may be added to the first fully connected layer and the second fully connected layer. The third activation function layer still uses a new activation function. For details, refer to the first activation function layer, which will not be described here.

[0116] Finally, the Softmax classifier is applied, with the input being the output dimension of the previous layer and the output being the number of categories, and the predicted probability of the predicted value can be obtained.

[0117] The classification rules of the Softmax classifier are:

[0118]

[0119] Among them, x is the weighted input, y j is the Softmax activation value of the jth neural node, and k is the number of classification categories.

[0120] Fig.10 A schematic diagram of a convolutional neural network provided for an embodiment of the present application, in which an input layer, convolutional layer 1, pooling layer 2, convolutional layer 2, pooling layer 2, fully connected layer 1, fully connected layer 2 and an output layer are shown.

[0121] Exemplarily, taking 6 states as an example, Table 5 is a schematic diagram of the structural parameters of each layer of a convolutional neural network provided in an embodiment of the present application.

[0122] Table 5 Structural parameters of each layer of convolutional neural network

[0123] name Number of layers Output Dimensions Conv convolutional layer 1 (20,32,32) Activate Layer 2 (20,32,32) MaxPool pooling layer 3 (20,16,16) Conv convolutional layer 4 (50,12,12) Activate Layer 5 (50,12,12) MaxPool pooling layer 6 (50,6,6) Fully connected layer 7 (500) Activate Layer 8 (500) Fully connected layer 9 (6) Softmax Classifier 10 (6)

[0124] In addition, in step S202, the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is the detection probability of different states of the device.

[0125] In S203, according to the multiple groups of first detection results, the weights corresponding to the multiple neural networks are determined. The process can be specifically carried out through steps A1-A2:

[0126] A1: Determine a target tag for representing a real state of the first device.

[0127] The first device is the device to which the first detection data belongs. The real state of the first device at the time of collecting the first detection data can be used to determine the difference between other detection results and the real state. The real state can be high speed, medium speed or low speed operation, or the number of parameters used to evaluate the fault g1 is 1, 2, 3 or 4.

[0128] A2: For each group of first detection results, determine the similarity between the first detection results and the actual state according to the first detection results and the target label.

[0129] Exemplarily, the process of calculating the similarity can be implemented through steps A1 to A5.

[0130] A2-1: Determine a first detection result set.

[0131] Wherein, N is an integer greater than or equal to 2. In this example, there are N neural networks, that is, there are N first detection results. Each neural network corresponds to one detection result, and the first detection result set includes N groups of first detection results.

[0132] A2-2: Apply the first optimization function to determine a result vector that is most similar to the target label among the N result vectors corresponding to the N groups of first detection results.

[0133] In this example, the first optimization function is the argmax function, that is, the first optimization function is applied to find a result vector that is most similar to the target label among the N result vectors. In this example, the target label is a vector with the same dimension as the result vector.

[0134] A2-3: Apply the second optimization function to find a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results.

[0135] The second optimization function may be the same as the first optimization function or different from the first optimization function. In this example, the second optimization function is an argmax function, which finds a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results.

[0136] A2-4: Repeat the search process of applying the first optimization function and the second optimization function until all N result vectors are searched.

[0137] The search process is repeated, so that when all N result vectors are searched, the search order can be determined.

[0138] A2-5: The found order is used as the similarity between the first detection result of the N groups and the true state.

[0139] Exemplarily, the similarities between the N groups of first detection results and the true state are determined in the order in which they are found, that is, the earlier they are found, the higher the similarity.

[0140] A3 determines the weights corresponding to the plurality of neural networks according to the similarities corresponding to the plurality of groups of first detection results.

[0141] In actual situations, by multiplying the weights by the set coefficients, the weights corresponding to the multiple neural networks can be determined.

[0142] To make the weight determination process more complete, let's use a specific example:

[0143] A classifier ranking method based on joint mutual information is used for feature selection to obtain the fault diagnosis results of each neural network for each device. Through iteration, the result vectors most similar to the target label Y are continuously searched and stored in Yr. The result vector most similar to the target label Y is found in the first detection result set (prediction value set) Y'. The steps are as follows:

[0144] B1: Initialize Y r is an empty set, and is the result set most similar to the target Y.

[0145] B2: Y S1 =argmax(MI(Y i ; Y)), Y i ∈Y′, find the result vector Y that is most similar to the target Y in the current Y′ S .

[0146] B3: Y S Join Y r In the set, Y S Remove from Y'.

[0147] B4:Y S2 =argmax(JMI(Y i ; Y|Y j )), Y i ∈Y′,Y j ∈Y r , find the result vector Y that is most similar to Y in the current Y′ S .

[0148] B5: Y S Join Y r In the set, Y S Remove from Y'.

[0149] B6: Repeat steps B2 to B5 until Y′ is an empty set.

[0150] B7: Returns the final result Y r , which is the set of results most similar to the target label Y.

[0151] The formulas involved in steps B2 and B4 are as follows:

[0152] Information entropy is used to measure the information available in each feature variable:

[0153] H(x)=-∑p(x i )log(p(x i )),

[0154] Where p(x i ) is the probability mass function, which is the probability that x takes the value xi. Mutual information MI is the amount of information shared between two random variables:

[0155] MI(x;y)=MI(y;x)=H(x)-H(x|y)

[0156] Where H(x|y) is the conditional entropy, which is used to measure the amount of information remaining in x after introducing y, and is defined as:

[0157] H(x|y)=-∑∑p(x i ,y j )log(p(x i |y j ))

[0158] The joint mutual information JMI is:

[0159] JMI(x,y;z)=JMI(x;y|z)+MI(z;y);

[0160] JMI(x;y|z)=H(x|y)-H(x|y,z);

[0161] H(x|y,z)=H(x,y,z)-H(y,z);

[0162] H(x,y,z)=-∑∑∑p(x i ,y j ,z t )log(p(x i ,y j ,z t )).

[0163] In S204, the DS evidence theory is used to integrate the output of the convolutional neural network, and the final classification performance is improved through group fusion. The output results of each convolutional neural network are used as the evidence body of the evidence theory.

[0164] For each group of second detection results, the evidence support corresponding to the second detection results is determined according to the weights corresponding to the neural networks to which the second detection results belong.

[0165] The evidence support indicates the accuracy of the second test result. The relationship between the evidence support and the weight may be multiplied by a set coefficient, and the set coefficient may be related to the weight.

[0166] In S205, evidence fusion rules are applied to fuse evidence according to multiple evidence supports to obtain the evidence credibility corresponding to the second detection data. This process can be implemented through steps B1-B6.

[0167] B1: Determine the permutation and combination results of M states.

[0168] Each evidence support corresponds to a state. For example, if M is 5, A1, A2, and A3, then there are 7 possible results after permutation and combination.

[0169] B2: When the intersection of the first set, the second set and the third set is any one of the M states, multiply the evidence support of the first set, the evidence support of the second set and the evidence support of the third set to obtain the first target support.

[0170] Among them, the first set, the second set and the third set are any results of the permutation and combination results, represented by B, C and D respectively.

[0171] B3: Determine the sum of M first target support degrees as the first data.

[0172] In this example, the first target support is and is the first data:

[0173]

[0174] B4: When the intersection of the fourth set, the fifth set and the sixth set is not empty, multiply the evidence support of the fourth set, the evidence support of the fifth set and the evidence support of the sixth set to obtain the second target support.

[0175] Among them, the fourth set, the fifth set and the sixth set are any one of the permutation and combination results.

[0176] B5: Determine the sum of the M second target support degrees as the second data.

[0177] In this example, the sum of the second target support is the second data K,

[0178] B6: Determine the credibility of the evidence corresponding to the second detection data based on the first data and the second data.

[0179] Among them, the credibility of evidence corresponding to the second detection data is:

[0180]

[0181] In S206, based on the credibility of the evidence, it is determined whether the target device is abnormal at the target time, and an early warning is issued when an abnormality occurs.

[0182] The target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0183] This step can be achieved through steps C1-C3:

[0184] C1: Data for each dimension of evidence credibility.

[0185] The data of the evidence credibility is in vector form, and each data represents the probability that the target state is one of the M states.

[0186] C2: Determine the probability of the target device being abnormal at the target time based on the data in each dimension.

[0187] For example, the difference between 1 and the data in each dimension is the probability that the target cup is abnormal at the target time.

[0188] C3: If the probability of anomaly is greater than the probability threshold, it is determined that the target device is abnormal at the target time.

[0189] Among them, the probability threshold is pre-set, and Table 6 is a schematic table of decision results provided in an embodiment of the present application.

[0190] In a specific example, fusion evidence is obtained, and a fusion decision is made based on the fusion evidence. Taking a recognition framework with 2 states as an example, when the evidence of the abnormal state is not 0, an early warning is issued.

[0191] Table 6 Schematic diagram of decision results

[0192] evidence State 1 State 2 Abnormal state Θ Decision results m1(A) 1 0 0 No warning m2(A) 0 1 0 No warning m3(A) 0.7 0.3 0 No warning m4(A) 0.4 0.6 0 No warning m5(A) 0 0 1 Early Warning

[0193] Here, Θ is the recognition frame, and a subset of the frame is called a proposition A. The basic probability distribution function on the recognition frame is a 2 Θ →[0,1].

[0194] It should be noted that the number of states, parameters of the convolutional neural network, etc. in this application are only for illustration and are not intended to be a specific limitation.

[0195] In the embodiment of the present application, the knowledge-based diagnosis method extracts and uses basic knowledge variables representing system dependencies from a large amount of historical data, and makes fault diagnosis decisions in the classifier by checking the consistency between the observed behavior of the system process and the knowledge base.

[0196] Fig.11 A schematic diagram of a knowledge-based diagnosis principle provided in an embodiment of the present application. Fig.12 A flowchart of a fault diagnosis method based on convolutional neural network and DS evidence theory is provided in an embodiment of the present application.

[0197] The present application also provides an early warning system. Fig.13 This is a schematic diagram of an early warning result output provided in an embodiment of the present application. The early warning system is divided into four parts, and the functions of each part are as follows:

[0198] Warning result output: Output warning information through sound, text message, interface, etc., which belongs to the HMI interface part.

[0199] Data persistence: The user stores the device status data and algorithm model data using a time series history database and knowledge graph storage.

[0200] Real-time data processing: collect the status data of the equipment in real time, store the status data in the time series history library, and make early warning judgments in conjunction with the evidence theory model, and publish the early warning results to the outside.

[0201] Early warning algorithm core: responsible for generating and evolving algorithm models based on existing historical data, and providing evidence theory models for real-time data processing and early warning combined with real-time data.

[0202] The embodiment of the present application ensures the accuracy of fault classification and solves the problems of low fault warning accuracy and difficulty in extracting sensitive features. It simplifies the feature extraction and feature selection process of state data, which is automatically realized through convolutional neural networks and has strong self-learning and self-adaptive capabilities; it draws diagnostic conclusions through comprehensive decision-making based on DS evidence theory, and has strong robustness to uncertain information, which can realize cross-device system-level fault detection and warning.

[0203] Optionally, the core processing flow of the early warning system is respectively the early warning flow that is handled by the real-time data processing module and the algorithm flow that is handled by the early warning algorithm core.

[0204] Fig.14 A schematic diagram of a real-time data warning process provided in an embodiment of the present application.

[0205] Real-time data collection of equipment: It is the data access point of the entire system. All data in the system are collected by the collection module into the system.

[0206] Historical data storage: The collected real-time data is stored in a time series historical database in chronological order, and the saved historical data is used for model training of the algorithm process.

[0207] Data service: The data service determines the device of the data based on the collected real-time data, and pulls the matching model from the algorithm evidence library based on the device.

[0208] Data fusion decision: Calculate the equipment status based on the algorithm model and real-time data in the previous step.

[0209] Early warning service: Determine whether an early warning is needed based on the result of the calculation. If necessary, proceed to the next step; otherwise, the process terminates.

[0210] Data access: Send the determined warning results to other required systems and modules, such as the front-end display interface, App terminal, etc.

[0211] Fig.15 A schematic diagram of a model algorithm execution process provided in an embodiment of the present application.

[0212] Algorithm evolution engine: drives the entire algorithm training and operation at a fixed cycle, usually in weeks, to ensure that the algorithm model can be trained to produce a suitable model based on the latest device status.

[0213] Historical data extraction: Extract data used to build and train models at a fixed period, usually extracting data from the most recent month.

[0214] Data quality: Adjust the quality of data to ensure that the data can meet the needs of model training.

[0215] Convolutional Neural Network: Automatically extract features and obtain preliminary fault diagnosis results.

[0216] Evidence theory model construction: Based on the previous information, a model is constructed and stored in the algorithm evidence library with the device as the index for use in real-time data processes.

[0217] In the embodiment of the present application, the detection data (operation status data) of different devices are input into the convolutional neural network for automatic feature extraction, and the output results are obtained. The results of the outputs of multiple convolutional neural networks are used as evidence of evidence theory, and are fused based on the fusion rules of evidence theory to obtain fusion evidence and make fusion decisions to obtain the final system diagnosis results. According to the operation status data of the equipment, the degraded operation and hidden fault operation status are diagnosed, and early warnings are issued in time to remind users to perform maintenance and repairs, reduce operation and maintenance costs, and prevent the impact of equipment failure and downtime.

[0218] like Fig.16As shown, based on the same inventive concept, an embodiment of the present application provides a device for detecting an abnormality of an equipment, including a data acquisition unit 161, an identification unit 162, a weight determination unit 163, an evidence support determination unit 164, a fusion unit 165 and an early warning unit 166.

[0219] The data acquisition unit 161 is used to: acquire first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times;

[0220] The identification unit 162 is used to: apply multiple neural networks to identify the first detection data respectively to obtain multiple groups of first detection results; and apply multiple neural networks to identify the second detection data respectively to obtain multiple groups of second detection results; wherein the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is a detection probability of different states of the device;

[0221] The weight determination unit 163 is used to determine the weights corresponding to the plurality of neural networks according to the plurality of groups of first detection results;

[0222] The evidence support determination unit 164 is used to: for each group of second detection results, determine the evidence support corresponding to the second detection result according to the weight corresponding to the neural network to which each second detection result belongs; wherein the evidence support indicates the degree of accuracy of the second detection result;

[0223] A fusion unit 165 is used to: apply an evidence fusion rule to fuse the evidence according to the multiple evidence supports to obtain the evidence credibility corresponding to the second detection data;

[0224] The warning unit 166 is used to determine whether the target device is abnormal at the target time according to the credibility of the evidence, and issue a warning when it is abnormal; wherein the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0225] In an optional implementation manner, the weight determination unit 163 is specifically configured to:

[0226] Determine a target tag for representing a real state of a first device; wherein the first device is a device to which the first detection data belongs;

[0227] For each group of first detection results, determine the similarity between the first detection result and the actual state according to the first detection result and the target label;

[0228] According to the similarities corresponding to the multiple groups of first detection results, the weights corresponding to the multiple neural networks are determined.

[0229] In an optional implementation manner, the weight determination unit 163 is specifically configured to:

[0230] Determine a first detection result set; wherein the first detection result set includes N groups of first detection results; N is an integer greater than or equal to 2;

[0231] Applying the first optimization function, determining a result vector that is most similar to the target label among the N result vectors corresponding to the N groups of first detection results;

[0232] Apply the second optimization function to find a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results;

[0233] Repeat the search process of applying the first optimization function and the second optimization function until all N result vectors are searched;

[0234] The found order is used as the similarity between the first detection result of the N groups and the actual state.

[0235] In an optional implementation, the fusion unit 165 is specifically configured to:

[0236] Determine the permutation and combination results of M states, where each evidence support corresponds to one state;

[0237] When the intersection of the first set, the second set and the third set is any one of the M states, the evidence support of the first set, the evidence support of the second set and the evidence support of the third set are multiplied to obtain the first target support; wherein the first set, the second set and the third set are any one of the permutation and combination results;

[0238] Determine the sum of M first target support degrees as first data;

[0239] When the intersection of the fourth set, the fifth set and the sixth set is not empty, the evidence support of the fourth set, the evidence support of the fifth set and the evidence support of the sixth set are multiplied to obtain the second target support; wherein the fourth set, the fifth set and the sixth set are any one of the permutation and combination results;

[0240] Determine the sum of the M second target support degrees as the second data;

[0241] The credibility of evidence corresponding to the second detection data is determined based on the first data and the second data.

[0242] In an optional implementation, the early warning unit 167 is specifically used for:

[0243] Determine the data of each dimension of the credibility of the evidence; wherein each data represents the probability that the target state is one of the M states;

[0244] Based on the data from each dimension, determine the probability of the target device being abnormal at the target time;

[0245] If the probability of the abnormality is greater than the probability threshold, it is determined that the target device is abnormal at the target time.

[0246] In an optional embodiment, the structure of each neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a second fully connected layer, a third activation layer, a second fully connected layer and a classifier.

[0247] In an optional implementation, the activation functions of the first activation layer, the second activation layer, and the third activation layer are:

[0248]

[0249] Among them, act(x) is the activation function; α is the learnable adaptive decay parameter; x is the output of the network layer of the neural network.

[0250] Since the device is the device in the method in the embodiment of the present application, and the principle of solving the problem by the device is similar to that of the method, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0251] like Fig.17 As shown, based on the same inventive concept, an embodiment of the present application provides a detection device, which includes: a data receiving unit 171 and a processor 172;

[0252] The data receiving unit 171 is configured to execute:

[0253] Receive first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times;

[0254] The processor 172 is configured to execute:

[0255] Applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results; wherein the number of groups of first detection results and the number of groups of second detection results are respectively the same as the number of neural networks, each group of first detection results and each group of second detection results are respectively represented by a vector, and each element in a vector is a detection probability of different states of the device;

[0256] Determining weights corresponding to each of the plurality of neural networks according to the plurality of groups of first detection results;

[0257] For each group of second detection results, determining the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results belong; wherein the evidence support indicates the degree of accuracy of the second detection results;

[0258] Applying the evidence fusion rule, fusing the evidence according to the support of multiple pieces of evidence, and obtaining the credibility of the evidence corresponding to the second detection data;

[0259] According to the credibility of the evidence, determine whether the target device is abnormal at the target time, and issue an early warning when it is abnormal; wherein the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

[0260] In an optional implementation, the processor 172 is specifically configured to execute:

[0261] Determine a target tag for representing a real state of a first device; wherein the first device is a device to which the first detection data belongs;

[0262] For each group of first detection results, determine the similarity between the first detection result and the actual state according to the first detection result and the target label;

[0263] According to the similarities corresponding to the multiple groups of first detection results, the weights corresponding to the multiple neural networks are determined.

[0264] In an optional implementation, the processor 172 is specifically configured to execute:

[0265] Determine a first detection result set; wherein the first detection result set includes N groups of first detection results; N is an integer greater than or equal to 2;

[0266] Applying the first optimization function, determining a result vector that is most similar to the target label among the N result vectors corresponding to the N groups of first detection results;

[0267] Apply the second optimization function to find a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results;

[0268] Repeat the search process of applying the first optimization function and the second optimization function until all N result vectors are searched;

[0269] The found order is used as the similarity between the first detection result of the N groups and the actual state.

[0270] In an optional implementation, the processor 172 is specifically configured to execute:

[0271] Determine the permutation and combination results of M states, where each evidence support corresponds to one state;

[0272] When the intersection of the first set, the second set and the third set is any one of the M states, the evidence support of the first set, the evidence support of the second set and the evidence support of the third set are multiplied to obtain the first target support; wherein the first set, the second set and the third set are any one of the permutation and combination results;

[0273] Determine the sum of M first target support degrees as first data;

[0274] When the intersection of the fourth set, the fifth set and the sixth set is not empty, the evidence support of the fourth set, the evidence support of the fifth set and the evidence support of the sixth set are multiplied to obtain the second target support; wherein the fourth set, the fifth set and the sixth set are any one of the permutation and combination results;

[0275] Determine the sum of the M second target support degrees as the second data;

[0276] The credibility of evidence corresponding to the second detection data is determined based on the first data and the second data.

[0277] In an optional implementation, the processor 172 is specifically configured to execute:

[0278] Determine the data of each dimension of the credibility of the evidence; wherein each data represents the probability that the target state is one of the M states;

[0279] Based on the data from each dimension, determine the probability of the target device being abnormal at the target time;

[0280] If the probability of the abnormality is greater than the probability threshold, it is determined that the target device is abnormal at the target time.

[0281] In an optional embodiment, the structure of each neural network includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a second fully connected layer, a third activation layer, a second fully connected layer and a classifier.

[0282] In an optional implementation, the activation functions of the first activation layer, the second activation layer, and the third activation layer are:

[0283]

[0284] Among them, act(x) is the activation function; α is the learnable adaptive decay parameter; x is the output of the network layer of the neural network.

[0285] An embodiment of the present application further provides a computer storage medium, in which computer program instructions are stored. When the instructions are executed on a computer, the computer executes the steps of the above-mentioned device abnormality detection method.

[0286] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0287] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0288] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0289] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0290] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for detecting device abnormality, characterized in that: include: Acquire first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times; Applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying the multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results; wherein the number of groups of the first detection results and the number of groups of the second detection results are respectively the same as the number of neural networks, each group of the first detection results and each group of the second detection results are respectively represented by a vector, and each element in the one vector is a detection probability of different states of the device; Determining weights corresponding to each of the plurality of neural networks according to the plurality of groups of first detection results; For each group of second detection results, determining the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results respectively belong; wherein the evidence support indicates the degree of accuracy of the second detection results; Applying the evidence fusion rule to fuse the evidence according to the multiple evidence supports, and obtaining the evidence credibility corresponding to the second detection data; According to the credibility of the evidence, determine whether the target device is abnormal at the target time, and issue an early warning when it is abnormal; wherein, the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

2. The method according to claim 1, characterized in that: The step of determining the weights corresponding to the plurality of neural networks according to the plurality of groups of first detection results comprises: Determine a target tag for representing a real state of a first device; wherein the first device is a device to which the first detection data belongs; For each group of first detection results, determining the similarity between the first detection result and the actual state according to the first detection result and the target label; According to the similarities corresponding to the multiple groups of first detection results, the weights corresponding to the multiple neural networks are determined.

3. The method according to claim 2, characterized in that The determining, for each group of first detection results, the similarity between the first detection result and the actual state according to the first detection result and the target label, includes: Determine a first detection result set; wherein the first detection result set includes N groups of first detection results; wherein N is an integer greater than or equal to 2; Applying the first optimization function, determining a result vector that is most similar to the target label among the N result vectors corresponding to the N groups of first detection results; Applying the second optimization function, finding a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results; Repeat the search process of applying the first optimization function and the second optimization function until all the N result vectors are searched; The found order is used as the similarity between the N groups of first detection results and the true state.

4. The method according to claim 1, characterized in that: The applying of the evidence fusion rule to fuse the evidence according to the multiple evidence supports to obtain the evidence credibility corresponding to the second detection data includes: Determine the permutation and combination results of M states, where each evidence support corresponds to one state; When the intersection of the first set, the second set and the third set is any one of the M states, multiplying the evidence support of the first set, the evidence support of the second set and the evidence support of the third set to obtain a first target support; wherein the first set, the second set and the third set are any one of the permutation and combination results; Determine the sum of M first target support degrees as first data; When the intersection of the fourth set, the fifth set and the sixth set is not empty, multiplying the evidence support of the fourth set, the evidence support of the fifth set and the evidence support of the sixth set to obtain a second target support; wherein the fourth set, the fifth set and the sixth set are any one of the permutation and combination results respectively; Determine the sum of the M second target support degrees as the second data; Determine the credibility of evidence corresponding to the second detection data based on the first data and the second data.

5. The method according to claim 4, characterized in that The determining, based on the credibility of the evidence, whether the target device is abnormal at the target time includes: Determine data for each dimension of the credibility of the evidence; wherein each data represents the probability that the target state is one of the M states; Determine, based on the data of each dimension, the probability that the target device is abnormal at the target time; If the probability of the abnormality is greater than the probability threshold, it is determined that the target device is abnormal at the target time.

6. The method according to any one of claims 1 to 5, characterized in that: The structure of each of the neural networks includes a first convolutional layer, a first activation layer, a first maximum pooling layer, a second convolutional layer, a second activation layer, a second maximum pooling layer, a second fully connected layer, a third activation layer, a second fully connected layer and a classifier.

7. The method according to claim 6, characterized in that The activation functions of the first activation layer, the second activation layer and the third activation layer are: Wherein, act(x) is an activation function; α is a learnable adaptive decay parameter; and x is the output of the network layer of the neural network.

8. A detection device, characterized in that: comprising a data receiving unit and a processor; The data receiving unit is configured to execute: Receive first detection data and second detection data; wherein the first detection data and the second detection data are detection data of two devices respectively, or detection data of the same device at different times; The processor is configured to perform: Applying multiple neural networks to respectively identify the first detection data to obtain multiple groups of first detection results; and applying the multiple neural networks to respectively identify the second detection data to obtain multiple groups of second detection results; wherein the number of groups of the first detection results and the number of groups of the second detection results are respectively the same as the number of neural networks, each group of the first detection results and each group of the second detection results are respectively represented by a vector, and each element in the one vector is a detection probability of different states of the device; Determining weights corresponding to each of the plurality of neural networks according to the plurality of groups of first detection results; For each group of second detection results, determining the evidence support corresponding to the second detection results according to the weights corresponding to the neural networks to which the second detection results respectively belong; wherein the evidence support indicates the degree of accuracy of the second detection results; Applying the evidence fusion rule to fuse the evidence according to the multiple evidence supports, and obtaining the evidence credibility corresponding to the second detection data; According to the credibility of the evidence, determine whether the target device is abnormal at the target time, and issue an early warning when it is abnormal; wherein, the target device is the device to which the second detection data belongs, and the target time is the time when the second detection data is generated.

9. The detection device according to claim 8, characterized in that: The processor is specifically configured to execute: Determine a target tag for representing a real state of a first device; wherein the first device is a device to which the first detection data belongs; For each group of first detection results, determining the similarity between the first detection result and the actual state according to the first detection result and the target label; According to the similarities corresponding to the multiple groups of first detection results, the weights corresponding to the multiple neural networks are determined.

10. The detection device according to claim 8, characterized in that: The processor is specifically configured to execute: Determine a first detection result set; wherein the first detection result set includes N groups of first detection results; wherein N is an integer greater than or equal to 2; Applying the first optimization function, determining a result vector that is most similar to the target label among the N result vectors corresponding to the N groups of first detection results; Applying the second optimization function, finding a result vector that is most similar to the target label among the N-1 result vectors corresponding to the remaining N-1 groups of first detection results; Repeat the search process of applying the first optimization function and the second optimization function until all the N result vectors are searched; The found order is used as the similarity between the N groups of first detection results and the true state.

Images