A fault detection method, device, household appliance, and computer storage medium

Abstract

The present invention provides a fault detection method, apparatus, and home appliance. The fault detection method may include the following steps. Obtain an input signal from a device to be detected, that is, the signal is emitted by the device to be detected. Process the input signal using a compressed deep residual shrinkage network to output a classification result. Among them, the compressed deep residual shrinkage network includes, but is not limited to, a first residual unit and a second residual unit, and the second residual unit is a first residual unit without a soft threshold unit. Determine whether there is a fault in the device to be detected according to the classification result. The present invention can classify the sound or image signal from the device to be detected to quickly detect faults when there are slight fault signs, and has the advantages of good timeliness, high intelligence level, and strong reliability. It can be seen that the present invention can preferably solve the problem of fault detection lag in conventional technologies, avoid the spread or development of faults, and improve the user experience and satisfaction with home appliances.

Description

Technical Field

[0001] The present invention relates to the technical field of home appliance fault detection. More specifically, the present invention can provide a fault detection method, device and home appliance. Background Art

[0002] With the progress of society and the improvement of people's production and living quality, users' requirements for product quality are becoming increasingly strict. The increase in the number of product uses and the growth of product usage time often cause product wear and tear. For example, any home appliance product has a reasonable service life; it can be seen that an accurate fault detection system directly determines the quality of home appliance products, so as to effectively solve it as early as possible before the fault spreads or develops, thereby improving the user experience and satisfaction with home appliance products. However, traditional equipment fault detection schemes often can only find problems when the fault phenomenon is relatively obvious, and there is a large lag. Summary of the Invention

[0003] The main purpose of the present invention is to provide a fault detection method, device and home appliance to solve at least one problem existing in the conventional equipment fault detection technology.

[0004] To achieve the above technical purpose, the present invention can specifically provide a fault detection method; the method includes but is not limited to the following at least one step.

[0005] First, obtain an input signal from the device to be detected, that is, the signal is sent by the device to be detected. Secondly, process the input signal using a compressed deep residual shrinkage network to output a classification result. Among them, the compressed deep residual shrinkage network includes but is not limited to a first residual unit and a second residual unit, and the second residual unit is a first residual unit without a soft threshold unit. Finally, determine whether the device to be detected has a fault according to the classification result.

[0006] To achieve the above technical purpose, the present invention can also provide a fault detection device. The device may include but is not limited to an input signal acquisition module, an input signal processing module, and a device fault judgment module.

[0007] The input signal acquisition module is used to obtain an input signal from the device to be detected.

[0008] The input signal processing module is used to process the input signal using a compressed deep residual shrinkage network to output a classification result. Among them, the compressed deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is a first residual unit without a soft threshold unit;

[0009] The device fault judgment module is used to determine whether the device to be detected has a fault according to the classification result.

[0010] To achieve the above technical objectives, the present invention can also provide an appliance, which includes but is not limited to the fault detection device provided by the present invention.

[0011] To achieve the above technical objectives, the present invention can provide a computer storage medium, on which a fault detection program is stored; when the fault detection program is executed by a processor, the fault detection method in the embodiments of the present invention is implemented.

[0012] The beneficial effects of the present invention are as follows:

[0013] The present invention can classify the sound or image signals from the device to be detected, and thus quickly detect faults when there are signs of minor faults according to the classification results, having the advantages of good timeliness, high intelligence level, and strong reliability. It can be seen that the present invention can better solve the problem of fault detection lag in conventional technologies, avoid the spread or development of faults, and greatly improve the user experience and satisfaction with devices such as appliances.

[0014] The present invention classifies the input signal through the compressed deep residual shrinkage network, which can improve the detection effect while reducing the high cost generated by the application of complex deep learning technologies, and thus can reduce the design and processing costs of home appliance products, meet the user needs, and improve the user satisfaction level.

[0015] In addition, the present invention can dynamically simplify the deep residual shrinkage network according to the output threshold of the soft thresholding unit, specifically by pruning operations to remove some soft thresholding units in the network, so as to effectively reduce the waste in the calculation of the attention mechanism of the deep residual shrinkage network. Moreover, on the basis of simplifying the deep residual shrinkage network (model), the present invention can significantly improve the nonlinearity of the soft threshold function of the entire deep residual shrinkage network, making the deep residual shrinkage network obtained by the present invention have higher accuracy and precision.

[0016] The present invention simplifies the composition structure of the deep residual shrinkage network, effectively solves the problem of redundant model design existing in the conventional deep residual shrinkage network, and thus reduces the complexity, the number of parameters, and the amount of calculation of the deep residual shrinkage network.

[0017] The training cost of the deep residual shrinkage network optimized by the present invention is greatly reduced, so that it can be deployed in embedded devices and low-cost devices, and thus is more likely to be selected by technicians, which is beneficial to the further optimization and development of the deep residual shrinkage network. It can be seen that the present invention has a wide range of applications. Description of the Drawings

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0019] Figure 1 It shows a schematic flowchart of the fault detection method in the embodiment of the present invention.

[0020] Figure 2 It shows a schematic diagram of the overall structure of the deep residual shrinkage network in the embodiment of the present invention.

[0021] Figure 3 It shows a schematic diagram of the overall structure of the residual unit in the embodiment of the present invention.

[0022] Figure 4 It shows a schematic diagram of the relationship between the input and output of the soft thresholding function in the embodiment of the present invention.

[0023] Figure 5 It shows a schematic diagram of the relationship between the input and output of the derivative of the soft thresholding function in the embodiment of the present invention.

[0024] Figure 6 It shows a schematic diagram of the relationship between the input and output of the ReLU function and its derivative in the embodiment of the present invention.

[0025] Figure 7 It shows a schematic flowchart of the specific implementation process of the deep residual shrinkage network compression method in the embodiment of the present invention.

[0026] The realization of the purpose of the present invention, functional features and advantages will be further described in conjunction with the embodiments with reference to the drawings. Detailed implementation manners

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments of the present invention are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0028] As Figure 1 shown, the embodiment of the present invention can provide a fault detection method, which may include but is not limited to one or more of the following steps.

[0029] First, obtain the input signal from the device to be detected. The input signal includes at least one of a sound signal and an image signal.

[0030] Secondly, the compressed deep residual shrinkage network is used to process the input signal to output the classification result. Among them, the compressed deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is the first residual unit without a soft threshold unit.

[0031] Specifically, the embodiment of the present invention uses the compressed deep residual shrinkage network to process the input signal, including: performing batch normalization, rectified linear unit (ReLU) processing, convolution (Con) processing, and soft thresholding processing on the input signal by using the first residual unit, and performing batch normalization processing, rectified linear unit processing, and convolution processing on the input signal by using the second residual unit.

[0032] Finally, it is determined whether the device to be detected has a fault according to the classification result, so as to complete the device fault identification work based on the acquired sound and / or image signals.

[0033] Taking the bearing wear detection as an example, bearings generally contain noise and redundant signals during rotation, so it is possible to judge whether the bearing has a fault and the loss degree during the fault through the sound signal. The bearing is, for example, a bearing inside a washing machine or other household electrical appliances.

[0034] Optionally, the embodiment of the present invention determines whether the device to be detected has a fault according to the classification result, including: identifying the mapping value included in the classification result, and the mapping value can be, for example, a mapping value generated by a softmax operation. Judging the numerical interval where the mapping value is located, and the numerical interval can include, for example, but not limited to, (0, 0.5) and (0.5, 1). It is determined whether the device to be detected has a fault according to the numerical interval. In this embodiment, when the mapping value is in the interval (0.5, 1), it is determined that the device has a fault, and when the mapping value is in the interval (0, 0.5), it is determined that the device has no fault for the time being. It should be understood that this embodiment can also lock the fault type according to a more detailed interval and can perform intelligent reminders such as early warning or alarm according to the judgment result.

[0035] The deep residual shrinkage network (Residual Shrinkage Network) involved in the present invention is formed on the basis of the deep residual network (ResNet), and this deep residual shrinkage network can automatically derive the threshold τ to be set through the squeeze-and-excitation network.

[0036] In the embodiments of the present invention, an improved deep residual shrinkage network is used. Before obtaining the input signal from the device to be detected, the following compression process of the deep residual shrinkage network is further included to process the acquired signal using the compressed deep residual shrinkage network.

[0037] Construct the original deep residual shrinkage network. The original deep residual shrinkage network of the present invention includes, but is not limited to, convolutional units, multiple stacked residual units, fully connected units, etc.

[0038] As Figure 2 shown, the original deep residual shrinkage network includes multiple sequentially connected original residual units. Specifically, the deep residual shrinkage network of the present invention may include dozens or even hundreds of continuously stacked residual units, such as 50 or 100 residual units, etc. Of course, the number s of residual units can also be more.

[0039] Specifically, in this embodiment, the feature dimension of the input signal is C×W×1; where C represents the number of channels of the feature (channel), W represents the width of the feature, and 1 represents the height of the feature. Of course, the feature of the input signal can also be in the form of W×1×C, etc.

[0040] The input signal contains noise signals. Specifically, for the input feature with a dimension of C×W×1, in this embodiment, semantic information is extracted through convolution processing, and the dimension of the feature after extraction is still C×W×1. Then, a large number of stacked residual units in the network are used to remove noise redundancy, and the output feature is passed through BN (Batch Normalization), ReLU (Rectified Linear Unit), and GAP (Global Average Pooling) to obtain a feature with a dimension of N×1×1; where N represents a hyperparameter, specifically the channel dimension; in the present invention, these residual units have the same slope. Then, the feature with a dimension of N×1×1 can be subjected to a matrix transformation (reshape) operation to become a vector with a length of N. Then, FC (fully connected) processing and a logistic regression (softmax) operation are performed on the vector with a length of N, and the fault classification result can be directly output.

[0041] As Figure 3 shown, each original residual unit respectively has an original soft thresholding unit (which can be represented by ASSU for example). In the embodiments of the present invention, the soft thresholding unit is specifically a soft thresholding unit under the attention mechanism.

[0042] Soft Threshlding, that is, the soft thresholding function, is a function that shrinks the input data towards zero.

[0043] As Figure 4As shown in the figure, the mathematical expression of the soft thresholding function as an activation function is as follows:

[0044]

[0045] Among them, τ represents the threshold, x represents the input, y represents the output, and both x and y are real numbers.

[0046] As Figure 5 shown in the figure, the mathematical expression of the derivative of the soft thresholding function is as follows:

[0047]

[0048] Among them, τ represents the threshold, x represents the input, y represents the output, and both x and y are real numbers.

[0049] As Figure 4 、 5 shown in the figure, a schematic diagram of the relationship between the derivative of the soft thresholding function and its input and output is specifically given.

[0050] As Figure 6 shown in the figure, the present invention gives a schematic diagram of the relationship between the input and output of the ReLU function and its derivative. The ReLU function is a non-linear activation function, and in the present invention's Figure 4 、 5 it is used as a rectified linear unit, and the mathematical expression of the ReLU function is as follows:

[0051]

[0052] Among them, x represents the input, y represents the output, and both x and y are real numbers.

[0053] The mathematical expression of the derivative of the ReLU function is as follows:

[0054]

[0055] Among them, x represents the input, y represents the output, and both x and y are real numbers.

[0056] It can be understood that the soft thresholding function and the ReLU function have similar characteristics: the slope value is 0 within a certain threshold range, and the slope value is 1 outside the threshold range. Therefore, both the soft thresholding function and the ReLU function can suppress noise interference within the threshold range, eliminate redundant features, and have good anti-interference performance.

[0057] As Figure 3As shown, for the feature with the dimension of C×W×1, operations of BN (Batch Normalization) + ReLU (Rectified Linear Unit) + Con (convolution processing) are performed. Then, another operation of BN + ReLU + Con is performed on the obtained output to extract higher-level and more useful semantic information. For the output obtained thereby (i.e., the input x of the soft thresholding unit), the present invention obtains the output threshold τ based on the attention mechanism; specifically, it includes operations of absolute (taking the absolute value) + GAP (Global Average Pooling) + FC (Fully Connected) + BN (Batch Normalization) + ReLU (Rectified Linear Unit) + FC (Fully Connected) + sigmoid (activation function). Through the above operations, the sigmoid output threshold τ is obtained, and then the soft thresholding function can be constructed through the threshold τ output by the sigmoid function. The result processed by the soft thresholding function is superimposed on the result of the identity shortcut of the input feature to obtain the output of the current residual unit, which helps to reduce the training difficulty of the deep residual shrinkage network.

[0058] Train the original deep residual shrinkage network to obtain the first deep residual shrinkage network. Corresponding to the original deep residual shrinkage network, the first deep residual shrinkage network specifically includes a plurality of first residual units connected in sequence. Based on the deep learning method, the parameter features of the perturbed signal can be automatically learned during the training process, and the correct and reasonable parameters can be automatically deduced. Of course, these parameters may include the threshold τ involved in the embodiments of the present invention. During the process of training the original deep residual shrinkage network, the first residual unit is obtained by converting the original residual unit, and the first soft thresholding unit is obtained by converting the original soft thresholding unit.

[0059] The training of the original deep residual shrinkage network in this embodiment may include: continuously training the original deep residual shrinkage network repeatedly until the test accuracy of the original deep residual shrinkage network can reach a relatively high test accuracy, such as the second test accuracy = 95%; this embodiment also sets a preset test accuracy p according to the second test accuracy and makes the preset test accuracy p less than the second test accuracy. The preset test accuracy p in this embodiment is, for example, 80%.

[0060] Obtain the output thresholds of each first soft thresholding unit, and perform pruning operations on the obtained first deep residual shrinkage network according to the range of the output thresholds to remove at least one corresponding first soft thresholding unit, that is Figure 3 the soft thresholding unit part shown in; it can be seen that the embodiments of the present invention can provide a method for compressing a deep residual shrinkage network based on soft thresholding.

[0061] Specifically, the pruning operation on the current first deep residual shrinkage network according to the range of the output threshold in the embodiments of the present invention may include: comparing the output thresholds of each first soft thresholding unit with the current specified threshold respectively, and determining the first soft thresholding units to be removed according to the comparison results, and then removing the first soft thresholding units to be removed through the pruning operation. More specifically, determining the first soft thresholding units to be removed according to the comparison results in the embodiments of the present invention may include: counting the output thresholds less than the current specified threshold according to the comparison results, and taking the first soft thresholding units corresponding to the statistically obtained output thresholds as the first soft thresholding units to be removed, that is, locking the first soft thresholding units where they are located according to the output thresholds; if all the output thresholds are statistically greater than or equal to the current specified threshold according to the comparison results, no pruning is performed based on this judgment, and the judgment of the next specified threshold (the currently specified threshold selected again) is performed.

[0062] Optionally, comparing the output thresholds of each first soft thresholding unit with the current specified threshold in the present invention may include: selecting a threshold from a preset threshold list as the current specified threshold; the preset threshold list includes multiple thresholds, and all the multiple thresholds are greater than 0 and less than 1; comparing the output thresholds of each first soft thresholding unit with the current specified threshold respectively. Optionally, the preset threshold list includes a set of thresholds in an arithmetic progression; specifically, the preset threshold list includes a set of distinct thresholds, a set of thresholds consists of multiple thresholds, and there is a fixed difference between two adjacent thresholds in a set of thresholds. The threshold list in this embodiment includes a set of thresholds from small to large, the smallest threshold may be 0.01, the fixed difference may be 0.01, and the largest threshold is, for example, 0.5.

[0063] It can be seen that when the threshold τ derived based on the attention mechanism tends to 0, the present invention can effectively avoid the problem that the soft threshold function of the entire deep residual shrinkage network degenerates into a linear function of y = x (equivalent to no change in the input data itself). The selective and targeted removal of the soft thresholding units based on the threshold τ in the present invention significantly improves the nonlinearity of the soft threshold function of the entire network, making the performance of the deep residual shrinkage network based on deep learning better.

[0064] Optionally, in this embodiment, a threshold (for example, represented by a) is sequentially selected from small to large. Selecting a threshold a from the preset threshold list as the current specified threshold includes: traversing the preset threshold list to determine the unselected threshold for the first deep residual shrinkage network, that is, the threshold not used in the current network compression process; on this basis, selecting the smallest threshold from the unselected thresholds as the current specified threshold.

[0065] In an embodiment of the present invention, the first deep residual shrinkage network after pruning operation is trained, that is, the simplified new network is re-input with training data for training, and the first test accuracy is calculated. In an embodiment of the present invention, the training of the new network can be terminated according to that the first test accuracy is less than the preset test accuracy p; or according to that the first test accuracy is greater than the preset test accuracy p, return to the step of traversing the preset threshold list described above to continue the pruning operation of the deep residual shrinkage network with a new threshold. If all the thresholds in the preset list have been selected, the loop is terminated.

[0066] After ending this training or terminating the threshold selection loop, save the first deep residual shrinkage network after pruning operation, and then use the current first deep residual shrinkage network as the second deep residual shrinkage network. The second deep residual shrinkage network in this embodiment is the compressed deep residual shrinkage network to be used. Through the deep residual shrinkage network compression method provided by the present invention, the second deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is the first residual unit after removing the first soft thresholding unit.

[0067] Based on the same technical concept as the fault detection method, an embodiment of the present invention can also provide a fault detection device; the fault detection device may include, but is not limited to, an input signal acquisition module, an input signal processing module, and a device fault judgment module.

[0068] The input signal acquisition module is used to acquire an input signal from a device to be detected, and the input signal may include at least one of a sound signal and an image signal.

[0069] The input signal processing module is used to process the input signal by using the compressed deep residual shrinkage network to output a classification result. Among them, the compressed deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is the first residual unit without a soft threshold unit. Specifically, the input signal processing module is used to perform batch normalization processing, rectified linear processing, convolution processing, and soft thresholding processing on the input signal by using the first residual unit, and is used to perform batch normalization processing, rectified linear processing, and convolution processing on the input signal by using the second residual unit.

[0070] The device fault judgment module is used to determine whether there is a fault in the device to be detected according to the classification result.

[0071] Based on the same technical concept as the fault detection method in an embodiment of the present invention, an embodiment of the present invention can also provide a household appliance, and the household appliance includes, but is not limited to, the fault detection device in an embodiment of the present invention.

[0072] It is understandable that the household appliances in the embodiments of the present invention may be, for example, one or more of household appliances such as dishwashers, range hoods, refrigerators, washing machines, air conditioners, and microwave ovens. The present invention does not limit this.

[0073] Based on the same technical concept as the fault detection method, the embodiments of the present invention can also provide a computer storage medium, on which a fault detection program is stored. When the fault detection program is executed by a processor, the fault detection method in the embodiments of the present invention is implemented.

[0074] The specific steps of the fault detection method of the present invention have been described in the foregoing content and will not be elaborated herein.

[0075] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a predefined sequence of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable storage medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable storage medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable storage media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM, or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable storage medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other suitable processing as necessary, and then stored in a computer memory.

[0076] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logic functions on data signals, application specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0077] In the description of this specification, the descriptions with reference to the terms "this embodiment", "an embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0078] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, the meaning of "a plurality" is at least two, such as two, three, etc., and the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise specifically defined.

[0079] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc. are based on the orientation or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention.

[0080] In the present invention, unless otherwise clearly defined and limited, terms such as "installed", "connected", "joined", "fixed", etc. shall be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or integrated; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two components or the interaction relationship between two components, unless otherwise clearly defined. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0081] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural transformation made under the concept of the present invention by using the content of the specification and drawings of the present invention, or directly / indirectly applied in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A fault detection method, characterized in that, Including: Obtain an input signal from a device to be detected; Process the input signal using a compressed deep residual shrinkage network to output a classification result; wherein, the compressed deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is a first residual unit without a soft threshold unit; Determine whether the device to be detected has a fault according to the classification result; Before obtaining the input signal from the device to be detected, it further includes: Construct an original deep residual shrinkage network, which includes a plurality of sequentially connected original residual units, and each original residual unit has an original soft thresholding unit respectively; Train the original deep residual shrinkage network to obtain a first deep residual shrinkage network; wherein, the first residual unit is obtained through the original residual unit, and the first soft thresholding unit is obtained through the original soft thresholding unit; Obtain the output thresholds of the first soft thresholding units; Perform a pruning operation on the first deep residual shrinkage network according to the range of the output thresholds to remove at least one of the first soft thresholding units; Save the first deep residual shrinkage network after the pruning operation as a second deep residual shrinkage network, and the second deep residual shrinkage network is a compressed deep residual shrinkage network; The performing a pruning operation on the first deep residual shrinkage network according to the range of the output thresholds includes: Compare the output thresholds of the first soft thresholding units with a current specified threshold respectively, including: select a threshold from a preset threshold list as the current specified threshold; the preset threshold list includes a plurality of thresholds, and the plurality of thresholds are all greater than 0 and less than 1; compare the output thresholds of the first soft thresholding units with the current specified threshold respectively; Determine the first soft thresholding unit to be removed according to the comparison result; Remove the first soft thresholding unit to be removed through a pruning operation.

2. The fault detection method according to claim 1, wherein, The determining whether the device to be detected has a fault according to the classification result includes: Identify the mapping value included in the classification result; Judge the numerical interval where the mapping value is located; Determine whether the device to be detected has a fault according to the numerical interval.

3. The fault detection method according to claim 1 or 2, wherein The input signal includes at least one of a sound signal and an image signal.

4. The fault detection method according to claim 1, characterized in that, The processing the input signal using the compressed deep residual shrinkage network includes: Use the first residual unit to perform batch normalization processing, rectified linear processing, convolution processing, and soft thresholding processing on the input signal; Use the second residual unit to perform batch normalization processing, rectified linear processing, and convolution processing on the input signal.

5. The fault detection method according to claim 1, wherein The determining the first soft thresholding unit to be removed according to the comparison result includes: Count the output thresholds less than the current specified threshold according to the comparison result; Take the first soft thresholding unit corresponding to the statistically obtained output threshold as the first soft thresholding unit to be removed.

6. The fault detection method according to claim 1, wherein The preset threshold list includes a set of thresholds in an arithmetic progression.

7. The fault detection method according to claim 1, characterized in that Selecting a threshold from the preset threshold list as the current specified threshold includes: Traversing the preset threshold list to determine the unselected thresholds for the first deep residual shrinkage network; Selecting the smallest threshold from the unselected thresholds as the current specified threshold.

8. The fault detection method according to claim 7, characterized in that, Before saving the first deep residual shrinkage network after the pruning operation as the second deep residual shrinkage network, it further includes: Training the first deep residual shrinkage network after the pruning operation and calculating the first test accuracy; Terminating this training according to the first test accuracy being less than the preset test accuracy; Or, according to the first test accuracy being greater than the preset test accuracy, returning to the step of traversing the preset threshold list.

9. The fault detection method according to claim 8, characterized in that, Training the original deep residual shrinkage network includes: Repeatedly training the original deep residual shrinkage network until the test accuracy of the original deep residual shrinkage network reaches the second test accuracy; Setting the preset test accuracy according to the second test accuracy and making the preset test accuracy less than the second test accuracy.

10. A fault detection device, characterized in that, It includes: An input signal acquisition module for acquiring an input signal from a device to be detected; An input signal processing module for processing the input signal by using the compressed deep residual shrinkage network to output a classification result; wherein, the compressed deep residual shrinkage network includes a first residual unit and a second residual unit, and the second residual unit is a first residual unit lacking a soft threshold unit; Before acquiring the input signal from the device to be detected, it further includes: constructing an original deep residual shrinkage network, the original deep residual shrinkage network includes a plurality of sequentially connected original residual units, and each original residual unit respectively has an original soft thresholding unit; training the original deep residual shrinkage network to obtain a first deep residual shrinkage network; wherein, the first residual unit is obtained through the original residual unit, and the first soft thresholding unit is obtained through the original soft thresholding unit; acquiring the output thresholds of each of the first soft thresholding units; pruning the first deep residual shrinkage network according to the range of the output thresholds to remove at least one of the first soft thresholding units; saving the first deep residual shrinkage network after the pruning operation as the second deep residual shrinkage network, and the second deep residual shrinkage network is the compressed deep residual shrinkage network; Pruning the first deep residual shrinkage network according to the range of the output thresholds includes: comparing the output thresholds of each of the first soft thresholding units with the current specified threshold respectively, including: selecting a threshold from the preset threshold list as the current specified threshold; the preset threshold list includes a plurality of thresholds, and the plurality of thresholds are all greater than 0 and less than 1; comparing the output thresholds of each first soft thresholding unit with the current specified threshold respectively; determining the first soft thresholding unit to be removed according to the comparison result; removing the first soft thresholding unit to be removed through the pruning operation; A device fault judgment module for determining whether there is a fault in the device to be detected according to the classification result.

11. An appliance, characterized in that, Including the fault detection device described in claim 10.

12. A computer storage medium, characterized in that, A fault detection program is stored on the computer storage medium; when the fault detection program is executed by a processor, the fault detection method described in any one of claims 1 to 9 is implemented.

Images