Deep intelligent fault diagnosis model dynamic bipolarization pruning method

Abstract

The application provides a deep intelligent fault diagnosis model dynamic two-polarization pruning method. The method comprises the following steps: designing a mechanical equipment intelligent fault diagnosis model according to the actual situation of a diagnosis task, and determining a network layer in the mechanical equipment intelligent fault diagnosis model that needs to be pruned; embedding a two-polarization pruning module into the mechanical equipment intelligent fault diagnosis model, and connecting the two-polarization pruning module with the network layer that needs to be pruned, and taking an output feature map of the network layer that needs to be pruned as an input of the two-polarization pruning module; and synchronously training the mechanical equipment intelligent fault diagnosis model and the two-polarization pruning module by using a mechanical equipment fault diagnosis data set, and outputting a feature map of the network layer that needs to be pruned after pruning and channel enhancement by the two-polarization pruning module. The method can synchronously complete dynamic structural pruning of a large network granularity unit in the training process of the diagnosis model, the pruning mode is structural pruning, and no specific hardware device is needed; and the method has the advantages of not affecting diagnosis accuracy and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment fault diagnosis, and in particular to a deep intelligent fault diagnosis model dynamic two-polarization pruning method. BACKGROUND

[0002] Due to complex and variable working conditions, human operation influence, component natural degradation and other reasons, mechanical equipment inevitably occurs various faults in the running process, which may cause shutdown and production stoppage, or even personnel casualties. Therefore, timely diagnosis of mechanical equipment faults and taking corresponding maintenance measures have important practical significance for ensuring the safe, reliable and efficient operation of the equipment. Traditional mechanical equipment maintenance usually maintains or replaces each component of the equipment according to an experiential time period, which may easily lead to "over-maintenance" or "under-maintenance" problems, resulting in unnecessary resource waste, rising enterprise operation and maintenance costs, and being unable to effectively avoid major safety accidents. With the rapid development of new generation information technologies such as Internet of Things, big data, 5G communication and artificial intelligence, intelligent fault diagnosis based on deep belief network, convolutional neural network and stacked auto-encoding network and other deep learning technologies is favored, and has become a "sharp blade" for ensuring the safe operation of mechanical equipment.

[0003] At present, the intelligent fault diagnosis system framework used in industrial sites is usually a star-shaped radiation structure, that is, a server is used as a data processing and calculation center, and is connected with each sensing node through a gateway, a data acquisition board card and other hardware. Obviously, this mode has problems such as high time consumption of data transmission and poor real-time performance of diagnosis results. Therefore, in order to reduce communication costs and obtain better real-time diagnosis results, researchers are trying to develop various lightweight deep intelligent diagnosis models for sensing nodes, edge board cards and other weak computing power terminals, so as to complete fault diagnosis on the end-edge equipment, and then only upload the diagnosis results and a small amount of auxiliary information to the server. However, due to the constraints of the existing deep learning theory learning paradigm, iterative training cannot optimize the structure of the model, so even the lightweight model inevitably has the problem of over-parameterization. Structured network pruning can solve the above problems, but the traditional structured network pruning is performed after the model iterative training is completed, and it needs to rely on expert experience or domain knowledge to develop a pruning strategy in advance, and also needs to try repeatedly according to the evaluation index. This process not only consumes time and effort, but also may cause network collapse if the pruning strategy, evaluation index and the like are not properly designed and selected.

[0004] In summary, there is currently a lack of an adaptive network pruning method synchronized with model iterative training, so as to efficiently, quickly and accurately reduce the complexity of the model, so that the deep intelligent fault diagnosis model can be adapted to a wider range of weak computing power terminal equipment. SUMMARY

[0005] The embodiment of the present application provides a deep intelligent fault diagnosis model dynamic two-polarization pruning method to realize dynamic network pruning synchronized with training in the intelligent fault diagnosis model training process of a mechanical equipment, so that redundant parameters in the intelligent fault diagnosis model of the mechanical equipment are effectively removed.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme.

[0007] A deep intelligent fault diagnosis model dynamic two-polarization pruning method, comprising:

[0008] Designing a mechanical equipment intelligent fault diagnosis model according to the actual situation of a diagnosis task, and determining a network layer in the mechanical equipment intelligent fault diagnosis model that needs to be pruned;

[0009] Embedding a two-polarization pruning module into the mechanical equipment intelligent fault diagnosis model, connecting the two-polarization pruning module with the network layer that needs to be pruned, and taking an output feature map of the network layer that needs to be pruned as an input of the two-polarization pruning module;

[0010] Obtaining a mechanical equipment fault diagnosis data set, and synchronously training the mechanical equipment intelligent fault diagnosis model and the two-polarization pruning module by using the mechanical equipment fault diagnosis data set, wherein the two-polarization pruning module outputs a feature map of the network layer that needs to be pruned after pruning and channel enhancement.

[0011] Preferably, the two-polarization pruning module indirectly evaluates the contribution of a network granularity unit to a task by means of a correspondence between the network granularity unit and a channel through a learnable evaluation vector, and generates a channel-level two-polarization vector in cooperation with a dynamic adaptive threshold operator, so as to cut off the influence of the pruned network granularity unit on network forward calculation and back propagation, and realize dynamic redundant parameter pruning of the network layer that needs to be pruned in synchronization with training.

[0012] Preferably, the execution steps of the two-polarization pruning module are as follows:

[0013] 3.1) Let the output feature map of the network layer that needs to be pruned be denoted as W, H, and C respectively represent the width, height, and channel number of the feature map; X is first calculated according to the channel respectively, and the mean value μ n and the variance σ n of each channel feature map X 2 n wherein n = 1, 2, 3...C, and the μ n and σ 2 of each channel are based on Gaussian distribution random sampling, and the sampling value β n of each channel is obtained. nand β n β is spliced into a sample vector β in the order of channels, and the specific calculation expression is as follows:

[0014]

[0015]

[0016] β n = Sampling(N~(μ n ,σ n 2 )), n = 1, 2...C

[0017] β = concat(β1, β2...β C )

[0018] In the formula, sampling(·) represents a sampling operation, N~(μ n ,σ 2 n ) represents a Gaussian distribution with mean μ n and variance σ 2 n , and concat(·) represents a channel splicing operation.

[0019] 3.2) The two-polarization vector γ in the two-polarization pruning module is a binary vector, 0 corresponds to pruning, and 1 corresponds to retention, γ is obtained by comparing α with a dynamic adaptive threshold operator s(e t-T -1), and the mathematical expression is as follows:

[0020]

[0021] In the formula, s(e t-T -1) is a dynamic adaptive threshold operator that generates a dynamic threshold, t represents the training algebra where the current training is located, T represents the total training algebra, s is a sensitive adaptive coefficient, and the function property of the dynamic adaptive threshold operator is that the operator is dynamically raised from -s to 0 from t = 1 to t = T, and the second derivative is positive, and the dynamic adaptive threshold increases with the increase of the training algebra.

[0022] 3.3) The two-polarization vector γ is multiplied with the feature map in the channel dimension, the channels that are pruned are forced to be zero, based on the correspondence between the channels and the network granularity units, the channel zero operation is used to indirectly realize the pruning of the network granularity units, after the channel zero operation, the network granularity units that are pruned are also indirectly disconnected from the network, and no longer affect the subsequent training process, and the above mathematical expression is as follows:

[0023]

[0024] In the formula, represents the pruned feature map, represents the channel-wise multiplication;

[0025] 3.4) α and β are multiplied by the channel and then activated by tanh, obtaining the weighting vector ξ, which is used to enhance the unpruned channels to different degrees, and the expression is as follows:

[0026]

[0027] In the formula, tanh represents the hyperbolic tangent activation function;

[0028] 3.5) Enhance each channel feature map after pruning, represents the output of the two-polarization pruning module, and the calculation expression is as follows:

[0029]

[0030] α is learned by iteration, and dynamic adaptive threshold operator γ is generated to realize adaptive dynamic pruning, and the multi-class cross-entropy loss function is minimized, and the mathematical expression is as follows:

[0031]

[0032]

[0033] In the formula, represents the target function of training, I is the input of the deep intelligent fault diagnosis model, and y is the corresponding label, and the model training process is optimized by unpruned network parameters so that the output of the mapping F constructed by the deep model and the cross-entropy loss function of the sample label y is minimized, and the expression of CE is p k represents the probability that the model predicts that the current sample belongs to the kth class, y k represents the value of the kth class corresponding to the one-hot encoding of the sample label y, and cl represents the class of the sample to be identified in the diagnosis task.

[0034] Preferably, when the method is applied to a weak computing power terminal device supporting sparse acceleration operation, no operation is required on the trained mechanical equipment intelligent fault diagnosis model embedded with the two-polarization pruning module; when applied to a hardware device that does not support sparse acceleration operation, the two-polarization pruning module is deleted from the mechanical equipment intelligent fault diagnosis model after the training of the mechanical equipment intelligent fault diagnosis model embedded with the two-polarization pruning module is completed.

[0035] It can be seen from the technical solutions provided by the above-mentioned embodiments of the application that the application provides a two-polarization pruning method of an intelligent fault diagnosis model for weak computing power terminals, which can dynamically complete the structural pruning of a large network granularity unit (such as a convolution kernel) in the diagnosis model training process.

[0036] Additional aspects and advantages of the application will be described in the description that follows, and will become apparent from the description, or will be learned from practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0038] Figure 1 A processing flow chart of the deep intelligent fault diagnosis model dynamic two-polarization pruning method provided by the embodiments of the application.

[0039] Figure 2 A structure diagram of the two-polarization pruning module provided by the embodiments of the application.

[0040] Figure 3 A working principle diagram of the two-polarization pruning module provided by the embodiments of the application. DETAILED DESCRIPTION

[0041] The embodiments of the application are described in detail below, and examples of the embodiments are shown in the drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below are exemplary and are only used to explain the application, and cannot be interpreted as a limitation on the application.

[0042] As those skilled in the art will readily appreciate from the disclosure herein, terms used herein are intended to be construed in a non-limiting fashion. As used herein, the following terms have the following meanings.

[0043] As those skilled in the art will readily appreciate from the disclosure herein, terms used herein are intended to be construed in a non-limiting fashion. As used herein, the following terms have the following meanings.

[0044] To facilitate the understanding of the embodiments of the present application, further explanation will be made in conjunction with the accompanying drawings and several specific embodiments as examples, and each embodiment does not constitute a limitation to the embodiments of the present application.

[0045] The processing flow of the dynamic two-polarization pruning method for a weak computing power terminal-oriented deep intelligent fault diagnosis model provided by the embodiments of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0046] Step 1, obtain a mechanical equipment fault diagnosis dataset, design a reasonable mechanical equipment intelligent fault diagnosis model according to the actual situation of the diagnosis task, and determine the network layer that needs to be pruned in the mechanical equipment intelligent fault diagnosis model.

[0047] Step 2, for a certain network layer that needs to be pruned, its output feature map is denoted as X, W, H, and C represent the width, height, and channel number of the feature map, respectively. In particular, for the output of a network layer such as one-dimensional convolution, H is equal to 1.

[0048] The two-polarization pruning module is embedded in the mechanical equipment intelligent fault diagnosis model, and the two-polarization pruning module is connected with the above-mentioned certain network layer that needs to be pruned, and X will serve as the input of the two-polarization pruning module.

[0049] ​Step 3: The present invention trains the above-mentioned intelligent fault diagnosis model for mechanical equipment and the above-mentioned bipolar pruning module simultaneously. A structural diagram of a bipolar pruning module provided in this embodiment of the invention is shown below. Figure 2 As shown in the diagram, the working principle is as follows: Figure 3 As shown, the input to the polarization pruning module is X, and the output is a feature map that has undergone pruning and channel enhancement. The basic logic behind pruning lies in the correspondence between network granularity units and output channels. Based on this correspondence, the importance of network granularity units can be indirectly evaluated by assessing the importance of channels. Therefore, less important network granularity units can be pruned. This evaluation is implemented using a learnable evaluation vector α within the module, which can also enhance the preserved channels.

[0050] The execution steps of the polarization pruning module are as follows:

[0051] 3.1) First, obtain the feature map X for each channel. n mean μ n and variance σ 2 n (n = 1, 2...C), and based on μ for each channel n and σ 2 n Gaussian random sampling is performed to obtain the sampled value β of the feature map for each channel. n The samples are then concatenated in channel order to form a sampling vector β. This vector introduces random factors into the polarization pruning module, increasing the learning difficulty and thus forcing the model to be more robust. The specific calculation expression is as follows:

[0052]

[0053]

[0054] β n =Sampling(N~(μ) n ,σ n 2 ), n=1,2...C

[0055] β=[β1,β2...β C ]

[0056] In the formula, sampling(·) represents the sampling operation, and N ~ (μ n , σ 2 n ) indicates that the mean is μ n The variance is σ 2 n The Gaussian distribution is used, and concat(·) represents the concatenation operation by channel;

[0057] 3.2) The key to implement end-to-end pruning is the two-polarization vector γ in the two-polarization pruning module. The vector is a binary vector (0 corresponds to pruning, 1 corresponds to keeping). γ is obtained by comparing α with a dynamic adaptive threshold operator s(e t-T -1) to obtain, the mathematical expression is as follows:

[0058]

[0059] In the formula, s(e t-T -1) is a dynamic adaptive threshold operator, which can generate a dynamic threshold, t represents the training generation number where the current training is located, T represents the total training generation number, s is a sensitive adaptive coefficient, if α is initialized using a normal distribution with a mean of 0 and a standard deviation of std, α can take 5std. The functional property of the dynamic adaptive threshold operator is that from t = 1 to t = T, the operator is dynamically increased from -s to 0, and the second derivative is positive. With the sensitive adaptive coefficient, the initialization will not blindly perform pruning, and the threshold is adjusted to the appropriate sensitive adaptability. With the increase of the training generation number, the dynamic adaptive threshold increases, which means that the pruning standard becomes more and more stringent;

[0060] 3.3) The two-polarization pruning is realized by multiplying γ with the feature map in the channel dimension to force the pruned channel to be zero. Based on the correspondence between the channel and the network granularity unit, the pruning of the network granularity unit can be indirectly realized by means of the channel zero operation. After the channel is zeroed, the pruned network granularity unit is also indirectly disconnected from the network and no longer affects the subsequent training process. The above mathematical expression is as follows:

[0061]

[0062] In the formula, represents the pruned feature map, represents the multiplication by channel;

[0063] 3.4) α and β are multiplied by channel and then activated by tanh to obtain the weighting vector ξ. The role of the vector is to enhance the unpruned channel to different degrees (similar to the channel attention mechanism), and the expression is as follows:

[0064]

[0065] In the formula, tanh represents the hyperbolic tangent activation function;

[0066] 3.5) The pruned channel feature map is enhanced, represents the output of the two-polarization pruning module, and the calculation expression is as follows:

[0067]

[0068] The two-polarization pruning module is embedded in the intelligent fault diagnosis model of the mechanical equipment, and is trained synchronously with the intelligent fault diagnosis model of the mechanical equipment. The optimizer can select Adam, stochastic gradient descent algorithm, etc. After setting the training algebra, learning rate and other hyperparameters, a can learn through iteration, and a dynamic adaptive threshold operator is generated to realize adaptive dynamic pruning. Training is consistent with the traditional model, and the most accurate diagnostic accuracy is obtained as the target, and the multi-class cross-entropy loss function is minimized. The mathematical expression is as follows:

[0069]

[0070]

[0071] In the formula, The target function of training is represented by I, the input of the deep intelligent fault diagnosis model, and y is the corresponding label. The model training process optimizes the network parameters of the unpruned network So that the output of the mapping F constructed by the deep model The cross-entropy loss function of the sample label y is minimized. The expression of CE is p k The probability that the model predicts that the current sample belongs to the kth class is represented by y k The value of the kth class corresponding to the one-hot encoding of the sample label y is represented by cl, and cl represents the kind of sample to be identified in the diagnosis task.

[0072] Step 4, in the case of supporting sparse acceleration operation in weak computing terminal equipment, no operation is needed for the trained intelligent fault diagnosis model of the mechanical equipment; and for hardware devices that do not support sparse acceleration operation, only the pruning network granularity unit needs to be deleted from the model file after the training of the intelligent fault diagnosis model of the mechanical equipment is completed.

[0073] Example: Take rolling bearing in mechanical equipment fault diagnosis as an example, based on the data of bearing test bench, the effectiveness of the method is verified. The test bench used in this example is composed of driving motor, transmission device, axle, supporting bearing, loading device (horizontal and vertical) and cooling fan, etc. The test bearing is SKF197726 type double-row tapered roller bearing. In order to simulate different working conditions, the horizontal load is set to 0kN and 20kN, the vertical load is set to 56kN and 146kN, the rotating speed is set to 589rpm, 786rpm and 983rpm, and the experiment involves 12 kinds of working conditions. The vertical acceleration of the test bearing is collected by using single-axis acceleration sensor, and the sampling frequency is 12.8kHz. The test bearing includes 7 kinds of fault bearings such as inner ring spalling, multiple spalling of inner ring, outer ring spalling, rolling element pitting, rolling element spalling, rolling element crack and retainer crack, and there is one normal bearing. As described above, the data set is divided into 8 categories of labels. Each bearing is sampled for 10 seconds under each working condition, which is divided into 125 samples without overlap (i.e. each sample contains 1024 sampling points), and a total of 12000 samples (1500 samples per category) are obtained. After the order of the samples is randomly shuffled, they are evenly divided into 3 groups as the training set, validation set and test set of the implementation.

[0074] The network parameters of the basic model used are shown in Table 1. The training related hyperparameter settings are shown in Table 2. The training is tested 5 times. Using the method of the application, the three convolutional layers (one-dimensional convolutional layer #1, one-dimensional convolutional layer #2 and one-dimensional convolutional layer #3) of the model are pruned, and the diagnostic accuracy, floating point operation number FLOPs and parameter size of the pruned model are compared with the model without pruning using the same training related hyperparameters, and the results are summarized in Table 3. The results of the example show that the deep intelligent fault diagnosis model training synchronous dynamic two-polarization network pruning method proposed by the application can significantly reduce the complexity of the model while not losing the performance of the deep intelligent fault diagnosis model, and due to its ingenious design, the pruning can be completed synchronously with the model training, overcoming the drawbacks of traditional network pruning methods which rely too much on expert knowledge and have a complicated process, and can provide strong support for the deployment of deep intelligent fault diagnosis models to more extensive edge devices, further improving the real-time performance of the fault diagnosis system.

[0075] Table 1 Basic network parameter settings of the example

[0076]

[0077]

[0078] Table 2 Training related hyperparameter settings of the example

[0079]

[0080] Table 3 comparison table of experimental results of examples

[0081]

[0082] In summary, the application can realize dynamic network pruning synchronized with training in the training process of the intelligent fault diagnosis model of mechanical equipment, thereby effectively removing redundant parameters in the intelligent fault diagnosis model of mechanical equipment, compressing the size of the intelligent fault diagnosis model of mechanical equipment, improving the calculation efficiency, and finally enabling the deep learning-based intelligent fault diagnosis model of mechanical equipment to be deployed on a wider range of weak computing terminal devices. The application overcomes the drawbacks of traditional network pruning methods that excessively rely on expert knowledge and have a complicated process, and has the following advantages:

[0083] ① The pruning effect is obvious, and the complexity of the model can be significantly reduced without affecting the accuracy of the deep intelligent fault diagnosis model;

[0084] ② Strong universality, applicable to current mainstream standard convolutional layers, separable convolutional layers, atrous convolutional layers, point convolutional layers, etc.

[0085] ③ The method is a structured pruning, without the need for specific hardware device support.

[0086] The application can significantly reduce the complexity of the model without affecting the accuracy of the deep intelligent fault diagnosis model; it has strong universality and is applicable to current mainstream standard convolutional layers, separable convolutional layers, atrous convolutional layers, point convolutional layers, etc.; the method is a structured pruning, without the need for specific hardware device support, and many other advantages. It has important technical value for reducing the complexity of the mechanical equipment fault diagnosis model, so that it can be deployed on a wider range of weak computing terminal devices.

[0087] Those skilled in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or flows in the drawings are not necessarily necessary for implementing the application.

[0088] From the above description of the embodiments, those skilled in the art can clearly understand that the application can be implemented by means of software and the necessary general hardware platform. Based on this understanding, the technical solutions of the application can be embodied in the form of a software product, which can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments.

[0089] The various embodiments described in this specification are described with reference to a particular sequence or order, but the order of the steps can be modified so that particular steps can be performed in an order other than that described. Many of the embodiments described in this specification are implemented as a method, apparatus, or means for practicing those steps. The steps of a method or algorithm described in connection with the embodiments described herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of non-transitory computer-readable storage medium known in the art. An exemplary non-transitory computer-readable storage medium is coupled to the processor such that the processor can read information from, and write information to, the non-transitory computer-readable storage medium. In the alternative, the non-transitory computer-readable storage medium can be integral to the processor. The processor and the non-transitory computer-readable storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the non-transitory computer-readable storage medium can reside as discrete components in a user terminal. The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. The various embodiments shown and described above are only examples. Therefore, the above description should not be interpreted as a limitation on the scope of the application, and it is not intended to limit the scope of the application. Therefore, the scope of the application should be determined by the following claims.

[0090] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements within the technical scope disclosed by the present application can be easily thought by those skilled in the art, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A dynamic bipolar pruning method for a deep intelligent fault diagnosis model, characterized in that, include: Design an intelligent fault diagnosis model for mechanical equipment based on the actual situation of the diagnostic task, and determine the network layers that need to be pruned in the intelligent fault diagnosis model for mechanical equipment. The bipolar pruning module is embedded into the intelligent fault diagnosis model of the mechanical equipment, the bipolar pruning module is connected to the network layer that needs to be pruned, and the output feature map of the network layer that needs to be pruned is used as the input of the bipolar pruning module. Obtain a mechanical equipment fault diagnosis dataset, and use the mechanical equipment fault diagnosis dataset to synchronously train the mechanical equipment intelligent fault diagnosis model and the polarization pruning module. The polarization pruning module outputs the feature map of the network layer that needs to be pruned after pruning and channel enhancement. The execution steps of the polarization pruning module are as follows: 3.1) Let the output feature map of the network layer that needs pruning be denoted as X = {X1, X2, X3...X}. C }, W, H, and C represent the width, height, and number of channels of the feature map, respectively; X is first calculated for each channel separately. n mean μ n and variance σ 2 n Where n = 1, 2, 3...C, based on μ for each channel n and σ 2 n Perform Gaussian distribution random sampling to obtain the sampled value β for each channel. n and β n The samples are concatenated into a sampling vector β in channel order, and the specific calculation expression is as follows: b n =Sampling(N~(μ n ,s n 2 )),n=1,2...C β=concat(β1,β2...β C ) In the formula, sampling(·) represents the sampling operation, and N ~ (μ n , σ 2 n ) indicates that the mean is μ n The variance is σ 2 n The Gaussian distribution is used, and concat(·) represents the concatenation operation by channel; 3.2) The polarization vector γ within the polarization pruning module is a binary vector, where 0 corresponds to pruning and 1 corresponds to retention. γ is determined by α and a dynamic adaptive threshold operator s(e t-T -1) is obtained through comparison, and its mathematical expression is as follows: In the formula, s(e t-T -1) is the dynamic adaptive threshold operator, which generates a dynamic threshold. t represents the current training generation, T represents the total number of training generations, m represents the channel index, and s is a sensitive adaptive coefficient. The function property of the dynamic adaptive threshold operator is that from t=1 to t=T, the operator dynamically increases from -s to 0, and the second derivative is positive. As the number of training generations increases, the dynamic adaptive threshold increases. 3.3) The polarization vector γ is multiplied by the feature map along the channel dimension, forcing the pruned channels to be zeroed. Based on the correspondence between channels and network granularity units, the pruning of network granularity units is indirectly achieved through the channel zeroing operation. After the channels are zeroed, the pruned network granularity units are also indirectly disconnected from the network and no longer affect the subsequent training process. The mathematical expression above is as follows: In the formula, This represents a characteristic diagram after pruning. This indicates multiplication by channel; 3.4) α and β are multiplied side-by-side according to channel alignment and then activated by tanh to obtain a weighted vector ξ. This weighted vector ξ is used to enhance the unpruned channels to different degrees, as shown in the following expression: In the formula, tanh represents the hyperbolic tangent activation function; 3.5) Enhance the feature maps of each channel after pruning. The output of the bipolar pruning module is expressed as follows: α is learned iteratively and combined with a dynamic adaptive threshold operator to generate γ, achieving adaptive dynamic pruning and minimizing the multi-class cross-entropy loss function, the mathematical expression of which is as follows: In the formula, Let I represent the objective function for training, y be the input to the deep intelligent fault diagnosis model, and y be its corresponding label. The model training process optimizes the unpruned network parameters. This makes the output of the mapping F constructed by the deep model... The cross-entropy loss function with sample label y is minimized. For the expression of CE, p k y represents the probability that the model predicts the current sample belongs to the k-th class. k represents the value of the k-th class corresponding to the one-hot encoding of the sample label y, and cl represents the type of sample to be identified in the diagnostic task.

2. The method according to claim 1, characterized in that, The polarization pruning module leverages the correspondence between network granularity units and channels to indirectly evaluate the contribution of network granularity units to the task through learnable evaluation vectors. It also generates channel-level polarization vectors in conjunction with a dynamic adaptive threshold operator. These polarization vectors are used to cut off the influence of pruned network granularity units on the network's forward computation and backpropagation, thereby achieving dynamic redundant parameter pruning of the network layers that need pruning, synchronized with training.

3. The method according to any one of claims 1 to 2, characterized in that, When the method is applied to a terminal device with weak computing power that supports sparse accelerated computing, no operation is required on the trained intelligent fault diagnosis model of mechanical equipment with embedded polarization pruning module; when applied to a hardware device that does not support sparse accelerated computing, the polarization pruning module is removed from the intelligent fault diagnosis model of mechanical equipment after the training of the intelligent fault diagnosis model with embedded polarization pruning module is completed.

Images