Lightweight method and system of intelligent fault diagnosis model based on dynamic trade-off tailoring

Abstract

The present invention relates to the field of fault diagnosis technology, and specifically to a method and system for lightweight intelligent fault diagnosis models based on dynamic trade-off pruning, wherein the method comprises: determining a mechanical equipment fault diagnosis model to be pruned; constructing a target optimization function; obtaining the output distortion of the fault diagnosis model after different pruning of a single layer; transforming the target optimization function according to the additivity of the output distortion of the fault diagnosis model; solving the transformed target optimization function by dynamic programming; in the solution process, constructing a state transfer equation, searching for the output distortion of all layers of the fault diagnosis model; iterating the pruning process until a preset number of pruning iterations is reached. The present invention comprehensively considers the inter-layer influence, can search for the optimal pruning strategy in a short time, and realizes the lightweight of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and more particularly to a lightweight method and system for an intelligent fault diagnosis model based on dynamic trade-off tailoring. Background Art

[0002] Large-scale mechanical equipment is an important part of modern industry and plays a key role in the production work of many industries. Mechanical equipment is prone to failure when it is in high-intensity service for a long time, which will cause equipment downtime, economic losses and even casualties. Equipment operation and maintenance personnel hope to find the cause of the failure in the early stage and carry out maintenance work to ensure the safe and reliable operation of mechanical equipment.

[0003] With the advancement of artificial intelligence technology, many scholars have applied it to the field of mechanical equipment fault diagnosis. Intelligent fault diagnosis technology not only gets rid of the dependence on maintenance experience, reduces operation and maintenance costs, but also greatly improves the accuracy of diagnosis. Traditional cloud computing-based fault diagnosis methods rely on centralized data processing systems to upload data to cloud data centers or servers for processing and analysis to obtain diagnostic results. However, this method is constrained by data transmission delays and bandwidth limitations, and faces challenges such as high network latency and difficulty in large-scale data processing. It is difficult to adapt to fault diagnosis scenarios with extremely strict response time requirements, such as aircraft fault diagnosis, self-driving car fault diagnosis, and high-speed train fault diagnosis. This limitation limits the application potential of intelligent diagnosis methods in actual engineering.

[0004] Edge computing reduces the reliance on centralized cloud systems by moving data processing and analysis tasks to edge devices, and can achieve high-timeliness fault diagnosis. However, edge devices are usually limited in computing power and storage capacity, making it difficult to deploy models with high complexity. Therefore, scholars have customized lightweight networks for performance-constrained edge devices, aiming to reduce the number of parameters and computational complexity of network models while maintaining high fault diagnosis accuracy. In the process of lightweighting models, layer adaptive pruning is a common method, which dynamically adjusts the structure of the network according to the importance of each layer of the model and removes unimportant parameters, thereby reducing the size and amount of computation of the model and improving the speed and efficiency of reasoning. Existing layer adaptive pruning strategies have many defects. On the one hand, they only consider the performance changes after pruning a single layer, ignoring the mutual influence between different layers; on the other hand, the time complexity of the pruning strategy search method is exponential, requiring a large amount of computing resources. These limitations prevent lightweight networks from flexibly and efficiently carrying out real-time fault diagnosis on edge devices. Therefore, there is an urgent need for a dynamic pruning method that comprehensively considers the influence between layers, searches for the optimal pruning strategy in a short time, and establishes a fault diagnosis network with high accuracy and excellent lightweight effect. Summary of the invention

[0005] In view of this, the present invention provides a lightweight method and system for an intelligent fault diagnosis model based on dynamic trade-off pruning, which comprehensively considers the impact between layers and can search for the optimal pruning strategy in a short time to achieve lightweight model.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a lightweight method for an intelligent fault diagnosis model based on dynamic trade-off tailoring, characterized in that it comprises the following steps:

[0008] S1. Determine the fault diagnosis model of the mechanical equipment to be pruned;

[0009] S2, under the premise of specifying the minimum cropping rate constraint, with the goal of minimizing the distortion after cropping, construct a target optimization function;

[0010] S3, obtaining the output distortion of the fault diagnosis model after different trimmings of a single layer;

[0011] S4. Determine the additivity of the output distortion of the fault diagnosis model through the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence; transform the target optimization function according to the additivity of the output distortion of the fault diagnosis model;

[0012] S5. Solving the transformed target optimization function using a dynamic programming method;

[0013] S6. During the solving process, a state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model; each search for output distortions of all layers of the fault diagnosis model is regarded as a trimming process;

[0014] S7, iterating the trimming process until a preset number of trimming iterations is reached.

[0015] Furthermore, in S1, the fault diagnosis model is expressed as: ;in, represents the diagnosed sample, a represents the signal sampling length, and b represents the number of signal channels; represents the weight parameters of each layer, represents the weight of the i-th layer of the fault diagnosis model F, , l represents the total number of layers of the fault diagnosis model.

[0016] Furthermore, in S2, the weight parameter of the fault diagnosis model F after trimming is ,use and The Euclidean space distance between represents the impact of clipping, and the expression of the objective optimization function is:

[0017] ;

[0018] ;

[0019] in, Represents a parameter set norm, which is equal to the number of non-zero elements in the parameter set; Indicates the specified minimum clipping rate.

[0020] Furthermore, S3 includes:

[0021] S31, traverse each layer of the fault diagnosis model, obtain the weight of each layer, and press The norm determines the order of weight pruning, giving priority to pruning Weights with smaller norms;

[0022] S32, output distortion of different levels of clipping of each layer ; It represents the distortion caused by clipping k weights of the i-th layer. k is used to measure the degree of clipping. The larger k is, the greater the degree of clipping.

[0023] Further, S41, a parameter perturbation is given to each layer weight of the fault diagnosis model F , the weight is changed to , the output after weight modification is approximately expressed by the first-order Taylor expansion:

[0024] ;

[0025] in, Represents the weight The partial derivative of

[0026] S42. Assume parameter perturbations between different layers Independent of each other, there exists: ;

[0027] Among them, i and m represent different layers respectively. Indicates expectation, represents the parameter perturbation of the i-th layer weight, represents the parameter perturbation of the mth layer weight;

[0028] S43, distortion after clipping According to the first-order Taylor expansion formula, it can be expanded into:

[0029] ;

[0030] Where T represents the transpose of the matrix, represents the parameter perturbation of the i-th layer weight, represents the parameter perturbation of the mth layer weight;

[0031] Taking the expected value on both sides of the formula, we get:

[0032] ;

[0033] S44. According to the independence assumption in S42, when When , the cross term is 0, that is: , then the distortion is expressed as:

[0034] ;

[0035] S45. According to the first-order Taylor expansion formula of S41, we get:

[0036] ;

[0037] S46. Substitute the formula of S45 into the distortion representation equation in S44. The left side of the equation becomes the total output distortion when all layers are clipped. , the right side becomes the sum of the output distortions caused by individually trimming the weights of each single layer, transforming the objective optimization function into:

[0038] .

[0039] Furthermore, in S5, a dynamic programming model is constructed to solve the transformed target optimization function. The expression of the dynamic programming model is:

[0040] ;

[0041] in, represents the total number of weights to be pruned, represents the number of weights to be pruned in layer i; represents the output distortion after pruning the i-th layer.

[0042] Furthermore, in S6, we define is the minimum distortion caused by pruning j weights in the first i layers, and the state and The transfer equation between is:

[0043] ;

[0044] in, , ;

[0045] During the search, through the array Store pruning decisions for each layer .

[0046] In a second aspect, the present invention provides a lightweight system of an intelligent fault diagnosis model based on dynamic trade-off tailoring, which adopts the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described above, including:

[0047] An objective function building module is used to build an objective optimization function with the goal of minimizing the distortion after clipping under the premise of specifying the minimum clipping rate constraint;

[0048] A model distortion acquisition module is used to obtain the output distortion of the fault diagnosis model after different trimmings of a single layer;

[0049] An objective function conversion module is used to determine the additivity of the output distortion of the fault diagnosis model through the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence; and to convert the objective optimization function according to the additivity of the output distortion of the fault diagnosis model;

[0050] The solution module is used to solve the transformed target optimization function by dynamic programming; during the solution process, the state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model.

[0051] In a third aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off trimming as described above are implemented.

[0052] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described above.

[0053] It can be seen from the above technical solutions that, compared with the prior art, the present invention has the following beneficial effects:

[0054] The present invention first proves the additivity of the output distortion of the network model through the two assumptions of Taylor first-order expansion approximation and independent weight pruning between layers, and concludes that the total distortion is the accumulation of the distortions of each layer; then uses this conclusion to formulate the pruning problem as a combinatorial optimization problem, and uses dynamic programming to effectively solve it. By decomposing the complex decision problem into sub-problems, linear time complexity is achieved, and finally unstructured pruning is completed, achieving the best balance between FLOPs and accuracy.

[0055] The present invention comprehensively considers the mutual influence between layers, reduces the fault diagnosis model parameters and the computational complexity; searches for the pruning strategy through the dynamic programming optimization algorithm, greatly reduces the time complexity of the optimization strategy search algorithm, and reduces the dependence on computing resources.

[0056] The present invention can perform unstructured pruning on the trained fault diagnosis network to achieve model lightweighting, which solves the problem of lightweighting of mechanical equipment fault diagnosis network model in edge computing environment. The lightweight neural network effectively reduces the consumption of computing resources, improves computing efficiency and processing speed, is suitable for computing power-constrained devices in engineering environments, and can realize real-time fault diagnosis, which is of great significance to the promotion and application of fault diagnosis networks based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0058] Figure 1 A flow chart of a lightweight method for an intelligent fault diagnosis model based on dynamic trade-off tailoring provided by the present invention;

[0059] Figure 2 A schematic diagram of lightweighting the fault diagnosis model provided by the present invention;

[0060] Figure 3 A schematic diagram of a lightweight method for an intelligent fault diagnosis model based on dynamic trade-off tailoring provided by the present invention;

[0061] Figure 4 This is a performance comparison chart between the method of the present invention and the existing cutting method. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] like Figure 1-Figure 3 As shown, the embodiment of the present invention discloses a lightweight method of an intelligent fault diagnosis model based on dynamic trade-off tailoring, comprising the following steps:

[0064] S1. Determine the fault diagnosis model of the mechanical equipment to be pruned;

[0065] S2, under the premise of specifying the minimum cropping rate constraint, with the goal of minimizing the distortion after cropping, construct a target optimization function;

[0066] S3, obtaining the output distortion of the fault diagnosis model after different trimmings of a single layer;

[0067] S4. Determine the additivity of the output distortion of the fault diagnosis model through the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence; transform the target optimization function according to the additivity of the output distortion of the fault diagnosis model;

[0068] S5. Solving the transformed target optimization function using a dynamic programming method;

[0069] S6. During the solving process, a state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model; each search for output distortions of all layers of the fault diagnosis model is regarded as a trimming process;

[0070] S7, iterating the trimming process until a preset number of trimming iterations is reached.

[0071] The above steps are further explained below.

[0072] S1. Determine the fault diagnosis model of the mechanical equipment to be pruned.

[0073] Before pruning, the mechanical equipment fault diagnosis model needs to be trained. The fault diagnosis model is expressed as , which is generally composed of multiple convolutional layers, pooling layers and fully connected layers, and finally uses the Softmax function to predict the health category. represents the diagnosed sample, a represents the signal sampling length, and b represents the number of signal channels; represents the weight parameters of each layer, represents the weight of the i-th layer of the fault diagnosis model F, , l represents the total number of layers of the fault diagnosis model.

[0074] After the model is trained, set the hyperparameters of pruning, including pruning rate, number of pruning iterations, etc.

[0075] S2. Construct the target optimization function.

[0076] The specific process is: Assume that the weight parameter of the fault diagnosis model F after trimming is ,use and The Euclidean distance between them represents the effect of clipping. Under the constraint, the objective optimization function of the clipping strategy is to minimize the distortion after clipping, and its expression is:

[0077] ;

[0078] ;

[0079] in, Represents a parameter set norm, which is equal to the number of non-zero elements in the parameter set, that is, the constraint function is that the ratio of the total number of parameters after pruning to the total number of parameters before pruning is less than the minimum specified pruning rate .

[0080] S3. Obtain the output distortion of the fault diagnosis model after different trimmings of a single layer, including:

[0081] S31, traverse each layer of the fault diagnosis model, obtain the weight of each layer, and gradually increase the weight of each layer. The norm determines the order of weight pruning, giving priority to pruning Weights with smaller norms;

[0082] S32, output distortion of different levels of clipping of each layer ; It represents the distortion caused by clipping k weights of the i-th layer. k is used to measure the degree of clipping. The larger k is, the greater the degree of clipping.

[0083] S4. Transformation of the target optimization function, specifically including:

[0084] S41, give a parameter disturbance to each layer weight of the fault diagnosis model F , the weight is changed to , the output after weight modification is approximately expressed by the first-order Taylor expansion:

[0085] ;

[0086] in, Represents the weight The partial derivative of .

[0087] S42. Assume parameter perturbations between different layers Independent of each other, there exists: ;

[0088] Among them, i and m represent different layers respectively. Indicates expectation, represents the parameter perturbation of the i-th layer weight, represents the parameter perturbation of the m-th layer weights.

[0089] S43, distortion after clipping According to the first-order Taylor expansion formula, it can be expanded into:

[0090] ;

[0091] Where T represents the transpose of the matrix, represents the parameter perturbation of the i-th layer weight, represents the parameter perturbation of the m-th layer weights.

[0092] Taking the expected value on both sides of the formula, we get:

[0093] .

[0094] S44. According to the independence assumption in S42, when When , the cross term is 0, that is: , then the distortion is expressed as:

[0095] .

[0096] S45. According to the first-order Taylor expansion formula of S41, we get:

[0097] .

[0098] S46. Substitute the formula of S45 into the distortion representation equation in S44. The left side of the equation becomes the total output distortion when all layers are clipped. , the right side becomes the sum of the output distortions caused by individually trimming the weights of each single layer, transforming the objective optimization function into:

[0099] .

[0100] in, represents the output distortion when only the i-th layer is pruned. The new objective optimization function reflects the additivity of the model output distortion, that is, the total distortion is the cumulative result of the distortion of each layer.

[0101] S5. Build a dynamic programming model.

[0102] Using the additivity of the model output distortion, the objective optimization function can be reformulated as a combinatorial optimization problem. This problem can be effectively solved using a dynamic programming model, which is expressed as:

[0103] ;

[0104] in, represents the total number of weights to be pruned, represents the number of weights to be pruned in layer i; represents the output distortion after pruning the i-th layer.

[0105] S6. Construct the state transfer equation.

[0106] Define dp as a state transfer function, and define The minimum distortion caused by pruning j weights in the first i layers is the final goal of the search, which is the distortion of all layers. . For initialization, the state transfer equation is:

[0107] ;

[0108] in, represents the distortion caused when pruning j weights at layer 1.

[0109] state and The transfer equation between is:

[0110] ;

[0111] in, , , represents the distortion caused by clipping k weights in the i-th layer.

[0112] During the search, through the array Store pruning decisions for each layer .

[0113] After searching the output distortions of all layers of the fault diagnosis model once, it is regarded as a pruning process.

[0114] S7, iterating the trimming process until a preset number of trimming iterations is reached.

[0115] In other embodiments, the present invention further provides a lightweight system of an intelligent fault diagnosis model based on dynamic trade-off tailoring, which adopts the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described above, including:

[0116] An objective function building module is used to build an objective optimization function with the goal of minimizing the distortion after clipping under the premise of specifying the minimum clipping rate constraint;

[0117] A model distortion acquisition module is used to obtain the output distortion of the fault diagnosis model after different trimmings of a single layer;

[0118] An objective function conversion module is used to determine the additivity of the output distortion of the fault diagnosis model through the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence; and to convert the objective optimization function according to the additivity of the output distortion of the fault diagnosis model;

[0119] The solution module is used to solve the transformed target optimization function by dynamic programming; during the solution process, the state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model.

[0120] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off trimming as described above are implemented.

[0121] In other embodiments, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described above are implemented.

[0122] Next, taking the fault diagnosis of the electromechanical hybrid transmission system of a rail train as an example, the effectiveness of the method of the present invention was verified through the fault simulation data of the test bench.

[0123] The experimental bench used in this experiment is controlled by a frequency converter to control different motor speeds, and the electro-hydraulic load device applies different lateral loads. The experiment collects 24 channels and 4 types of signals with a sampling frequency of 64kHz, including the three-axis acceleration of the motor drive end, the motor non-drive end, the gearbox input end, the gear output end, the left and right axle boxes, the three-phase current and speed of the motor, and the sound intensity of the left and right axle boxes. In the experiment, 17 health states of the electromechanical hybrid transmission system were considered, including the normal state, 4 axle box bearing faults (inner ring fault, outer ring fault, rolling element fault and cage fault), 4 pinion support bearing faults, 4 pinion faults (cracks, wear, broken teeth and missing teeth), and 4 motor faults (bearing fault, short circuit, rotor bending and rotor broken bars). Each health state includes 2040 samples, and each sample includes 3200 sampling points.

[0124] The experiment uses the method proposed in the present invention to perform fault diagnosis on the above signal samples. The experiment compares the method proposed in the present invention with four commonly used methods: Lamp, Erk, Glob and Unif. In the iterative pruning scheme, it starts with the same pre-trained fault diagnosis network model. During fine-tuning, a certain proportion of parameters are gradually pruned at each iteration stage, and the pruning proportion decreases exponentially with each iteration of 10 iterations. The basic structure of the network is shown in Table 1, and the training related parameters are shown in Table 2. Table 3 and Figure 4 The average values ​​of classification accuracy, sparsity, and FLOPs for 10 experimental diagnoses on 17 health status sample datasets are listed.

[0125] Table 1. Summary of basic network structure

[0126]

[0127] Table 2. Summary of training related parameters

[0128]

[0129] Table 3. Summary of accuracy at the final stage

[0130]

[0131] It can be seen from the experimental results that the method of the present invention utilizes the additivity of output distortion in unstructured pruning to transform the layer-by-layer sparse optimization problem into a dynamic joint fast pruning problem, which can achieve the best balance between FLOPs and accuracy.

[0132] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0133] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A lightweight method for intelligent fault diagnosis model based on dynamic trade-off tailoring, characterized in that: The following steps are involved: S1. Determine the fault diagnosis model of the mechanical equipment to be pruned; S2, under the premise of specifying the minimum cropping rate constraint, with the goal of minimizing the distortion after cropping, construct a target optimization function; S3, obtaining the output distortion of the fault diagnosis model after different trimmings of a single layer; S4. The additivity of the output distortion of the fault diagnosis model is determined by the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence. According to the additivity of the output distortion of the fault diagnosis model, the target optimization function is transformed; S4 include: S41, give a parameter disturbance Δθ to each layer weight of the fault diagnosis model F (1：l) , the weight is changed to The output after weight modification is approximately expressed by first-order Taylor expansion: in, Represents the weight θ (i) The partial derivative of S42, assuming that the parameter perturbations Δθ between different layers (i) Independent of each other, there exists: Among them, i and m represent different layers, E represents the expected value, Δθ (i) represents the parameter perturbation of the i-th layer weight, Δθ (m) represents the parameter perturbation of the mth layer weight; S43, distortion after clipping According to the first-order Taylor expansion formula, it can be expanded into: Where T represents the transpose of the matrix, Δθ (i) represents the parameter perturbation of the i-th layer weight, Δθ (m) represents the parameter perturbation of the mth layer weight; Taking the expected value on both sides of the formula, we get: S44. According to the independence assumption in S42, when i≠m, the cross term is 0, that is: The distortion is expressed as: S45. According to the first-order Taylor expansion formula of S41, we get: S46. Substitute the formula of S45 into the distortion representation equation in S44. The left side of the equation becomes the total output distortion φ when all layers are pruned, and the right side becomes the sum of the output distortions caused by pruning the weights of each single layer separately. The objective optimization function is transformed into: Among them, φ i represents the distortion of the output when only the i-th layer is pruned; S5, solving the transformed target optimization function by using dynamic programming; S6. During the solving process, a state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model; each search for output distortions of all layers of the fault diagnosis model is regarded as a trimming process; S7, iterating the trimming process until a preset number of trimming iterations is reached.

2. The lightweight method of intelligent fault diagnosis model based on dynamic trade-off tailoring according to claim 1 is characterized in that: In S1, the fault diagnosis model is expressed as: F(X, θ (1：l) );in, represents the diagnosed sample, a represents the signal sampling length, b represents the number of signal channels; θ (1：l) represents the weight parameters of each layer, θ (i) represents the weight of the i-th layer of the fault diagnosis model F, 1≤i≤l, and l represents the total number of layers of the fault diagnosis model.

3. The lightweight method of intelligent fault diagnosis model based on dynamic trade-off tailoring according to claim 2 is characterized in that: In S2, the weight parameter of the fault diagnosis model F after trimming is Use F(X,θ (1：l) )and The Euclidean space distance between represents the impact of clipping, and the expression of the objective optimization function is: Among them, ||θ||0 represents the L0 norm of the parameter set, which is equal to the number of non-zero elements in the parameter set; α represents the specified minimum clipping rate.

4. The lightweight method of intelligent fault diagnosis model based on dynamic trade-off tailoring according to claim 3 is characterized in that S3 include: S31, traverse each layer of the fault diagnosis model, obtain the weight of each layer, determine the order of weight pruning according to the L1 norm, and give priority to pruning weights with smaller L1 norms; S32, output the distortion φ of each layer with different degrees of clipping i (k); φ i (k) represents the distortion caused by clipping the k weights of the i-th layer. k is used to measure the degree of clipping. The larger k is, the greater the degree of clipping.

5. The lightweight method of intelligent fault diagnosis model based on dynamic trade-off tailoring according to claim 1 is characterized in that: In S5, a dynamic programming model is constructed to solve the transformed target optimization function. The expression of the dynamic programming model is: min φ1+φ2+…+φ l s.t.p1+p2+...+p l =P; Where P represents the total number of weights to be pruned, p i represents the number of weights to be pruned in layer i.

6. The lightweight method of intelligent fault diagnosis model based on dynamic trade-off tailoring according to claim 5 is characterized in that: In S6, define is the minimum distortion caused by pruning j weights in the first i layers, and the state dp i-1 and dp i The transfer equation between is: Among them, 1≤j≤P,1≤k≤j; During the search, the pruning decision of each layer is stored in array s.

7. A lightweight system of intelligent fault diagnosis model based on dynamic trade-off tailoring, characterized in that: The method adopts a lightweight method of an intelligent fault diagnosis model based on dynamic trade-off tailoring as claimed in any one of claims 1 to 6, comprising: An objective function building module is used to build an objective optimization function with the goal of minimizing the distortion after clipping under the premise of specifying the minimum clipping rate constraint; A model distortion acquisition module is used to obtain the output distortion of the fault diagnosis model after different trimmings of a single layer; An objective function conversion module is used to determine the additivity of the output distortion of the fault diagnosis model through the two assumptions of Taylor first-order expansion approximation and inter-layer weight clipping independence; and to convert the objective optimization function according to the additivity of the output distortion of the fault diagnosis model; The solution module is used to solve the transformed target optimization function by dynamic programming; during the solution process, the state transfer equation is constructed to search for output distortions of all layers of the fault diagnosis model.

8. An electronic device, comprising: A memory, a processor and a computer program stored in the memory and executable on the processor; characterized in that when the processor executes the computer program, the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the lightweight method of the intelligent fault diagnosis model based on dynamic trade-off tailoring as described in any one of claims 1 to 6 are implemented.

Images