A permanent magnet wheel hub motor demagnetization fault diagnosis method considering multi-source signal fusion

Abstract

The application discloses a kind of permanent magnet wheel hub motor demagnetization fault diagnosis methods considering multi-source signal fusion, belong to motor fault diagnosis technical field, utilize the fusion method of multi-source heterogeneous signal feature to carry out permanent magnet wheel hub motor demagnetization fault diagnosis, solve the dependence on additional sensor.The method mainly includes: collecting current, back electromotive force, vibration and other multi-source signals under different conditions;Multi-source signals are preliminarily processed by symmetry point mode, and SDP parameters are adaptively optimized and fused according to image fitness, and image processing is carried out again;The image database of multi-source signals is used for fault diagnosis model training, and the diagnosis accuracy of each signal is obtained;Dynamic decision environment is built by combining environmental conditions, image fitness and fault diagnosis accuracy, and the optimal selection of multi-source heterogeneous signals under different conditions is realized by training model.The application improves the accuracy and efficiency of motor fault diagnosis by signal selection and fusion under different conditions.

Description

Technical Field

[0001] This invention belongs to the field of motor fault diagnosis technology, specifically relating to a method for diagnosing demagnetization faults in permanent magnet hub motors that considers multi-source signal fusion. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) have become the most widely used drive motors in electric vehicles due to their outstanding advantages such as high energy density, strong overload capacity, fast response, low cost, and simple maintenance. They are also undergoing a shift from centralized, wheel-side to hub drive modes, with PMSMs gradually becoming the mainstream in electric vehicle drive systems. However, this new drive form faces complex operating conditions such as road impacts, high ambient temperatures, and sudden load changes. Electrical faults induced by these conditions, such as winding short circuits, inter-turn short circuits, and permanent magnet demagnetization, directly affect their performance and may even jeopardize the reliability and safety of electric vehicles. Therefore, monitoring the operating status of PMSMs and diagnosing motor faults under various operating conditions is essential. When a motor fault occurs, the status signals also change. Signals commonly used for motor fault diagnosis include vibration signals, back electromotive force signals, and current signals.

[0003] CN114264953A proposes a method, system, and device for diagnosing demagnetization faults in permanent magnet synchronous motors (PMSMs). The method first acquires three-phase current signals, then performs two-dimensional image preprocessing on these signals and inputs the processed images into a sample library. Next, the sample library is divided into training and testing samples, and a convolutional neural network model is trained using the training samples to obtain a model capable of diagnosing demagnetization faults. However, this method relies solely on a single three-phase current signal for fault diagnosis. Since a single signal is easily affected by environmental factors such as electromagnetic interference and temperature changes, the diagnostic accuracy may decrease under different environments. To address this issue, CN116597167A proposes a small-sample demagnetization fault diagnosis method for PMSMs. This method acquires the motor's vibration and current signals, converts the processed results into grayscale images, fuses them, and then uses an improved meta-learning network to construct a fault classifier. This method can achieve multi-sensor information fusion without increasing the number of data samples, improving the stability and accuracy of the diagnosis. However, this method also introduces some new problems when fusing multiple signals. On the one hand, the fusion of multiple signals significantly increases the complexity of image information, posing a greater challenge to the fault diagnosis process and increasing the difficulty of diagnosis. On the other hand, noise in the signals is not effectively filtered out. After the fusion of multiple signals, a single image often contains multiple noise signals, which further interfere with the accuracy of fault diagnosis. Furthermore, although this method has made some optimizations and improvements to the selection of the optimal signal among multiple signals and the parameter matching and adjustment of SDP, the parameter adjustment process requires a significant amount of time. Moreover, to ensure high applicability under different data and conditions, continuous parameter adjustment and optimization are necessary, which undoubtedly increases the difficulty and cost in practical applications. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a method for diagnosing demagnetization faults in permanent magnet hub motors that considers multi-source signal fusion. First, the signal is converted into an SDP image. Second, the parameters of the SDP are optimized and adjusted using the whale optimization algorithm to obtain a highly adaptive image. Reinforcement learning is then employed to select and fuse single signals and signal combinations. The design is reasonable, overcomes the shortcomings of existing technologies, and achieves good results.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for diagnosing demagnetization faults in permanent magnet hub motors that considers multi-source signal fusion includes the following steps:

[0007] Step 1: Under different environmental conditions, collect multi-source signals of the permanent magnet hub motor under different operating conditions, including current, back electromotive force, and vibration signals;

[0008] Step 2: Perform preliminary SDP image processing on each signal, calculate the image fitness, and use the whale algorithm to optimize the SDP parameters. Based on the optimization results, perform image processing on each signal again.

[0009] Step 3: Based on the image results of multi-source signals, establish fault diagnosis models corresponding to different signals respectively, and test the accuracy of each model.

[0010] Step 4: Based on environmental conditions and accuracy, build a dynamic decision-making environment. Based on image fitness and fault diagnosis accuracy, obtain the optimal solution for signal selection under different environmental conditions for fault diagnosis.

[0011] Further, in step 2, calculating the image fitness specifically involves:

[0012] First, the grayscale images of different fault types are converted into a pixel matrix G, expressed as:

[0013]

[0014] In the formula, L is the number of rows in the pixel matrix; W is the number of columns in the pixel matrix; G(h,y) is the pixel value corresponding to point (h,y), h=1,...,L,y=1,...,W;

[0015] make and Let H and Y represent the feature values ​​of any two images within the range N×N at position (h, y); and let Diff represent the Euclidean distance between the k-th image and the u-th image. k,u Represented as:

[0016]

[0017] Considering the impact of noise on the image, a smoothness function is introduced to neutralize the image discrimination function. Let Ω be the neighborhood of the pixel at point (h,y) within the N×N range. Let (sh, sy) represent the smoothness value at pixel (sh, sy). Then, the formula for the smoothness of the image at that point is:

[0018]

[0019] make Let Ω be the smoothness matrix corresponding to the neighborhood Ω, and the total number of elements in the matrix is ​​(L-N+1)×(W-N+1). Then, the formula for calculating the intra-class smoothness S of the image used to quantize the image is denoted as:

[0020]

[0021] The fitness Diff(·) of the overall image is calculated using the following formula:

[0022]

[0023] Among them, a s As a composite factor, N z S represents the number of images. k and S u denoted as the intra-class smoothness of the k-th image and the u-th image, respectively.

[0024] Furthermore, in step 2, the whale algorithm specifically comprises:

[0025] Whale predation and search behavior is represented as:

[0026]

[0027] In the formula, X rand (t) represents the target location of a randomly selected whale individual from the current group, X(t) represents the current location of the whale individual, and w rand To enclose the step size, representing the distance between the current whale individual and a randomly selected individual, the mathematical expressions for coefficients A and C are:

[0028]

[0029] In the formula, the control parameter a gradually decreases from 2 to 0 as the number of iterations increases; r1 and r2 are random numbers distributed in [0, 1], and the smaller the algebraic value of A, the shorter the step size.

[0030] The mathematical model of the contraction encirclement mechanism is expressed as follows:

[0031]

[0032] In the formula, X best (t) represents the current optimal feeding position for the whale, w best To achieve the optimal encirclement step size;

[0033] The mathematical model for the spiral predation position update phase is expressed as:

[0034] X(t+1)=|X best (t)-X(t)|·e bl ·cos(2πl)+X best (t); (9)

[0035] In the formula, |X best (t)-X(t)| represents the distance between the current whale position and the optimal predation position; b is the constant coefficient of the spiral encirclement route; l is a random number between [-1, 1];

[0036] In the algorithm, the whale's search for encirclement and spiral predation behaviors occur simultaneously. The search for encirclement mechanism includes two parts: shrinking encirclement and spiral predation. Here, the probability of both behaviors is set to 0.5. The formula for updating the position of an individual whale is expressed as follows:

[0037]

[0038] In addition, the value of |A| is the main factor that determines the behavior of whales. When |A|>1, a global search is performed and the target position is updated by formula (6); when |A|<1, a local search is performed and the position is updated by formula (8) by referring to the optimal individual position, where p is a random probability factor.

[0039] Furthermore, step 2, optimizing the SDP parameters using the whale algorithm, includes the following sub-steps:

[0040] Step 2.1: Input is a partial diagnostic signal Z = [E1, E2, ..., E n ];

[0041] Step 2.2: Randomly generate multiple sets of initial parameters, each set including two parameters τ and κ. Perform SDP image processing on the input diagnostic signal to obtain a set of initial feature maps Z. p =[P1,P2...,P n ];

[0042] Step 2.3: Calculate the fitness Diff() for each group of images and calculate the current optimal solution;

[0043] Step 2.4: Update the position according to formula (6), find the global optimal solution, and perform local optimization according to formulas (8) and (9). Use l as a random number to perform local optimal iterative solution.

[0044] Step 2.5: Substitute the optimal solution from this step into the next optimization process;

[0045] Step 2.6: Repeat steps 2.2 to 2.5 multiple times to obtain the optimal parameters τ′ and κ′;

[0046] Based on the solution results of step 2.6, the diagnostic signals are processed into images to obtain a database with high image discrimination.

[0047] Furthermore, in step 3, an image classification algorithm based on a convolutional neural network (CNN) is used to classify faults in the database. The activation function in the convolutional neural network is the ReLU activation function, the Softmax activation function is used for multi-class classification, the Adam optimizer is used for weight updates, and the loss function is sparse classification cross-entropy.

[0048] Furthermore, in step 4, the overall state environment S is established. t This includes temperature T, load L, rotational speed R, and battery power status B, represented by a vector S. t =[T t ,L t ,R t B t Action A t To select signals from the candidate signal set, specifically including: motor no-load back electromotive force signal EMF, vibration signal Vib, motor current signal I, back electromotive force signal fused with vibration signal [EMF,Vib], vibration signal fused with motor current signal [Vib,I], and back electromotive force signal fused with motor current signal [EMF,I]; the action space is defined as the combination A of signals. t ∈{EMF,Vib,I,[EMF,Vib],[Vib,I],[EMF,I]};Reward function R t Taking into account both the accuracy of signal selection (Acc) and the image fitness function (Fit), it is defined as R. t =α·Acc + β·Fit, where α and β are weighting coefficients used to balance the influence of the two indicators;

[0049] The goal of reinforcement learning is to learn the optimal signal selection policy by maximizing cumulative reward, expressed as:

[0050]

[0051] Among them, G t For steps from step 1 to step n d The cumulative reward for each step, γ∈[0,1] is the discount factor, and K is the number of iterations;

[0052] Deep Q-Network (DQN) approximates the action-value function Q(S) through a neural network. t A t This allows for the evaluation of the value of state-action pairs. The Q-function update formula is:

[0053] Q(S t A t )←Q(S t A t )+δ·[R t +γ·max A′ Q(S t+1 ,A′)-Q(S t A t (12)

[0054] Where δ is the learning rate, R t For the current reward, max A′ Q(St+1 ,A′) is the maximum Q value of the next state.

[0055] The beneficial technical effects of this invention are as follows:

[0056] This invention can systematically and dynamically learn the optimal signal selection strategy, adapting to different environmental conditions (temperature, load, speed, battery power status). The DQN algorithm combined with a deep neural network can efficiently approximate Q(S) t A t This enables signal optimization in complex state spaces. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the fault diagnosis method for permanent magnet hub motors based on data analysis according to the present invention.

[0058] Figure 2 This is a schematic diagram of data processing based on reinforcement learning according to the present invention.

[0059] Figure 3 This is a schematic diagram of the database processing based on the WOA algorithm optimization of the present invention. Detailed Implementation

[0060] The specific embodiments of the present invention will be further described below with reference to specific examples:

[0061] A method for diagnosing demagnetization faults in permanent magnet hub motors considering multi-source signal fusion, such as... Figure 1 As shown, it includes the following steps:

[0062] Step 1: Under different environmental conditions, collect multi-source signals of the permanent magnet hub motor under different operating conditions, including current, back electromotive force, and vibration signals;

[0063] Step 2: Perform preliminary SDP image processing on each signal, calculate image fitness, and fine-tune the SDP parameters using the whale algorithm. Based on the fine-tuning results, re-process each signal into an image, such as... Figure 3 As shown;

[0064] SDP (Simulated Graph Processing) is a mapping algorithm from Cartesian to polar coordinates. It can intuitively describe the amplitude variations and frequency characteristics of time-series signals. Through simple calculations, it can convert the signal between time i and time i+τ in a time-domain time series x into points with corresponding radii and rotation angles in polar coordinates. The specific conversion formula can be expressed as:

[0065]

[0066] In the formula, x max With x minα represents the maximum and minimum values ​​in the time series x; α0 is the initial angle for plotting; κ is the angular domain gain factor; and τ is the time delay coefficient.

[0067] To quantify the discriminative power of SDP images, a fitness function that reflects this discriminative power needs to be established. First, grayscale images of different fault types are converted into pixel matrices. The pixel matrix features at any point in the image are represented by the average value, and the representational differences between different images are defined based on Euclidean distance. Simultaneously, considering the impact of noise on the image, a smoothness function is introduced for control. Finally, the image discriminative power and smoothness are combined to obtain the image fitness.

[0068] The specific steps for calculating image fitness are as follows:

[0069] First, the grayscale images of different fault types are converted into a pixel matrix G, expressed as:

[0070]

[0071] In the formula, L is the number of rows in the pixel matrix; W is the number of columns in the pixel matrix; G(h,y) is the pixel value corresponding to point (h,y), h=1,...,L,y=1,...,W;

[0072] make and Let H and Y represent the feature values ​​of any two images within a range of N×N at position (h, y); then the feature value of the k-th image is... It can be represented as:

[0073]

[0074] Image feature values ​​of the u-th image The corresponding formula can be expressed as:

[0075]

[0076] The pixel difference between the k-th image and the u-th image is calculated as follows:

[0077]

[0078] use Reference The elements in the matrix are... Reference The elements in the matrix. Then the Euclidean distance Diff between corresponding points in the two images. k,u It can be represented as:

[0079]

[0080] Since the SDP method discards pixels with essentially blank outer edges during image transformation when calculating Euclidean distance, it calculates the sum of differences between (L-N+1)×(L-N+1) pixel pairs. The Euclidean distance between the k-th and u-th images represents the difference based on the Euclidean distance definition, and its formula is:

[0081]

[0082] This formula quantifies the overall difference between images by calculating the feature value at each pixel location. In image processing, a smaller Euclidean distance value indicates a higher similarity between the features of two images, while a larger value indicates a greater difference in features between the two images. Ideally, the greater the difference between different types of images, the better, because a larger difference means better differentiation of different fault modes. However, in practical applications, noise is an unavoidable factor in image processing. The presence of noise may perturb image features, thus affecting the calculation result of Euclidean distance and reducing the distinguishability between different fault modes. Therefore, how to effectively suppress the influence of noise when calculating Euclidean distance has become an urgent problem to be solved in image difference measurement. To this end, in addition to calculating Euclidean distance, the influence of noise on the image is also considered, and a smoothness function is introduced to neutralize the image discrimination function. Assuming Ω is the neighborhood of the pixel at point (h,y) in the range of N×N, Let (sh, sy) represent the smoothness value at pixel (sh, sy). Then, the formula for the smoothness of the image at that point is:

[0083]

[0084] The smoothness value of all points in the entire image can be written as:

[0085]

[0086] make Let Ω be the smoothness matrix corresponding to the neighborhood Ω, and the total number of elements in the matrix is ​​(L-N+1)×(W-N+1). Then, the formula for calculating the intra-class smoothness S of the image used to quantize the image is denoted as:

[0087]

[0088] The formula for calculating image fitness, considering the image difference and intra-image smoothness between the k-th and u-th images, can be expressed as:

[0089]

[0090] Among them, Diff k,u (·) is the fitness calculation function between the k-th image and the u-th image, Sk and S u The intra-class smoothness of the k-th image and the u-th image are respectively; a s The composite factor adjusts the ratio between image discrimination and smoothness functions. The larger the composite factor, the greater the weight of smoothness, and the more sensitive the image is to data noise; the smaller the composite factor, the more emphasis is placed on image discrimination, and the easier it is to distinguish between different types of images.

[0091] After processing the motor signal into an image using the SDP algorithm, the fitness of the overall image is calculated using the formula described above. The formula can be written as:

[0092]

[0093] Where, N z This represents the number of images.

[0094] By calculating the image fitness, the quality of the angular domain gain factor κ and the time delay coefficient τ in the SDP parameters can be quantified as a whole. When the original motor signal is converted into a polar coordinate graph, if the angular domain gain factor κ and the time delay coefficient τ are too large or too small, the image symmetry will be poor, the fitness result will be poor, and the image features will not be obvious.

[0095] The whale algorithm is as follows:

[0096] Whale predation and search behavior is represented as:

[0097]

[0098] In the formula, X rand (t) represents the target location of a randomly selected whale individual from the current group, X(t) represents the current location of the whale individual, and w rand To enclose the step size, representing the distance between the current whale individual and a randomly selected individual, the mathematical expressions for coefficients A and C are:

[0099]

[0100] In the formula, the control parameter a gradually decreases from 2 to 0 as the number of iterations increases; r1 and r2 are random numbers distributed in [0, 1], and the smaller the algebraic value of A, the shorter the step size.

[0101] The mathematical model of the contraction encirclement mechanism is expressed as follows:

[0102]

[0103] In the formula, X best (t) represents the current optimal hunting position for the whale, w best To achieve the optimal encirclement step size;

[0104] The mathematical model for the spiral predation position update phase is expressed as:

[0105] X(t+1)=|X best (t)-X(t)|·e bl ·cos(2πl)+X best (t); (16)

[0106] In the formula, |X best (t)-X(t)| represents the distance between the current whale position and the optimal predation position; b is the constant coefficient of the spiral encirclement route; l is a random number between [-1, 1];

[0107] In the algorithm, the whale's search for encirclement and spiral predation behaviors occur simultaneously. The search for encirclement mechanism includes two parts: shrinking encirclement and spiral predation. Here, the probability of both behaviors is set to 0.5. The formula for updating the position of an individual whale is expressed as follows:

[0108]

[0109] In addition, the value of |A| is the main factor that determines the behavior of whales. When |A|>1, a global search is performed and the target position is updated by formula (13); when |A|<1, a local search is performed and the position is updated by formula (15) by referring to the optimal individual position, where p is a random probability factor.

[0110] Optimizing SDP parameters using the whale algorithm includes the following sub-steps:

[0111] Step 2.1: Input is a partial diagnostic signal Z = [E1, E2, ..., E n ];

[0112] Step 2.2: Randomly generate multiple sets of initial parameters, each set including two parameters τ and κ. Perform SDP image processing on the input diagnostic signal to obtain a set of initial feature maps Z. p =[P1,P2...,P n ];

[0113] Step 2.3: Calculate the fitness Diff() for each group of images and calculate the current optimal solution;

[0114] Step 2.4: Update the position according to formula (13), find the global optimal solution, and perform local optimization according to formulas (15) and (16). Use l as a random number to perform local optimal iterative solution.

[0115] Step 2.5: Substitute the optimal solution from this step into the next optimization process;

[0116] Step 2.6: Repeat steps 2.2 to 2.5 multiple times to obtain the optimal parameters τ′ and κ′;

[0117] Based on the solution results of step 2.6, the diagnostic signals are processed into images to obtain a database with high image discrimination. CNN fault diagnosis models are then trained, and a dynamic decision-making environment is established for reinforcement learning training.

[0118] Step 3: Based on the image results of the multi-source signals, establish corresponding fault diagnosis models and test their accuracy respectively, such as... Figure 2 As shown;

[0119] An image classification algorithm based on Convolutional Neural Networks (CNNs) is used to classify faults in a database. The CNN is the basic convolutional network structure, employing ReLU activation and Softmax activation for multi-class classification. The Adam optimizer is used for weight updates, and the loss function is sparse classification cross-entropy, suitable for integer labels.

[0120] Step 4: Based on environmental conditions and accuracy, build a dynamic decision-making environment. Based on image fitness and fault diagnosis accuracy, obtain the optimal solution for signal selection under different environmental conditions for fault diagnosis.

[0121] Establish the overall state environment S t This includes temperature T, load L, rotational speed R, and battery power status B, represented by a vector S. t =[T t ,L t ,R t B t Action A t To select signals from the candidate signal set, specifically including: motor no-load back electromotive force signal EMF, vibration signal Vib, motor current signal I, back electromotive force signal fused with vibration signal [EMF,Vib], vibration signal fused with motor current signal [Vib,I], and back electromotive force signal fused with motor current signal [EMF,I]; the action space is defined as the combination A of signals. t ∈{EMF,Vib,I,[EMF,Vib],[Vib,I],[EMF,I]};Reward function R t Taking into account both the accuracy of signal selection (Acc) and the image fitness function (Fit), it is defined as R. t =α·Acc + β·Fit, where α and β are weighting coefficients used to balance the influence of the two indicators;

[0122] The goal of reinforcement learning is to learn the optimal signal selection policy by maximizing the cumulative reward (expected discounted return), expressed as:

[0123]

[0124] Among them, G t For steps from step 1 to step n d The cumulative reward for each step, γ∈[0,1] is the discount factor, and K is the number of iterations;

[0125] Deep Q-Network (DQN) approximates the action-value function Q(S) through a neural network. t A t This allows for the evaluation of the value of state-action pairs. The Q-function update formula is:

[0126] Q(S t A t )←Q(S t A t )+α·[R t +γ·max A′ Q(S t+1 ,A′)-Q(S t A t (19)

[0127] Where δ is the learning rate, R t For the current reward, max A′ Q(S t+1 ,A′) is the maximum Q value of the next state.

[0128] The reinforcement learning scheme described above can systematically and dynamically learn the optimal signal selection strategy to adapt to different environmental conditions (temperature, load, speed, battery power status). The DQN algorithm combined with a deep neural network can efficiently approximate Q(S) t A t This enables signal optimization in complex state spaces.

[0129] The trained diagnostic model and classifier were used to extract features and classify the target domain dataset, verifying the reliability and accuracy of the trained model in detecting and diagnosing motor demagnetization faults under various conditions.

[0130] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for diagnosing demagnetization faults in permanent magnet hub motors considering multi-source signal fusion, characterized in that, Includes the following steps: Step 1: Under different environmental conditions, collect multi-source signals of the permanent magnet hub motor under different operating conditions, including current, back electromotive force, and vibration signals; Step 2: Perform preliminary SDP image processing on each signal, calculate the image fitness, and use the whale algorithm to optimize the SDP parameters. Based on the optimization results, perform image processing on each signal again. Step 3: Based on the image results of multi-source signals, establish fault diagnosis models corresponding to different signals respectively, and test the accuracy of each model. Step 4: Based on environmental conditions and accuracy, build a dynamic decision-making environment. Based on image fitness and fault diagnosis accuracy, derive the optimal solution for signal selection under different environmental conditions for fault diagnosis. In step 2, the whale algorithm specifically refers to: Whale predation and search behavior is represented as: ； (6) In the formula, To randomly select the target location of an individual whale from the current group, This indicates the current location of the individual whale. The bounding step size represents the distance between the current whale individual and a randomly selected individual, with a coefficient. and The mathematical expression is: ； (7) In the formula, the control parameters The value gradually decreases from 2 to 0 as the number of iterations increases; , All are random numbers distributed in [0, 1]. The smaller the algebraic value, the shorter the step size; The mathematical model of the contraction encirclement mechanism is expressed as follows: ； (8) In the formula, This is currently the optimal hunting location for whales. To achieve the optimal encirclement step size; The mathematical model for the spiral predation position update phase is expressed as: ； (9) In the formula, is the distance between the current whale position and the optimal predation position; b is the constant coefficient of the spiral encirclement route; l is a random number between [−1, 1]; In the algorithm, the whale's search for encirclement and spiral predation behaviors occur simultaneously. The search for encirclement mechanism includes two parts: shrinking encirclement and spiral predation. Here, the probability of both behaviors is set to 0.

5. The formula for updating the position of an individual whale is expressed as follows: ； (10) also, Numerical values ​​are the main factor determining whale behavior, when A global search is performed, and the target position is updated using formula (6); when Local search is performed, and the position is updated using formula (8) by referring to the optimal individual position. It is a random probability factor; Step 2, which optimizes the SDP parameters using the whale algorithm, includes the following sub-steps: Step 2.1: Input is a partial multi-source signal ; Step 2.2: Randomly generate multiple sets of initial parameters, each set including two parameters. , The input diagnostic signal is processed using SDP image processing to obtain a set of initial feature maps. ; Step 2.3: Calculate the fitness Diff() for each group of images and calculate the current optimal solution; Step 2.4: Update the position according to formula (6), find the global optimal solution, and perform local optimization according to formulas (8) and (9). Use l as a random number to perform local optimal iterative solution. Step 2.5: Substitute the optimal solution from this step into the next optimization process; Step 2.6: Repeat steps 2.2 to 2.5 multiple times to obtain the optimal parameters. and ; Based on the solution results of step 2.6, the diagnostic signals are processed into images to obtain a database with high image discrimination. In step 4, the overall state environment S is established. t This includes temperature T, load L, rotational speed R, and battery power status B, represented by a vector. Action A t To select signals from the candidate signal set, specifically including: motor no-load back electromotive force (EMF) signal, vibration signal Vib, motor current signal I, back EMF signal fused with vibration signal [EMF, Vib], vibration signal fused with motor current signal [Vib, I], and back EMF signal fused with motor current signal [EMF, I]; the action space is defined as a combination of signals. Reward function R t Taking into account both the signal selection accuracy Acc and the image fitness function Fit, it is defined as follows: ,in, and It is a weighting coefficient used to balance the influence of the two indicators; The goal of reinforcement learning is to learn the optimal signal selection policy by maximizing cumulative reward, expressed as: ；(11) in, From step 1 to step 2 Cumulative rewards for steps It is a discount factor. This represents the number of iterations. Deep Q-Network (DQN) approximates the action-value function Q(S) through a neural network. t A t This enables the evaluation of the value of state-action states; the Q-function update formula is: ；(12) in, R is the learning rate. t For the current reward, This represents the maximum Q value for the next state.

2. The method for diagnosing demagnetization faults in a permanent magnet hub motor considering multi-source signal fusion as described in claim 1, characterized in that, In step 2, the image fitness is calculated as follows: First, the grayscale images of different fault types after SDP image processing are converted into pixel matrices. The expression is: ； (1) In the formula, L is the number of rows in the pixel matrix; W is the number of columns in the pixel matrix. For point The corresponding pixel value, , ; make and Represent the positions of any two images. The eigenvalues ​​within an N×N range; the Euclidean distance between the k-th image and the u-th image. Represented as: ；(2) Considering the impact of noise on the image, a smoothness function is introduced to neutralize the image discrimination function. Assuming... For point The neighborhood of a pixel within an N×N range, Represents pixels If the smoothness value is given at a certain point, then the formula for the smoothness of the image at that point is: ； (3) make For the neighborhood The corresponding smoothness matrix has a total number of elements. The image is then used to quantify the intra-class smoothness of the image. The calculation formula is denoted as: ； (4) Adaptability to the overall image The calculation is performed, and the formula is denoted as: ；(5) in, It is a composite factor. For the number of images, and denoted as the intra-class smoothness of the k-th image and the u-th image, respectively.

3. The method for diagnosing demagnetization faults in a permanent magnet hub motor considering multi-source signal fusion as described in claim 1, characterized in that, In step 3, an image classification algorithm based on a convolutional neural network (CNN) is used to classify faults in the database. The activation function in the convolutional neural network is the ReLU activation function, the Softmax activation function is used for multi-class classification, the Adam optimizer is used for weight updates, and the loss function is sparse classification cross-entropy.

Images