Farm machinery equipment driving motor fault diagnosis method and system based on signal collaborative analysis

Abstract

The invention discloses an agricultural machinery equipment driving motor fault diagnosis method and system based on signal collaborative analysis, and belongs to the technical field of bearing fault diagnos.The method comprises the steps that a motor vibration signal is subjected to multi-stage preprocessing, the vibration signal is converted into a cochlea graph, and the preprocessed cochlea graph is obtained through preprocessing; calculating an auditory saliency map by counting effective values in the cochlea by adopting a self-adaptive cochlea saliency feature extraction method from the preprocessed cochlea to obtain a saliency feature cochlea; extracting channel level and frame level features based on the significant feature cochlea and fusing the channel level and frame level features to obtain multi-scale fusion features; and forming a fusion feature dictionary through the label samples based on the multi-scale fusion features, performing sparse representation on the test samples by using the constructed dictionary, and realizing fault diagnosis of the agricultural machinery equipment driving motor based on the sparse representation. The method has excellent diagnosis performance in a strong noise environment and a variable speed working condition, and the diagnosis robustness is remarkably improved.

Description

Technical Field

[0001] This application relates to the technical field of bearing fault diagnosis, specifically to a method and system for fault diagnosis of agricultural machinery drive motors based on signal collaborative analysis. Background Technology

[0002] With the widespread application of agricultural machinery in modern agricultural production, the fault diagnosis of drive motors, as their core components, has received increasing attention. Motor failures not only affect the efficiency and quality of agricultural machinery operations but can also lead to equipment downtime, resulting in operational delays and economic losses. Agricultural machinery drive motors consist of various components such as stators, rotors, bearings, and transmission connections. Common fault types include bearing damage, broken rotor bars, shaft imbalance, and gear damage. These faults are typically caused by factors such as overload operation, long-term wear, harsh operating environments, and manufacturing defects. Therefore, timely detection and accurate fault diagnosis are crucial to ensuring the safe and stable operation of agricultural machinery drive motors.

[0003] Currently, common methods for fault diagnosis of drive motors are mainly divided into three categories: analytical model-based methods, knowledge-based methods, and data-driven methods. Model-based fault diagnosis methods primarily establish mathematical models of the motor's operating state and utilize tools such as state observers and sliding mode observers to identify key parameters online. Fault detection is then achieved by analyzing the differences between the model output and actual measured values. However, model-based methods require constructing mathematical models that accurately reflect the target. Due to the complex characteristics of nonlinearity and strong coupling in actual engineering, it is difficult to form accurate mathematical models. Furthermore, even slight variations in the research object necessitate the creation of new models, making it difficult to generalize mechanistic model-based fault diagnosis methods. Knowledge-based methods effectively interpret, reason, diagnose, and predict current diagnoses based on historical data knowledge. They mimic human problem-solving by comprehensively considering multiple factors. For knowledge-based methods, expert experience is crucial; accumulating sufficient experience requires significant resources, and expert knowledge varies across different bearing types, resulting in limited universality. Data-driven fault diagnosis methods utilize data collected from the system for fault detection and diagnosis. Unlike model-based fault diagnosis, which relies on mathematical models of the system, data-driven fault diagnosis uses statistical methods, machine learning, and deep learning techniques to analyze data and identify anomalous patterns that may indicate faults. While some progress has been made in drive motor fault diagnosis, most work still focuses on detection under single operating conditions, with relatively little research on complex conditions such as variable speed and high noise levels. Considering practical engineering needs, data-driven fault diagnosis methods, due to their self-learning ability with massive amounts of data and high robustness, are more suitable as a research direction for fault detection of drive motors in agricultural machinery.

[0004] As the core power unit of agricultural machinery, the drive motor's operating status directly affects the overall machine's efficiency and safety. However, in the complex and ever-changing farmland environment, drive motors often operate at variable speeds, accompanied by strong background noise interference, posing numerous challenges to fault signal extraction and identification. Especially in the early stages of bearing failure, weak fault characteristics are easily masked by environmental noise, and the non-stationary characteristics caused by fluctuations in operating load further increase the difficulty of feature extraction, limiting the effectiveness of existing diagnostic methods. Furthermore, existing fault diagnosis methods based on a single signal source often struggle to balance accuracy and robustness under complex operating conditions, easily leading to misdiagnosis and missed diagnosis. Summary of the Invention

[0005] Purpose of the Invention: In existing technologies, subtle fault characteristics are easily masked by environmental noise, and the non-stationary characteristics caused by fluctuating operating loads further increase the difficulty of feature extraction, limiting the effectiveness of existing diagnostic methods. Existing fault diagnosis methods based on a single signal source often struggle to balance accuracy and robustness under complex operating conditions, easily leading to misdiagnosis and missed diagnosis. This invention provides a fault diagnosis method for agricultural machinery drive motors based on signal co-analysis. This method fully utilizes the complementary characteristics between current and vibration signals, estimating motor speed through current spectrum analysis to suppress the non-stationary effects of variable speed conditions. Simultaneously, it combines vibration signal data preprocessing, multi-scale feature extraction, and a fault diagnosis model based on sparse representation to effectively improve the accuracy and robustness of fault identification. Compared to existing methods, this method exhibits superior fault identification performance under complex farmland conditions, providing reliable technical support for intelligent health monitoring of agricultural machinery drive motors.

[0006] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0007] A fault diagnosis method for drive motors of agricultural machinery based on signal co-analysis includes the following steps:

[0008] Step S1: Based on the engineering background, determine the fault modes that need to be diagnosed, and collect vibration and current signals of the agricultural machinery drive motor during operation in each mode.

[0009] Step S2: By performing spectrum analysis on the collected motor current signal, the frequency components of the motor under variable speed conditions are extracted, and then the motor speed signal is estimated.

[0010] Step S3: After extracting the motor speed signal, perform multi-stage preprocessing on the motor vibration signal to convert the vibration signal into a cochlear image, and then preprocess to obtain the preprocessed cochlear image.

[0011] Step S4: By simulating the auditory selective attention mechanism, the auditory saliency map is calculated by statistically analyzing the effective values ​​in the preprocessed cochlear map using an adaptive cochlear map saliency extraction method, thus obtaining the saliency cochlear map.

[0012] Step S5: Extract channel-level and frame-level features based on the salient cochlear map, and then fuse the two features to obtain multi-scale fused features.

[0013] Step S6: Based on multi-scale fusion features, construct a fusion feature dictionary using labeled samples, use the constructed dictionary to perform sparse representation on the test samples, and realize fault diagnosis of agricultural machinery drive motors based on sparse representation.

[0014] Preferably, the adaptive cochlear map salient feature extraction method in step S4 includes the following steps:

[0015] Initialization: Input the cochlear map composed of valid channels. Significant threshold Th Sal .

[0016] Vectorization:

[0017]

[0018] Among them, M EC The cochlear map matrix, V, is composed of effective channels. EC From M EC The transformed vector, V EC The size is n C1 ·n F1 ×1. The function f1 is a reshaping function that transforms a matrix into its corresponding vector, n. C1 Representation matrix M EC The number of rows, n F1 The representation matrix M represents EC The number of columns.

[0019] Valid element retrieval:

[0020] [V NZ ,P NZ ]=f2(V EC )

[0021] Here, f2 is a non-zero element search function that finds V. EC The non-zero elements in V are returned, and the result is returned by the non-zero element V. NZ and its index P NZ The vector formed, assuming V EC The number of non-zero elements in n NZ Then V NZ and P NZ The size is nNZ ×1.

[0022] Significance value calculation:

[0023]

[0024] Among them, Sal(a k ) is V NZ The significance value of the k-th element, θ k This represents the significance value of the k-th element.

[0025] Significance vector construction:

[0026]

[0027] First, initialize the saliency vector as follows:

[0028]

[0029] Then, replace the significance value with V. Sal middle:

[0030] V Sal (p k )=θ k =Sal(a k ), k = 1, 2, ..., n NZ

[0031] Among them, V Sal (p k ) is V Sal p in k Each element.

[0032] Saliency plot generation:

[0033]

[0034] Among them, M ASM It is M EC The auditory saliency map, with a size of n C1 ×n F1 f1 -1 It is the inverse function of f1, which transforms a vector into its corresponding matrix. Indicates the nth C1 row n F1 The elements of the column.

[0035] Significant feature extraction:

[0036] All significance values ​​θ k k = 1, 2, ..., n NZ Sort in descending order, with the significance threshold being the k-th rank in the sequence. Th A value, denoted as

[0037] k Th =round(n NZ ·Th Sal )

[0038] Significant features of cochlear diagram M Sal for:

[0039]

[0040] Among them, M Sal The cochlear map shows the salient features. Indicates the nth C1 row n F1 The significance value of column elements, c ij It means, s ij Representation matrix M Sal The elements in the i-th row and the j-th row, The significance threshold is the k-th element in the permutation sequence. Th Values.

[0041] Preferred method: Step 6 involves constructing a fusion feature dictionary based on multi-scale fusion features using labeled samples, and then using the constructed dictionary to perform sparse representation of the test samples, which includes the following steps:

[0042] Atoms in the dictionary matrix are constructed based on multi-scale fusion features. An atom φ in the dictionary matrix is ​​represented as:

[0043] φ=V Mul

[0044] In this formula, V Mul The length is n C1 +n F1 Therefore, the length of each atom in the dictionary matrix Φ is also (n C1 +n F1 ).

[0045] Assuming there are p patterns to be identified, and q atoms in each pattern, then the size of the dictionary matrix Φ is (n C1 +n F1 The arrangement of atoms in the dictionary matrix Φ is as follows: (x) × p·q

[0046] Φ=(φ 1,1 ,…,φ 1,q ,φ 2,1 ,…,φ 2,q ,…,… p,1 ,…,φ p,q )

[0047] The test samples and label samples undergo the same processing, and their feature vectors have the same structure as the atoms in the dictionary matrix, denoted as y, with a size of (n C1 +n F1 The L1 norm is used for the solution, as shown below:

[0048]

[0049] in, Let be the sparse vector to be determined, with a size of pq×1, where ||y||1 represents the L1 norm, and It is a reconstructed signal.

[0050] Preferred: for sparse vectors The Block Matching Pursuit (BMP) method is used for solving the problem.

[0051] The steps of the Block Matching Tracking (BMP) method are as follows:

[0052] Initialization: Input the fusion features y of the test samples, the dictionary matrix Φ, and the preset sparsity S M And the error threshold e. Initialize the residual re0 = y, that is, the initial residual is equal to the signal itself, and set the selected mode block. Φ1,Φ2,...,φ p It is a sub-dictionary corresponding to each pattern in the dictionary, and each pattern has q atoms.

[0053] Iterative matching:

[0054] The correlation of the computation block: for each mode φ i Calculate its relationship with the current residual re t Correlation:

[0055] co i =φ i T re t

[0056] Then calculate the L1 norm of the pattern block:

[0057]

[0058] Select the pattern block that maximizes the L1 norm.

[0059] i * =arg max i ||co i ||1

[0060] That is, select the pattern block that best matches the current residual.

[0061] Updated support set:

[0062] Add the index of this pattern block to the support set:

[0063] Λ=Λ∪{i *}

[0064] Calculate the sparsity coefficient:

[0065] On the currently selected block set Λ, the sparse coefficients κ are solved by minimizing the L1 norm. Λ The objective function is:

[0066]

[0067] in It is the dictionary of the currently selected pattern blocks, κ Λ It is the sparsity coefficient associated with the selected pattern block.

[0068] Update residuals: Calculate new residuals

[0069] re t =y-Φ Λ κ Λ

[0070] This residual will be used for dictionary selection in the next iteration.

[0071] Stopping condition: Iteration stops when the following condition is met.

[0072] re t ≤e or Λ=S M

[0073] Assumption It is the sparse vector after the solution is obtained. The size is pq×1:

[0074]

[0075] Subsequently, from Extract the weight distribution vector W:

[0076] W p×1 =(w1 w2 … w) p ) T

[0077] w i yes The sum of the i-th pattern element:

[0078]

[0079] According to sparse representation theory, atoms in the dictionary matrix that share the same pattern as the test sample contribute the most. Given w... iThe maximum value w is determined by the contribution of the corresponding atom. i The pattern representing the test sample:

[0080] Pattern = i, stmax(w i ), i = 1, 2, ..., p

[0081] Here, Pattern represents the pattern of the test sample.

[0082] Preferably, the method for obtaining multi-scale fusion features in step S5 includes the following steps:

[0083] By performing row-wise and column-wise summation operations on the salient cochlear map matrix, the frequency domain features and time domain features of the signal are extracted, and further feature fusion is performed to obtain multi-scale fused features.

[0084] The formula for summing rows is:

[0085]

[0086] Among them, row S (j) is M Sal The summation value in the j-th row represents the total response of each frequency component on the j-th channel.

[0087] The formula for summing by column is:

[0088]

[0089] Among them, col S (l) is M Sal The summation in column l represents the total response of each channel in time frame l.

[0090] Multi-scale fusion features:

[0091] V Mul =(row) S ;col S )

[0092] =(row) S (1) row S (2) … row S (n C1 ) col S (1) col S (2) … col S (n F1 )) T

[0093] Among them, V Mul This is a multi-scale fusion feature.

[0094] Another object of the present invention is to provide an electronic device comprising: at least one processor, at least one memory, and a communication interface. The processor, memory, and communication interface communicate with each other. The memory stores program instructions executable by the processor, which invokes the program instructions to execute the method for diagnosing faults in agricultural machinery drive motors based on signal cooperative analysis.

[0095] Compared with the prior art, the present invention has the following advantages:

[0096] 1. This invention addresses the challenge of effectively extracting fault features under coupled conditions of variable speed, complex noise interference, and speed fluctuations by proposing a fault diagnosis framework based on signal collaborative analysis.

[0097] 2. This invention proposes a vibration signal reconstruction method based on the combination of current spectrum analysis and order ratio analysis, which can effectively suppress the influence of speed fluctuation.

[0098] 3. This invention proposes a data processing method based on auditory bionics, which provides a general framework for fault diagnosis in similar industrial and mining environments and broadens the cross-domain application of auditory mechanisms. Attached Figure Description

[0099] Figure 1 A test bench for simulating drive motor failures in agricultural machinery.

[0100] Figure 2 This is a flowchart of a method according to an embodiment of the present invention.

[0101] Figure 3 This is a schematic diagram of the drive motor speed signal according to an embodiment of the present invention.

[0102] Figure 4 This is a schematic diagram of the order reconstruction of the original vibration data in an embodiment of the present invention.

[0103] Figure 5 This is a schematic diagram of the spectrum of the original vibration signal and the order-reconstructed vibration signal according to an embodiment of the present invention.

[0104] Figure 6 This is a schematic diagram of the cochlear vibration signal under four modes according to an embodiment of the present invention.

[0105] Figure 7 This is a schematic diagram of the autocorrelation graph of the cochlear image according to an embodiment of the present invention.

[0106] Figure 8 This is a schematic diagram of effective channel selection based on channel validity marking according to an embodiment of the present invention.

[0107] Figure 9 This is a saliency diagram and a schematic diagram of retained elements in the cochlear image according to an embodiment of the present invention.

[0108] Figure 10 This is a schematic diagram illustrating the significant features of the effective cochlear map in an embodiment of the present invention.

[0109] Figure 11 This is a schematic diagram of the channel-level and frame-level features of the tag sample in an embodiment of the present invention.

[0110] Figure 12 This is a schematic diagram of the multi-scale fusion features of the label samples in an embodiment of the present invention.

[0111] Figure 13 A schematic diagram of the sparse vectors of the test samples after BMP solution in an embodiment of the present invention.

[0112] Figure 14 A schematic diagram of the weighted distribution vector of the fused sparse vector in an embodiment of the present invention.

[0113] Figure 15 A schematic diagram of the fault diagnosis results of the drive motor in an embodiment of the present invention. Detailed Implementation

[0114] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0115] Example 1

[0116] In fault detection, subtle fault features are easily masked by environmental noise, and the non-stationary characteristics caused by fluctuating workloads further complicate feature extraction, limiting the effectiveness of existing diagnostic methods. Existing fault diagnosis methods based on single signal sources often struggle to balance accuracy and robustness under complex operating conditions, leading to misdiagnosis and missed diagnosis. This embodiment provides a fault diagnosis method for agricultural machinery drive motors based on signal co-analysis. This method fully utilizes the complementary characteristics between current and vibration signals, estimating motor speed through current spectrum analysis to suppress the non-stationary effects of variable speed conditions. Simultaneously, by combining vibration signal data preprocessing, multi-scale feature extraction, and a fault diagnosis model based on sparse representation, the accuracy and robustness of fault identification are effectively improved. Figure 1 , 2 As shown, it includes the following steps:

[0117] Step S1: Based on the engineering background, determine the fault modes that need to be diagnosed, and collect vibration and current signals of the agricultural machinery drive motor during operation in each mode.

[0118] In another embodiment, step S1 includes the following steps:

[0119] Step S101: Based on the statistical analysis of the maintenance records of the agricultural machinery drive motors, determine the fault modes that need to be diagnosed for the agricultural machinery drive motors. Fault modes include shaft imbalance faults, bearing inner ring faults, and bearing outer ring faults, etc.

[0120] Step S102: Simultaneously collect vibration signals and current signals of the drive motor under different fault modes using a data acquisition device.

[0121] Step S2: By performing spectrum analysis on the collected motor current signal, the frequency components of the motor under variable speed conditions are extracted, and then the motor speed signal is estimated.

[0122] In another embodiment, estimating the motor speed signal in step S2 includes the following steps:

[0123] Step S201: Perform frequency domain filtering on the acquired motor current signal to remove low-frequency noise and high-frequency interference.

[0124] During operation, the current signal of a drive motor contains not only frequency components related to speed changes, but also low-frequency noise (such as power grid fluctuations and mechanical vibrations) and high-frequency interference (such as switching noise and instantaneous current fluctuations). To improve the accuracy of speed estimation, the current signal first needs to be frequency-domain filtered to remove these irrelevant frequency components.

[0125] In another embodiment, the frequency domain filtering method includes a combination of a low-pass filter and a high-pass filter. The aim is to remove irrelevant low-frequency and high-frequency components from the signal, thereby preserving frequency information related to the motor speed. The low-pass filter removes high-frequency noise above a certain threshold, and the high-pass filter removes low-frequency noise below a certain threshold. The transfer function H of the low-pass filter is... low (f) and the high-pass filter transfer function H high (f) is represented as:

[0126]

[0127] Where f is the frequency, f c n is the cutoff frequency. H denoted as the order of the filter.

[0128] Step S202: Perform frequency domain analysis on the current signal, extract the frequency components related to the rotational speed, and then calculate the rotational speed of the motor.

[0129] In a drive system, the frequency components of the current signal directly reflect the motor's operating state. The frequency converter controls the motor speed by changing the current frequency, so the motor speed changes with the frequency of the current signal. To estimate the motor speed, frequency domain analysis of the current signal is required to extract the speed-related frequency components, thereby calculating the motor speed.

[0130] Assume the motor's current signal is i M (t), by performing a Fourier transform, its spectrum is expressed as:

[0131]

[0132] Among them, I M (f) represents the amplitude spectrum of the current signal in the frequency domain, where f is the frequency and i M (t) is the current signal of the motor, where t is the time variable and j is the imaginary unit.

[0133] Under variable speed operation, the motor speed n M As time changes, the spectrum of the motor current also changes dynamically with the change in rotational speed. The fundamental frequency component f0(t) of the motor reflects the motor's rotational speed information. Assume the following relationship exists between the motor's rotational speed and the fundamental frequency of the current signal:

[0134]

[0135] Where f0(t) is the fundamental frequency of the current signal at time t, and n M (t) is the motor speed at time t (unit: rpm), p M It is the number of pole pairs of the motor.

[0136] Based on this formula, the fundamental frequency f0(t) can be extracted from the current signal, and then the instantaneous speed n of the motor can be calculated at each time t. M (t). This is especially important under variable speed conditions, where the speed fluctuates dynamically with changes in load, and changes in the fundamental frequency can effectively reflect these fluctuations.

[0137] By extracting the fundamental frequency f0(t) from the current signal, the instantaneous speed n of the motor is calculated using the following formula. M (t):

[0138]

[0139] Where, n M (t) represents the instantaneous speed of the motor.

[0140] Step S3: After extracting the motor speed signal, the motor vibration signal undergoes multi-stage preprocessing to convert the vibration signal into a cochlear image, and the preprocessed cochlear image is obtained. First, the order ratio analysis method is used to filter out the influence of speed fluctuations on the vibration signal. Second, based on Gammatone filtering, the vibration signal is converted into a cochlear image to simulate the frequency response characteristics of the human ear. Finally, auditory flow separation technology is used for effective channel selection, retaining key fault information and removing irrelevant noise.

[0141] In another embodiment, the method for multi-stage preprocessing of the motor vibration signal in step S3 includes the following steps:

[0142] Step S301: The rotational speed signal obtained in step S202 is used as the input for the order analysis. The vibration signal is reconstructed through the order analysis method, thereby effectively removing the influence of rotational speed fluctuations on the vibration signal.

[0143] First, the motor speed signal n M (t) is normalized to a standard frequency range to facilitate subsequent order calculations. The normalized speed signal n MN The formula for calculating (t) is:

[0144]

[0145] Where, n M_max and n M_min These are the maximum and minimum values ​​in the speed signal, respectively.

[0146] Then, the normalized speed signal n MN (t), calculate the order signal ω related to the rotational speed. M (t), its calculation formula is:

[0147] ω M (t)=n MN (t) / n base

[0148] Where, n base This is the reference rotational speed, representing the standard rotational speed set by the system. The vibration signal x after order reconstruction. VOR (t) is given by the following formula:

[0149] x VOR (t)=x V (t)·cos(2π·ω M (t))

[0150] Where, x VOR (t) represents the reconstructed vibration signal with the correct order, x V (t) represents the original vibration signal, ω M(t) represents the order signal related to the rotational speed.

[0151] Through this reconstruction method, the vibration signal is mapped to the order frequency of the rotational speed, effectively eliminating the influence of rotational speed fluctuations. The signal x after order analysis and filtering... VOR (t) has high stability and can be used as input data for subsequent data processing.

[0152] Step S302: The vibration signal processed in step S301 is converted into a cochlear image. Before generating the cochlear image, an overlapping sliding window is used to obtain signal segments to ensure time resolution. Here, the width t of the time window... W1 It should be much smaller than the time period t STP The window sliding step size t S1 It should be much smaller than the width of the time window.

[0153] t STP =1 / f max

[0154] Among them, t STP For the time period, f max This is the highest frequency of the signal.

[0155] Assume n C1 n is the number of frequency channels in the cochlear image. F1 If it is the number of time frames per channel, then the high-resolution cochlear map M HCG The size is n C1 ×n F1 .

[0156] Number of time frames per channel:

[0157]

[0158] Where, n F1 n is the number of time frames per channel. C1 If M is the number of frequency channels in the cochlear image, then the high-resolution cochlear image M... HCG The size is n C1 ×n F1 N signal t is the total length of the signal. W1 t is the width of the time window. S1 This is the step size for the window to slide.

[0159] Step S303: Perform effective channel selection on the cochlear image converted in step S302, assuming c i It is M HCG The i-th channel, and c i The length is n F1 i = 1, 2, ..., nC1 r i It is C i If the autocorrelation sequence is r, then r i The length is 2n F1 -1:

[0160]

[0161] Where, r i It is C i The autocorrelation sequence, where m is the sequence r i The m-th point in the array, c i,n It is the nth element in the i-th channel.

[0162] Based on this formula, the autocorrelation sequences of all channels are calculated and combined to form the autocorrelation matrix M. AC And M AC The size is n C1 ×(2n F1 -1), next, based on the autocorrelation matrix M AC Binarization is used for effective channel selection. The key parameter for binarization is the Global Image Threshold (GIT), which is determined based on the OtSu method. All M values ​​greater than the global image threshold are selected. AC The values ​​in the list will be replaced by the boolean constant TRUE, and other values ​​will be replaced by the boolean constant FALSE.

[0163] M BI =f BI (M AC ,th1)

[0164] Among them, M BI It is a binary matrix, M BI The size is n C1 ×(2n F1 -1), f BI th1 is the binarization function, and th1 is the global image threshold.

[0165] Next, calculate M. BI The sums of each row in the vector s are used to construct a vector s. BI Its size is n C1 ×1, where s BI The elements in represent M HCG The validity of the corresponding frequency channels. Based on the computational auditory scene analysis model, s BI The zero value in M ​​represents HCG The periodicity difference of the corresponding frequency channel is finally determined by M. HCG Invalid channels are set to zero, and the matrix that constitutes the valid channels is denoted as M. EC Its size is nC1 ×n F1 .

[0166] Step S4: By simulating the auditory selective attention mechanism, an adaptive cochlear map saliency map is calculated from the preprocessed cochlear map using an adaptive cochlear map saliency extraction method. This improves the identifiability of fault features and yields a saliency cochlear map.

[0167] Based on a bottom-up driven saliency attention mechanism, a saliency model can be constructed to simulate extrinsic attention, and time-frequency units can be filtered through saliency maps to extract key features. The cochlea of ​​the auditory system nonlinearly encodes signals of different frequencies and highlights saliency features based on the intensity and structure of the acoustic signal. The saliency attention mechanism relies on the statistical distribution characteristics of specific time-frequency regions, enhancing the perceptual effect of the target signal by calculating the contrast between local and global features. In engineering applications, vibration signals contain rich time-frequency features, and the high sensitivity of the auditory system to perceptual stimuli provides theoretical support for saliency feature extraction. Therefore, saliency feature extraction is also crucial in fault diagnosis tasks. The selective attention of the auditory system determines that only key information can be transmitted along the auditory pathway to higher cognitive areas for more refined analysis. This process can be analogous to feature extraction in fault diagnosis, i.e., using auditory saliency maps to extract key information from the effective channels of the cochlear image, improving feature separability, thereby enhancing the accuracy and stability of fault diagnosis. In this invention, the saliency map of the cochlear image is obtained through an improved global contrast method, with pixel I... k The significance value in image I is defined as:

[0168]

[0169] Among them, Sal(I k I represents the saliency value of the k-th pixel in the image. k I represents the k-th pixel in the image. i Let I represent the i-th pixel in the image, and let I represent the image.

[0170] Unlike speech signals, cochlear images consist of multiple effective frequency channels, contain a large number of zero-value elements, and exhibit significant differences in the distribution range of saliency values ​​among different samples. To achieve adaptive adjustment of the saliency threshold and reduce computational resource consumption, we designed a novel adaptive cochlear image saliency feature extraction method. This method calculates the auditory saliency map by statistically analyzing the effective values ​​(i.e., non-zero elements) in the cochlear image, thereby improving the accuracy and efficiency of feature extraction. The adaptive cochlear image saliency feature extraction method includes the following steps:

[0171] Initialization: Input the cochlear map composed of valid channels. Significant threshold Th Sal.

[0172] Vectorization:

[0173]

[0174] Among them, M EC The cochlear map matrix, V, is composed of effective channels. EC From M EC The transformed vector, V EC The size is n C1 ·n F1 ×1. The function f1 is a reshaping function that transforms a matrix into its corresponding vector, n. C1 Representation matrix M EC The number of rows, n F1 The representation matrix M represents EC The number of columns.

[0175] Valid element retrieval:

[0176] [V NZ ,P NZ ]=f2(V EC )

[0177] Here, f2 is a non-zero element search function that finds V. EC The non-zero elements in V are returned, and the result is returned by the non-zero element V. NZ and its index P NZ The vector formed, assuming V EC The number of non-zero elements in n NZ Then V NZ and P NZ The size is n NZ ×1.

[0178] Significance value calculation:

[0179]

[0180] Among them, Sal(a k ) is V NZ The significance value of the k-th element, θ k This represents the significance value of the k-th element.

[0181] Significance vector construction:

[0182]

[0183] First, initialize the saliency vector as follows:

[0184]

[0185] Then, replace the significance value with V. Salmiddle:

[0186] V Sal (p k )=θ k =Sal(a k ), k = 1, 2, ..., n NZ

[0187] Among them, V Sal (p k ) is V Sal p in k Each element.

[0188] Saliency plot generation:

[0189]

[0190] Among them, M ASM It is M EC The auditory saliency map, with a size of n C1 ×n F1 f1 -1 It is the inverse function of f1, which transforms a vector into its corresponding matrix. Indicates the nth C1 row n F1 The elements of the column.

[0191] Significant feature extraction:

[0192] All significance values ​​θ k k = 1, 2, ..., n NZ Sort in descending order, with the significance threshold being the k-th rank in the sequence. Th A value, denoted as

[0193] k Th =round(n NZ ·Th Sal )

[0194] Significant features of cochlear diagram M Sal for:

[0195]

[0196] Among them, M Sal The cochlear map shows the salient features. Indicates the nth C1 row n F1 The significance value of column elements, c ij It means, s ij Representation matrix M Sal The elements in the i-th row and the j-th row, The significance threshold is the k-th element in the permutation sequence. Th Values.

[0197] Step S5: Extract channel-level and frame-level features based on the salient cochlear map, and then fuse the two features to obtain multi-scale fusion features, so as to obtain a more comprehensive picture of the vibration signal at different scales.

[0198] The method for obtaining multi-scale fusion features in step S5 includes the following steps:

[0199] Step S5, following the generation of the effective salient feature cochlear map in step S4, further extracts and fuses multi-scale features. Specifically, this method extracts frequency domain and time domain features of the signal by performing row-wise and column-wise summation operations on the salient feature cochlear map matrix, respectively, and further fuses these features to construct a more discriminative feature vector. The row-wise summation operation is used to extract global features at the channel level, i.e., the energy distribution of different frequency channels, thereby reflecting the overall frequency components of the signal. On the other hand, the column-wise summation operation is used to extract temporal features at the frame level to capture the transient change patterns and temporal evolution trends of the signal.

[0200] By performing row-wise and column-wise summation operations on the salient cochlear map matrix, the frequency domain features and time domain features of the signal are extracted, and further feature fusion is performed to obtain multi-scale fused features.

[0201] Row-wise summation: Channel-level feature extraction

[0202] The significance matrix M Sal Row-wise summation focuses on the frequency dimension of the signal. Each row corresponds to a frequency channel in the cochlear image matrix, M. Sal Row-wise summation aggregates the responses of each frequency component across all channels into a scalar, providing global information about that frequency component across different channels. Frequency domain features of the signal are extracted at the channel level, revealing the global variations of the signal at each frequency. The specific formula is as follows:

[0203]

[0204] Among them, row S (j) is M Sal The summation value in the j-th row represents the total response of each frequency component on the j-th channel.

[0205] Summation by column: Frame-level feature extraction

[0206] Unlike row-wise summation, column-wise summation focuses on analyzing signal variations over time. Each column represents a time frame in the cochlear diagram matrix. Column-wise summation aggregates the responses of all frequency components within that time frame, providing an overall characteristic of the signal within that frame. Column-wise summation reveals the dynamic changes of the signal in the time domain, and is particularly helpful in identifying the timing of fault occurrence and the temporal characteristics of fault modes. The process of column-wise summation can be represented as:

[0207]

[0208] Among them, col S (l) is M Sal The summation in the l-th column represents the total response of each channel in the l-th time frame. This process reflects the global changes of the signal in the time dimension, revealing the evolution of the signal over time.

[0209] Feature fusion: Combining features at multiple scales

[0210] By using matrix M Sal The feature vectors obtained by summing rows and columns are concatenated into a new vector, thus achieving the fusion of features in both the frequency domain and the time domain. The resulting multi-scale fused feature V Mul It can be represented as:

[0211] V Mul =(row) S ;col S )

[0212] =(row) S (1) row S (2) … row S (n C1 ) col S (1) col S (2) … col S (n F1 )) T

[0213] Among them, V Mul This is a multi-scale fusion feature.

[0214] Step S6: Based on multi-scale fusion features, construct a fusion feature dictionary using labeled samples, use the constructed dictionary to perform sparse representation on the test samples, and realize fault diagnosis of agricultural machinery drive motors based on sparse representation.

[0215] The method based on multi-scale fusion features involves constructing a fusion feature dictionary from labeled samples and using the constructed dictionary to perform sparse representation of test samples, which includes the following steps:

[0216] Step S6, following the multi-scale fusion feature extraction in step S5, implements fault diagnosis of the agricultural machinery drive motor based on sparse representation. In this step, the dictionary matrix is ​​constructed based on the multi-scale feature vector extracted in step S5. This feature vector is composed of channel-level and frame-level features from the salient cochlear image, aiming to comprehensively capture the time-frequency characteristics of the signal. Atoms in the dictionary matrix are constructed based on the multi-scale fusion features, and an atom φ in the dictionary matrix is ​​represented as:

[0217] φ=V Mul

[0218] In this formula, V Mul The length is n C1 +n F1 Therefore, the length of each atom in the dictionary matrix Φ is also (n C1 +n F1 ).

[0219] Assuming there are p patterns to be identified, and q atoms in each pattern, then the size of the dictionary matrix Φ is (n C1 +n F1 The arrangement of atoms in the dictionary matrix Φ is as follows: (x) × p·q

[0220] Φ=(φ 1,1 ,…,φ 1,q ,φ 2,1 ,…,φ 2,q ,…,φ p,1 ,…,φ p,q )

[0221] The test samples and label samples undergo the same processing, and their feature vectors have the same structure as the atoms in the dictionary matrix, denoted as y, with a size of (n C1 +n F1 The L1 norm is used for the solution, as shown below:

[0222]

[0223] in, Let be the sparse vector to be determined, with a size of pq×1, where ||y||1 represents the L1 norm, and It is a reconstructed signal.

[0224] Because the extracted multi-scale features have obvious periodicity and multiple frequency components, they present complex structured signals. These signals not only fluctuate in local regions but may also exhibit certain regularities across multiple time scales. Furthermore, motor faults are usually not sudden but manifest through long-term accumulation and gradual changes in vibration patterns, resulting in signal variations with certain local correlations and repeatability. Therefore, when dealing with sparse vectors... In the solution process, the Block Matching Pursuit (BMP) method is employed. BMP captures these periodic and multi-frequency features through block-level matching, making it suitable for processing structured and periodic signals. Unlike the traditional OMP method that selects dictionary atoms one by one, BMP improves the efficiency of sparse representation by selecting multiple dictionary atoms (i.e., a block) simultaneously. This strategy not only better captures repetitive patterns in the signal but also effectively captures multi-scale features, thereby improving the accuracy and efficiency of signal representation.

[0225] The steps of the Block Matching Tracking (BMP) method are as follows:

[0226] Initialization: Input the fusion features y of the test samples, the dictionary matrix Φ, and the preset sparsity S M And the error threshold e. Initialize the residual re0 = y, that is, the initial residual is equal to the signal itself, and set the selected mode block. Φ1,Φ2,...,Φ p It is a sub-dictionary corresponding to each pattern in the dictionary, and each pattern has q atoms.

[0227] Iterative matching:

[0228] The correlation of the computation block: for each pattern Φ i Calculate its relationship with the current residual re t Correlation:

[0229] co i =Φ i T re t

[0230] Then calculate the L1 norm (sum of absolute values) of the pattern block:

[0231]

[0232] Select the pattern block that maximizes the L1 norm.

[0233] i * =arg max i ||co i ||1

[0234] That is, select the pattern block that best matches the current residual.

[0235] Updated support set:

[0236] Add the index of this pattern block to the support set:

[0237]

[0238] Calculate the sparsity coefficient:

[0239] On the currently selected block set Λ, the sparse coefficients κ are solved by minimizing the L1 norm. Λ The objective function is:

[0240]

[0241] in It is the dictionary of the currently selected pattern blocks, κ Λ It is the sparsity coefficient associated with the selected pattern block.

[0242] Update residuals: Calculate new residuals

[0243] re t =y-Φ Λ κ Λ

[0244] This residual will be used for dictionary selection in the next iteration.

[0245] Stopping condition: Iteration stops when the following condition is met.

[0246] re t ≤e or Λ=S M

[0247] Assumption It is the sparse vector after the solution is obtained. The size is pq×1:

[0248]

[0249] Subsequently, from Extract the weight distribution vector W:

[0250] W p×1 =(w1 w2 … w) p ) T

[0251] w i yes The sum of the i-th pattern element:

[0252]

[0253] According to sparse representation theory, atoms in the dictionary matrix that share the same pattern as the test sample contribute the most. Given w... i The maximum value w is determined by the contribution of the corresponding atom. i The pattern representing the test sample:

[0254] Pattern = i, stmax(w i ), i = 1, 2, ..., p

[0255] Here, Pattern represents the pattern of the test sample.

[0256] In another embodiment, an electronic device is provided, comprising: at least one processor, at least one memory, and a communication interface. The processor, memory, and communication interface communicate with each other. The memory stores program instructions executable by the processor, which invokes the program instructions to execute the fault diagnosis method for agricultural machinery drive motors based on signal cooperative analysis.

[0257] This embodiment addresses the problem of insufficient diagnostic accuracy of existing methods under complex loads and variable speed conditions in drive motors, proposing a highly robust diagnostic framework based on signal co-analysis. First, rotational speed estimation is performed using a current spectrum analysis method, avoiding reliance on additional speed sensors. Second, adaptive noise reduction of vibration signals is achieved by combining order analysis and a Gammatone filter, and an effective channel selection method based on the Otsu method for cochlear diagrams is used to improve signal quality. Next, key fault features are extracted based on a saliency-driven attention mechanism, and a multi-scale feature fusion strategy is employed to extract and fuse features at both the channel and frame levels, enhancing the discriminative power of fault modes while reducing data dimensionality and computational complexity. Finally, a sparse representation method based on the Block Matching Pursuit (BMP) algorithm is used to achieve high-precision classification of drive motor fault modes. Furthermore, this method exhibits excellent diagnostic performance in high-noise environments and variable speed conditions, significantly improving diagnostic robustness.

[0258] Simulation Example

[0259] like Figure 1 The table shows a simulation test bench for drive motor failures in agricultural machinery. The experimental data obtained from this test bench is shown in Table 1.

[0260] Table 1. Experimental Dataset for Motor Fault Diagnosis

[0261]

[0262]

[0263] like Figure 2 As shown, a fault diagnosis method for agricultural machinery drive motors based on signal collaborative analysis includes the following steps:

[0264] S1: Based on the engineering background, determine the fault modes that need to be diagnosed, and collect vibration and current signals of the drive motor during each mode of operation.

[0265] S2: As Figure 3As shown, this is an example diagram of the estimated motor speed signal after performing spectrum analysis on a current signal sample acquired in step S1.

[0266] S3: Using the rotational speed signal obtained in step S2 as input for the order ratio analysis, the comparison diagram between the reconstructed order ratio signal of the obtained vibration signal and the original signal is shown in the figure. Figure 4 As shown in the image, the blue original vibration signal exhibits complex and highly fluctuating characteristics, reflecting the motor's vibration under variable speed conditions. In contrast, the green reconstructed signal from the order analysis is more stable with smaller fluctuations. This indicates that the order analysis successfully removed the influence of speed variations from the reconstructed signal, resulting in a more regular waveform. This stable signal better matches the signal characteristics under uniform speed conditions, which is helpful for subsequent fault feature extraction and pattern recognition.

[0267] Meanwhile, to verify the advantages of the frequency domain reconstructed signal in suppressing the spectral blurring effect caused by velocity fluctuations, a graph was also plotted. Figure 4 The frequency spectra of the original vibration signal and the reconstructed vibration signal, such as Figure 5 As shown. From Figure 5 (a) and Figure 5 (b) The comparison shows that after the order analysis reconstruction, the spectral lines in the orange box area of ​​the spectrum are not only significantly improved in terms of clarity, but also have significantly reduced energy, effectively reducing the interference of velocity fluctuations and noise on frequency components.

[0268] S4: As Figure 6 As shown, the cochlear images plotted are obtained after transformation based on the first data sample from four modes: NM, IR, OR, and UB. The figures reveal that the OR and UB cochlear images exhibit significant differences in time-frequency distribution characteristics compared to the other cochlear images, demonstrating strong discriminability. However, due to the influence of uncertain mixed noise, the NM and IR cochlear images show only slight feature differences, making accurate differentiation difficult and limiting the accuracy and robustness of fault diagnosis. To address these issues, enhance the expressive power of time-frequency features, and effectively suppress noise interference, this paper introduces effective channel selection and salient feature extraction methods in subsequent processing to improve the overall diagnostic model's discriminative ability and generalization performance.

[0269] S5: Essentially, the basis for effective channel selection is the periodicity of the effective frequency channels, as shown in the autocorrelation plot of the cochlear image. Figure 7As shown, the autocorrelation plot quantitatively describes the periodicity of different channels. In the autocorrelation plots of the four modes, bright areas represent effective channels, while dark areas represent frequency channels containing a large amount of noise. To distinguish the effectiveness of these channels, the GIT value was determined using the Otsu method. The GIT values ​​for NM, IR, OR, and UB are 0.0152, 0.0136, 0.0144, and 0.0147, respectively. Then, a binarization method was applied based on the GIT value, such as... Figure 7 As shown in the middle, the binarized image provides channel validity markers for auditory flow separation. Figure 7 (Lower part)

[0270] In this process, based on predetermined channel validity markers, valid channels in the cochlear image are retained while invalid frequency channels are removed to improve signal representation and computational efficiency. The core idea is to use an effective channel selection attention mechanism to filter out the most representative frequency channels during fault feature extraction. These valid channels provide key signal features, helping to accurately capture the dynamic changes of fault signals. Simultaneously, invalid channels that contribute little to the classification or recognition task or may introduce noise interference are removed, effectively reducing computational complexity and avoiding interference from redundant information. Figure 8 This demonstrates the specific effects of the effective channel selection attention mechanism. The upper part of the original cochlear map contains all frequency channels, some of which may contain noise components, adversely affecting signal analysis. The lower part of the effective channel cochlear map removes redundant and invalid frequency channels by selecting effective channels. (Original cochlear map) Figure 8 (Upper part) and effective channel cochlear map ( Figure 8 The comparison in the lower part shows that auditory flow separation offers a significant advantage in reducing uncertain noise, providing more accurate data support for subsequent pattern recognition and fault diagnosis tasks.

[0271] S6: As Figure 9 As shown, the upper part displays a saliency map of each channel in the cochlear image, which allows for a visual observation of the importance of different channels in feature representation. Subsequently, a pre-set saliency level threshold Th is used... Sal =0.2, the saliency index of the effective channels in the cochlear image was extracted, and the result is as follows. Figure 9 As shown in the lower section, based on these salient feature element indices, the corresponding salient features M were further extracted from the cochlear image. Sal The extraction results are as follows Figure 10 As shown, compared to the original complete cochlear image, the extracted salient features based on effective channels significantly reduce the interference of redundant features while retaining key discriminative information, thus improving the compactness of feature representation and computational efficiency. Furthermore, pattern differences in different states can be clearly distinguished by the index positions and corresponding values ​​of salient features.

[0272] S7: After completing the above steps, based on the sample information in Table 1, extract the channel-level features and frame-level features of all label samples respectively. The results are as follows: Figure 11 As shown in (a) and (b). Subsequently, multi-scale fusion was performed on the two types of features, and the results are as follows. Figure 12 As shown in the figure, the fused features demonstrate stronger discriminative power in fault mode recognition compared to single features. The fusion result not only preserves key information at both the channel and frame levels, improving the completeness of feature representation, but also enhances the discriminative power of the channel dimension and maintains the temporal effectiveness of the frame-level features.

[0273] S8: Calculate the sparse vector of all test samples based on the block matching pursuit algorithm, where the sparsity S M =5, the error threshold is ε = 1 × 10 -5 The motor has 4 fault modes, and the number of labeled samples in each mode is q=20. Given 120 test samples (as listed in Table 1), the corresponding number of sparse vectors is also 120, which are displayed as column vectors. Figure 13 In the diagram, the horizontal axis represents the index of the test sample, and the vertical axis represents the index of the atom in the dictionary matrix. The index of the non-zero element in the sparse vector is also the column index of the selected support vector in the dictionary matrix. The value of the non-zero element represents the weight or contribution of the support vector in the reconstruction of the test sample. Therefore, based on the arrangement of the dictionary matrix, the index of the non-zero element represents the pattern of the test sample, such as... Figure 13 The vertical axis is shown in Table 1. Based on the test sample arrangement listed in Table 1, the non-zero elements in the sparse vectors of all features correctly represent the patterns of the test samples.

[0274] Finally, the weight distribution vector W is calculated. Since the number of patterns p = 4, the number of elements in the weight distribution vector is 4. Also, the number of test samples is 120. Figure 14 The display shows 120 weight distribution vectors. The maximum value of each weight distribution vector is marked with a red dot, and the position of the red dot indicates the pattern of the test sample. Figure 14 The index of the largest element in the weighted distribution vector perfectly matches the pattern of the test samples in Table 1. Therefore, the accuracy of fault diagnosis is 100%, and the confusion matrix of the diagnosis results is as follows. Figure 15 As shown, the reliability and effectiveness of the proposed method are fully demonstrated. The results prove that the fault diagnosis framework proposed in this study still has excellent robustness even under variable speed and strong vibration conditions, providing a high-precision and adaptive solution for intelligent fault diagnosis.

[0275] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A signal coordination analysis-based fault diagnosis method for a driving motor of agricultural equipment, characterized by, The method comprises the following steps: Step S1: According to the engineering background, determine the fault mode that needs to be diagnosed, and collect the vibration signal and current signal of each mode of the agricultural equipment driving motor; Step S2: By performing spectrum analysis on the collected motor current signal, the frequency components of the motor under variable speed working conditions are extracted, and the motor speed signal is estimated; Step S3: After extracting the motor speed signal, the motor vibration signal is preprocessed in multiple stages, the vibration signal is converted into a cochlea graph, and a preprocessed cochlea graph is obtained; Step S4: Through the simulation of the selective attention mechanism of hearing, the effective values in the preprocessed cochlea graph are calculated by using the adaptive cochlea graph salient feature extraction method to calculate the auditory saliency map, and a salient feature cochlea graph is obtained; Step S5: Based on the salient feature cochlea graph, channel level and frame level features are extracted, and then the two kinds of features are fused to obtain multi-scale fusion features; Step S6: Based on the multi-scale fusion features, a fusion feature dictionary is formed by using labeled samples, the test sample is sparsely represented by using the constructed dictionary, and the fault diagnosis of the agricultural equipment driving motor is realized based on the sparse representation.

2. The method according to claim 1, characterized in that: The adaptive cochlea graph salient feature extraction method in step S4 includes the following steps: Initialization: input cochleagram of valid channel composition Significance threshold Th Sal ; Vectorization: where M EC is a cochleagram matrix composed of valid channels, V EC is a vector converted from M EC , V EC has a size of n C1 · n F1 × 1; function f1 is a reshaping function that converts a matrix into a corresponding vector, n C1 represents the number of rows of matrix M EC , and n F1 represents the number of columns of matrix M EC . Effective element retrieval: [V NZ ,P NZ ]=f2(V EC ) where f2 is a non-zero element search function that finds the non-zero elements in V EC and returns a vector consisting of the non-zero elements V NZ and their indices P NZ , assuming that the number of non-zero elements in V EC is n NZ , then V NZ and P NZ are of size n NZ x 1; Significance value calculation: wherein Sal(a k ) is the saliency value of the kth element in V NZ k denotes the saliency value of the kth element.​ Significance vector construction: First, initialize the significance vector as: Then, the saliency value is substituted into V Sal : V Sal (p k )=θ k =Sal(a k ),k=1,2,…,n NZ where V Sal is the vector of values k ) is the p Sal th element of the vector V k . Saliency map generation: where M ASM is the auditory saliency map of M EC , of size n C1 x n F1 , f1 -1 is the inverse function of f1, which converts a vector into a corresponding matrix, denotes the element in the n C1 th row and n F1 th column; Salient feature extraction: all the significance values θ k k = 1, 2,..., n NZ in descending order, the significance threshold is the k Th th value in the ordered sequence, denoted as k Th = round(n NZ · Th Sal ) Significant feature cochleogram M Sal Is: where M Sal represents the salient feature cochleagram, represents the n C1 th element of the m F1 th row and n ij th column of the matrix M ij represents the saliency value of the i Sal th element of the i th row and j Th th column of the matrix M represents the saliency threshold value of the k th value in the permutation sequence.

3. The method according to claim 2, wherein the method further comprises: The method of forming a fusion feature dictionary based on multi-scale fusion features in step 6 includes the following steps: Based on the multi-scale fusion features, an atom in the dictionary matrix is constructed, and an atom in the dictionary matrix is represented as: φ = V Mul In this formula, V Mul is the length of the vector C1 n F1 , and thus the length of each atom in the dictionary matrix Φ is also (n C1 +n F1 ). Assuming the number of patterns to be identified is p, and the number of atoms in each pattern is q, the size of the dictionary matrix Φ is (n C1 + n F1 )×p·q, and the arrangement of atoms in the dictionary matrix Φ is: Φ = (φ 1,1 ,…,φ 1,q , φ 2,1 ,…,φ 2,q , φ p,1 ,…,φ p,q ) The test sample and the label sample are processed in the same way, and the feature vector structure is the same as the atom in the dictionary matrix, denoted as y, and the size of the vector y is (n C1 + n F1 ) ; L1 norm is used for solving, as follows: where is the sparse vector to be found, of size pq x 1, ||y||1 represents the L1 norm, and is the reconstructed signal.

4. The method according to claim 3, wherein the method further comprises: To a sparse vector In solving, block matching pursuit (BMP) method is adopted, The block matching pursuit (BMP) method steps are as follows: Initialization: input the fusion feature y of the test sample, the dictionary matrix Φ, the preset sparsity S, the error threshold e; initialize the residual re0=y, that is, the initial residual is equal to the signal itself, and set the selected mode block M Φ1, Φ2,..., Φ Φ1, Φ2,..., Φ p is the sub-dictionary corresponding to each mode in the dictionary, and each mode has q atoms; Iterative matching: Compute the correlation of the block: For each pattern Φ i , compute its correlation with the current residual re t . co i = Φ i T re t Then calculate the L1 norm of the mode block: selecting the mode block that maximizes the l1 norm i * = arg max i ‖co i ‖1 i.e. selecting the mode block that best matches the current residual Update the support set: Add the index of the mode block to the support set: Λ = Λ U {i *} Calculate the sparse coefficient: On the current selected set of blocks Λ, solve for the sparse coefficients κ by minimizing the L1 norm Λ The objective function is: wherein is the currently selected mode block dictionary, κ Λ is a sparse coefficient associated with the selected mode block; Update the residual: calculate the new residual re t = y - Φ Λ κ Λ The residual will be used for dictionary selection in the next iteration; Stopping condition: when the following conditions are met, the iteration stops re t ≤e or Λ=S M Assume is the solved sparse vector, of size pq x 1: Subsequently, the weight distribution vector W is extracted from the matrix M: W p×1 = (w1 w2... w p ) T w i is sum of the i-th mode element in the middle According to the sparse representation theory, the atoms in the dictionary matrix have the largest contribution to the part of the test sample that shares the same pattern, given w i The maximum value w i represents the pattern of the test sample: Pattern = i, s.t. max(w i ), i = 1, 2,..., p Where, Pattern represents the mode of the test sample.

5. The method according to claim 4, wherein the method further comprises: The method of obtaining multi-scale fusion features in step S5 includes the following steps: By summing the salient feature cochlea graph matrix by row and by column, the frequency domain features and time domain features of the signal are extracted, and further feature fusion is performed to obtain multi-scale fusion features; The formula for summing by row is: where row S (j) is M Sal the sum value of the jth row, representing the total response of each frequency component on the jth channel; The formula for summing by column is: where col S (l) is M Sal The sum of the values in column l represents the total response of each channel over the lth time frame. Multi-scale fusion features: V Mul = (row S ; col S ) = (row S (1) row S (2) … row S (n C1 ) col S (1) col S (2) … col S (n F1 )) T wherein V Mul is a multi-scale fusion feature.

6. The method according to claim 5, wherein the method further comprises: The method of multi-stage preprocessing of the motor vibration signal in step S3 includes the following steps: Step S301: The speed signal obtained in step S202 is taken as the input of order ratio analysis, and the vibration signal is reconstructed by order ratio analysis method, so as to effectively remove the influence of speed fluctuation on the vibration signal; First, the motor speed signal n M (t) is normalized to a standard frequency range, the normalized speed signal n MN (t) is calculated as follows: wherein n M_max and n M_min are the maximum and minimum values in the rotational speed signal, respectively. Afterwards, the speed signal n MN (t) is normalized and a speed-dependent order signal ω M (t) is calculated, which is calculated as follows: ω M (t) = n MN (t) / n base where n base is the reference rotational speed, which represents a standard rotational speed set by the system; the order ratio-reconstructed vibration signal x VOR (t) is given by the following equation: x VOR (t) = x V (t) · cos(2π · ω M (t)) where x VOR (t) represents the reconstructed vibration signal, x V (t) represents the original vibration signal, ω M (t) represents the order signal related to the rotational speed; Step S302: Convert the vibration signal processed in step S301 into a cochlea graph, and before generating the cochlea graph, in order to ensure the time resolution, a sliding window is used to obtain a signal segment; t STP = 1 / f max where t STP is the time period, f max is the highest frequency of the signal; The number of time frames of each channel: wherein n F1 is the number of time frames per channel, n C1 is the number of frequency channels in the cochleagram, then the size of the high resolution cochleagram M HCG is n C1 × n F1 , N signal is the total length of the signal, t W1 is the width of the time window, and t S1 is the step size of the window slide. Step S303: effective channel selection is performed on the cochleagram converted in step S302, assuming that c i is the ith channel of M HCG , and the length of c i is n F1 , i = 1, 2, …, n C1 , r i is the autocorrelation sequence of c i , and the length of r i is 2n F1 -1: where r i is a sequence of autocorrelations of c i , m is the mth point in the sequence r i , and c i,n is the nth element in the ith channel. The autocorrelation sequences of all channels are calculated and composed into an autocorrelation matrix M AC , and M AC has a size of n C1 × (2n F1 -1). Then, effective channel selection is performed based on the binarization of the autocorrelation matrix M AC , and the key parameter of the binarization is a global image threshold, which is determined based on the OtSu method. All values in M AC that are greater than the global image threshold are replaced by a Boolean constant TRUE, and other values are replaced by a Boolean constant FALSE. M BI = f BI (M AC ,th1) wherein M BI is a binary matrix, M BI has a size of n C1 x (2n F1 -1), f BI is a binarization function, and th1 is a global image threshold value; Afterwards, the sum over each row of M BI is calculated, which forms a vector s BI of size n C1 x 1, where the elements in s BI represent the effectiveness of the corresponding frequency channel in M HCG ; according to the calculation of the auditory scene analysis model, zero values in s BI represent the periodicity difference of the corresponding frequency channel in M HCG , finally, the ineffective channels in M HCG are set to zero, and the matrix of effective channels is denoted as M EC , which has a size of n C1 x n F1 .

7. The method according to claim 6, wherein the method further comprises: The method of estimating the motor speed signal in step S2 includes the following steps: Step S201: frequency domain filtering is performed on the collected motor current signal to remove low frequency noise and high frequency interference; Step S202: frequency domain analysis is performed on the current signal to extract the frequency component related to the speed, and then the speed of the motor is calculated; The current signal of the motor is i M (t), by Fourier transforming it, its frequency spectrum is obtained as: where I M (f) represents the amplitude spectrum of the current signal in the frequency domain, f is the frequency, i M (t) is the current signal of the electric machine, t is the time variable, j is the imaginary unit; Under variable speed operation condition, the speed of the motor and the fundamental frequency of the current signal have the following relationship: where f0(t) is the fundamental frequency of the current signal at time t, n M (t) is the rotational speed of the motor at time t, p M is the number of pole pairs of the motor; By extracting the fundamental frequency f0(t) from the current signal, the instantaneous rotational speed n of the motor is calculated using the following equation M (t): wherein n M (t) denotes the instantaneous rotational speed of the electric machine.

8. The method according to claim 7, wherein the method further comprises: The frequency domain filtering method in step S201 includes a combination of a low-pass filter and a high-pass filter, the low-pass filter being used to remove high-frequency noise higher than a certain threshold, and the high-pass filter being used to remove low-frequency noise lower than a certain threshold; the low-pass filter transfer function H low (f) and the high-pass filter transfer function H high (f) are represented as: where f is the frequency, f c is the cutoff frequency, n H is the order of the filter.

9. The method according to claim 8, wherein the method further comprises: The step S1 comprises the following steps: Step S101: according to the statistical analysis of the maintenance records of the agricultural equipment driving motor, the fault modes of the agricultural equipment driving motor that need to be diagnosed are determined; Step S102: the vibration signal and the current signal of the driving motor under different fault modes are collected simultaneously through the data acquisition device.

10. An electronic device, comprising: Comprise: At least one processor, at least one memory and a communication interface; the processor, memory and communication interface communicate with each other; The memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the agricultural equipment driving motor fault diagnosis method based on signal collaborative analysis in any one of claims 1 to 9.

Images