Wheel damage detection method based on interval value fuzzy reasoning

Abstract

The invention provides a wheel damage detection method based on interval value fuzzy reasoning, and the method comprises the steps: collecting initial sound data of different wheel states, and carrying out the preprocessing, so as to obtain sound data of different states; identifying the collected sound data by an expert, giving a plurality of empirical values under different wheel states, and forming a normal standard library; determining the membership degree of each sound feature in different wheel states, and establishing an interval value fuzzy set in different wheel states; calculating similarities between different sound states and a normal standard library; training the weighted vectors of which the similarity is calculated, and obtaining a sound characteristic expert database of different wheel states at the same time; the collected sound signals of the wheel to be detected are preprocessed and converted into digital signals to calculate the membership degree, similarity calculation is conducted on the digital signals and the established different sound characteristic expert databases, reasoning is conducted through a reasoning method, and the hurt degree of the wheel is obtained. The method has the advantages of higher identification precision, higher speed and higher robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of wheel driving environment detection, and particularly relates to a method for detecting wheel indentation damage. Background Art

[0002] Due to the long-term driving of high-speed trains, trains, and freight cars on tracks, the wear caused by the rails on the wheels; coupled with the influence of the external environment, for example, interference such as rain, snow, wind, sun, etc., and the corrosion ionization reaction between the wheels and the environment. In addition, the collision of external objects with the wheels may cause wheel damage, which may further lead to safety risks and economic losses. For the damage detection of railway and motor vehicle wheels, the existing methods mainly rely on manual inspection of the wheels, hammering, and watching videos. It is mainly manual on-site supervision or by manually viewing static photos. Although these methods can detect the potential risks of indentation damage, there are problems such as high labor intensity and strong subjectivity of the detection results. This makes the manual labor intensity very high. Moreover, due to the limited photo angle and accuracy, it is very difficult to identify some subtle problems with the human eye, and potential safety hazards cannot be discovered in time. Currently, there is an urgent need for a targeted intelligent inspection system to reduce the manual labor intensity, improve the efficiency of potential hazard detection, and promptly investigate and eliminate potential safety hazards. This patent is based on the existing problems and proposes a method for detecting wheel indentation damage by intelligent inspection.

[0003] The invention patent with the application number 202211303654.2 discloses a monitoring system and method for wheel wear of rail vehicles. The wheel flange monitoring unit is used to detect the wear degree of the wheel flange of the measured wheel, and the detection information of the wear degree of the wheel flange of the measured wheel is uploaded to the monitoring terminal through the wireless communication module. The wheel tread monitoring unit is used to detect the wear degree of the wheel tread of the measured wheel, and the detection information of the wear degree of the wheel tread of the measured wheel is uploaded to the detection terminal through the wireless communication module. After the monitoring terminal processes the detection information of the wear degree of the wheel flange of the measured wheel and the detection information of the wear degree of the wheel tread of the measured wheel, it is displayed through the display. By adopting the above system, when checking the wear degree of the wheels of rail vehicles, there is no need for manual inspection by staff, and the inspection efficiency is higher. The above patent determines the edge distance of the measured wheel through polynomial by linear regression to detect the wear degree of the wheel flange. The parameters of different indentation damage situations in the above patent are independent and cannot be generalized, and the setting is relatively complex. However, once the parameters are set, the detection accuracy is extremely high after machine learning training. Summary of the Invention

[0004] Aiming at the technical problems that the existing wheel detection methods mainly rely on manual operation and have problems such as high labor intensity and strong subjectivity of detection results, the present invention proposes a method for detecting wheel indentation damage based on interval-valued fuzzy reasoning, which has higher identification accuracy, faster speed, stronger robustness, and can also adapt to the detection of diverse wheel driving environments.

[0005] To achieve the above object, the technical solution of the present invention is implemented as follows: A method for detecting wheel bumps based on interval-valued fuzzy inference, and the steps are as follows:

[0006] Step 1: Collect the initial sound data of the vehicle wheel during normal driving and driving under the action of bumps, and preprocess the initial sound data to obtain the normal sound data of the wheel during normal driving and the bump sound data in different bump states;

[0007] Step 2: Have experts identify the collected normal sound data and bump sound data, and give multiple empirical values in different wheel states as standard membership degrees to form a normal standard library; determine the membership degree of each sound feature in different wheel states according to the sound characteristics and membership degree mapping relationship of the normal sound data and the bump sound data, and establish an interval-valued fuzzy set A in different wheel states;

[0008] Step 3: Calculate the similarity between the interval-valued fuzzy set A in different sound states and the interval-valued fuzzy set B in the normal standard library;

[0009] Step 4: Train the weighted vector for calculating the similarity according to the similarity calculated in Step 3 to obtain a sound characteristic expert library for different wheel states;

[0010] Step 5: Preprocess the sound signal of the wheel to be detected collected and convert it into a digital signal, calculate the interval-valued fuzzy degree set using the membership degree mapping relationship, calculate the similarity with the different sound characteristic expert libraries established in Step 4, and perform inference using the interval-valued fuzzy inference method based on weighted similarity to obtain the degree of wheel bump.

[0011] Preferably, the method for collecting the initial sound data is as follows: In a section of construction route, install pickups of different models at the intersection of each wheel and the wheel axle of the vehicle, install a sound collection device on the top of the wheel, simulate the setting of bumps on the test route, and let the vehicle run back and forth on the route in the experimental area to collect the sound data in different bump situations; the sound collection device includes a sound input interface, a sound signal conversion unit, and a memory connected in sequence, the pickup is connected to the sound input interface, the sound input interface receives the audio signal collected by the pickup and transmits the audio signal to the sound signal conversion unit, the sound signal conversion unit converts each index value of the sound characteristics of the sound analog signal into a corresponding digital signal, and the converted signal is stored in the memory; the sound characteristics generally refer to the amplitude, frequency, spectrum, direction, timbre, wavelength, cut-off frequency, bandwidth frequency, margin, amplitude, and crossing frequency of the sound;

[0012] The preprocessing is to prune or truncate the initial sound data with amplitudes or frequencies different from the mainstream audio, abnormally low and high;

[0013] Engineers who regularly inspect and detect the wheels on the railway distinguish the sounds collected under normal and different bruise conditions, give the empirical values of each sound feature under different wheel states, and construct an interval-valued fuzzy set based on the empirical values to form a normal standard library; Railway experienced masters classify the sounds collected in different states into several categories, and calculate the membership values of each sound feature in different sound states for each of these categories according to the membership mapping relationship to obtain all the memberships of several categories, and form an interval-valued fuzzy set of different wheel states;

[0014] According to the mapping relationship between the universe of discourse composed of different wheel states and the set I[0,1] of all closed sub-interval value sets on the interval [0,1], determine the maximum and minimum memberships of different sound states to obtain the interval-valued fuzzy set A.

[0015] Preferably, let X = {x1, x2, …, x i , … x n} be a universe of discourse, where x i is a different bruise state or normal state, then A: X → I[0,1] is called an interval-valued fuzzy set; where, I[0,1] represents the set of all closed sub-interval value sets on the interval [0,1]; for Denote the interval-valued fuzzy set A = {[A - (x i ), A + (x i )]|x i ∈ X}, A - (x i ) and A + (x i ) respectively represent the minimum membership and the maximum membership of the state x i belonging to the interval-valued fuzzy set A, and the membership A - : X → I[0,1], A + : X → I[0,1]; and for A - (x i ) ≤ A + (x i ); The set of all interval-valued fuzzy sets on the universe of discourse X is denoted as IF(X)

[0016] For interval-valued fuzzy sets A, B ∈ IF(X), the operations of intersection and union of interval-valued fuzzy sets A and B are as follows:

[0017] A ∩ B = {[A - (x i ) ∨ B- (x i ), [A + (x i ) ∧ B + (x i )] | x i ∈ X}

[0018] A ∪ B = {[A - (x i ) ∧ B - (x i ), [A + (x i ) ∨ B + (x i )] | x i ∈ X}

[0019] Among them, ∨ represents the logical operation of "or", and ∧ represents the logical operation of "and".

[0020] Preferably, the membership degree mapping relationship A(x) of the interval-valued fuzzy set A is or where x0 and σ 2 respectively represent the central value of the function and the extension degree of the function image.

[0021] Preferably, the calculation method of the similarity is as follows: If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2,..., x n} is a finite set, xi is different bruise states or normal states; for any given positive integer p ∈ N * , λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2,..., ω n ) is a weighted vector related to the universe of discourse X, where the weighting coefficient ω i ∈ [0, 1], n represents the number of wheel states, and the similarity between the interval-valued fuzzy sets A and B is:

[0022]

[0023] When the positive integer p = 1 or 2, the calculation method of the similarity is:

[0024] If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2,..., x n} is a finite set, λ i , μ i ∈ [0, 1] and λ i + μ i= 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where ω i ∈[0, 1], The similarity between the interval-valued fuzzy sets A and B is:

[0025]

[0026] If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, λ i , μ i ∈[0, 1] and λ i + μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where ω i ∈[0, 1], The similarity between the interval-valued fuzzy sets A and B is:

[0027]

[0028] Preferably, the weighted vector is an ordered weighted average operator, and the implementation method is: Let the decision data (a1, a2, …, a n ) ∈ R n , R is the set of real numbers, a i represents the membership degree of the sound characteristics of the wheel. Define the function F: R n →R. If then the function F is called an n-dimensional ordered weighted average operator, and the weighted coefficient ω i ∈[0, 1], i ∈ {1, 2, …, n}, ω = (ω1, ω2, … ω n ) is an n-dimensional weighted vector associated with the function F, and The sorted data

[0029] When the weighted vector ω takes the following special values, the function F is a special value, and the situations are as follows:

[0030] (2) When ω = (1, 0, 0, …, 0),

[0031] (2) When ω = (0, 0, 0, …, 1),

[0032] (3) When then

[0033] Preferably, the weighted vector ω = (ω1, ω2, … ω n ) is an ordered weighted vector, the weighting coefficients are combination numbers, and the weighting coefficients:

[0034]

[0035] and holds;

[0036] From the properties of combination numbers, we know that Then the weighting coefficients

[0037]

[0038] If the wheel is normal, i.e., not bruised, then the combination number The calculation result is 0; otherwise, according to different bruise degrees, the combination number The calculation result is any integer between 0 and 2 n-1 inclusive.

[0039] Preferably, the implementation method of step four is: making the calculated interval value fuzzy, calculating the similarity with the established normal standard library, obtaining whether the collected sound data belongs to the normal state or which bruise state from the similarity, training and learning both the bruise data and the normal data, and establishing different sound characteristic expert libraries;

[0040] The method of the training and learning is: randomly extracting all the collected sound data, calculating the membership degrees in the normal state and different bruise states by the membership mapping relationship, calculating the similarity between the obtained membership degrees and the interval value fuzzy sets corresponding to the empirical values of the sound characteristics in the normal standard library to obtain the similarity of the corresponding sound characteristic fuzzy interval values, comparing the calculated similarity with a given threshold. If the similarity is greater than the given threshold, then the weighting coefficient is decreased by 0.1 - 0.2; if the similarity is less than the given threshold, then the weighting coefficient is increased by 0.1 - 0.2, and the range of the weighting coefficient is [0, 1]; then, taking several values in the neighborhood of the membership degree calculated for each wheel state, when the given number of iteration steps is reached, calculating the mean value of all the membership degrees as the standard and storing it in the sound characteristic expert library as the true value of this sound characteristic;

[0041] After the training reaches the required accuracy, the remaining part of the voice data used as the test set is imported into the trained similarity calculation model for testing. The specific implementation method is as follows: Calculate the membership degrees of the normal state and different bruised states based on the voice data; Compare the membership values with the membership degrees of the same state in the voice characteristic expert library established after training with the training set. When the difference is less than the given error requirement, determine that this state is a normal or bruised state of a certain situation, and thus mark this state as a certain bruised or normal state; When the difference is greater than or equal to the given error requirement, continue to compare with the next state until the last similar state is compared.

[0042] Preferably, perform preprocessing of noise reduction, redundant voice truncation, and pruning on a wheel sound signal to be detected; Convert the preprocessed sound signal into a digital signal and store it in a memory; Then encode and compress the digital signal and store it in a processing unit; Calculate the membership degree of the wheel sound signal based on the digital signal, calculate each weight value of the antecedent of the interval-valued fuzzy rule using the weight assignment method of the ordered weighted average operator based on combination numbers, calculate the similarity with the interval-valued fuzzy set corresponding to the voice characteristic expert library, and use the state of the voice characteristic expert library with a similarity less than the preset error as the state of whether the wheel is bruised.

[0043] Taking multiple rules as an example, the interval-valued fuzzy reasoning method based on weighted similarity is given as follows:

[0044] Rule 1: If the membership degree corresponding to the normal state of the wheel is 0, then the similarity value obtained by comparing with the standard expert library is less than the given threshold δ1;

[0045] Rule 2: If the membership degree corresponding to the slightly bruised state of the wheel is 0.01, then the similarity value obtained by comparing with the standard expert library is less than the given threshold δ2;

[0046] Rule 3: If the membership degree corresponding to the moderately bruised state of the wheel is 0.02, then the similarity value obtained by comparing with the standard expert library is less than the given threshold δ3;

[0047] …

[0048] Rule n: If the membership degree corresponding to the extremely severely bruised state of the wheel is 1, then the similarity value obtained by comparing with the standard expert library is less than the given threshold δ n ;

[0049] Set the following three conditions:

[0050] (1) Assign an appropriate threshold τ i , τ i ∈[0,1], i = 1, 2, …, n, to determine whether the i-th rule can be used. If it can be used, activate the i-th rule; otherwise, do not activate it;

[0051] (2) Equip the antecedent of the \(i\)-th rule with an appropriate threshold vector \(\gamma\) i \(= (\gamma\) i1 , \(\gamma\) i2 , \(\cdots\), \(\gamma\) im ), where \(\gamma\) ij is the threshold equipped for the antecedent \(A\) ij of the \(i\)-th rule. Given the threshold value \(\gamma\) ij \(\in [0, 1]\), \(i = 1, 2, \cdots, n\), \(j = 1, 2, \cdots, m\); when the given fact to be measured and the similarity of the antecedent \(A\) ij is , then the membership degree corresponding to the fact to be measured is the membership degree of this voice feature to be measured; if the similarity , then take the similarity , then skip the current antecedent \(A\) ij , and make the fact to be measured calculate the interval-valued similarity with the next antecedent \(A\) ij in turn;

[0052] (3) Assign a weight \(\omega\) i to the antecedent of the \(i\)-th rule, \(\omega\) i1 \(= (\omega\) i2 , \(\omega\) im ), \(\omega\) ij \(\in [0, 1]\), \(i = 1, 2, \cdots, n\), \(j = 1, 2, \cdots, m\), and where, \(\omega\) ij represents the influence degree of the antecedent \(A\) ij of the \(i\)-th rule on the given consequent \(B\) i ; after introducing the threshold, the general form of interval-valued fuzzy modus ponens inference is:

[0053] Given \(R1: A\) 11 and \(A\) 12 and \(\cdots\) and \(A\) 1m \(\to B1, \gamma1, \tau1\)

[0054] \(R2: A\) 21 and \(A\) 22 and \(\cdots\) and \(A\) 2m \(\to B2, \gamma2, \tau2\)

[0055] \(\cdots\cdots\)

[0056] \(R\) n : \(A\) n1 and \(A\) n2 and \(\cdots\) and \(A\) nm \(\to B\) n , \(\gamma\) n , \(\tau\) n

[0057] and given facts and and … and

[0058] Find B *

[0059] where A i1 is an interval-valued fuzzy set on the universe of discourse A i2 is an interval-valued fuzzy set on the universe of discourse A ij is an interval-valued fuzzy set on the universe of discourse B i is an interval-valued fuzzy set on the universe of discourse Y = {y1, y2, …, y q}, m1, m2, m j respectively represent the parameter index of the m1-th in the first case of the antecedent, the parameter index of the m2-th in the second case of the antecedent, and the parameter index of the m j -th in the j-th case of the antecedent; y q represents the index of the q-th parameter of the consequent.

[0060] Preferably, the implementation steps of the interval-valued fuzzy reasoning method based on weighted similarity are as follows:

[0061] Step 1: Calculate the similarity between each antecedent of each rule and the antecedent corresponding to the given fact to be measured:

[0062]

[0063] where m and n represent the maximum upper limit values of the index numbers of different categories;

[0064] Compare the obtained similarity α ij with the given corresponding threshold γ ij . If α ij ≥ γ ij , then record If α ij <γ ij , then record represents the optimal similarity of the interval; record the similarity vector after comparison as

[0065] Let the weighting coefficient w i =(w i1 , w i2 , …, w im ), i = 1, 2, …, n is an OWA operator based on combination numbers, and calculate the comprehensive similarity π i of the i-th rule, then

[0066] Among them, (ε i1 , ε i2 , …, ε im ) is the vector obtained by sorting the sub-vectors of the similarity vector in descending order;

[0067] According to the given threshold τ i , the following conclusion can be drawn:

[0068] If π i ≥ τ i , then activate this rule;

[0069] If π i < τ i , then this rule is not activated;

[0070] Step 2: When only the i-th rule is activated, calculate the output conclusion:

[0071] When p1 rules are all activated, the conclusions deduced from these activated rules are added with appropriate weighting coefficients to obtain the final conclusion;

[0072] Assume that the parameter β is involved in the weights. Use the comprehensive similarity π i to determine the weight vector. The parameter β is Let s be the number of rules with the maximum value of the comprehensive similarity among the activated rules;

[0073] Let I = {i | the i-th rule is activated and π i = β, 1 ≤ i ≤ n}, and assign the weight i to B Then the conclusion of the rule with the maximum value of the comprehensive similarity is

[0074] Assign the weight to the conclusions deduced from the remaining activated rules

[0075] Step 3: Take the union of the conclusions obtained in Step 2 to get the actual output

[0076] Compared with the prior art, the beneficial effects of the present invention are as follows: A method for calculating the similarity of interval-valued fuzzy sets is proposed: interval-valued fuzzy sets are defined; a method for calculating the degree of similarity is given; a method for assigning weights to the OWA (Ordered Weighted Averaging operator) operator is studied, which can better weaken the adverse effects brought by subjective factors; for multi-dimensional fuzzy reasoning, a double-threshold interval-valued fuzzy reasoning method is proposed. By assigning an appropriate threshold vector to the antecedent of each rule, the influence of some secondary factors is filtered out, so that some unnecessary rules are filtered out, thereby assigning high weights to key features and low weights to noise or redundant features at the same time, which can speed up, accurately identify, and highlight the bruised parts; applying this interval-valued fuzzy reasoning method to distinguish the sound of wheel bruising, the effectiveness and feasibility of the proposed interval-valued fuzzy reasoning method are shown through example tests and analyses; compared with other existing wheel detection methods, the present invention has the advantages of higher identification accuracy, faster speed, and stronger robustness, and can also adapt to the detection of diverse wheel driving environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0078] Figure 1 It is a flowchart of the present invention.

[0079] Figure 2 It is a schematic diagram of a data acquisition device used in different scenarios of the present invention.

[0080] Figure 3 It is a schematic diagram of an interval-valued fuzzy reasoning detection system developed by the present invention.

[0081] Figure 4 It is a curve graph comparing the calculated value of the interval-valued similarity of the present invention with the membership degree value of an actual expert system.

[0082] Figure 5 It is a stability curve graph of the present invention.

[0083] Figure 6 It is a correct detection rate curve graph comparing the present invention with existing detection methods. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0084] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0085] As Figure 1 shown, a method for detecting wheel bruising based on interval-valued fuzzy inference includes the following steps:

[0086] Step 1: Data collection: Collect the initial sound data of the vehicle wheel during normal driving and driving under the action of bruising objects in different bruising situations, and preprocess the initial sound data to obtain the normal sound data of the wheel during normal driving and the bruising sound data after being bruised in different situations.

[0087] In a construction route, pickups of different models are respectively installed at the intersection of each wheel and the wheel axle of the vehicle. At the same time, a corresponding sound collection device (5 cm × 6 cm × 0.5 cm in size) is installed on the top of the wheel. Bruising objects are simulated and set on the test route, and the vehicle is allowed to run back and forth on the route in the experimental area to collect sound data in different bruising situations. The collected sound data is transmitted to a small data server in real time. The process of sound data collection is as Figure 2 shown, Figure 2 The regulated DC power supply in it is 5V or 12V, which provides power supply for the pickup. The sound collection device includes a sound input interface, a sound signal conversion unit and a memory connected in sequence. The pickup is connected to the sound input interface. The sound input interface receives the audio signal collected by the pickup and transmits the audio signal to the sound signal conversion unit. The sound signal conversion unit converts the sound analog signal into a digital signal, and the converted signal is stored in the memory. The memory is a read-write and storage TF card.

[0088] The noise reduction preprocessing is to prune or truncate the audio different from the mainstream audio, with different low and high amplitudes, or different low and high frequencies, and remove redundant information through the depth-first algorithm, leaving only the information with sound characteristics to reduce noise interference. If there is bruising, the corresponding wheel driving sound is within the range of 10% to 50% lower or higher than the normal sound frequency, and its tone chromatogram deviates more than 15° from the normal tone chromatogram. The present invention has collected the sounds of the wheel during normal driving and after being bruised in different situations through the pickup, and stored them in the processor as a database for ready use at any time.

[0089] Step 2: Have experts identify the collected normal sound data and bruised sound data, and give the empirical values under different wheel states as standard membership degrees to form a normal standard library; determine the membership degrees of each sound feature under different wheel states according to the sound feature and membership degree mapping relationship between the normal sound data and the bruised sound data, and establish an interval-valued fuzzy set A under different wheel states.

[0090] Have engineers who have been inspecting and detecting wheels on the railway for many years identify the normal and bruised sounds collected, give the empirical values of each sound feature under different wheel states, construct an interval-valued fuzzy set according to the empirical values, and form a normal standard library. Have railway experienced masters classify the sounds in different states collected, that is, divide them into several categories, and calculate the membership degree values of each sound feature in different sound states according to the mapping of the membership degree function for these several categories respectively, and obtain all the fuzzy degrees of several categories respectively. The fuzzy degree set of the first category is A, the fuzzy degree set of the second category is B, and so on.

[0091] According to the sound signal collected by the pickup, through a digital converter, convert each index value of the sound feature of the sound signal into a corresponding digital signal. The sound feature generally refers to the amplitude, frequency, spectrum, direction, timbre, wavelength, cut-off frequency, bandwidth frequency, margin, amplitude, and crossing frequency of the sound. Determine the maximum and minimum membership degrees of different sound states according to the mapping relationship between the universe of discourse composed of different wheel states and the set I[0,1] of all closed sub-interval value sets on the interval [0,1], and obtain the interval-valued fuzzy set A.

[0092] Definition 1. Let X = {x1, x2, …, x i , … x n} be a universe of discourse, where x i is different bruised states or normal states. Then A: X → I[0,1] is called an interval-valued fuzzy set; where, I[0,1] represents the set of all closed sub-interval value sets on the interval [0,1]; for Denote the interval-valued fuzzy set A = {[A - (x i ), A + (x i )] | x i ∈ X}, A - (x i ) and A + (x i ) respectively represent the minimum membership degree and the maximum membership degree of the state x i belonging to the interval-valued fuzzy set A. These membership degrees are all calculated by the same membership degree function. Calculate the membership degrees of all states within the interval according to the same membership degree function, arrange them from small to large, the smallest is the minimum membership degree, the largest is the maximum membership degree, and the membership degree A- : X → I[0, 1], A + : X → I[0, 1]; and for A - (x i ) ≤ A + (x i ); All interval-valued fuzzy sets on the universe of discourse X are denoted as IF(X).

[0093] Definition 2. For interval-valued fuzzy sets A, B ∈ IF(X), the operations of intersection and union of interval-valued fuzzy sets A and B are as follows:

[0094] A ∩ B = {[A - (x i ) ∨ B - (x i ), [A + (x i ) ∧ B + (x i )] | x i ∈ X}

[0095] A ∪ B = {[A - (x i ) ∧ B - (x i ), [A + (x i ) ∨ B + (x i )] | x i ∈ X}

[0096] Among them, ∨ represents the logical operation of "or", and ∧ represents the logical operation of "and".

[0097] Step 3: Calculate the similarity between the interval-valued fuzzy set A in different sound states and the interval-valued fuzzy set B in the normal standard library.

[0098] To measure the situation between defects and normal components, a metric index is needed. Then, for the research on detecting wheel gouges using interval-valued fuzzy sets, the index to be used is the similarity of interval-valued fuzzy sets. The similarity of interval-valued fuzzy sets is defined below.

[0099] Definition 3. Take the membership function of the interval-valued fuzzy set A (A ∈ IF(X)) as A(x), where IF(X) represents the set formed by the interval-valued fuzzy membership degrees corresponding to all states X, and X represents the states of the wheel sound signals under normal or different gouge conditions, including the gouged component, the degree of gouging, and the position of the component. The membership function A(x) = [A - (x), A +(x)], assume there is a mapping N: IF(X)×IF(X)→I, for interval-valued fuzzy sets A, B ∈ IF(X), if the following conditions are satisfied:

[0100] (1) N(A, A) = 1;

[0101] (2) X = (1, 1), 0 indicates no relevance to this state, and 1 indicates complete relevance to this state.

[0102] (3) N(A, B) = N(B, A)

[0103] (4)

[0104] Then N(A, B) is called the similarity degree between interval-valued fuzzy set A and interval-valued fuzzy set B. Among them, I represents the set of all interval fuzzy values obtained by converting all sounds into corresponding numbers, and C has the same meaning as the interval-valued fuzzy set A\B, both referring to the interval-valued fuzzy sets of different sound segments. The membership function A(x) generally takes Here, x0, σ 2 respectively represent the central value of the function and the degree of extension of the function image.

[0105] Considering the different influence degrees of the upper and lower limits of the interval values on calculating the similarity degree, and the different importance of each factor or attribute in the universe of discourse X for the conclusion, different weights are assigned to them according to the actual situation, and the following similarity calculation formula is obtained:

[0106] Conclusion 1. If interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, for any given positive integer p ∈ N * , λ i , μ i ∈[0, 1] and λ i +μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where the weighting coefficient Define the similarity degree between interval-valued fuzzy sets A and B:

[0107]

[0108] Among them, x i are different bruise states or normal states. n represents the number of sound states.

[0109] When the positive integer p = 1 or 2, the following inferences can be obtained from the above theorem:

[0110] Corollary 1. If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where Define the similarity degree of the interval-valued fuzzy sets A and B as

[0111]

[0112] Corollary 2. If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where Define

[0113]

[0114] Then N1(A, B) is the similarity degree of the interval-valued fuzzy sets A and B.

[0115] Transform the whole process of the above interval-valued fuzzy calculation into machine language and load it into the hardware device processor to obtain the fuzzy interval-valued algorithm model. Calculate the fuzzy interval values of the bruised sound and the normal sound by this fuzzy interval-valued algorithm model. According to the sound signal collected by the pickup, convert each index value of the sound into the corresponding number. First, calculate the membership degree of the sound state according to the mapping relationship given by the membership function, and then calculate the fuzzy set of the interval value corresponding to each membership degree, that is, the similarity degree, by formula (1).

[0116] Step 4: Train the weighted vector for calculating the similarity degree according to the similarity degree calculated in Step 3 to obtain the expert library of sound characteristics for different wheel states.

[0117] Calculate the similarity degree of formula (2) with the calculated fuzzy interval values and the standard library established by the interval-valued fuzzy sets of the normal wheels composed of expert experience values and the interval-valued fuzzy sets of wheels with different bruised degrees. Determine whether the collected sound signal belongs to the normal state or which bruised degree according to the similarity degree. Train and learn both the normal data and the data of different bruised degrees to establish different expert libraries of sound characteristics.

[0118] Bring the collected sound data back to the laboratory, calculate the fuzzy interval values between the bruised sound and the normal sound, classify and analyze according to the interval values, compare the similarity calculated by formula (2) with the given threshold, so as to find out the different characteristics between the bruised state and the normal sound, and write the difference algorithm for the data in the standard library and the test data. The different characteristics of sound are mainly reflected by the amplitude, frequency, and spectrum of the sound. For example, judged by the amplitude, when the wheel is damaged by an external object, the amplitude of the sound vibration waveform will mutate and increase significantly. Randomly extract all the collected data, which may include normal data and data in different bruised states. Use a part of the collected sound data as a data set for training. Remove the redundant information from the collected sound data, convert the sound signal into a digital signal with a digital-to-analog converter, and then calculate the membership degrees in the normal state and different bruised states by the membership function A(x). Substitute the fuzzy set of the interval values corresponding to the empirical values of the sound characteristics corresponding to the normal standard library into formula (3) for the calculated membership degrees, and obtain the similarity of the corresponding sound characteristic fuzzy interval values, so as to obtain the similarity between the normal state of the wheel and different bruised stages. Although there are multiple states in the standard library, each state in the standard library is labeled, so the corresponding sound characteristics are carried out according to the labels. Judge whether this similarity is greater than the given threshold according to the given threshold, so as to obtain the sound characteristics in the normal or different bruised degree states. If the similarity is greater than the given threshold, the weighting coefficient ω i+1 = ω i - 0.1; if it is less than the given threshold, then adjust the weighting coefficient ω i+1 = ω i + 0.1, and the weight range is 0 to 1. The membership degree is the numerical value of a single sound signal calculated by Definition 3. The fuzzy interval value similarity refers to the calculation of two index factors of the sound characteristics to be detected and the corresponding sound characteristics in the expert library. The sound characteristics include amplitude, frequency, timbre, etc. Then, take several values in the neighborhood of the membership degree calculated for each wheel state. When the given number of iteration steps is reached, the iteration ends, and calculate the mean value of all membership degrees as the standard and store it in the sound characteristic expert library as the true value of this sound characteristic. Here, the membership degree neighborhood refers to a circle with each membership degree as the center and a given r (r < 1) as the radius. The data inside the circle are the neighborhoods of this membership degree.

[0119] After the training reaches the required accuracy, import the remaining part of the voice data used as the test set into the trained model for testing. The specific implementation method is as follows: Similar to the previous step, first convert the voice signal into a digital signal, and then calculate the membership degrees in the normal state and different gouge states by the defined membership function A(x); then, compare the calculated membership degree values with the membership degrees of the same state in the standard library established after training with the training set (the standard library is already marked. Compare the state to be measured with the states in the standard library and determine according to the marked number). When the difference is less than the given error requirement, determine that this state is the normal state or a gouge state of a certain situation, and thus mark this state as a certain gouge state or the normal state; when the difference is greater than or equal to the given error requirement, continue to compare with the next state until the last similar state is compared (the states in the standard library are complete, and there must be a state that matches it). Test the entire model of the wheel to obtain whether the wheel is in the normal state or in different degrees of gouge states. Such a test can quickly and accurately obtain the determination result of the wheel state (normal state or different degrees of gouge states), and has strong anti-noise ability; there will be no noise interference. The interval value fuzzy set in the voice characteristic expert library is perfect and fixed; all new data are for testing.

[0120] Step 5: Preprocess the collected voice signal of the wheel to be detected and convert it into a digital signal, calculate its interval value fuzzy degree set using the membership function, calculate the similarity with different voice characteristic expert libraries established in Step 4, and perform reasoning using the interval value fuzzy reasoning method based on weighted similarity to obtain the degree of wheel gouge.

[0121] For a wheel sound signal to be detected, noise reduction, redundant sound truncation, and pruning preprocessing are first performed. Then, the preprocessed sound signal is converted into a digital signal and stored in a memory. Secondly, the digital signal is encoded and compressed and stored in a processing unit for ready access. The encoding and compression method here is as follows: Assign a value of 0 to those with membership values greater than a given threshold, and assign a value of 1 to those with membership values less than the given threshold. First, count the total number of 0s and 1s, and determine whether the total number is a prime number. If it is not a prime number, use the largest factor among all the decomposed factors as the length of the segment; if it is a prime number, segment it into segments of length 5. For each segment, check if it is all 0s. If it is all 0s, assign the segment a single-digit 0; if 1s appear in the segment, the first digit is 1, and the numbers after 1 are the numbers of the segment itself, that is, 0 for 0s and 1 for 1s for encoding. Because in the experiment, a certain threshold is set according to experience, compared with all the calculated membership values, this threshold is less than most of the membership values. In this way, the situation where the membership value is greater than a given threshold is the most common. Therefore, the sound length obtained by this encoding and compression method is less than the length of the original sound, achieving data encoding and compression. Compare the magnitudes of the digital signals corresponding to the sound characteristics of the wheel after conversion from the one to be measured and the sound digital signals in the sound characteristic expert database, and substitute them into formula (3) to calculate the similarity. Then, obtain the similarity of the matching between the characteristics of the wheel to be measured and the standard database. Use the OWA operator weighting method based on combination numbers to calculate the weights of the antecedents of the interval-valued fuzzy rules. Compare the difference between this similarity and a given threshold to obtain the matching degree. Increase the weight with a small matching degree by 0.2 and decrease the weight with a large matching degree by 0.1 (obtained according to the experience of machine learning weight adjustment. These weight adjustments are during the model training process. Once the model is trained, the weighting coefficients are no longer adjusted during its detection process). Here, the matching degree refers to the value obtained by taking the difference between the similarity and the given threshold. Match this similarity with the sound characteristic expert database, and determine whether the wheel is bruised according to the calculated matching degree. This process is the fuzzy interval-valued reasoning process. Highlight the features of interest and weaken the features of no interest through the matching degree.

[0122] Definition 4. Let the decision-making data (a1, a2, …, a n ) ∈ R n , where R is the set of real numbers, and ai represents a specific real number that can determine the judgment result, which is the membership value obtained by the mapping calculation in Definition 3 and represents a certain performance index of the wheel, such as the amplitude, direction, frequency, and sound tone on the wheel sound signal. Define the function F: R n → R. If then the function F is called an n-dimensional ordered weighted average operator (OWA operator). Here, ω i ∈ [0, 1], i ∈ {1, 2, …, n}, and ω = (ω1, ω2, … ω n ) is the n-dimensional weighted vector associated with the function F, and where the sorted data The weighted operator is obtained by the scoring of masters with railway experience. The weights greater than the threshold are increased by 0.1 - 0.2, and the weights less than the threshold are decreased by 0.1.

[0123] The obvious characteristic of the OWA operator is that it first re - sorts the given decision - making data (a1, a2, …, a n ) in descending order to obtain new data (b1, b2, …, b n ) and weights the new data with the given weighted vector. The weight ω i has nothing to do with the element a i and is only related to the i - th position in the decision - making process.

[0124] The OWA operator is a weighted method for multi - attribute decision - making information between the maximum operator and the minimum operator. When the weighted vector ω takes the following special values, the function F is a special value, and the situations are as follows:

[0125] (1) When ω=(1, 0, 0, …, 0), At this time, the OWA operator is equivalent to the "∨" operator in fuzzy operations. Because the first element of the weight vector ω is the value 1, which is a maximum operation, that is, an "or" operation.

[0126] (2) When ω=(0, 0, 0, …, 1), At this time, the OWA operator is equivalent to the "∧" operator in fuzzy operations. Because the first element of the weight vector ω is the value 0, which is a minimum operation, that is, an "and" operation.

[0127] (3) When , At this time, the OWA operator is equivalent to the arithmetic - mean operator.

[0128] In the process of decision - making or reasoning, some experts may be influenced by emotions and make unreasonable evaluations of the research object according to their own likes and dislikes. Therefore, in the process of weighting the experimental data, it is necessary to minimize the unfairness caused by such subjective factors so that the evaluation results can reflect fairness and justice as much as possible. The weights given from this perspective are relatively reasonable because whether it is a high score given by an expert due to preference or a low score due to dislike, they are all arranged in positions with smaller weights, which can better weaken the adverse effects brought by subjective factors. Based on this consideration, a method for assigning weights to the OWA operator based on different weights is given, which is the total similarity of each component of the comprehensive vector.

[0129] Definition 5. Assume that the components of the weighted vector ω=(ω1, ω2, … ω n ) are defined according to the following combination numbers and satisfy the conditions of Definition 4, then the weight at this time is the ordered weighted weight.

[0130]

[0131] Obviously, it holds. represents the total number of combinations of taking \(i - 1\) from \(n - 1\) numbers, has the same meaning as that is, it represents the total number of taking \(k\) numbers from \(n - 1\) numbers.

[0132] From the properties of combination numbers, we know that then

[0133]

[0134] According to the natural random sequence, the above-defined formula (4) is sorted in an orderly manner according to objective phenomena and conforms to the natural law of the development of things.

[0135] If the wheel is normal, that is, not bruised, then the calculation result is 0; otherwise, according to different degrees of bruising, the calculation result is any integer between 0 and 2 n-1 including 2 n-1 .

[0136] The wheel bruising is affected by different factors such as the operating environment, road conditions, load, vehicle type, etc. To calculate the similarity of different factors, multiple decision-making inferences are required. Taking multiple rules as an example below, an interval-valued fuzzy inference method based on weighted similarity is given to make the discussion closer to reality and more general.

[0137] Rule 1: If the membership degree corresponding to the normal state of the wheel is 0, then the similarity value obtained by comparing with the standard expert database is less than the given threshold \(\delta_1\);

[0138] Rule 2: If the membership degree corresponding to the slightly bruised state of the wheel is 0.01, then the similarity value obtained by comparing with the standard expert database is less than the given threshold \(\delta_2\);

[0139] Rule 3: If the membership degree corresponding to the moderately bruised state of the wheel is 0.02, then the similarity value obtained by comparing with the standard expert database is less than the given threshold \(\delta_3\);

[0140] …

[0141] Rule \(n\): If the membership degree corresponding to the extremely severely bruised state (cannot work normally) of the wheel is 1, then the similarity value obtained by comparing with the standard expert database is less than the given threshold \(\delta\) n ;

[0142] Set the following three conditions:

[0143] (1) Assign an appropriate threshold τ to the i-th rule i , τ i ∈[0,1], i=1,2,…,n, is used to determine whether this rule can be used. If it can be used, then this rule is activated; otherwise, it is not activated.

[0144] (2) Equip the antecedent of the i-th rule with an appropriate threshold vector γ i =(γ i1 ,γ i2 ,…,γ im ), where γ ij is the antecedent A assigned to the i-th rule ij The threshold value, γ ij ∈[0,1], i=1,2,…,n,j=1,2,…,m, the antecedent refers to the premise of judging whether the wheel is damaged, which is the condition of rule i, i.e., the degree of damage such as slight damage and moderate damage. With the antecedent A ij Similarity When The corresponding membership is the membership of the sound feature to be tested; if the similarity Then take the similarity Then skip the current antecedent A ij , so that the fact to be tested Sequentially with the next antecedent A ij Calculate the similarity of interval values. This method has certain practical significance, because when the facts to be tested are With the antecedent A ij Similarity When the time is too small, it is considered that the facts to be tested at this time have no effect on the result.

[0145] (3) According to the actual situation, the impact of each antecedent on the result is different. The antecedent of the i-th rule is given a weight ω i =(ω i1 ,ω i2 ,…,ω im ),ω ij ∈[0,1], i=1,2,…,n, j=1,2,…,m, and ω ij A represents the antecedent of the i-th rule ij For a given rule consequent B i The aftermath refers to the judgment result, that is, whether the wheel is damaged.

[0146] After the threshold is introduced, the general form of interval-valued fuzzy formula reasoning is:

[0147] Known R1: A 11 And A12 and... and A 1m → B1, γ1, τ1

[0148] R2: A 21 and A 22 and... and A 2m → B2, γ2, τ2

[0149] ......

[0150] R n : A n1 and A n2 and... and A nm → B n , γ n , τ n

[0151] and given facts and and... and

[0152] Find B *

[0153] where Ai1 is an interval-valued fuzzy set on the universe of discourse and Ai2 is an interval-valued fuzzy set on the universe of discourse and A ij is an interval-valued fuzzy set on the universe of discourse and B i is an interval-valued fuzzy set on the universe of discourse Y = {y1, y2,..., y q}. m1, m2, m j respectively represent the parameter indexes of the first case of the antecedent, i.e., the qth of the normal state of the conditional attribute, the parameter indexes of the second case of the conditional attribute, i.e., the qth of the mild bruise state 1, and the parameter indexes of the jth case of the conditional attribute, i.e., the qth of the severe bruise state j. y j represents the index of the qth parameter of the consequent, i.e., the decision attribute. In the case of normal wheels and different degrees of bruising, the sound indexes of different situations collected by the pickup are marked, and the set composed of these different attributes is obtained as different universes of discourse X1, X2, X q . The universe of discourse X is the antecedent of rule i (X is the membership degree set composed of the membership degree values corresponding to different classification interval segments of the wheel), and the universe of discourse Y is the consequent of rule i (the set composed of the similarity values corresponding to the reasoning conclusions), obtained according to the rules. j

[0154]

[0155] ​Then, the similarity-based interval-valued fuzzy inference algorithm given above is as follows:

[0156] Step 1: First, calculate the similarity between each antecedent of each rule and the antecedent corresponding to the given fact to be measured:

[0157]

[0158] Among them, α ij can be calculated using the above similarity calculation formula. m and n represent the maximum upper limit values of the index numbers of different types of indicators.

[0159] First, compare the obtained similarity α ij with the given corresponding threshold γ ij . If α ij ≥γ ij , then record If α ij <γ ij , then record represents the optimal similarity of the interval. Then, record the similarity vector after comparison as

[0160] Then, let w i =(w i1 , w i2 ,…, w im ), i = 1, 2,…, n is the OWA operator based on the combination number, and calculate the comprehensive similarity of the i-th rule, denoted as π i , then

[0161]

[0162] Among them, (ε i1 , ε i2 ,…, ε im ) is the vector obtained by sorting the sub-vectors of the similarity vector from largest to smallest.

[0163] According to the given threshold τ i , there are the following conclusions:

[0164] If π i ≥τ i , then activate this rule;

[0165] If π i <τ i , then this rule is not activated.

[0166] Step 2: When only the i-th rule is activated, calculate the output conclusion according to the following calculation method: When p1 (p1 > 1) rules are activated, consider the conclusions deduced from these activated rules and add appropriate weights to obtain the final conclusion. Assume that the parameter β is involved in the weights, and use the comprehensive similarity π i to determine the weight vector, and β is defined as Let s be the number of rules with the maximum comprehensive similarity value among the activated rules.

[0167] Let I = {i | the i-th rule is activated and π i = β, 1 ≤ i ≤ n}, and assign the weight i to B Then the conclusion of the rule with the maximum comprehensive similarity value is denoted as

[0168] Assign weights to the conclusions deduced from the remaining activated rules Then the conclusions of the remaining rules are denoted as

[0169] Step 3: Take the union of the conclusions obtained in Step 2 to get the actual output The similarity value of a certain consequent, and "union" means taking the maximum of the two similarities.

[0170] Discriminate the differences in sound waveforms according to the interval value fuzzy inference method of weighted similarity, send out an alarm signal, and locate the position of the wheel with a bruise. According to the comparison result of the calculated similarity with the threshold, increase the weight value for those greater than the threshold and decrease the weight value for those less than the threshold, and then perform weighted calculation according to Definition 4. The process of increasing and decreasing weights adheres to the principle that the sum of all weights is 1. The difference between the similarity and the threshold gives the matching degree. Use machine language commands to draw waveforms in different sound states, so as to discriminate different sound waveforms. Obtain the membership degree values corresponding to the sound waves through the inverse operation of each process of calculating the matching degree before, and then convert the digital signal into a sound wave signal by a digital-to-analog converter. When collecting sound by a pickup, the obtained sound has been numbered for position at the collection end. During the machine learning training process, the position number is used as a parameter for learning at the same time. Therefore, when the machine language discriminates different sound waveforms, at the same time, it will give the position signal of the wheel with a bruise.

[0171] Step Six: Calibrate the bruised wheel, bruise time, location, and bruising object.

[0172] (1) Bruise time

[0173] By analyzing the amplitude or frequency data of the sound waves collected under normal and bruised conditions, calculate the fuzzy interval value to determine whether there is a bruise, establish an identification algorithm model, then write the program for this fuzzy interval value algorithm model to form a bruise identification system. When the system is running, the system will record the time point when the bruise appears according to its own time clock. When the sound is collected by the pickup, while the position of the obtained sound is numbered at the collection end, the time is also recorded. During the machine learning training process, the recorded time point is also used as a parameter for learning. The machine learning process is a multi-dimensional parameter learning process. Therefore, when the machine language identifies the bruise sound waveform, at the same time, the time point when the bruised wheel appears will also be given.

[0174] From the established algorithm model, when a certain wheel in a certain wheel is bruised, the algorithm model will automatically calibrate the specific bruised wheel. Because the processes of calculation, judgment, reasoning, etc. of the algorithm model are processes in which multi-dimensional parameter vectors are implemented simultaneously, in each step of calculation, judgment, and reasoning of this algorithm model, the position, time, etc. of the bruised wheel are respectively used as a parameter for implementation. Therefore, when this algorithm model identifies the bruise sound waveform, at the same time, it will automatically calibrate the specific time and position when the bruised wheel appears.

[0175] (2) Bruise location

[0176] From the established fuzzy interval value algorithm model software system, and based on the moving speed of the vehicle and the value of the time point when the system records the bruise, the system will calculate the line position when the bruise occurs, that is, the system will give a specific position on the line when the bruise occurs. For the same reason as above, because the processes of calculation, judgment, reasoning, etc. of the fuzzy interval value algorithm model are processes in which multi-dimensional parameter vectors are implemented simultaneously, in each step of calculation, judgment, and reasoning at each node of this algorithm model, the position, time, etc. of each node of the bruised wheel are respectively used as a parameter for implementation. Therefore, when this algorithm model identifies the bruise sound waveform at each node, at the same time, it will automatically calibrate the specific time and position of the bruised wheel at each node, and thus, this system will calculate the line position when the bruise occurs.

[0177] (3) Bruise object judgment

[0178] Before data analysis and during data collection, it is necessary to conduct a test on the wheel being bruised by different possible obstacles, collect the acoustic wave data of different bruising objects, and calculate the fuzzy interval value of the acoustic wave. Through different fuzzy interval values, the acoustic wave characteristics of different bruising objects on the wheel are extracted to form different acoustic wave expert databases. The parameters such as the amplitude, frequency, sound chromatogram, and the corresponding difference values, median values, mean values, variances, etc. of the converted digital sound signal whose bruising sound is different from the normal sound are written as a vector into the index value area for marking. Once a certain wheel is bruised by a certain bruising object, the proposed fuzzy interval value algorithm system will automatically match the currently occurring bruising acoustic wave with the acoustic wave expert database, thereby obtaining what the bruising object is. Calculate the similarity between the fuzzy sets A and B of the bruising acoustic wave interval value according to the previous conclusion 1, inference 1, or inference 2. Match this similarity with the acoustic wave expert database, calculate its corresponding matching degree according to definition 4 or definition 5, and determine whether the wheel is bruised and the degree of bruising based on the matching degree.

[0179] Example: Because of the different environments, different parts, different degrees, etc. of the wheel bruising, when calculating the similarity to measure different situations, there are multi - rule interval values. Then, when applying wheel bruising detection, a multi - rule fuzzy reasoning method is required. Here, define the multi - rule interval value fuzzy production rule as follows:

[0180] Given R1: A 11 and A 12 and A 13 →B1, γ1, τ1

[0181] R2: A 21 and A 22 and A 23 →B2, γ2, τ2

[0182] R3: A 31 and A 32 and A 33 →B3, γ3, τ3

[0183] and given fact and and

[0184] Find B *

[0185] where, R i represents a pre - defined rule, which includes the premise A i and the corresponding conclusion B i , γ i , τ i . A ij represents the quantization value of the wheel attribute, B i represents the membership degree value of the amplitude of the sound, γi Membership value representing the sound wavelength, τ i Membership values representing the sound frequency, which are the results used to judge the degree of wheel damage. Ai* represents the magnitude of the membership corresponding to the input, i.e., the state of the wheel to be detected, and B* represents the output corresponding to Ai*. These are mainly applied in the "inference and decision-making layer" of the last step of the algorithm system, i.e., used for judging whether the wheel has a dent and the degree of the dent.

[0186] Suppose the membership of each component in a certain interval of wheel dent (one method is obtained by scoring with an experienced expert scoring system, and the other method is calculated by a given initial value and the membership function defined by Definition 3) is calculated as follows:

[0187] A 11 ={[0.1,0.2],[0.1,0.3]},A 12 ={[0.1,0.3],[0.4,0.5],[0.7,0.8]},

[0188] A 13 ={[0.2,0.3],[0.1,0.3],[0.4,0.6],[0.8,0.9]};

[0189] A 21 ={[0.4,0.6],[0.6,0.7]},A 22 ={[0.2,0.4],[0.7,0.9],[0.4,0.5]},

[0190] A 23 ={[0.4,0.6],[0.5,0.6],[0.7,0.8],[0.3,0.5]};

[0191] A 31 ={[0.7,0.9],[0.8,0.9]},A 32 ={[0.6,0.8],[0.1,0.3],[0.5,0.7]},

[0192] A 33 ={[0.7,0.8],[0.6,0.8],[0.2,0.4],[0.5,0.7]};

[0193] B1 = {[0.2,0.4],[0.5,0.6]}, B2 = {[0.5,0.7],[0.1,0.3]}, B3 = {[0.6,0.8],[0.9,0.95]};

[0194] Assume the membership value when the wheel is normal is:

[0195]

[0196] For simplicity, take γ1 = γ2 = γ3 = (0.6, 0.6, 0.6), τ1 = 0.65, τ2 = 0.65, τ3 = 0.70.

[0197] Next, the reasoning will be carried out according to the aforementioned reasoning steps to obtain the reasoning result of the system.

[0198] First, use the similarity calculation formula given in the above Corollary 1 to calculate the similarity α between the given fact to be measured and the i-th rule. Among them, λ ij = 0.4, μ i = 0.6, and the value of ω i is determined by the OWA operator based on combinations, that is, formula (5). After calculation, the similarities are α1 = (0.57, 0.965, 0.74), α2 = (0.83, 0.77, 0.78), α3 = (0.78, 0.675, 0.865). After comparing with γ i = (0.6, 0.6, 0.6), it can be obtained that j

[0199] Then, let the influence degree weight vector ω i of the antecedent of the i-th rule on the conclusion be ω i1 , ω i2 , …, ω im ), where i = 1, 2, …, n, and the values are obtained by the OWA operator based on combinations. Then the comprehensive similarity π * between the given fact A i and the i-th rule are respectively: π1 = 0.61125, π2 = 0.79, π3 = 0.775. According to the given threshold τ i , the following multi-rule interval-valued fuzzy production rules of the fuzzy inference system can be obtained:

[0200] π1 = 0.61125 < τ1 = 0.62, so the first rule is not activated;

[0201] π2 = 0.79 > τ2 = 0.65, so the second rule is activated;

[0202] π3 = 0.775 > τ3 = 0.70, so the third rule is activated.

[0203] Since more than one rule is activated, calculate the final result according to the method in Step 2 given. From the known,

[0204] ​

[0205] Then, the final actual output result

[0206] If in the above example, the threshold γ is cancelled i , i = 1, 2, 3, and others remain unchanged. Then

[0207] π1 = 0.75375 > τ1 = 0.65, then the first rule is also activated.

[0208] Then all three rules are activated, so

[0209]

[0210] Then, at this time, the final actual output result

[0211] Comparing the above two final output results, it can be seen that there are differences between the two results. The set threshold γ i , i = 1, 2, 3 is effective and can filter out some unnecessary rules and affect the final output result.

[0212] To verify the feasibility and reliability of the interval - valued similarity defined in the present invention, through the calculations of the above formulas (1)-(3), compared with the membership degrees given by the expert system for the actual wheel bruises as Figure 4 shown. First, the collected sound data is converted from the sound signal into a digital signal by an analog - to - digital converter, and then the membership degrees in the normal state and different bruise states are calculated by the membership degree function defined in Definition 3 above; then, according to the conditions satisfying Conclusion 1, Inference 1 or Inference 2, it is directly calculated by their corresponding formulas. From Figure 4 it can be known that the values calculated by the defined formulas (1)-(3) are almost close to the true values, indicating the effectiveness and feasibility of the defined interval - valued fuzzy similarity values.

[0213] By the above interval - valued fuzzy inference method, a program is written to establish a wheel bruise detection system. From this detection system, when the sound situation of the wheel is input, the bruise state of the wheel can be automatically detected. The developed detection system is as Figure 3As shown in the figure, this detection system includes five modules, namely the interval value fuzzy inference function architecture editor, whose function is to layout each calculation and inference function module according to actual needs and write function codes for each function module; the membership degree, similarity degree, and matching degree calculation editor, whose function is to calculate the specific values of the membership degree, similarity degree, and matching degree; the fuzzy inference rule editor, whose function is to infer the judgment result from the antecedent, that is, the conditional attribute; the inference rule observation and debugging tool, whose function is to display the changes in the entire inference process and view the inference process from different angular indicators; the result output debugging and observation tool, whose function is to display the final inference result and view the result from different angular indicators. The stability of the present invention is as Figure 5 shown, and through Figure 5 it can be seen that when faced with interferences such as different road surface unevenness, slight track deformation, and slight loosening of railway tracks, this inference detection system still exhibits good robustness.

[0214] According to the application of the interval value fuzzy inference method given above, and the existing vehicle wheel detection methods such as the GAN network, Yolox, and the STC89C51 processor of the 1MCU unit, the detection results are as Figure 6 shown. After the collected sound data is converted from analog to digital by the analog-to-digital converter, this digital signal is respectively input to the input ends of different detectors, and then the output ends of each detector can be observed. Conduct 200 experiments, with each experiment consisting of 1000 groups of sound wave signals. Each time, different 1000 groups of sound wave signals are respectively input to the input ends of different detectors, and then the correct detection numbers obtained by observing the output ends of each detector are recorded. The correct detection rate for each time is obtained by dividing the correct detection number by 1000. Thus, the correct detection rates for all 200 experiments are obtained. Then, these detection rates are arranged in ascending order to obtain the results as Figure 6 shown. From the experimental process and Figure 6 it can be known that compared with the existing detection methods, the interval value fuzzy inference method of the present invention performs inference by calculating the similarity degree, and it has the following distinct advantages:

[0215] (1) It is close to the natural attributes of things, can reflect objective facts, has higher accuracy, faster detection speed, and is less affected by external interference;

[0216] (2) It can accurately locate which wheel is bruised; from the above discussion on the location of the bruise, that is, the interval value fuzzy inference method is a process in which a multi-dimensional parameter vector is simultaneously implemented through processes such as calculation, judgment, and inference. Therefore, when implementing the calculation, judgment, and inference processes at each node, the position, time, etc. of each node of the bruised wheel are respectively used as a parameter for implementation. Therefore, the present invention will automatically calibrate the specific position of the bruised wheel at each node when it is bruised, and thus, it can accurately locate which wheel is bruised.

[0217] (3) Bruising time;

[0218] (4) Bruising location;

[0219] (5) Calibration of the bruising object.

[0220] The present invention provides an interval-valued fuzzy reasoning method based on the OWA operator. The weights of the antecedents of the interval-valued fuzzy rules are calculated by the OWA operator weight assignment method based on combination numbers. When calculating the similarity of interval-valued fuzzy sets, the different influence degrees of the upper and lower limits of the interval values on the calculation of similarity are considered, as well as the different importance degrees of various factors or attributes in the universe of discourse for the conclusion, and a new calculation method for the similarity of interval-valued fuzzy sets is obtained. For multi-rule fuzzy reasoning, a double-threshold fuzzy reasoning method is proposed. This method assigns an appropriate threshold vector to each rule antecedent to filter out the influence of some secondary factors, thereby filtering out some unnecessary rules and affecting the final output result. Therefore, the reasoning algorithm given on this basis is closer to actual reasoning and the results are convenient for application. At the same time, this interval-valued fuzzy reasoning method is applied to the sound discrimination of wheel bruising. The experimental results show the effectiveness and feasibility of the proposed interval-valued fuzzy reasoning method. And the comparison between the interval-valued fuzzy reasoning method and other existing wheel detection methods is discussed, and its advantages and disadvantages are given.

[0221] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wheel bruise detection method based on interval-valued fuzzy reasoning, characterized in that, The steps are as follows: Step 1: Collect the initial sound data of the vehicle wheels during normal driving and driving under the action of a bump, and preprocess the initial sound data to obtain the normal sound data of the wheels during normal driving and the bump sound data in different bump states; Step 2: Have experts identify the collected normal sound data and bump sound data, and give multiple empirical values in different wheel states as standard membership degrees to form a normal standard library; Determine the membership degree of each sound feature in different wheel states according to the sound features and membership degree mapping relationship of the normal sound data and the bump sound data, and establish an interval-valued fuzzy set A in different wheel states; Step 3: Calculate the similarity between the interval-valued fuzzy set A in different sound states and the interval-valued fuzzy set B in the normal standard library; Step 4: Train the weighted vector for calculating the similarity according to the similarity calculated in Step 3 to obtain a sound characteristic expert library for different wheel states; Step 5: Preprocess the sound signal of the wheel to be detected collected and convert it into a digital signal, calculate the interval-valued fuzzy degree set using the membership degree mapping relationship, calculate the similarity with the different sound characteristic expert libraries established in Step 4, and perform reasoning using the interval-valued fuzzy reasoning method based on weighted similarity to obtain the degree of wheel bump.

2. The wheel bruise detection method based on interval-valued fuzzy reasoning according to claim 1, wherein The method for collecting the initial sound data is: In a section of construction route, install pickups of different models at the intersection of each wheel and the axle of the vehicle respectively. There is a sound collection device installed on the top of the wheel. Simulate and set bumps on the test route, and let the vehicle run back and forth on the route in the experimental area to collect sound data under different bump conditions; The sound collection device includes a sound input interface, a sound signal conversion unit, and a memory connected in sequence. The pickup is connected to the sound input interface. The sound input interface receives the audio signal collected by the pickup and transmits the audio signal to the sound signal conversion unit. The sound signal conversion unit converts each index value of the sound characteristics of the sound analog signal into a corresponding digital signal, and the converted signal is stored in the memory; The sound characteristics generally refer to the amplitude, frequency, spectrum, direction, timbre, wavelength, cut-off frequency, bandwidth frequency, margin, amplitude, and crossover frequency of the sound; The preprocessing is to prune or truncate the initial sound data with amplitudes different from the mainstream audio, abnormally low and high amplitudes, or abnormally low and high frequencies; Have engineers who have been inspecting and detecting wheels on the railway all year round identify the normal and bump sound data in different bump situations, give the empirical values of each sound feature in different wheel states and construct an interval-valued fuzzy set according to the empirical values to form a normal standard library; Have railway experienced masters classify the sounds in different states collected, and calculate the set of membership degree values of each sound feature in different sound states for each of these categories according to the membership degree mapping relationship to obtain all the membership degrees of several categories, and form an interval-valued fuzzy set for different wheel states; Determine the maximum and minimum membership degrees of different sound states according to the mapping relationship between the universe of discourse composed of different wheel states and the set \(I_{[0,1]}\) of all closed sub-interval value sets on the interval \([0,1]\), and obtain the interval-valued fuzzy set \(A\).

3. The method for detecting wheel bruising based on interval-valued fuzzy inference according to claim 2, wherein Let \(X = \{x_1, x_2, \ldots, x i , \ldots, x n \}\) be a universe of discourse, where \(x i \) are different bruise states or normal states. Then \(A: X \to I[0, 1]\) is called an interval-valued fuzzy set; where \(I[0, 1]\) represents the set of all closed sub-interval-valued sets on the interval \([0, 1]\); for Denote the interval-valued fuzzy set \(A=\{[A - (x i ), A + (x i )]|x i \in X\}\), \(A - (x i )\) and \(A + (x i )\) respectively represent the minimum membership degree and the maximum membership degree of the state \(x i \) belonging to the interval-valued fuzzy set \(A\), and the membership degree \(A - : X \to I[0, 1]\), \(A + : X \to I[0, 1]\); and for A - (x j ) \leq A + (x i )\); all interval-valued fuzzy sets on the universe of discourse \(X\) are denoted as \(IF(X)\) For interval-valued fuzzy sets \(A,B\in IF(X)\), the operations of intersection and union of interval-valued fuzzy sets \(A\) and \(B\) are as follows: A ∩ B = {[A - (x i ) ∨ B - (x i ), [A + (x i ) ∧ B + (x i )] | x i ∈ X} A ∪ B = {[A - (x i ) ∧ B - (x i ), [A + (x i ) ∨ B + (x i )] | x i ∈ X} Among them, \(\vee\) represents the logical operation of "or", and \(\wedge\) represents the logical operation of "and".

4. The wheel bruise detection method based on the interval-valued fuzzy reasoning method according to any one of claims 1-3, characterized in that The membership mapping relationship A(x) of the interval-valued fuzzy set A is or where x0 and σ 2 represent the central value of the function and the extension degree of the function image, respectively.

5. The method for detecting wheel bruising based on interval-valued fuzzy inference according to claim 4, wherein The calculation method of the similarity is as follows: If the interval-valued fuzzy sets A and B ∈ IF(X), when the universe of discourse X = {x1, x2,..., x n} is a finite set, x i represents different bruise states or normal states; for any given positive integer p ∈ N * , λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2,..., ω n ) is a weighted vector related to the universe of discourse X, where the weighting coefficient ω i ∈ [0, 1], n represents the number of wheel states, and the similarity between the interval-valued fuzzy sets A and B is: When taking the positive integer \(p = 1\) or \(2\), the calculation method of the similarity is: If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where ω i ∈ [0, 1], The similarity degree between the interval-valued fuzzy sets A and B is: If the interval-valued fuzzy sets A, B ∈ IF(X), when the universe of discourse X = {x1, x2, …, x n} is a finite set, λ i , μ i ∈ [0, 1] and λ i + μ i = 1, ω = (ω1, ω2, …, ω n ) is a weighted vector related to the universe of discourse X, where ω i ∈ [0, 1], The similarity degree between the interval-valued fuzzy sets A and B is:

6. The method for detecting wheel bruising based on interval-valued fuzzy inference according to claim 5, characterized in that The weighted vector is an ordered weighted average operator, and the implementation method is as follows: Let the decision-making data (a1, a2, …, a n ) ∈ R n , where R is the set of real numbers, and a i represents the membership degree of the sound characteristics of the wheel. Define the function F: R n → R. If , then the function F is called an n-dimensional ordered weighted average operator, and the weighted coefficient ω i ∈ [0, 1], i ∈ {1, 2, …, n}, and ω = (ω1, ω2, … ω n ) is the n-dimensional weighted vector associated with the function F, and the sorted data When the weighted vector \(\omega\) takes the following special values, the function \(F\) is a special value, and the situation is as follows: (1) When ω = (1, 0, 0, …, 0), (2) When ω = (0, 0, 0, …, 1), (3) When then 7. The method for detecting wheel bruising based on interval-valued fuzzy reasoning according to claim 6, wherein The weighted vector ω = (ω1, ω2, … ω n ) is an ordered weighted vector, the weighting coefficients are combination numbers, and the weighting coefficients are: and is established; As known from the properties of combination numbers Then the weighting coefficient If the wheel is normal, i.e., not damaged by a bump, the combination number is calculated to be 0; otherwise, depending on the degree of damage caused by the bump, the combination number is calculated to be any integer between 0 and 2 n-1 .

8. The method for detecting wheel bruising based on interval-valued fuzzy inference according to any one of claims 5-7, characterized in that The implementation method of the fourth step is: calculate the interval-valued fuzzy by calculation, calculate the similarity with the established normal standard library, and obtain whether the collected sound data belongs to the normal state or which kind of bruise state from the similarity. Train and learn both the bruise data and the normal data to establish different sound characteristic expert libraries; The method of the training and learning is: randomly extract all the collected sound data, calculate the membership degrees in the normal state and different bruise states according to the membership mapping relationship, calculate the similarity between the obtained membership degrees and the interval-valued fuzzy set corresponding to the empirical value of the sound characteristics in the normal standard library to obtain the similarity of the corresponding sound characteristic fuzzy interval value, compare the calculated similarity with the given threshold. If the similarity is greater than the given threshold, then the weighting coefficient is decreased by \(0.1 - 0.2\); if the similarity is less than the given threshold, then the weighting coefficient is increased by \(0.1 - 0.2\), and the range of the weighting coefficient is \([0,1]\); Then, take several values in the neighborhood of the membership degree calculated for each wheel state. When the given number of iteration steps is reached, calculate the mean value of all membership degrees as the standard and store it in the sound characteristic expert library as the true value of this sound characteristic; After the training reaches the required accuracy, then import the remaining part of the sound data used as the test set into the trained similarity calculation model for testing. The specific implementation method is: calculate the membership degrees in the normal state and different bruise states according to the sound data; compare the membership values with the membership degrees of the same state corresponding in the sound characteristic expert library established after training with the training set. When the difference is less than the given error requirement, judge that this state is normal or a certain kind of bruise state, and thus mark this state as a certain bruise or normal state; when the difference is greater than or equal to the given error requirement, continue to compare with the next state until the last similar state is compared.

9. The method for detecting wheel bruising based on interval-valued fuzzy inference according to claim 8, characterized in that, Perform preprocessing of noise reduction, redundant sound truncation, and pruning on a wheel sound signal to be detected; convert the preprocessed sound signal into a digital signal and store it in a memory; then encode and compress the digital signal and store it in a processing unit; Calculate the membership degree of the wheel sound signal according to the digital signal, calculate the respective weights of the antecedents of the interval-valued fuzzy rules by using the weight assignment method of the ordered weighted average operator based on combination numbers, calculate the similarity with the interval-valued fuzzy set corresponding to the sound characteristic expert library, and use the state of the sound characteristic expert library with a similarity less than the preset error as the state of whether the wheel is bruised; Taking multiple rules as an example, the interval-valued fuzzy reasoning method based on weighted similarity is given as: Rule 1: If the membership degree corresponding to the normal state of the wheel is 0, then the similarity value obtained by comparing with the standard expert database is less than the given threshold δ1; Rule 2: If the membership degree corresponding to the slightly bruised state of the wheel is 0.01, then the similarity value obtained by comparing with the standard expert database is less than the given threshold δ2; Rule 3: If the membership degree corresponding to the moderately bruised state of the wheel is 0.02, then the similarity value obtained by comparing with the standard expert database is less than the given threshold δ3; … Rule n: If the membership degree corresponding to the extremely severe bruise state of the wheel is 1, then the similarity value obtained by comparing with the standard expert database is less than the given threshold δ n ; Set the following three conditions: (1) Assign an appropriate threshold τ to the i-th rule i , τ i ∈ [0, 1], i = 1, 2, …, n, which is used to determine whether the i-th rule can be used. If it can be used, activate the i-th rule; otherwise, do not activate it. (2) Equip the antecedent of the $i$-th rule with an appropriate threshold vector $\gamma$ i $ = (\gamma$ i1 , $\gamma$ i2 , …, $\gamma$ im ), where $\gamma$ ij is the threshold equipped for the antecedent $A$ ij of the $i$-th rule. Given the threshold value $\gamma$ ij $\in [0, 1]$, $i = 1, 2, \ldots, n$, $j = 1, 2, \ldots, m$; when the given fact to be measured and the similarity ij of the antecedent $A$ are considered, then the membership degree corresponding to the fact to be measured is the membership degree of this sound feature to be measured; If the similarity then take the similarity then skip the current antecedent A ij , and make the fact to be measured calculate the interval value similarity with the next antecedent A in turn ij ; (3) Assign a weight ω to the antecedent of the i-th rule i =(ω i1 , ω i2 , …, ω im ), ω ij ∈[0, 1], i = 1, 2, …, n, j = 1, 2, …, m, and where ω ij represents the influence degree of the antecedent A ij of the i-th rule on the given consequent B i ; after introducing the threshold, the general form of interval-valued fuzzy modus ponens inference is: Given R1: A 11 and A 12 and... and A 1m → B1, γ1, τ1 R2: A 21 and A 22 and... and A 2m → B2, γ2, τ2 …… R n : A n1 and A n2 and... and A nm →B n , γ n , τ n and given fact and and... and Find B * Among them, A i1 is an interval-valued fuzzy set on the universe of discourse above, and A i2 is an interval-valued fuzzy set on the universe of discourse above, and A ij is an interval-valued fuzzy set on the universe of discourse above, and B i is an interval-valued fuzzy set on the universe of discourse Y = {y1, y2, …, y q}, and m1, m2, m j respectively represent the parameter index of the m1-th in the first case of the antecedent, the parameter index of the m2-th in the second case of the antecedent, and the parameter index of the m j -th in the j-th case of the antecedent; y q represents the index of the q-th parameter of the consequent.

10. The method for detecting wheel bruising based on interval-valued fuzzy reasoning according to claim 9, characterized in that, The implementation steps of the interval-valued fuzzy reasoning method based on weighted similarity are as follows: Step 1: Calculate the similarity between each antecedent of each rule and the antecedent corresponding to the given measured fact: where m and n represent the maximum upper limit values of the index numbers of different types of indicators; Using the obtained similarity α ij Compare with the given corresponding threshold γ ij If α ij ≥γ ij , then record If α ij <γ ij , then record Denote the optimal similarity of the interval; Denote the similarity vector after comparison as Let the weighting coefficient be w i =(w i1 , w i2 , …, w im ), where i = 1, 2, …, n is the OWA operator based on the combination number, and calculate the comprehensive similarity π i , then Among them, (ε i1 , ε i2 , …, ε im ) is the vector obtained by sorting the sub-vectors of the similarity vector in descending order; According to the given threshold τ i , the following conclusion can be drawn: If π i ≥ τ i , then activate this rule; If π i < τ i , then this rule is not activated; Step 2: When only the i-th rule is activated, calculate the output conclusion: When p1 rules are all activated, the conclusions deduced from these activated rules are combined with appropriate weighting coefficients to obtain the final conclusion; Assume that the parameter β is involved in the weights, and use the comprehensive similarity π i to determine the weight vector, where the parameter β is Let s denote the number of rules with the maximum comprehensive similarity value among the activated rules; Let \(I = \{i|\) the \(i\) - th rule is activated and \(\pi i =\beta, 1\leq i\leq n\}\), assign weight to \(B i Then the conclusion of the rule with the maximum value of the comprehensive similarity is ​ Assign weights to the conclusions derived from the remaining activated rules Then, denote the conclusions of the remaining rules as Step 3: Take the union of the conclusions obtained in Step 2 to get the actual output

Images