Method for realizing fault diagnosis through sound wave detection

Abstract

The invention discloses a method for realizing fault diagnosis through sound wave detection, and the method specifically comprises the following steps: firstly, collecting sound through a microphone, directly inputting the sound into a sound decoding chip in an analog signal form, collecting a sound wave signal through the sound decoding chip, and transmitting the sound wave signal into a dual-core processor; then, specific sound wave signals are monitored in real time, an improved wavelet transform algorithm is used for effectively filtering and removing background noise, then feature value extraction is carried out on the denoised sound signals through a Meisson cepstrum MFCC algorithm, and an effective fault sound wave classifier is constructed according to feature values; and finally, an improved automatic vector machine pattern recognition algorithm is utilized to realize effective recognition of the sound wave characteristic signals. The safety valve has the advantages that specific sound signals generated when the safety valve is opened are monitored in real time and dynamically recognized, and early warning of thermal runaway is effectively achieved.

Description

Technical Field

[0001] The present invention relates to the technical field related to energy storage safety, and in particular to a method for realizing fault diagnosis by acoustic wave detection. Background Art

[0002] As the proportion of renewable energy in the energy structure increases, energy storage technology will be responsible for ensuring the safe and efficient operation of the power grid in the future, and the potential market is huge. At present, there are many application scenarios for energy storage technology, covering various fields such as new energy power generation, grid side, microgrid, etc., and it is still expanding. The scale of energy storage power stations is constantly increasing under the impetus of market demand, and the highest has increased to hundreds of megawatts. Electrochemical energy storage has a long cycle life and high energy density, so it can well meet the needs of modern energy storage, and occupies a large proportion in energy storage projects. Lithium-ion battery technology has become the mainstream electrochemical energy storage technology due to its advantages in cost and energy density.

[0003] Lithium battery energy storage systems are usually container-type, consisting of a number of lithium-ion battery clusters and electrical equipment placed in a confined space. Because the entire system is relatively closed and heat dissipation is poor, lithium batteries are prone to heat accumulation during operation, especially under extreme working conditions (overcharging, short circuit, etc.). The large amount of heat accumulated inside the battery will cause thermal runaway of the battery, thereby causing fire and explosion accidents. Existing safety protection facilities cannot completely prevent the occurrence of fire accidents in lithium battery energy storage systems. In July 2018, a lithium battery energy storage in South Korea caught fire due to aging equipment, resulting in the destruction of more than 3,000 lithium batteries and corresponding economic losses of millions of dollars.

[0004] Energy storage safety issues frequently occur. In addition to the activity problem of the battery itself, it also reflects the insufficiency of the intelligent operation and maintenance level, fault warning, state perception and other technical means of the existing energy storage system. It is imperative to implement energy storage safety assurance. Domestic and foreign scholars have studied the problem of early identification of thermal runaway from different angles and proposed the following methods: ① A method of collecting information such as battery body temperature, surface pressure, voltage and discharge rate through the battery management system (BMS) and judging thermal runaway based on this; ② A thermal runaway warning method based on the prediction of the battery internal temperature based on the battery thermal model or the internal chemical impedance of the battery; ③ A method of early warning of thermal runaway based on the monitoring of the corresponding gas composition in the battery compartment when the battery produces characteristic gas during thermal runaway. However, due to the battery quality problem and the lack of rationality of system integration, the single battery, especially the retired lithium-ion battery, has strong inconsistency and many potential fire risks. Most of these warning methods are in the research and exploration stage, and are complex to apply and have harsh conditions. It is difficult to meet the safety requirements of energy storage systems. There is an urgent need for a warning method with a wide range of applications and high sensitivity. Summary of the invention

[0005] The present invention aims to overcome the above-mentioned deficiencies in the prior art and provides a method for realizing fault diagnosis by acoustic wave detection which can detect and identify dynamically in real time.

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

[0007] A method for realizing fault diagnosis by acoustic wave detection specifically comprises the following steps:

[0008] (1) Acquisition of sound wave signals: The microphone collects sound and directly inputs it into the sound decoding chip in the form of analog signals. The clock signal of the sound decoding chip comes from the dual-core processor, and the external sound wave signals are synchronously collected in a synchronous signal collection manner. After the sound decoding chip collects the sound wave signals, it transmits them to the dual-core processor in the form of digital signals through the IIS channel;

[0009] (2) Acoustic signal preprocessing: Based on the quadratic threshold determination method of unbiased likelihood function estimation, the threshold is adaptively determined and the wavelet coefficients are dynamically calibrated to effectively deal with various background noises in the prefabricated cabin of the energy storage station and the white noise caused by the microphone sensor;

[0010] (3) Extraction of characteristic parameters of acoustic wave signals: The pre-processed acoustic wave signals are subjected to characteristic parameter extraction using Mel-frequency cepstral coefficients to form a classifier capable of identifying the acoustic wave signals of the safety valves, thereby dynamically identifying the acoustic wave signals of the safety valves;

[0011] (4) Safety valve acoustic signal pattern recognition based on characteristic parameters: For the characteristic coefficient vector of the safety valve acoustic signal constructed by the Mel-frequency cepstral coefficient algorithm, it is necessary to optimize it by feature matching through the acoustic signal recognition classifier. According to the specific application environment of the prefabricated cabin of the energy storage station, the safety valve acoustic signal is dynamically recognized based on the improved support vector machine method to ensure that the acoustic signal with the smallest error is recognized.

[0012] The present invention proposes a method for real-time acoustic wave monitoring and dynamic parameter identification of lithium batteries in energy storage stations. The implementation principle of the method is: in the early stage of thermal runaway of lithium batteries, a large amount of flammable gas will break through the battery pack safety valve and generate a specific sound signal. By real-time monitoring of the specific sound wave signal, an improved wavelet transform algorithm is used to effectively filter and remove the background noise, and the denoised sound signal is extracted by the Mersenne Cepstrum MFCC algorithm. An effective fault sound wave classifier is constructed based on the eigenvalues, and finally an improved automatic vector machine pattern recognition algorithm is used to effectively identify the sound wave characteristic signal. Since the thermal runaway of the battery is not severe at the moment the safety valve is opened and it is still in the early stage of thermal runaway, real-time monitoring and dynamic identification of the specific sound signal generated when the safety valve is opened will be of great significance to the effective implementation of early warning of thermal runaway.

[0013] Preferably, in step (2), the specific operation method of the acoustic wave signal preprocessing is as follows:

[0014] (21) Set the mother wavelet function and determine the scaling and translation factors:

[0015] Set wavelet transform L 2 (R) represents a signal with finite energy, whose Fourier transform is Ψ(w); when Requirements:

[0016]

[0017] say is the mother wavelet, Ψ(ω) is the mother wavelet spectrum function; A wavelet sequence can be obtained through appropriate linear transformation:

[0018]

[0019] In the formula, a and b are the scaling and translation factors of the mother wavelet respectively;

[0020] (22) Discretize the above formula to obtain the discrete wavelet function:

[0021]

[0022] Where a = 2 j ,b=2 j k (j, k is a positive integer);

[0023] (23) For any input sound wave signal f(t)∈L 2 The discrete wavelet transform of (R) is:

[0024]

[0025] (24) The discrete wavelet is decomposed into four layers according to the approximation value and the noise value, as shown in the following formula:

[0026] w f (j,k)=C A4 +C D4 +C D3 +C D2 +C D1

[0027] Where C Ai is the approximate part of wavelet decomposition, C Di It indicates the part where the noise is located;

[0028] (25) Determine the threshold value based on sample estimation: The selection and setting of the threshold is the key in the wavelet denoising process. The quadratic threshold value method based on unbiased likelihood function estimation removes the wavelet coefficients that do not meet the set threshold requirements and retains the remaining coefficients, thereby achieving a good denoising effect.

[0029] Preferably, the quadratic threshold value determination method based on unbiased likelihood function estimation refers to determining the distribution of the threshold using an unbiased likelihood estimation function, and then determining the specific value of the threshold using the quadratic threshold value determination method. The specific operation steps are:

[0030] (a) Set the coefficient threshold of the wavelet transform to satisfy the unbiased likelihood estimation function:

[0031]

[0032] Where α and β are pre-constraints that satisfy: α + β = 1, and ensure that the value of T is the minimum value; ξ is the standard deviation of the noise, n is the number of wavelet coefficients at each scale, and w is the b is the square value corresponding to a specific wavelet coefficient;

[0033] (b) Perform threshold determination on the wavelet coefficients and determine the final wavelet coefficients through secondary threshold determination.

[0034] Preferably, in step (b), the threshold determination process is as follows:

[0035] (b1) First processing: keep the wavelet coefficients greater than the set threshold T to construct a new coefficient set S1, and return the rest to zero, as shown in the following formula:

[0036]

[0037] In the formula, S1 is the wavelet coefficient set after judgment, x is the wavelet coefficient, and T is the threshold;

[0038] (b2) Second processing: Since the wavelet coefficients after the first processing still have large errors and contain large noise influences, the wavelet coefficients are processed for the second time, and the wavelet coefficients greater than the threshold are further judged in the direction of zero, and the rest are returned to zero:

[0039]

[0040] Where sig is the sign function, S2 is the wavelet coefficient set after secondary determination, s1 is the wavelet coefficient after primary threshold determination, and T is the threshold.

[0041] Preferably, in step (3), the specific process of extracting characteristic parameters of the collected sound wave signal using Mel-cepstral coefficients is as follows: first, the sound wave signal is framed, the overlapping parts between frames are retained, and the frames are multiplied by a Hamming window for windowing processing; after pre-processing, the corresponding spectrum parameters are determined by fast Fourier transform, and the corresponding spectrum energy distribution is obtained on this basis; then the spectrum is filtered through a triangular filter group to eliminate the influence of harmonics; the logarithmic energy of each filter group output is calculated, and the logarithmic energy is subjected to discrete cosine transform to obtain an MFCC characteristic coefficient vector with a dimension equal to the number of triangular filters.

[0042] Preferably, in step (4), the specific process of the improved support vector machine method is as follows:

[0043] (41) Preliminary setting of optimization function:

[0044] For the input m samples X={x1,x2,...,x i}, i=1,2,...,m to make a weighted optimization setting, that is:

[0045] f(X)=W T X+B

[0046] Where W = {ω1,ω2,...,ω i}, i=1,2,...,m is a vector, B={b1,b2,...,b i}, i = 1, 2, ..., m is a scalar; for the above weighted sample feature vector x i To divide, if f(x i )<0, the sample is divided into -1 category; when f(x i )>0, divided into +1 category; use a category label k i To express;

[0047]

[0048] So the feature vector x i Construct a new associated sample (xi ,k i );

[0049] (42) Determine the Euclidean distance from the sample to the hyperplane:

[0050]

[0051] There is a positive correlation between the distance of the sample from the hyperplane and the geometric interval. The larger the distance, the greater the possibility that the sample is classified into a certain category. For a specific training set, the role of training the support vector machine is to determine the optimal value of the weighted entropy ω and the scalar b so that the distance between the support vector and the hyperplane is the largest.

[0052] (43) Introduce relaxation factor to remove interference effects:

[0053] In order to reduce the interference of other sound waves in the energy storage battery prefabricated cabin on the specific sound waves of the safety valve, the relaxation factor ξ is introduced according to the objectives and constraints. i , in order to minimize the interference of bad data, after convex optimization, the problem is transformed and expressed as the following optimization problem:

[0054]

[0055] The constraints are:

[0056]

[0057] In the formula, the relaxation factor ξ i represents the edge and a data point x on the wrong side of the edge i distance;

[0058] (44) The Lagrange multiplier is introduced to obtain the optimal classification surface:

[0059] Then, the Lagrange multiplier a is introduced into the above model. i =(1,2...,m), we get

[0060]

[0061] Taking partial derivatives of the parameters ω and b, so that the above Lagrange multipliers return to zero, we can get the corresponding Lagrange multiplier a * and parameter b * , get the optimal function (optimal classification surface):

[0062]

[0063] (45)Introduce the kernel function to determine the classifier model:

[0064] Since most of the data involved in the actual application of energy storage prefabricated cabins are linearly inseparable, the kernel function ψ(x i ,x j )Optimization function calculation:

[0065]

[0066] In the formula, σ is the kernel width parameter, and is the minimum mean square error of the feature vector;

[0067] After the kernel function is introduced, the final model of the acoustic signal feature parameter recognition classifier is:

[0068]

[0069] The beneficial effect of the present invention is that the specific sound signal generated when the safety valve is opened is monitored in real time and dynamically identified, thereby effectively realizing early warning of thermal runaway. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of the method of the present invention;

[0071] Figure 2 It is a flowchart of acoustic signal preprocessing;

[0072] Figure 3 It is a flow chart of extracting characteristic parameters of acoustic wave signals. DETAILED DESCRIPTION

[0073] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0074] like Figure 1 In the embodiment described, a method for implementing fault diagnosis by acoustic wave detection specifically comprises the following steps:

[0075] (1) Acquisition of sound wave signals: The microphone collects sound and directly inputs it into the sound decoding chip in the form of analog signals. The clock signal of the sound decoding chip comes from the dual-core processor, and the external sound wave signals are synchronously collected in a synchronous signal collection manner. After the sound decoding chip collects the sound wave signals, it transmits them to the dual-core processor in the form of digital signals through the IIS channel;

[0076] Among them, the ARM+DSP dual-core processor chip is used. Based on the software and hardware combined collaborative technology, task scheduling, communication, data acquisition, etc. are completed by the ARM core, while the processing of computationally intensive operation tasks is delivered to the DSP core, and real-time data acquisition, processing and analysis are completed through dual-core collaboration. The sound wave signal acquisition module consists of a microphone and a UDM1341TS sound decoding chip. After the microphone collects the sound, it directly inputs it into the UDM1341TS sound decoding chip in the form of an analog signal. The clock signal of the UDM1341TS sound decoding chip comes from the dual-core processor, and the external sound signal is synchronously collected in a synchronous signal acquisition mode. The sampling frequency is 100KHZ and the sampling accuracy is 16-bit depth. After the UDM1341TS sound decoding chip collects the sound wave signal, it transmits it to the dual-core processor in the form of a digital signal through the IIS channel. After the dual-core processor performs preliminary filtering and denoising, it extracts and identifies the feature values.

[0077] (2) Acoustic signal preprocessing: Based on the quadratic threshold determination method of unbiased likelihood function estimation, the threshold is adaptively determined and the wavelet coefficients are dynamically calibrated to effectively deal with various background noises in the prefabricated cabin of the energy storage station and the white noise caused by the microphone sensor;

[0078] In addition to the specific sound waves of the battery pack safety valve opening, there are also many background noises such as air conditioning noise, BMS system operation noise, PCS noise, battery charging and discharging current reflux noise, maintenance personnel activity noise, and prefabricated cabin door opening noise in the prefabricated cabin of the energy storage station. These noises will be collected by the microphone together with the sound waves of the battery pack safety valve opening. In addition, the quality of the microphone itself or the analog signal line under the influence of the magnetic field must be considered. There is a certain degree of white noise interference in the collected sound wave signal. Therefore, the sound signal collected by the microphone needs to be preprocessed before entering the recognition. By filtering and background denoising, the noise interference is first removed, thereby improving the accuracy of the detection result. The preprocessing of the sound wave signal here adopts the wavelet denoising method. Since the traditional threshold-based wavelet denoising coefficient is affected by the threshold value and the error is large, the present invention proposes a secondary threshold value method based on unbiased likelihood function estimation, adaptively determining the threshold, and dynamically calibrating the wavelet coefficient. The wavelet transform after adopting this method can effectively deal with various background noises in the prefabricated cabin of the energy storage station and the white noise caused by the microphone sensor, with high denoising efficiency and good effect. The complete process of sound signal preprocessing based on wavelet transform, such as Figure 2 The specific operation method is as follows:

[0079] (21) Set the mother wavelet function and determine the scaling and translation factors:

[0080] Set wavelet transform L 2(R) represents a signal with finite energy, whose Fourier transform is Ψ(w); when Requirements:

[0081]

[0082] say is the mother wavelet, Ψ(ω) is the mother wavelet spectrum function; A wavelet sequence can be obtained through appropriate linear transformation:

[0083]

[0084] In the formula, a and b are the scaling and translation factors of the mother wavelet respectively;

[0085] (22) Discretize the above formula to obtain the discrete wavelet function:

[0086]

[0087] Where a = 2 j ,b=2 j k (j, k is a positive integer);

[0088] (23) For any input sound wave signal f(t)∈L 2 The discrete wavelet transform of (R) is:

[0089]

[0090] (24) The discrete wavelet is decomposed into four layers according to the approximation value and the noise value, as shown in the following formula:

[0091] w f (j,k)=C A4 +C D4 +C D3 +C D2 +C D1

[0092] In the formula, C Ai is the approximate part of wavelet decomposition, C Di It indicates the part where the noise is located;

[0093] (25) Determine the threshold value based on sample estimation: The selection and setting of the threshold is the key in the wavelet denoising process. By removing the wavelet coefficients that do not meet the set threshold requirements and retaining the remaining coefficients, a good denoising effect can be achieved. In view of the mixed sounds in the energy storage prefabricated cabin, the similarity in the spectral distribution of the background noise and the specific sound waves of the safety valve, if the existing threshold setting method is used directly, it will cause poor denoising effect. Therefore, the present invention proposes a quadratic threshold value determination method based on unbiased likelihood function estimation, that is, after determining the distribution of the threshold using the unbiased likelihood estimation function, the quadratic threshold value method is used to determine the specific value of the threshold. The specific operation steps are:

[0094] (a) Set the coefficient threshold of the wavelet transform to satisfy the unbiased likelihood estimation function:

[0095]

[0096] Where α and β are pre-constraints that satisfy: α + β = 1, and ensure that the value of T is the minimum value; ξ is the standard deviation of the noise, n is the number of wavelet coefficients at each scale, and w is the b is the square value corresponding to a specific wavelet coefficient;

[0097] (b) Threshold determination is performed on the wavelet coefficients, and the final wavelet coefficients are determined by secondary threshold determination; the secondary threshold determination process is as follows:

[0098] (b1) First processing: keep the wavelet coefficients greater than the set threshold T to construct a new coefficient set S1, and return the rest to zero, as shown in the following formula:

[0099]

[0100] In the formula, S1 is the wavelet coefficient set after judgment, x is the wavelet coefficient, and T is the threshold;

[0101] (b2) Second processing: Since the wavelet coefficients after the first processing still have large errors and contain large noise influences, the wavelet coefficients are processed for the second time, and the wavelet coefficients greater than the threshold are further judged in the direction of zero by the second threshold, and the rest are returned to zero:

[0102]

[0103] Where sig is the sign function, S2 is the wavelet coefficient set after secondary determination, s1 is the wavelet coefficient after primary threshold determination, and T is the threshold.

[0104] (3) Extraction of characteristic parameters of acoustic wave signals: The pre-processed acoustic wave signals are subjected to characteristic parameter extraction using Mel-frequency cepstral coefficients to form a classifier capable of identifying the acoustic wave signals of the safety valves, thereby dynamically identifying the acoustic wave signals of the safety valves;

[0105] The pre-processed acoustic signal needs to be extracted through feature parameters and form a classifier that can identify the acoustic signal of the safety valve in order to dynamically identify the acoustic signal of the safety valve. This method uses Mel-frequency cepstral coefficients (MFCC) to extract feature parameters of the collected acoustic signal. The MFCC algorithm is mainly used for the recognition of speech signals and has achieved good application results. The MFCC algorithm can simulate the human cochlear vocal tract model and realize the function of auditory masking. Therefore, it shows obvious applicability in situations where the human ear needs to distinguish sounds. Since the frequency range of the acoustic signal generated by the safety valve is within the hearing range of the human ear, it is very appropriate to use the MFCC algorithm for feature value extraction.

[0106]

[0107] like Figure 3 As shown in the figure, the specific process of extracting characteristic parameters of the collected sound wave signal using Mel-cepstral coefficients is as follows: first, the sound wave signal is framed, the overlapping parts are retained between frames, and the frames are multiplied by a Hamming window for windowing processing; after preprocessing, the corresponding spectrum parameters are determined by fast Fourier transform, and the corresponding spectrum energy distribution is obtained on this basis; then the spectrum is filtered through a triangular filter group to eliminate the influence of harmonics; the logarithmic energy of the output of each filter group is calculated, and the logarithmic energy is subjected to discrete cosine transform to obtain the MFCC characteristic coefficient vector with a dimension equal to the number of triangular filters.

[0108] (4) Safety valve acoustic signal pattern recognition based on characteristic parameters: For the characteristic coefficient vector of the safety valve acoustic signal constructed by the Mel-frequency cepstral coefficient algorithm, it is necessary to optimize it by means of feature matching through the acoustic signal recognition classifier. According to the specific application environment of the energy storage station prefabricated cabin, the safety valve acoustic signal is dynamically recognized based on the improved support vector machine method to ensure that the acoustic signal with the smallest error is recognized;

[0109] For the characteristic coefficient vector of the safety valve acoustic signal constructed by the MFCC algorithm, it is necessary to use the recognition classifier of the acoustic wave signal to search for the best in a feature matching manner to ensure that the acoustic signal with the smallest error is identified. The current pattern recognition method is basically used for the dynamic recognition of acoustic signals. The application effect of the pattern recognition method is closely related to the specific application scenario. There is currently no standardized recognition method for all pattern recognition problems. When dealing with different types of pattern recognition problems, it is necessary to compare the characteristics of the problem and determine a targeted processing method. The common pattern recognition methods for acoustic signals currently used are the K nearest neighbor method KNN, the probabilistic neural network method PNN and the support vector machine SVM method. Due to the large number of types of acoustic wave signals in the battery prefabricated cabin of the energy storage station, there is a relatively large interference in the feature recognition of the specific safety valve acoustic signal. If the common pattern recognition method is directly used, there is a problem of local convergence but global non-convergence. Therefore, the present invention proposes an improved support vector machine method based on the specific application environment of the prefabricated cabin of the energy storage station. The specific process is as follows:

[0110] (41) Preliminary setting of optimization function:

[0111] For the input m samples X={x1,x2,...,x i}, i=1,2,...,m to make a weighted optimization setting, that is:

[0112] f(X)=W T X+B

[0113] Where W = {ω1,ω2,...,ω i}, i=1,2,...,m is a vector, B={b1,b2,...,b i}, i = 1, 2, ..., m is a scalar; for the above weighted sample feature vector x i To divide, if f(x i )<0, the sample is divided into -1 category; when f(x i )>0, divided into +1 category; use a category label k i To express;

[0114]

[0115] So the feature vector x i Construct a new associated sample (x i ,k i );

[0116] (42) Determine the Euclidean distance from the sample to the hyperplane:

[0117]

[0118] There is a positive correlation between the distance of the sample from the hyperplane and the geometric interval. The larger the distance, the greater the possibility that the sample is classified into a certain category. For a specific training set, the role of training the support vector machine is to determine the optimal value of the weighted entropy ω and the scalar b so that the distance between the support vector and the hyperplane is the largest.

[0119] (43) Introduce relaxation factor to remove interference effects:

[0120] In order to reduce the interference of other sound waves in the energy storage battery prefabricated cabin on the specific sound waves of the safety valve, the relaxation factor ξ is introduced according to the objectives and constraints. i , in order to minimize the interference of bad data, after convex optimization, the problem is transformed and expressed as the following optimization problem:

[0121]

[0122] The constraints are:

[0123]

[0124] In the formula, the relaxation factor ξ i represents the edge and a data point x on the wrong side of the edge i distance;

[0125] (44) The Lagrange multiplier is introduced to obtain the optimal classification surface:

[0126] Then, the Lagrange multiplier a is introduced into the above model. i =(1,2...,m), we get

[0127]

[0128] Taking partial derivatives of the parameters ω and b, so that the above Lagrange multipliers return to zero, we can get the corresponding Lagrange multiplier a * and parameter b * , get the optimal function (optimal classification surface):

[0129]

[0130] (45)Introduce the kernel function to determine the classifier model:

[0131] Since most of the data involved in the actual application of energy storage prefabricated cabins are linearly inseparable, the kernel function ψ(x i ,x j )Optimization function calculation:

[0132]

[0133] In the formula, σ is the kernel width parameter, and is the minimum mean square error of the feature vector;

[0134] After the kernel function is introduced, the final model of the acoustic signal feature parameter recognition classifier is:

[0135]

[0136] In summary, the present invention proposes a method for real-time acoustic wave monitoring and dynamic parameter identification of lithium batteries in energy storage stations. The implementation principle of the method is: in the early stage of thermal runaway of lithium batteries, a large amount of flammable gas will break through the battery pack safety valve and generate a specific sound signal. By real-time monitoring of the specific sound wave signal, an improved wavelet transform algorithm is used to effectively filter and remove the background noise, and the denoised sound signal is extracted by the Mersenne Cepstrum MFCC algorithm. An effective fault sound wave classifier is constructed based on the eigenvalues, and finally an improved automatic vector machine pattern recognition algorithm is used to effectively identify the sound wave characteristic signal. Since the thermal runaway of the battery is not severe at the moment the safety valve is opened and it is still in the early stage of thermal runaway, real-time monitoring and dynamic identification of the specific sound signal generated when the safety valve is opened will be of great significance to the effective implementation of early warning of thermal runaway.

Claims

1. A method for fault diagnosis by acoustic wave detection, characterized in that: The specific steps include: (1) Acquisition of sound wave signals: The microphone collects sound and directly inputs it into the sound decoding chip in the form of analog signals. The clock signal of the sound decoding chip comes from the dual-core processor, and the external sound wave signals are synchronously collected in a synchronous signal collection manner. After the sound decoding chip collects the sound wave signals, it transmits them to the dual-core processor in the form of digital signals through the IIS channel; (2) Acoustic signal preprocessing: Based on the quadratic threshold determination method of unbiased likelihood function estimation, the threshold is adaptively determined and the wavelet coefficients are dynamically calibrated to effectively deal with various background noises in the prefabricated cabin of the energy storage station and the white noise caused by the microphone sensor; (3) Extraction of characteristic parameters of acoustic wave signals: The pre-processed acoustic wave signals are subjected to characteristic parameter extraction using Mel-frequency cepstral coefficients to form a classifier capable of identifying the acoustic wave signals of the safety valves, thereby dynamically identifying the acoustic wave signals of the safety valves; (4) Safety valve acoustic signal pattern recognition based on characteristic parameters: For the characteristic coefficient vector of the safety valve acoustic signal constructed by the Mel-frequency cepstral coefficient algorithm, it is necessary to optimize it by feature matching through the acoustic signal recognition classifier. According to the specific application environment of the prefabricated cabin of the energy storage station, the safety valve acoustic signal is dynamically recognized based on the improved support vector machine method to ensure that the acoustic signal with the smallest error is recognized.

2. The method for realizing fault diagnosis by acoustic wave detection according to claim 1, characterized in that: In step (2), the specific operation method of the acoustic wave signal preprocessing is as follows: (21) Set the mother wavelet function and determine the scaling and translation factors: Set wavelet transform L 2 (R) represents a signal with finite energy, whose Fourier transform is Ψ(w); when Requirements: say is the mother wavelet, Ψ(ω) is the mother wavelet spectrum function; A wavelet sequence can be obtained through appropriate linear transformation: In the formula, a and b are the scaling and translation factors of the mother wavelet respectively; (22) Discretize the above formula to obtain the discrete wavelet function: Where a = 2 j ,b=2 j k (j, k is a positive integer); (23) For any input sound wave signal f(t)∈L 2 The discrete wavelet transform of (R) is: (24) The discrete wavelet is decomposed into four layers according to the approximation value and the noise value, as shown in the following formula: w f (j,k)=C A4 +C D4 +C D3 +C D2 +C D1 Where C Ai is the approximate part of wavelet decomposition, C Di It indicates the part where the noise is located; (25) Determine the threshold value based on sample estimation: The selection and setting of the threshold is the key in the wavelet denoising process. The quadratic threshold value method based on unbiased likelihood function estimation removes the wavelet coefficients that do not meet the set threshold requirements and retains the remaining coefficients, thereby achieving a good denoising effect.

3. A method for fault diagnosis by acoustic wave detection according to claim 1 or 2, characterized in that: The quadratic threshold determination method based on unbiased likelihood function estimation refers to using the unbiased likelihood estimation function to determine the distribution of the threshold, and then using the quadratic threshold determination method to determine the specific value of the threshold. The specific steps are: (a) Set the coefficient threshold of the wavelet transform to satisfy the unbiased likelihood estimation function: In the formula, α and β are pre-constraints, satisfying: α + β = 1, and ensuring that the value of T is the minimum value; ξ is the standard deviation of the noise, n is the number of wavelet coefficients at each scale, and w b is the square value corresponding to a specific wavelet coefficient; (b) Perform threshold determination on the wavelet coefficients and determine the final wavelet coefficients through secondary threshold determination.

4. The method for realizing fault diagnosis by acoustic wave detection according to claim 3 is characterized in that: In step (b), the secondary threshold determination process is as follows: (b1) First processing: keep the wavelet coefficients greater than the set threshold T to construct a new coefficient set S1, and return the rest to zero, as shown in the following formula: In the formula, S1 is the wavelet coefficient set after judgment, x is the wavelet coefficient, and T is the threshold; (b2) Second processing: Since the wavelet coefficients after the first processing still have large errors and contain large noise influences, the wavelet coefficients are processed for the second time, and the wavelet coefficients greater than the threshold are further judged in the direction of zero by the second threshold, and the rest are returned to zero: Where sig is the sign function, S2 is the wavelet coefficient set after secondary determination, s1 is the wavelet coefficient after primary threshold determination, and T is the threshold.

5. The method for realizing fault diagnosis by acoustic wave detection according to claim 1, characterized in that: In step (3), the specific process of extracting characteristic parameters of the collected sound wave signal using Mel-frequency cepstrum coefficients is as follows: first, the sound wave signal is framed, the overlapping parts between frames are retained, and the signal is multiplied by a Hamming window for windowing; after pre-processing, the corresponding spectrum parameters are determined by fast Fourier transform, and the corresponding spectrum energy distribution is obtained on this basis; then, the spectrum is filtered by a triangular filter group to eliminate the influence of harmonics; the logarithmic energy of the output of each filter group is calculated, and the logarithmic energy is subjected to discrete cosine transform to obtain an MFCC characteristic coefficient vector with a dimension equal to the number of triangular filters.

6. The method for realizing fault diagnosis by acoustic wave detection according to claim 1, characterized in that: In step (4), the specific process of the improved support vector machine method is as follows: (41) Preliminary setting of optimization function: For the input m samples X={x1,x2,...,x i }, i=1,2,...,m to make a weighted optimization setting, that is: f(X)=W T X+B Where W = {ω1,ω2,...,ω i }, i=1,2,...,m is a vector, B={b1,b2,...,b i }, i=1,2,...,m is a scalar; For the above weighted sample feature vector x i To divide, if f(x i )<0 divides the sample into -1 category; When f(x i )>0, divided into +1 category; use a category label k i To express; So the feature vector x i Construct a new associated sample (x i ,k i ); (42) Determine the Euclidean distance from the sample to the hyperplane: There is a positive correlation between the distance of the sample from the hyperplane and the geometric interval. The larger the distance, the greater the possibility that the sample is classified into a certain category. For a specific training set, the purpose of training the support vector machine is to determine the optimal value of the weighted entropy ω and the scalar b so that the distance between the support vector and the hyperplane is maximized; (43) Introduce relaxation factor to remove interference effects: In order to reduce the interference of other sound waves in the energy storage battery prefabricated cabin on the specific sound waves of the safety valve, the relaxation factor ξ is introduced according to the objectives and constraints. i , in order to minimize the interference of bad data, after convex optimization, the problem is transformed and expressed as the following optimization problem: The constraints are: In the formula, the relaxation factor ξ i represents the edge and a data point x on the wrong side of the edge i distance; (44) The Lagrange multiplier is introduced to obtain the optimal classification surface: Then, the Lagrange multiplier a is introduced into the above model. i =(1,2...,m), we get Taking partial derivatives of the parameters ω and b, so that the above Lagrange multipliers return to zero, we can get the corresponding Lagrange multiplier a * and parameter b * , get the optimal function (optimal classification surface): (45)Introduce the kernel function to determine the classifier model: Since most of the data involved in the actual application of energy storage prefabricated cabins are linearly inseparable, the kernel function ψ(x i ,x j )Optimization function calculation: In the formula, σ is the kernel width parameter, and is the minimum mean square error of the feature vector; After the kernel function is introduced, the final model of the acoustic signal feature parameter recognition classifier is:

Images