TWI395963B - State-of-health estimation method for lead-acid batteries - Google Patents

Description

The present invention relates to a battery life estimation method, and more particularly to an estimation method for a lead acid battery life state using a modified Extension Neural Network (MENN).

Battery is undoubtedly a very popular electric energy storage tool, so the development of battery application technology has become an important issue in various industrial developments.

A lead-acid battery is a device that stores electrical energy in a chemical energy manner, and has the characteristics of being portable, resistant to vibration, impact, low internal resistance, simple structure, good cycle life and mature manufacturing technology, and is recyclable. Use the DC power supply. Although its energy density and life are not optimal, but the price is the lowest and its own electromotive force, operating temperature is wide and the material is stable, the application of lead-acid batteries is quite extensive.

For the user, the most desirable time to use the battery is the number of times the battery can be used (ie, its life state), so as to avoid the situation of insufficient power storage.

Therefore, the estimation of the state of charge and life of the battery is a very important part. There are many factors affecting the state of charge and life of lead-acid batteries, including the structural composition of the battery itself, the ambient temperature, the previous discharge depth, the discharge current, the method of charging the battery, and the self-discharge of the battery itself. Especially when the lead-acid battery is in the process of discharging, its battery state of charge is not a linear function, which makes the estimation of the state of charge of the battery deeper.

The battery life state is judged based on the battery available capacity estimation. By measuring the available capacity of the lead-acid battery after charging, the remaining life of the battery is known. In many of the previous cases, the characteristics used to estimate the life of a lead-acid battery, in addition to the direct use of the Coulomb method, include the voltage at the end of the battery, the resistance within the battery, the operating temperature of the battery, or the alternating current. An estimation method of inputting a battery to measure the phase of the response voltage. However, these characteristics generally change with the aging of the battery, thus deepening the difficulty of estimation.

At present, in order to improve the accuracy of estimating the life state of lead-acid batteries, many related researches use intelligent algorithms to estimate the life state of lead-acid batteries, such as neural network, fuzzy (Fuzzy). And algorithms such as neural-fuzzy. And with a variety of characteristics representing the life status of lead-acid batteries, by discussing the data distribution relationship between various characteristics and various life states, the accuracy of identification can be improved.

However, due to the characteristics of the state of charge of the lead-acid battery, the numerical distribution range tends to be non-linear between the battery life cycle and the battery usable capacity. Therefore, if you want to estimate the life state of the battery by the change of the available capacity of the battery, in addition to the characteristics of the state of charge and life of the battery, it is necessary to apply an additional test method to improve the accuracy of the life state estimation. degree.

The object of the present invention is to provide an estimation method for the life state of a lead-acid battery for improving the accuracy of the conventional estimation method and the slow estimation speed.

According to the above object of the present invention, an estimation method for the life state of a lead-acid battery is proposed, and the life of the battery is estimated by analyzing various characteristics of the lead-acid battery, thereby prolonging the service life of the lead-acid battery, and It is possible to properly plan the work that the battery can perform and provide enough power to avoid uninterrupted system interruption.

According to an embodiment of the invention, a method for estimating the life state of a lead-acid battery is characterized in that the internal resistance of the lead-acid battery after entering the floating state is matched with the peak voltage of the discharge dip voltage of the lead-acid battery, and the two are added. The ratio of the characteristics is another feature, and an extension matter element model is established by using these features, and then the improved extension type neural network is used for training. After the training, the improved type neural network is used for the lead-acid battery. The life state is estimated to achieve an estimate of continuity.

The method for estimating the life state of the lead-acid battery proposed by the present invention adopts an improved extension type neural network, and solves the conventional extension method, which can only classify the estimation result into a fixed category and has poor accuracy. The shortcomings can be further improved in the accuracy of identification.

Therefore, the accuracy of estimating the life state of the lead-acid battery by the improved extension type neural network proposed by the present invention is better than that of using other conventional methods, and the correlation function is used due to the proposed method. By mapping the output values, the categories of the data to be measured are relatively reduced, so the number of memory required by the computer system is relatively reduced, which can speed up the estimation.

In order to make the reviewers easy to understand the technical features and effects of the present invention, the extension theory and the extension-like neural network are first described as follows:

(1) Evaluation method of extension theory

Extension theory is applied to practice and usually needs to be implemented by appropriate mathematical tools. The extension evaluation method is based on the "extension set" and "association function" as a tool for applying theory to practice. The main concept of the extension evaluation method is to divide the data accumulated by one thing through various experiments into several levels, and the data range of each level set is given by the database or expert opinions, and the data to be assessed is substituted into the data range of each level. Perform the correlation calculation. The evaluation result is compared with the degree of association with each level set. The greater the degree of relevance, the better the degree of compliance of the data to be assessed with the level set.

The assessment steps are detailed below:

(1) Determine the classic domain and the local domain

If there is a thing R, it can be divided into j levels, called R _0j , and there are i features C _i representing the thing, and the distribution range of features is called X, and the definition P means that all features in j levels are included. A set of distribution ranges, X _{P i} can be defined as all possible ranges of values for the feature, and a _{P i} and b _{P i} represent the maximum and minimum values of the feature in all possible ranges of values, respectively. The local area R _P that can define the thing is as shown in equation (1).

The set of values that divide the thing R into j levels in the section is called the classical domain, where N _{0 j} represents the feature range of each of the j hierarchical sets that are divided, where C _{i is} within the level Each group of features, X _{0 ji} is the distribution range of the i-th feature at the j- th level, and a _{0 ji} and b _{0 ji} respectively represent the maximum and minimum values of the feature within the level, so that the thing can be defined the classical field j th group R _{0 j,} it can be divided into a thing R j (2) a set of section _P of the formula R domain of classical field group R _{0 j.}

(2) Determining the object to be evaluated

For a matter to be evaluated, a set of eigenvalues belonging to the matter element R is called a matter R, and the object element can be expressed as

Where q represents the set of characteristic values, x _i is the magnitude of the characteristic C _i of q, that is, the specific data obtained by the detection of the object to be evaluated, so a thing R may have a plurality of sets of characteristic values q.

(3) Determine the weight coefficient

The importance degree of each group of features C _i for the thing R is called the weight coefficient w _i , and the weight coefficients of each group are represented by values between 0 and 1 according to their importance degrees, and the sum value thereof is 1, that is,

(4) Determine the degree of association between the data to be evaluated and each level

Calculate distance

The definition of the distance refers to the distance between a feature value x _i for the center point of the feature region R _P (or each group of classical domains R _{0 j} ) and the center point and the top region of the region R _P (or the classical domain R _0j ) The difference between the lower limit distances, so the eigenvalue x _i is defined as the distance between the regions R _P as

And the eigenvalue x _i is defined as the distance between each classical domain R _{0 j} as

b. Calculate the correlation function

After the calculation of the distance is completed, the calculation of the correlation function can be performed. The correlation function k _j (x _i ) actually describes the degree of attribution of each feature of the item to be evaluated q to each evaluation category j. The calculation method of the correlation function k _j ( x _i ) is listed in the equation (7). It includes the initial classification of the data to be tested, that is, if the distance of the classical domain calculated by equation (6) is greater than 0, it means that the data to be tested does not belong to the category of the charging state at all. In order to avoid the fact that the distance between the domain and the classical domain is the same, the value of the correlation function approaches infinity, which makes it impossible to resolve. Therefore, the distance between the classical domain and the classical domain is less than 0. When you define the calculation of the associated function.

(5) Calculating the degree of relevance

Finally, the calculated correlation function k _j ( x _i ) is accumulated according to the weight value of each of its categories multiplied by its characteristics, that is, the degree of association λ _{j of} each evaluation category. Finally, the category j to which the maximum value of each degree of association belongs is obtained by the equation (8) as the evaluation result.

(2) Improved extension theory

Although the above traditional extension theory can obtain discrete output estimation results based on continuous input data, if it is desired to make the output result close to a continuous function, it is necessary to increase the classical domain class of the extension matter element model. According to formula (7), if each category of the matter element model is added, the calculation of the extension correlation function needs to be performed once in the estimation, so that the calculation time and memory of the estimation system must increase. Therefore, in the present embodiment, an improved extension theory identification architecture is adopted, and the value of the output category is adjusted by the extension correlation function generated in the identification process, and then based on the identification architecture of the extension theory, A more accurate identification result is achieved with the same number of classical domains.

Please refer to Figure 1a, which is a schematic diagram showing the relationship between test data and estimated results using traditional extension theory. Please refer to Figure 1b, which is a schematic diagram showing the relationship between test data and estimated results using the improved extension theory.

Figures 1a and 1b consider the number of experiments in the battery life test as the input test data, and the available capacity measured by the experiment is regarded as the estimated result of the output. If four sets of classical domains are also used to depict the relationship between the test data and the experimental results, it is apparent from the figure that the error between the data and the actual values (Fig. 1b) depicted by the improved extension theory is more traditional. The extension theory (Fig. 1a) is much reduced.

The main difference between the improved extension identification architecture and the traditional extension evaluation method, in addition to adding the numerical range of the output identification results of each category in the matter-element model, is to calculate the extension correlation function of each category. Then, using the size of the correlation function, the final output identification result is calculated based on the defined range of output values to calculate a most likely clear value.

(3) Application of extension correlation function

If a thing R can be represented by a single feature C, and this thing only has a set of classical domains R ₀₁ and a region R _P , then according to the definition of the matter-element model described above, the matter element of the thing can be expressed as

R =( N , C , x ) (9)

Its classic domain is

R ₀₁ =( N ₀₁ , C , X ₀₁ =< a ₀₁ , b ₀₁ >) (10)

The section is

R _p =( P , C , X _p =< a _p , b _p >) (11)

Please refer to Figure 2, which is a diagram of the relationship between feature values and associated functions. If the formula is calculated by the extension correlation function of equation (7), and the characteristic value x of the matter element is calculated from 0~∞, the feature value x and the extension correlation function k of Fig. 2 can be drawn. The relationship between ( x ). It can be seen from the figure that the closer the feature value is to the center point of the classical domain R ₀₁ , the closer the correlation function value k( x ) can be obtained to 1, and if the feature value does not fall within the range of the classical domain, it is calculated. The associated function value will be negative. Therefore, in the architecture of the extension theory, the magnitude of the correlation function k( x ) can be used to determine the extent to which the eigenvalue belongs to a certain interval.

(4) Identification structure of improved extension theory

The main difference between the identification architecture of the improved extension theory and the traditional extension evaluation method is that the definition of the output range of each category is increased, and the output value is calculated by using the size of the correlation function. The definition of the output range for each category can be viewed as adding a range of data to the matter-element model to represent all possible output results within that category. Therefore, the formula (2) can be expanded into

The set of Y represents the range of possible output values of the Y-axis in each set of feature data, so it can also be regarded as the classic domain of the test data output category. In addition, the test object and the representation of the section are the same as the equations (1) and (3). After the test data is calculated by the correlation function of equation (7), the relationship between each category j in the test data and the correlation function k _j ( x _i ) can be obtained. At this time, the maximum correlation degree is calculated, and the category j belongs to the evaluation result of the traditional extension theory. However, if the estimated result of the output can reach the numerical output of continuity, if the feature value is closer to the center point of the classical domain, the value of the correlation function will be closer to 1. Therefore, a more explicit output value can be calculated from the correlation function k _j ( x _i ) of each class by using the center point of the classical domain Y _ji of each output in the equation (12) and the range of the classical domain Y _ji . The calculation method firstly subtracts the correlation function k _j ( x _i ) value of each category from the maximum value 1 that the correlation function may generate. The function value obtained at this time can be regarded as the correlation function of the output layer.

Output class and output layer correlation function after calculation by equation (13) The relationship between the two can be represented in Figure 3, which is a relationship diagram between the output layer correlation function and the matter element class j . Output layer correlation function The minimum value of the category j belongs to the output calculation, and the output value range is calculated by the formula (14) to obtain a clear output data out _i . Output layer correlation function The relationship diagram depicted by the data range of the output category j is as shown in Fig. 4, which is a relationship diagram between the output layer correlation function and the output classic domain.

Where v = x _i -( a _ji + b _ji )/2 is the distance between the test data and the center point of the classical domain, and the judgment of the ± symbol in equation (14) can be linearized by the data of the test data and the available capacity. The regression method determines the positive or negative slope of the equation once. And if a thing has the characteristics of n groups, the estimated results obtained by the respective sets of features after being calculated by the formula (14) can be multiplied by the weight coefficient of the formula (4) and then summed up. (15) Identification results.

(5) Extension-like neural theory

Using the advantages of the simple calculation of the extension theory, combined with the neural network to develop the extension neural network (ENN). By transforming the matter-element model into two layers of neurons, the input layer and the output layer, the method of supervised learning is used to train the weight value between the two layers of neurons, and by adjusting the upper and lower limits of the weight value, finally The category of the data is determined by the minimum value of the correlation value of the output layers belonging to each category.

The most important concept is to optimize the range of the classical domain by adjusting the weight of the neural-like weights. Based on this concept, the classical domain of the constructed matter-element model is adjusted by the neural network-based training architecture, and the final estimation result is the same as the known input data in the form of supervised learning. output value.

Please refer to Figure 5 for an architectural diagram of a supervised neurological training program. The identification architecture in the figure is a matter element R having a set of characteristic values in representing the x of the matter element model of equation (9), and a set of classical domains R ₀₁ . The classical domain range of the input is X ₁₁ =< a , b>, and the classical domain range of the output is Y ₁₁ =<c,d>. The relationship between the input value in and the output estimation result out is as shown in the upper left of Fig. 5. The part of the center of the figure is the relationship between the input data calculated by the input data in from 0 to the equation (7) and the correlation function k( in ). On the upper right side of the figure, the relationship between the output layer correlation function k ^out ( in ) and the data range Y _{11 of} the output category is calculated via equations (13) and (14). If the value of the input data in is exactly a or b, the output outout identified by equation (14) will be c and d.

In this embodiment, the training data is expressed as T={T ₁ , T ₂ , . . . , T _np }, where np is the total number of training materials, and the matter-element model of the training data can be defined as

Where out _np is the known output value of the training data in the supervised learning program, and The magnitude of the ith set of features C _i of the training data. Therefore, after inputting the np group training data to the identification structure of FIG. 5, the identification result out can be compared with the known out _np to obtain the identification error e. According to formula (16), two training data T ₁ and T _{2 are} defined. The identification results of the two pieces of data before training are out, and the error e of the output out _np of the two training materials is 1 and 0. The learning procedure of the supervised neurological training architecture constructed in this embodiment will be described below, wherein the training data T ₁ and T ₂ are respectively defined as

Where T ₂ input data Therefore, the result of the identification architecture estimated in Fig. 5 is out=d, and its error e=out-out _np =0 is obtained by comparing it with the known output result out _np =d of T ₂ . And if the input data of T ₁ is Estimating by the identification architecture of Figure 5, the estimation result will get out=c, which is compared with the known output result out _np = c+1 of T _{1 to} obtain the error e=1. The error e is multiplied by a learning rate η and then calculated by the adjustment of the equation (18) to adjust the size of the classical domain output range Y _{11 to} change the estimation result of the output.

Please refer to Figure 6, which is a schematic diagram of the adjustment of the classic domain of the output during training. In Fig. 6, the identification result obtained by the training data T _{1 is} due to out <out _np and the error e=1 is generated. At this time, the lower limit of the classical domain of the output is adjusted by the equation (18), and finally the training framework calculates a new one. The classical domain lower limit value c _{ij , new} = c+1 at the output end, so the identification result obtained by the last training data T ₁ is out=c+1=out _np , so that the output error e=0, indicating that the output classic domain is adjusted through the output. The output has been changed to the known output value of the training data.

Finally, the training data is repeatedly input for the training program until the output recognition result out of all the training data meets the target value out _np . At this time, the classical domain range of the output calculated by the equation (18) is no longer changed, indicating that the identification is performed at this time. The results have been optimized. Figure 7 is a graph showing the relationship between the input data in and the estimated output out of the training program before and after completion. Corresponding to the input value of the training data T ₁ in FIG. 7, the recognition result which has been changed after the original training procedure of c to c + 1, the identification error e is also decreased from 1 to 0, the adjustment of the training program represents the passage, The classical domain range Y _{11 of the} output has been optimized with the corresponding training data.

Please refer to FIG. 8 , which is a flow chart of a method for estimating the life state of a lead-acid battery according to a preferred embodiment of the present invention. The method for estimating the life state of a lead-acid battery comprises the following steps: firstly establishing a matter-element model of the life state of the lead-acid battery through the improved extension theory, as shown in step 210. An improved extension-like neural network architecture is established, and an improved extension-type neural network training program is performed, and the matter-element model is trained as an input sample of the improved extension-type neural network, as shown in step 220. An improved extension-type neural network architecture that has been trained is used to estimate the lifetime of the lead-acid battery, as shown in step 230.

In a preferred embodiment of the present invention, a method for estimating the life state of a lead-acid battery is further described as follows:

(1) Establishing a matter-element model of the life state of the lead-acid battery (step 210)

The peak voltage of the discharge dip voltage in the life cycle of the lead-acid battery (hereinafter referred to as the discharge dip voltage), the internal resistance of the battery in the floating state of the battery (hereinafter referred to as the internal resistance of the battery), and the instantaneous current of the discharge of the lead-acid battery are obtained through experiments. That is, the ratio of the discharge dip voltage to the internal resistance of the battery) is used as an input feature of the estimation method.

In this embodiment, the matter element of the life state of the lead-acid battery of 12V-13Ah is expressed as shown in the formula (19).

Where R is the matter element, N is the name of the thing, C is the feature vector of the thing, and x is the feature vector value. In the present embodiment, the name of the thing (N) is the life of the battery (the percentage of the usable capacity (%)), and the characteristic vector (C) is [C ₁ : the peak voltage of the discharge dip voltage (V _plateau ), C ₂ : The internal resistance (R _{i n} ) of the battery, the C ₃ : discharge current (V _plateau /R _{i n} ) of the lead-acid battery, and the eigenvector value ( x ) are [12.74V, 40.38mΩ, 1.02kA].

In the present embodiment, in order to make the estimation result easy to calculate and compare, the uniform use percentage % indicates the available capacity of the battery. That is to say, the rated capacity (10Ah) indicated by the battery manufacturer is regarded as 100% of the usable capacity of the battery. If the total discharge amount of the battery is reduced to one-half of the original rated capacity (5Ah) when the battery ages, the battery is available. The capacity is considered to be 50%, and the life status of the lead-acid battery is expressed as a percentage of the available capacity.

Please refer to Figure 9 for 14 discharge dip voltage measurements during the entire battery life and peak discharge voltage and battery usable capacity of the battery dip voltage test using the Coulomb method. Diagram of the relationship.

In the range where the available capacity of the battery is reduced from 100% to 90%, the distribution relationship between the available capacity and the peak voltage is quite linear, so the peak voltage range (12.69V~12.74V) in this interval is set according to equation (12). The upper limit of the classic domain X _{11 at} the input of category R ₀₁ , and the corresponding available capacity range (90%~100%) are set to the upper and lower limits of the classic domain Y ₁₁ of the output of the category R ₀₁ , so that better identification can be obtained. result.

Therefore, in Figure 9, according to the linearity of the relationship between the available capacity of the battery and the peak voltage distribution, the classical domains of the four categories are respectively established, wherein the peak voltage of the battery available capacity is 110%, but the available capacity is 90%. If the time is low, if it is classified into the category R ₀₁ according to the available capacity of 110%, the classical domain range of the input end of the categories R ₀₁ and R ₀₂ will overlap, resulting in an estimation error, so according to the available capacity 110 It is more appropriate to classify the peak voltage as % to class R ₀₂ . According to formula (12), the above three sets of input features can be constructed according to the linearity of their numerical distribution to establish an improved extension matter element model.

The domain R _{P of} each feature, as well as the input classic domain X _ji and the output classic domain Y _ji are listed in Table 1, respectively.

(2) Supervised learning training program for improved extension type neural network (step 220)

After the matter-element model of Table 1 is established according to equation (12), it can be trained in a modified neural network architecture to adjust the range of the classical domain of the output to be optimized.

According to the definition of matter-element theory, the training data established by the characteristic data of 14 lead-acid batteries taken in various life states can be expressed as formula (16). Detailed data of each of its materials is listed in Table 2.

Then, the 14 test data are sequentially input into the training program mentioned in the above extensional neural theory, and the classic domain Y _{ji of the} three characteristics of the output is trained. The overall structure of the training program is shown in Figure 10, which is a training architecture diagram of the lead-acid battery life state matter element model. The three sets of feature data of the training data respectively calculate the correlation function k _j ( x _i ) of the respective i-th class by equation (7), and then obtain the output layer correlation function by the calculation of equations (13) and (14). And then know the value of the output category out _i .

Therefore, by comparing the characteristic output values out _i with the known output category values out _{np of the} training data, the classical domain range of the output can be adjusted by the error e in between, and then the features can be changed by the equation (18). Estimated result out _i . After the 14 training materials are sequentially input to the training program of FIG. 10, it means that the training is completed. However, the actual training procedure for the matter-element model is to repeatedly input the training data into the training structure, and whether the estimated error e of each feature is less than the set target value, or whether the training times exceed the set value. The maximum number of trainings to determine if the training program is stopped.

In equation (18), the adjustment of the classical domain of the output of the matter-element model, in addition to the estimation error e to determine the increase or decrease of the classical domain, the magnitude of the learning rate η also affects the convergence speed of the output error and the estimation Accuracy, higher learning rate can make the output error converge faster, but the accuracy of the estimation may be reduced. On the contrary, the smaller learning rate, although the error convergence is more accurate, the relative number of learning and time will increase a lot.

Therefore, the setting of the learning rate is another important key in the training process. In the present embodiment, the learning rate η of each feature is divided into a learning rate of 500 types from 0 to 5 and 0.01 per unit, and the minimum estimation of each training rate for 14 training materials is recorded separately. Absolute average error under error.

In this embodiment, the absolute average error me used is defined as

And the characteristics (C ₁ : peak voltage, C ₂ : internal resistance and C ₃ : instantaneous current), the minimum absolute average error before and after training and its optimal learning rate, error convergence speed (learning error convergence times) ) are listed in Table 3, respectively.

The matter-element model established after the completion of the training is shown in Table 4. Compared with the matter-element model established by Table 1 based on human experience, the classical domain range of the output is adjusted after appropriate adjustment. The accuracy has been greatly improved.

After completing the training program of the matter-element model, the present embodiment selects one of the four sets of weight values according to the formula (21) according to the test result of the training data, and sets the four sets of weight values according to the formula (15). According to the distribution of the characteristic values C ₁ (peak voltage) and C ₂ (internal resistance) of the test data, one of the four preferred weight values is selected, and the selection range of the four sets of preferred weight values is as shown in FIG. 11 . A combination of preferred weight values selected for peak voltage and internal resistance in different distribution regions.

Take the characteristic data of the battery available capacity of 110% as an example. At this time, the peak voltage of the battery available capacity is 110.7% when viewed from Figure 9, and the internal resistance and available capacity of the lead-acid battery are shown in Figure 12. In the measured relationship diagram, the internal resistance of the battery is 13.29mΩ when the available capacity of the battery is 110%. In the interval between 90% and 110% of the available capacity of the battery, the internal resistance of the battery is more related to the distribution of the peak voltage and the available capacity. Linear, so according to the available capacity estimated by the peak voltage and the internal resistance, the identification error obtained by the internal resistance is necessarily small. At this time, if the weight value w _{2 of the} feature C _{2 is} given, the weight value group is selected. For W ₁ , the accuracy obtained by the formula (15) is necessarily higher.

Therefore, when the peak voltage of the input estimation program is greater than 12.67V, it can be determined whether the internal resistance of the battery is less than 20mΩ to determine the weight value group W ₁ that gives a larger specific gravity to the C ₂ internal resistance (ie, w ₁ =0.1, w ₂ = 0.8, w ₃ = 0.1), or a weight value group W ₂ giving a larger specific gravity of the C ₁ peak voltage (i.e., w ₁ = 0.8, w ₂ = 0.1, w ₃ = 0.1). The estimation program can determine the combination of weight values selected according to the characteristic data of the test data, so that the three characteristics used in the estimation program can be matched with each other according to the linearity of the estimation results, thereby improving the overall Estimate accuracy.

Finally, according to the formula (15), the estimation result of each feature is multiplied by the weight value category to which it belongs, and the estimated value of the three sets of features can be combined to obtain the identification result out. Please refer to Figure 13 for the relationship between the estimated life of lead-acid batteries before and after training. In this embodiment, the relationship between the number of experiments performed by the 14 training materials according to the life cycle of the battery and its available capacity is depicted as shown in Fig. 13, and the estimated results of the data before and after the training are listed in the figure. In order to compare.

In Figure 13, the results of the unconstructed matter-element model (as shown in Table 1) are estimated. The absolute average error between the test and the test data is 8.43%, and the matter-element model after training is completed (Table 4). As a result of the estimation, the absolute average error is reduced to 2.15%.

Therefore, it is proved that the improved extension type neural network training architecture proposed in a preferred embodiment of the present invention can optimize the matter element model of the training data and improve the accuracy of the identification.

It is indeed a linear relationship between the discharge dip voltage of the lead-acid battery and the available capacity of the battery. Therefore, the life state of the lead-acid battery can be estimated by the improved extension-type neural network proposed in this embodiment, and the estimated result is compared with the traditional extension theory, and the number of the same classical domain can be Further improve the accuracy of lead acid battery life state estimation.

Moreover, combined with the improved extension theory and the neural network, the supervised learning architecture only needs the classical domain of each class of matter element model as the construction data, so that the identification result of the output can be optimized according to the training data. If you have more learning materials, you will be able to construct a more adaptive matter-element model, making the estimation system more robust.

(3) Estimation of the life status of the lead-acid battery, as shown in step 230.

In order to verify the identification accuracy and robustness of the life expectancy estimation method of the lead-acid battery mentioned in this embodiment, the characteristic values of the 14 training materials are respectively added according to the range of the section in the matter-element model, and 0% and 5 are added. Each of the 350 test data was obtained with % and 10% disturbance. The amount of disturbance of the test data is defined as

Where " rand " represents a random function created by a random number generator between -1 and 1. The test data of each of the three disturbance levels are estimated according to the matter-element model of Table 4 after training, and the available capacity of the battery is estimated by the formula (14), and the estimated results and the known materials are available. The capacities were compared and their average absolute errors were listed in Table 5 for comparison according to equation (20).

It is known from Table 5 that when the disturbance of the feature data is increased to ±10%, the estimated average error is about 8.01%, which is similar to the accuracy estimated by the untrained matter element model. Moreover, the number of trainings can be greatly reduced by setting an appropriate learning rate, and the identification architecture of the embodiment can optimize the range of the matter-element model after up to 4 learning procedures, and has about ±10 for the input feature data. % anti-noise ability.

Therefore, the accuracy of estimating the life state of the lead-acid battery by the improved extension type neural network in the present embodiment is better than that of other methods, and the correlation function is used to map the output value. Therefore, the classification of test data is relatively reduced, and the number of memory required for the identification system is relatively reduced, which can speed up the estimation. Compared with the extensional neural identification method used in the past, the results of the same test data are recognized by Table 5. Since the extensional neural network cannot make the corresponding input data reach the continuous output, it is the same. Under the classical domain number, the identification accuracy of the traditional method is relatively decreased, which also highlights the value of the present invention.

210~230. . . step

The above and other objects, features, advantages and embodiments of the present invention will become more apparent and understood.

Figure 1a shows a schematic diagram of the relationship between test data and estimated results using traditional extension theory.

Figure 1b is a schematic diagram showing the relationship between the test data and the estimated results using the improved extension theory.

Figure 2 is a diagram showing the relationship between feature values and associated functions.

Figure 3 is a diagram showing the relationship between the output layer correlation function and the matter element class j .

Figure 4 shows the relationship between the output layer correlation function and the classic domain of the output.

Figure 5 is a diagram showing the architecture of a supervised neurological training program.

Figure 6 is a schematic diagram showing the adjustment of the classical domain of the output during training.

Figure 7 is a graph showing the relationship between the input data in and the estimated output out of before and after the completion of the training program.

FIG. 8 is a flow chart showing a method for estimating the life state of a lead-acid battery according to a preferred embodiment of the present invention.

Figure 9 is a graph showing the peak voltage of the discharge dip voltage obtained by the battery usable capacity test and the available capacity of the battery.

Figure 10 is a diagram showing the training architecture of the lead-acid battery life state matter element model.

Figure 11 is a diagram showing the combination of preferred weight values selected for peak voltage and internal resistance in different distribution regions.

Figure 12 is a graph showing the measured relationship between the internal resistance and the available capacity of a lead-acid battery.

Figure 13 is a graph showing the relationship between the estimated life of lead-acid batteries before and after training.

210~230. . . step

Claims (5)

A method for estimating the life state of a lead-acid battery comprises: (a) obtaining a peak voltage of a discharge dip voltage during a life cycle of a lead-acid battery, a battery internal resistance of a battery floating state, and an instantaneous current of a lead-acid battery discharge, as Input characteristics; (b) establishing an extension matter element model of each life state of the lead-acid battery according to the input characteristics; (c) training the extension element model with an extension type neural network to adjust The classical domain of the output of the extension element model; and (d) estimating the lifetime state of the lead-acid battery using the trained extension-type neural network; wherein the lead-acid battery is discharged according to the step (a) The instantaneous current is the ratio of the discharge dip voltage to the internal resistance of the battery; wherein the modified extension element model described in the step (b) is expressed as follows: Where R is the matter element of the thing, N is the name of the thing, C is the feature vector of the thing, and x is the feature vector value; wherein the name N of the thing is the percentage of the available capacity of the lead-acid battery, and the feature vector C of the thing C ₁ is the peak voltage of the discharge dip voltage (V _plateau ), C ₂ is the battery internal resistance (R _in ) of the battery floating state, and C _{3 is} the discharge current of the lead acid battery (V _plateau /R _in And the characteristic vector value x has x ₁ of 12.74V, x ₂ of 40.38mΩ, and x ₃ of 1.02kA. The method for estimating the life state of a lead-acid battery according to claim 1, wherein the improved extension element model includes a domain R _P and a j-th classical domain R _{0 j} , and the region R _P represents to make Where P denotes a set containing all feature distribution ranges within j ranks, C _i represents a feature having i such things, and X _{P i} is all possible range of values of the feature, and a _{P i} and b _{P i} respectively Representing the maximum and minimum values of the feature in all possible ranges of values; and the j-th classical domain R _{0j is} expressed as Where R _0j is the j-th classical domain of the thing R, and N _j represents the feature range of each of the j hierarchical sets divided, wherein C _{i is} the set of features within the rank, and X _ji is the j- th the X-axis range of the i-th feature hierarchy pen, a _ji and b _ji respectively represent the maximum and minimum of the feature within this class, i.e., an input terminal classical field, Y _ji is the j in the i th hierarchy The Y-axis distribution range of the pen feature, c _ji and d _ji respectively represent the maximum and minimum values of the feature within this level, that is, the classical domain of the test data output category. The method for estimating the life state of a lead-acid battery according to claim 1, wherein the step (c) comprises the step of inputting training data to the modified extension-type neural network, the training data can be expressed as T= {T ₁ , T ₂ ,...,T _np }; where np is the total number of training materials, and the matter-element model of the training data is defined as Where out _np is the known output value of the training data in the supervised learning program, and The magnitude of the ith set of features C _i of the training data. For example, the method for estimating the life state of a lead-acid battery according to item 3 of the patent application scope, wherein the step of inputting the training data to the modified extension type neural network is to change the learning rate η of each feature from 0 to 5, respectively. The learning rate is divided into 500 units for each unit of 0.01, and the absolute average error of each learning rate for the minimum estimated error of the training data is recorded. The absolute average error me is defined as Where out _i is the output value of each feature, out _np is the known output category value of the training data, and n is the ratio of the training data. For example, the method for estimating the life state of a lead-acid battery according to claim 1 of the patent application, wherein the weight of the modified extension type neural network of the training completed in the step (d) is W ₁ = < w ₁ = 0.1, w ₂ =0.8, w ₃ =0.1〉, W ₂ =< w ₁ =0.8, w ₂ =0.1, w ₃ =0.1〉, W ₃ =< w ₁ =0.6, w ₂ =0.1, w ₃ = 0.3>, W ₄ = < w ₁ = 0.1, w ₂ = 0.1, w ₃ = 0.8 >; the above four sets of weight values are selected according to the distribution of the peak voltage characteristic value C ₁ and the internal resistance characteristic value C ₂ of the test data. One of the four weight values is selected, and the selection range of the four preferred weight values is the preferred weight value selected by the peak voltage and the internal resistance in different distribution regions.