Engine production line assembly quality control method and system based on large model

Abstract

The invention provides an engine production line assembly quality control method and system based on a large model, and the method comprises the steps: obtaining historical assembly data of all parts of an engine, and constructing an assembly technology detection sample library based on the historical assembly data; adopting a combined intelligent algorithm to improve the large model, and using the improved large model to carry out classified training on the assembly process detection sample library; and acquiring target assembly data of each part of the engine which is being assembled, processing the target assembly data by using the large model after classification training, determining the assembly state of each part of the engine, and performing assembly quality early warning on an abnormal state. The assembling quality of the engine production line is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent detection technology, and in particular relates to an engine production line assembly quality control method and system based on a large model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The engine is a complex system of multiple mechanisms and machines. Assembly is the final step in the engine manufacturing process. The assembly quality has a very important impact on the vehicle's fuel consumption, torque, power and other performance indicators. If problems are found in the assembly process, it will directly affect the engine's power output and may even seriously damage the working engine.

[0004] Therefore, improving assembly quality is the key to improving production efficiency. Accurately identifying assembly quality is crucial. Identifying assembly quality is generally done by comparing assembly data of perfectly assembled engines and providing timely warnings for unqualified equipment.

[0005] Traditional quality control in the assembly process of engine production lines mainly relies on manual experience judgment and random inspection mechanisms, as well as conventional big data methods such as probability statistics. These methods suffer from large detection blind spots and poor real-time performance. Assembly quality control methods based on large models (such as the Transformer model) can effectively monitor the multi-source heterogeneous data generated by the production line in real time and provide timely warnings for products with unqualified assembly quality. Summary of the Invention

[0006] In order to solve the above problems, the present invention proposes a large-scale model-based engine production line assembly quality control method and system. The present invention improves the large-scale model by constructing an assembly process detection sample library, and uses the improved large-scale model to perform classification training on the assembly process detection sample library, determine the normal or abnormal status of the assembly of each engine system, and issue an assembly quality warning, thereby greatly improving the assembly quality of the engine production line.

[0007] According to some embodiments, the present invention adopts the following technical solutions:

[0008] A large-scale model-based engine production line assembly quality control method comprises the following steps:

[0009] Obtain historical assembly data of each engine component and build an assembly process inspection sample library based on the historical assembly data;

[0010] A combined intelligent algorithm is used to improve the large model, and the improved large model is used to perform classification training on the assembly process inspection sample library;

[0011] Obtain target assembly data for each engine component being assembled, process the target assembly data using a large model trained through classification, determine the assembly status of each engine component, and issue assembly quality warnings for abnormal status;

[0012] Among them, the process of using combined intelligent algorithms to improve the large model includes using the whale migration optimization algorithm to improve the exploration stage and development stage of the exponential triangle optimization algorithm, using the Levy flight process to improve the judgment factors used to judge the exploration stage and development stage and jump out of local optimization to obtain the global optimal solution, iterate the optimization process until the iteration requirements are met, form a combined intelligent algorithm, and obtain the final optimal solution. The final optimal solution is used as the scaling factor and bias of the large model to achieve improvement of the large model.

[0013] As an optional implementation method, in the process of obtaining historical assembly data of various engine components, the intake pressure, early opening angle and late closing angle of the intake and exhaust valves, camshaft operating torque, camshaft axial clearance, cylinder head pair nut tightening torque, flywheel housing bolt tightening torque and flywheel housing bolt tightening angle of multiple engines are obtained, and the corresponding crankshaft torque histogram is drawn. The normal distribution is used for judgment. The data within the set range is normal value, and the data outside the set range is abnormal value.

[0014] As an optional implementation, in the process of using the improved large model to perform classification training on the assembly process detection sample library, the assembly process detection sample library is divided into a normal value library and an abnormal value library, wherein the assembly data within the normal value range is the normal value library, and the assembly data outside the normal value range is the abnormal value library.

[0015] As an optional implementation method, the process of using the whale migration optimization algorithm to improve the exploration phase and development phase of the exponential triangulation optimization algorithm includes: using the young whales in the whale migration optimization algorithm to follow the movements of the whales around them to improve the first exploration phase of the exponential triangulation optimization algorithm, using the young whales influenced by the movements of experienced whales to improve the second exploration phase of the exponential triangulation optimization algorithm, using experienced whales to improve the first development phase of the exponential triangulation optimization algorithm, and using whale migration groups to improve the second development phase of the exponential triangulation optimization algorithm.

[0016] As a further limited embodiment, the process of exploring the first phase of the exponential triangular optimization algorithm using the young whales following the whales around them in the whale migration optimization algorithm and exploring the second phase of the exponential triangular optimization algorithm using the young whales influenced by the experienced whales includes:

[0017] The first phase of the exploration of using the whale migration optimization algorithm to improve the exponential triangulation optimization algorithm:

[0018]

[0019] in, and They represent the positions of the i-th solution in the n-th dimension of the current iteration and the iteration respectively. The function rand(1,D) generates a random number of vectors from the interval [0,1] of dimension D. The operator ⊙ represents the symbolic operation of the Hadamard product of two vectors. Each element of the resulting vector is obtained by multiplying the corresponding elements of the two original vectors. q1 represents a random value between the interval [0,1], and α1 is the weight coefficient of the current stage. is the position of the optimal solution obtained so far in the jth dimension, t is the current number of iterations, N L is the number of experienced whales, N pop For individual juvenile whales, W i-1 is the influence of the neighbor whale on the current whale, that is, the previous baby whale, W i The current whale in this iteration.

[0020] As the first phase of exploration iterates, the whale group will continue to approach the optimal solution. At this time, the young whales will be influenced by the experienced whales. The second phase of exploration of the exponential triangulation optimization algorithm is improved by using the young whales to follow the experienced whales:

[0021]

[0022] Among them, rand() represents a random value between the interval [0,1], q2 represents a random value between the interval [0,1], α2 is the weight coefficient of the current stage, W Mean is the average of the current positions of all experienced whales, W best is the current best whale position, which is the best position among experienced whales, or the best position of the whale algorithm;

[0023] In both stages, half of the population is affected by the individuals around it, and the other half is affected by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained;

[0024] The parameter T1 is used to control the transition between the first and second stages of exploration, specifically:

[0025]

[0026] Among them, when t≤T1, the first stage of exploration is carried out, and when t>T1, the second stage of exploration is carried out. Floor() is a floor function, a1 is a coefficient used to adjust the starting position of the threshold, b1 is an empirical coefficient used to balance the time allocation of the first and second stages of exploration, and Max_Iter represents the maximum number of iterations.

[0027] As an optional implementation, the process of developing the first phase of the exponential triangulation optimization algorithm using experienced whales and the second phase of the exponential triangulation optimization algorithm using whale migration groups includes:

[0028] After entering the development phase, experienced whales were used to improve the first phase of the development of the exponential triangulation optimization algorithm, specifically:

[0029]

[0030] in, is the position of the nth experienced whale, q3 represents a random value between the interval [0,1], α3 is the weight coefficient of this stage, q4 represents a random value between the interval [0,1], is the optimal solution obtained so far at the position of the jth dimension, t is the current number of iterations, and Represents the position of the i-th solution in the n-th dimension in the current iteration and the iteration after, W best is the current best whale position;

[0031] Half of the population is influenced by experienced whales, and the other half is influenced by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained;

[0032] The same T2 control is used to develop the transition between the first and second stages, specifically:

[0033]

[0034] Among them, when t≤T2, the first stage of development is carried out, and when t>T2, the second stage of development is carried out. Floor() is a floor function, a2 is a coefficient used to adjust the starting position of the threshold, b2 is an empirical coefficient used to balance the time allocation of the first and second stages of development, and Max_Iter represents the maximum number of iterations;

[0035] After entering the second stage of development, the intensity of local optimization will continue to increase with the iteration. The development formula is improved using the whale migration group. Specifically:

[0036]

[0037] Where rand() represents a random value between the interval [0,1], c is a coefficient, and d1 and d2 are both symmetric functions whose value range is [-0.1,0.1].

[0038] As an optional implementation, the process of using the Levy flight process to improve the judgment factors for judging the exploration phase and the exploitation phase and to escape from the local optimization includes:

[0039] Levy flies as:

[0040]

[0041] Where η is a constant, u and v are random numbers that obey the standard normal distribution, and the calculation formula for σ is:

[0042]

[0043] Where, Γ represents the gamma function;

[0044] make It represents the i-th solution in the j-th dimension space after t+1 iterations and the i-th solution in the j-th dimension space after t iterations The position difference of divided by the position of the i-th solution in the j-th dimension space at t iterations;

[0045] When the number of consecutive set values ​​Z(t+1) is less than the set value during the optimization process, the position at this time is recorded as Use Levy flight to improve the judgment factor CM in the exploration and development phases and jump out of the current position:

[0046]

[0047] Both d1 and d2 are symmetric functions with a value range of [-0.1, 0.1]. rand() represents a random value between the interval [0, 1]. Max_Iter represents the maximum number of iterations.

[0048] CM≥1, in the exploration stage, move the individual away from the current optimal solution, use the new position and the new current number of iterations t to re-optimize the exploration stage, and then re-optimize the exploration stage and the development stage during the iteration process.

[0049] As a further limited implementation method, in the optimization process, T n When the number of occurrences of the set value Z(t+1) is less than the set value, the new optimal solution is compared with the previous optimal solution, the best of the two is selected, and the number of jumps is recorded at the same time. When the number of jumps reaches the set value or the number of iterations reaches the maximum, the iteration stops and the final optimal position X is obtained. best .

[0050] As an optional implementation, the process of using the final optimal solution as the scaling factor and bias of the large model is:

[0051] The large model is based on the Transformer model framework. The Transformer model includes an encoder and a decoder. The encoder is composed of multiple identical layers stacked together, each layer consisting of a multi-head attention layer and a fully connected feedforward neural network layer. The decoder is composed of multiple layers stacked together, each layer consisting of two multi-head attention layers and a fully connected feedforward neural network layer, where the first multi-head attention layer adopts a masked operation.

[0052] After each multi-head attention layer and fully connected feedforward neural network layer, there is a residual connection and layer normalization layer; the output of the encoder after multiple layers is passed to each layer of the decoder;

[0053] The scaling factor of the multi-head attention layer is determined by optimizing the combined intelligent algorithm;

[0054] The bias of the fully connected feedforward neural network layer is determined by optimizing the combined intelligent algorithm.

[0055] An engine production line assembly quality control system based on a large model, including:

[0056] An assembly inspection library module is configured to obtain historical assembly data of various engine components and build an assembly process inspection sample library based on the historical assembly data;

[0057] A combination algorithm improvement module is configured to improve the large model using a combination intelligent algorithm, and use the improved large model to perform classification training on the assembly process inspection sample library;

[0058] An assembly warning module is configured to obtain target assembly data of various engine components being assembled, process the target assembly data using a large model trained through classification, determine the assembly status of various engine components, and issue assembly quality warnings for abnormal status;

[0059] Among them, the process of using combined intelligent algorithms to improve the large model includes using the whale migration optimization algorithm to improve the exploration stage and development stage of the exponential triangle optimization algorithm, using the Levy flight process to improve the judgment factors used to judge the exploration stage and development stage and jump out of local optimization to obtain the global optimal solution, iterate the optimization process until the iteration requirements are met, form a combined intelligent algorithm, and obtain the final optimal solution. The final optimal solution is used as the scaling factor and bias of the large model to achieve improvement of the large model.

[0060] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps in the above method are completed.

[0061] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the above method are completed.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] The present invention proposes an engine production line assembly quality control method based on a large model, and proposes to improve the first exploration stage of the exponential triangular optimization algorithm by having the young whales in the whale migration optimization algorithm follow the movements of the whales around them, to improve the second exploration stage of the exponential triangular optimization algorithm by having the young whales be influenced by the movements of experienced whales, to improve the first development stage of the exponential triangular optimization algorithm by experienced whales, to improve the second development stage of the exponential triangular optimization algorithm by improving the whale migration group, and to improve the judgment factor CM of the exploration stage and the development stage of the Levy flight process to jump out of the local optimization and obtain the global optimal solution, so that the exponential triangular optimization algorithm can find the optimal point locally and can jump out of the optimal point to a certain extent, thereby realizing the ability of the algorithm to jump out of the local optimal point in complex, multi-peak and other data problems.

[0064] The present invention combines intelligent algorithms to optimize the scaling factor and bias of the Transformer model, thereby improving the Transformer model. The improved model is used to perform classification training on the assembly process detection sample library, determine the normal or abnormal status of the assembly of each engine system, and issue assembly quality warnings, greatly improving the assembly quality of the engine production line.

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0067] Figure 1 The present invention is a flow chart of an engine production line assembly quality control method based on a large model in an embodiment. DETAILED DESCRIPTION

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0070] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0071] In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0072] Example 1

[0073] As described in the background technology, the existing engine production line assembly quality control has problems such as large detection blind spots and poor real-time performance. In the engine production line assembly quality control method based on a large model, the present invention makes the young whales in the whale migration optimization algorithm follow the whales around them to move in the first exploration stage of the improved exponential triangular optimization algorithm, makes the young whales subject to the experienced whales to move in the second exploration stage of the improved exponential triangular optimization algorithm, makes the experienced whales improve the first development stage of the exponential triangular optimization algorithm, makes the whale migration group improve the second development stage of the exponential triangular optimization algorithm, improves the Levy flight process to judge the judgment factor CM in the exploration stage and the development stage to jump out of the local optimization and obtain the global optimal solution, iterates the optimization process until the iteration requirements are met, forms a combined intelligent algorithm, obtains the final optimal solution, and uses the final optimal solution as the scaling factor and bias of the Transformer model to realize the improvement of the Transformer model. The improved model is used to classify and train the assembly process detection sample library, determines the normal or abnormal state of the assembly of each system of the engine, and makes an assembly quality warning for it, which greatly improves the assembly quality of the engine production line.

[0074] like Figure 1 As shown, this embodiment provides an engine production line assembly quality control method based on a large model, comprising the following steps:

[0075] Step 1: Obtain historical assembly data values ​​of each system and component of the engine and build an assembly process inspection sample library, specifically:

[0076] The intake pressure, early opening and late closing angles of the intake and exhaust valves, camshaft operating torque, camshaft axial clearance, cylinder head nut tightening torque, flywheel housing bolt tightening torque, flywheel housing bolt tightening angle, etc. after assembling multiple engines (for example, tens of thousands of engines) are collected. The corresponding crankshaft torque histogram is judged using normal distribution. In this embodiment, data within 2.5σ is a normal value and is marked as 1. Data greater than or less than 2.5σ is an abnormal value and is marked as 0. In this way, an assembly process detection sample library is constructed.

[0077] Step 2: Use the whale migration algorithm to improve the exponential triangle optimization algorithm and use the Levy flight process to improve the judgment factor CM in the exploration and development stages to jump out of the local optimization to obtain the global optimal solution. Specifically:

[0078] In order to enable those skilled in the art to better understand the overall process, the exponential triangulation optimization algorithm and the whale migration algorithm are first introduced.

[0079] The exponential triangular optimization algorithm is:

[0080] A constrained search strategy is adopted to ensure that the exploration space is dynamically adapted. The specific equation is:

[0081]

[0082] Where U j and L j They represent the upper and lower bounds of the expected exploration space of the jth dimension, r1 and r2 are coefficients randomly selected between 0 and 1, is the optimal solution obtained so far at the position of the jth dimension, Indicates the position of the suboptimal solution in the jth dimension, t is the current number of iterations, and Max_Iter indicates the maximum number of iterations.

[0083] The equation for initializing the population is:

[0084] i i,j =(U j -L j )×rand()+L j

[0085] Among them, rand() represents a random value between the interval [0,1], i is the i-th individual, i = 1, 2, ..., N, j is the dimension, j = 1, 2, ..., d.

[0086] After initializing the population and determining the boundaries of the exploration space, individuals begin early exploration of the search domain, specifically:

[0087]

[0088] Where, and Represent the positions of the i-th solution in the n-th dimension of the current iteration and the next iteration, respectively. rand() and q1 represent random values ​​between the interval [0,1]. The weight coefficient at this stage is α1, and the formula is:

[0089]

[0090] Where d1 and d2 are both symmetric functions with a value range of [-0.1, 0.1]. The calculation formula is:

[0091]

[0092] In the later stages of exploration, the individual's position is updated as:

[0093]

[0094] In the formula, q2 represents a random value between the interval [0,1]. The weight coefficient at this stage is α2, which will continue to decrease and converge to zero to achieve the optimal solution. The calculation formula is:

[0095]

[0096] When the number of consecutive optimizations is less than the set value, the local optimization process will be exited and global optimization will be performed.

[0097] In the first phase of development, focus on exploration Nearby, specifically:

[0098]

[0099] Where q3 represents a random value between the interval [0,1]. The weight coefficient at this stage is α3, which will increase from a small number. The calculation formula is:

[0100]

[0101] In the second stage of development, as iterations progress, the intensity of local exploration increases, specifically:

[0102]

[0103] In the formula, c helps to maintain the diversity of solutions and is calculated as:

[0104]

[0105] The transition between the exploration phase and the development phase is determined by the CM:

[0106]

[0107] When CM≥1, it is in the exploration stage, and when CM<1, it is in the development stage.

[0108] The whale migration optimization algorithm is:

[0109] The population of humpback whales is divided into juvenile whales and experienced whales, and the position of each whale is from W1 to Indicates that the nth whale uses W n Indicates that,

[0110] W n =(w n,1 ,w n,2 ,...,w n,d ,...,w n,D )

[0111] n=1,2,...N pop

[0112] d=1,2,...,D

[0113] N pop This is the total number of whales, including all experienced whales and juveniles.

[0114] The whale population is an initial population generated in the D-dimensional space from the lower limit L to the upper limit U, specifically:

[0115] W n =L+rand(1,D)⊙(UL),n=1,2,...,N pop

[0116] The function rand(1,D) generates a random number of vectors from the interval [0,1] of dimension D. The operator ⊙ represents the symbolic operation of the Hadamard product of two vectors. Each element of the resulting vector is obtained by multiplying the corresponding elements of the two original vectors.

[0117] As the least experienced whales, the young whales will move closer to the whales closest to them. Assuming that all whale group members are sorted in descending order according to fitness value or position:

[0118]

[0119] Among them, W1 is the best member, Is the worst member. Each less experienced whale Will be affected by the previous member in the sequence, that is, W n-1 , the relationship is:

[0120]

[0121] in, represents the number of experienced whales, and rand(1,D) is a vector that generates random numbers from the interval [0,1] of dimension D.

[0122] In this article, the superscript Le represents young whales, and Me represents experienced whales. The corresponding numbers in parentheses are the ranges of the number of young whales and the number of experienced whales, respectively.

[0123] The position of the entire whale group in the ocean is the average of the current positions of all experienced whales, W Mean :

[0124]

[0125] in, is the position of the nth experienced whale, if W Mean Points and the best whale position W Best The distance between the points begins to decrease, and the entire group of experienced whales is approaching W Best At this time, the baby whale must also start to move along rand(1, D)⊙(W Best -W Mean ), the new position of the nth baby whale can be expressed as:

[0126]

[0127] The new position is only when the fitness value of the new position is better than the fitness value of the old position. Will replace the old position W i , otherwise, keep the old position unchanged.

[0128] Experienced whales are responsible for identifying and selecting the best route to the destination. The equation for the nth whale to find a suitable path to the destination is:

[0129]

[0130] Among them, r1 and r2 represent the [0,1] random number vectors in dimension D. Similarly, the new position is only when the fitness value of the new position is better than the fitness value of the old position. Will replace the old position W j , otherwise, keep the old position unchanged.

[0131] The process of using the whale migration algorithm to improve the exponential triangulation optimization algorithm includes:

[0132] The search space for exponential trigonometric functions is improved using young and experienced whales, specifically:

[0133]

[0134] in, is the optimal solution currently obtained by experienced whales, W i best This is the optimal solution currently obtained by the young whale. It is the optimal solution currently obtained by the exponential triangular optimization algorithm.

[0135] The first phase of the exponential triangulation optimization algorithm was improved by using the baby whale's movement to follow the whales around it. Specifically:

[0136]

[0137] W i-1 is the influence of the neighbor whale on the current whale, that is, the previous baby whale, W i The current calf in this iteration;

[0138] Half of the population is affected by the individuals around it, and the other half of the individuals are affected by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained.

[0139] Use T1 control to explore the transition between the first and second phases, specifically:

[0140]

[0141] Among them, when t≤T1, the first stage of exploration is carried out, and when t>T1, the second stage of exploration is carried out. floor() is a floor function, a1 is a coefficient used to adjust the starting position of the threshold, and b1 is an empirical coefficient to balance the time allocation of the first and second stages of exploration. A larger b1 will reduce the value of T1, so that the second stage of exploration will start earlier and accelerate convergence.

[0142] As the first phase of exploration iterates, the whale group will continue to approach the optimal solution. At this time, the young whales will be influenced by the experienced whales. The second phase of exploration of the exponential triangulation optimization algorithm is improved by using the young whales to follow the experienced whales. Specifically:

[0143]

[0144] Similarly, half of the population is affected by the individuals around it, and the other half of the individuals are affected by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained.

[0145] Use CM to control the transition between the exploration phase and the development phase, specifically:

[0146]

[0147] When CM≥1, it is in the exploration stage, and when CM<1, it is in the development stage.

[0148] After entering the development phase, experienced whales were used to improve the first phase of the development of the exponential triangulation optimization algorithm, specifically:

[0149]

[0150] q1-q4 all represent random values ​​in the interval [0,1].

[0151] Half of the population is influenced by experienced whales, and the other half of the individuals are influenced by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained.

[0152] Use T2 controls to develop the transition between Phase I and Phase II, specifically:

[0153]

[0154] Among them, when t≤T2, the first stage of development is carried out, and when t>T2, the second stage of development is carried out. floor() is a floor function, a2 is a coefficient used to adjust the starting position of the threshold, and b2 is an empirical coefficient to balance the time allocation of the first and second stages of development. A larger b2 will reduce the value of T2, so that the second stage of development can be started earlier and the convergence will be accelerated.

[0155] After entering the second stage of development, the intensity of local optimization will continue to increase with the iteration. The development formula is improved using the whale migration group. Specifically:

[0156]

[0157] Where c is a coefficient that helps maintain the diversity of solutions.

[0158] Levy Flying:

[0159]

[0160] Where η is a constant with a value of 1.5, u and v are random numbers that obey the standard normal distribution, and the calculation formula for σ is:

[0161]

[0162] Where Γ represents the gamma function.

[0163] make It represents the difference between the position of the i-th solution in the j-dimensional space after t+1 iterations and t iterations divided by the position of the i-th solution in the j-dimensional space at t iterations. is the current location, is the position at the next moment.

[0164] In the optimization process, continuous T n When the number of times the set value occurs Z(t+1) is less than the set value, such as T n When the value is between 25 and 40, the position at this time is recorded as Use Levy flight to improve the judgment factor CM in the exploration and development stages and jump out of the current position. The formula is:

[0165]

[0166] At this time, CM≥1, which means it is in the exploration stage. The individual can be moved away from the current optimal solution and the new position and the new current number of iterations t can be used to re-optimize the exploration stage. The exploration stage and the development stage can be optimized again during the iteration process.

[0167] In the optimization process, continuous T n When the number of occurrences of the set value Z(t+1) is less than the set value, for example, T n When the number of times is 25 to 40, the new optimal solution is compared with the previous optimal solution, and the best one is selected. n Record the number of times. When the number of bounces reaches T nn times, for example, T nn The iteration stops when it is 5 to 10 times or when the number of iterations reaches the maximum, and the final optimal position X is obtained. best , thus obtaining a combined intelligent algorithm.

[0168] Step 3: Improve the Transformer model, specifically:

[0169] The structure of Transformer mainly consists of an encoder and a decoder. The encoder is composed of M (M=1,2,...,n) identical stacked layers, each of which consists of a multi-head attention layer and a fully connected feedforward neural network layer; the decoder is also composed of the same M stacked layers, but each of its layers consists of two multi-head attention layers and a fully connected feedforward neural network layer, where the first multi-head attention layer adopts a masked operation; after each multi-head attention layer and fully connected feedforward neural network layer, there is an Add&Norm layer; the output results of the encoder after N layers are passed to each layer of the decoder.

[0170] For a single-head attention layer, the input vector will generate three vectors, specifically, W Q is the query vector weight matrix, W K is the key vector weight matrix, W Vis the value vector weight matrix. Multiply the input by the three matrices to obtain three vectors: query vector Q, key vector K, and value vector V. After multiplying the query vector and the key vector to calculate the score, divide the score by the scaling factor The scaling factor is given by the combined intelligent algorithm, that is, the value is obtained by optimizing the combined intelligent algorithm, and then the result is transferred using the softmax function. After obtaining the softmax function value of each individual, it is summed up. The calculation formula is:

[0171]

[0172] In the formula, the role of the softmax function is to normalize the individual scores of all inputs so that the obtained scores are all positive and the sum is 1.

[0173] Compared with the single-head attention layer, the calculation of the multi-head attention layer is:

[0174] MuliHead(Q,K,V)=Concat(head1,...,head h )W O

[0175] head i =Attention(QW i Q ,KW i K ,VW i V )

[0176] Where, and is the parameter matrix, W O is the additional weight matrix, d model is a dimension vector, d k is the key vector dimension, h is

[0177] After the calculation of the multi-head attention layer, the self-attention is obtained, and the obtained vector is passed to the fully connected feedforward neural network layer. This layer includes two linear transformations and a ReLU activation function, which is calculated as:

[0178] FFN(x)=max(0,ReLU(xW1+Bias1))W2+Bias2

[0179] Where Bias1 and Bias2 are biases, and the biases are given by the combined intelligent algorithm. That is, the two values ​​are obtained by optimizing the combined intelligent algorithm. There is no relationship between the two values. The first bias is to enhance the expressiveness of the model, and the second bias is to further adjust the output distribution, which can enhance the ability to fit complex patterns, improve learning ability, and improve the adaptability of the model.

[0180] The Add&Norm layer consists of two parts, Add and Norm, which prevent network degradation and is calculated as:

[0181] LayerNorm=(X+MulitHeadAttention(x))

[0182] LayerNorm=(X+FeedForward(x))

[0183] Where x represents the input of the multi-head attention layer and the fully connected feedforward neural network layer, and MulitHeadAttention(x) and FeedForward(x) represent the input of the multi-head attention layer and the fully connected feedforward neural network layer.

[0184] The Masked operation can make the output result only affected by the previous output result, but not by the value to be output in the future.

[0185] In order for Transformer to utilize the order information of the input sequence, a trigonometric function is used to generate a unique code for each position. The generation formula is:

[0186]

[0187] Where pos is the position and i is the dimension index.

[0188] After the decoder completes the calculation, it will undergo a linear transformation and a softmax function to obtain the value with the highest probability, search the corresponding database, and obtain the final output.

[0189] During the training of the improved Transformer model, F β The score is used as an evaluation indicator and the calculation formula is:

[0190]

[0191] Where P is precision, R is recall, TP represents the number of samples correctly detected as positive, FP represents the number of samples incorrectly detected as negative, FN represents the number of samples incorrectly detected as positive, and β is a fixed value.

[0192] The fitness function is maximized, that is, F β The maximum score is the objective function of the combined intelligent algorithm optimization.

[0193] Step 4: Use historical data sets for classification training, and then use the trained model to process the assembly data values ​​of the engine component systems being assembled based on the engine production line in real time, and then determine the normal or abnormal status of the assembly of each engine system and issue an assembly quality warning.

[0194] The specific process of this embodiment can also be described as:

[0195] The first stage of exploration of the improved exponential triangular optimization algorithm is to have the young whales in the whale migration optimization algorithm follow the whales around them, the second stage of exploration of the improved exponential triangular optimization algorithm is to have the young whales follow the experienced whales, the first stage of development of the improved exponential triangular optimization algorithm is to have the experienced whales, the second stage of development of the improved exponential triangular optimization algorithm is to have the whale migration group improve the Levy flight process. The judgment factor CM in the exploration stage and the development stage is improved to jump out of the local optimization and obtain the global optimal solution. The optimization process is iterated until the iteration requirements are met, a combined intelligent algorithm is formed, and the final optimal solution is obtained. The final optimal solution is used as the scaling factor and bias of the Transformer model to realize the improvement of the Transformer model. The improved model is used to classify and train the assembly process detection sample library, and then the trained model is used to process the real-time data to determine the normal or abnormal status of the assembly of each engine system and make an assembly quality warning.

[0196] Example 2

[0197] An engine production line assembly quality control system based on a large model, including:

[0198] An assembly inspection library module is configured to obtain historical assembly data of various engine components and build an assembly process inspection sample library based on the historical assembly data;

[0199] A combination algorithm improvement module is configured to improve the large model using a combination intelligent algorithm, and use the improved large model to perform classification training on the assembly process inspection sample library;

[0200] An assembly warning module is configured to obtain target assembly data of various engine components being assembled, process the target assembly data using a large model trained through classification, determine the assembly status of various engine components, and issue assembly quality warnings for abnormal status;

[0201] Among them, the process of using combined intelligent algorithms to improve the large model includes using the whale migration optimization algorithm to improve the exploration stage and development stage of the exponential triangle optimization algorithm, using the Levy flight process to improve the judgment factors used to judge the exploration stage and development stage and jump out of local optimization to obtain the global optimal solution, iterate the optimization process until the iteration requirements are met, form a combined intelligent algorithm, and obtain the final optimal solution. The final optimal solution is used as the scaling factor and bias of the large model to achieve improvement of the large model.

[0202] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0203] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0204] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0205] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0206] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made by those skilled in the art that fall within the spirit and principles of the present invention and do not require creative effort are intended to be within the scope of protection of the present invention.

Claims

1. A large-scale model-based engine production line assembly quality control method, characterized by: The following steps are involved: Obtain historical assembly data of each engine component and build an assembly process inspection sample library based on the historical assembly data; A combined intelligent algorithm is used to improve the large model, and the improved large model is used to perform classification training on the assembly process inspection sample library; Obtain target assembly data for each engine component being assembled, process the target assembly data using a large model trained through classification, determine the assembly status of each engine component, and issue assembly quality warnings for abnormal status; Among them, the process of using combined intelligent algorithms to improve the large model includes using the whale migration optimization algorithm to improve the exploration stage and development stage of the exponential triangle optimization algorithm, using the Levy flight process to improve the judgment factors used to judge the exploration stage and development stage and jump out of local optimization to obtain the global optimal solution, iterate the optimization process until the iteration requirements are met, form a combined intelligent algorithm, and obtain the final optimal solution. The final optimal solution is used as the scaling factor and bias of the large model to achieve improvement of the large model.

2. The large model-based engine production line assembly quality control method according to claim 1, characterized in that: In the process of obtaining the historical assembly data of various engine components, the intake pressure, early opening and late closing angles of the intake and exhaust valves, camshaft operating torque, camshaft axial clearance, cylinder head auxiliary nut tightening torque, flywheel housing bolt tightening torque and flywheel housing bolt tightening angle of multiple engines are obtained, and the corresponding crankshaft torque histogram is drawn. The normal distribution is used for judgment. The data within the set range is normal value, and the data outside the set range is abnormal value.

3. The large model-based engine production line assembly quality control method according to claim 1 or 2, characterized in that: In the process of using the improved large model to classify and train the assembly process detection sample library, the assembly process detection sample library is divided into a normal value library and an abnormal value library. Among them, the assembly data within the normal value range is the normal value library, and the assembly data outside the normal value range is the abnormal value library.

4. The large model-based engine production line assembly quality control method according to claim 1, characterized in that: The process of using the whale migration optimization algorithm to improve the exponential triangular optimization algorithm in the exploration phase and development phase includes: using the young whales in the whale migration optimization algorithm to follow the movements of the whales around them to improve the first exploration phase of the exponential triangular optimization algorithm, using the young whales influenced by the movements of experienced whales to improve the second exploration phase of the exponential triangular optimization algorithm, using experienced whales to improve the first development phase of the exponential triangular optimization algorithm, and using whale migration groups to improve the second development phase of the exponential triangular optimization algorithm.

5. The large model-based engine production line assembly quality control method according to claim 4, characterized in that: The first phase of the exploration of the exponential triangulation optimization algorithm is to use the young whales in the whale migration optimization algorithm to follow the movements of the whales around them. The second phase of the exploration of the exponential triangulation optimization algorithm is to use the young whales to follow the movements of the experienced whales. The process includes: The first phase of the exploration of using the whale migration optimization algorithm to improve the exponential triangulation optimization algorithm: i=N L +1,...,N pop in, and Represent the positions of the i-th solution in the n-th dimension of the current iteration and the iteration respectively. The function rand(1,D) generates a random number of vectors from the interval [0,1] of dimension D. The operator ⊙. represents the symbolic operation of the Hadamard product of two vectors. Each element of the resulting vector is obtained by multiplying the corresponding elements of the two original vectors. q1 represents a random value between the interval [0,1], and α1 is the weight coefficient of the current stage. is the position of the optimal solution obtained so far in the jth dimension, t is the current number of iterations, N L is the number of experienced whales, N pop For the individual of the juvenile whale, W i-1 is the influence of the neighbor whale on the current whale, that is, the previous baby whale, W i The current calf in this iteration; As the first phase of exploration iterates, the whale group will continue to approach the optimal solution. At this time, the young whales will be influenced by the experienced whales. The second phase of exploration of the exponential triangulation optimization algorithm is improved by using the young whales to follow the experienced whales: Among them, rand() represents a random value between the interval [0,1], q2 represents a random value between the interval [0,1], α2 is the weight coefficient of the current stage, W Mean is the average of the current positions of all experienced whales, W best is the current best whale position; In both stages, half of the population is affected by the individuals around it, and the other half is affected by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained; The parameter T1 is used to control the transition between the first and second stages of exploration, specifically: Among them, when t≤T1, the first stage of exploration is carried out, and when t>T1, the second stage of exploration is carried out. Floor() is a floor function, a1 is a coefficient used to adjust the starting position of the threshold, b1 is an empirical coefficient used to balance the time allocation of the first and second stages of exploration, and Max_Iter represents the maximum number of iterations.

6. The large model-based engine production line assembly quality control method according to claim 4, characterized in that: The process of developing the first phase of the exponential triangulation optimization algorithm using experienced whales and the second phase of the exponential triangulation optimization algorithm using whale migration groups includes: After entering the development phase, experienced whales were used to improve the first phase of the development of the exponential triangulation optimization algorithm, specifically: in, is the position of the nth experienced whale, q3 represents a random value between the interval [0,1], α3 is the weight coefficient of this stage, q4 represents a random value between the interval [0,1], is the optimal solution obtained so far at the position of the jth dimension, t is the current number of iterations, and Represents the position of the i-th solution in the n-th dimension in the current iteration and the iteration after, W best is the current best whale position; Half of the population is influenced by experienced whales, and the other half is influenced by the original exponential triangular optimization algorithm, and the positions whose fitness values ​​of the new positions are better than the fitness values ​​of the old positions are retained; Use T2 controls to develop the transition between Phase I and Phase II, specifically: Among them, when t≤T2, the first stage of development is carried out, and when t>T2, the second stage of development is carried out. Floor() is a floor function, a2 is a coefficient used to adjust the starting position of the threshold, b2 is an empirical coefficient used to balance the time allocation of the first and second stages of development, and Max_Iter represents the maximum number of iterations; After entering the second stage of development, the intensity of local optimization will continue to increase with the iteration. The development formula is improved using the whale migration group. Specifically: Where rand() represents a random value between the interval [0,1], c is a coefficient, and d1 and d2 are both symmetric functions whose value range is [-0.1,0.1].

7. The large model-based engine production line assembly quality control method according to claim 1, characterized in that: The process of using the Levy flight process to improve the judgment factors used to determine the exploration phase and the development phase and to escape from local optimization includes: Levy flies as: Where η is a constant, u and v are random numbers that obey the standard normal distribution, and the calculation formula for σ is: Where, Γ represents the gamma function; make It represents the i-th solution in the j-th dimension space after t+1 iterations and the i-th solution in the j-th dimension space after t iterations The position difference of divided by the position of the i-th solution in the j-th dimension space at t iterations; When the number of consecutive set values ​​Z(t+1) is less than the set value during the optimization process, the position at this time is recorded as Use Levy flight to improve the judgment factor CM in the exploration and development phases and jump out of the current position: Both d1 and d2 are symmetric functions with a value range of [-0.1, 0.1]. rand() represents a random value between the interval [0, 1]. Max_Iter represents the maximum number of iterations. CM≥1, in the exploration stage, move the individual away from the current optimal solution, use the new position and the new current number of iterations t to re-optimize the exploration stage, and then re-optimize the exploration stage and the development stage during the iteration process.

8. The large model-based engine production line assembly quality control method according to claim 7, characterized in that: In the optimization process, continuous T n When the number of occurrences of the set value Z(t+1) is less than the set value, the new optimal solution is compared with the previous optimal solution, the best of the two is selected, and the number of jumps is recorded at the same time. When the number of jumps reaches the set value or the number of iterations reaches the maximum, the iteration stops and the final optimal position X is obtained. best .

9. The large model-based engine production line assembly quality control method according to claim 1, characterized in that: The process of using the final optimal solution as the scaling factor and bias of the large model is: The large model is a Transformer model, which includes an encoder and a decoder. The encoder is composed of multiple identical layers stacked together, each layer consisting of a multi-head attention layer and a fully connected feedforward neural network layer. The decoder is composed of multiple layers stacked together, each layer consisting of two multi-head attention layers and a fully connected feedforward neural network layer, where the first multi-head attention layer adopts a masked operation. After each multi-head attention layer and fully connected feedforward neural network layer, there is a residual connection and layer normalization layer; the output of the encoder after multiple layers is passed to each layer of the decoder; The scaling factor of the multi-head attention layer is determined by optimizing the combined intelligent algorithm; The bias of the fully connected feedforward neural network layer is determined by optimizing the combined intelligent algorithm.

10. An engine production line assembly quality control system based on a large model, characterized by: include: An assembly inspection library module is configured to obtain historical assembly data of various engine components and build an assembly process inspection sample library based on the historical assembly data; A combination algorithm improvement module is configured to improve the large model using a combination intelligent algorithm, and use the improved large model to perform classification training on the assembly process inspection sample library; An assembly warning module is configured to obtain target assembly data of various engine components being assembled, process the target assembly data using a large model trained through classification, determine the assembly status of various engine components, and issue assembly quality warnings for abnormal status; Among them, the process of using combined intelligent algorithms to improve the large model includes using the whale migration optimization algorithm to improve the exploration stage and development stage of the exponential triangle optimization algorithm, using the Levy flight process to improve the judgment factors used to judge the exploration stage and development stage and jump out of local optimization to obtain the global optimal solution, iterate the optimization process until the iteration requirements are met, form a combined intelligent algorithm, and obtain the final optimal solution. The final optimal solution is used as the scaling factor and bias of the large model to achieve improvement of the large model.

Images