Warehouse management method and system

Abstract

The invention discloses a warehouse management method and system, and the method comprises the steps: S1, collecting warehouse management data, and extracting a feature vector of the warehouse management data; s2, carrying out coding reconstruction on the collected feature vectors through an optimized variational auto-encoder to obtain potential feature variables, and obtaining an initial reconstruction sequence through a decoder part of the optimized variational auto-encoder based on the potential feature variables; and S3, through a generative adversarial hybrid network, taking the initial reconstruction sequence as the input of a generator of the generative adversarial hybrid network, carrying out joint training through gradient penalty to optimize the generative adversarial hybrid network, obtaining a real reconstruction sequence, and obtaining an abnormal score through a training model and a calculation probability. Whether the input cargo data is abnormal or not can be more accurately judged, and the risk that the data is abnormal and missed is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of warehouse data management, and in particular to a warehouse management method and system. Background Art

[0002] In existing warehouse management, traditional data management methods face many challenges. For example, plush toys have the characteristics of various styles, different sizes, and rich colors, which makes the data dimensions of warehouse management complex and diverse. In the data entry process, manual operations are usually relied on to record the information such as the entry, exit, and inventory adjustment of goods. However, due to the high similarity of plush toy products, staff may make mistakes when entering the quantity, specifications (such as size, material category), batch and other information of goods due to visual fatigue or negligence. For example, plush toys from different batches but with similar appearance may be mistakenly recorded as the same batch when entering the warehouse, or the quantity may be counted incorrectly when leaving the warehouse.

[0003] At present, some warehouse management systems usually take certain verification measures in the data entry link. However, due to the uncontrollability of human operation and the complexity of the business, data entry errors are still difficult to completely avoid. For example, in the warehouse operation process, the warehousing and outbound operations of goods should be balanced in data, that is, for a certain product, within a certain time range, the difference between the total inbound and outbound volume should be equal to the change in current inventory. However, due to some human operations or other reasons, there may be a mismatch in inbound and outbound data, which affects the accuracy of warehouse management. Therefore, a warehouse management method and system are proposed herein. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art that the in-and-out warehouse data do not match, and to propose a warehouse management method and system.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] Warehouse management methods, including:

[0007] S1: Collect warehouse management data and extract feature vectors of warehouse management data;

[0008] S2: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables;

[0009] S3: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the true reconstructed sequence;

[0010] S4: Obtain a series of behavior sequences of the latest incoming data, and obtain the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it means that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

[0011] The steps for obtaining the feature vector of the warehouse management data are as follows:

[0012] When goods enter the warehouse, the quantity and specification information of the goods are obtained through the warehouse scanning equipment, and the entry time and batch are recorded at the same time to form a complete feature vector, expressed as X=(Q, S, B, T), where Q is the quantity, S is the specification, B is the batch, and T is the time.

[0013] Furthermore, this vector form of data can effectively reflect the details of warehouse operations. Warehouse management data contains various key information about the goods entering and leaving the warehouse, including quantity, specifications, batches and time, which can be represented in vector form, providing a basis for subsequent data processing and comprehensively capturing the actual situation of warehouse operations.

[0014] The step of encoding and reconstructing the collected feature vectors through an optimized variational autoencoder to obtain potential feature variables is as follows:

[0015] The vector feature X is used as the input of an autoencoder, and the autoencoder is optimized to a multi-layer autoencoder structure to obtain an optimized variational autoencoder;

[0016] The input data of the input layer of the optimized variational autoencoder is X;

[0017] The encoder layer of the optimized variational autoencoder is composed of two encoders, and the decoder layer is composed of two decoders;

[0018] The encoding process of the first layer encoder:

[0019] The input data X is input to the first layer encoder, which maps X to an intermediate representation through a neural network: Among them, W1 represents the weight matrix of the neural network in the first layer, b1 represents the bias vector of the neural network in the first layer, and f1 is the activation function of the neural network in the first layer. This intermediate representation is the encoding result of the first layer encoder;

[0020] The encoding process of the second layer encoder:

[0021] The intermediate representation Input to the second layer encoder, the second layer encoder will It is further mapped to the mean μ and standard deviation σ of the latent variable;

[0022] The calculation formula for the mean μ of the latent variable is: Wherein, W2 represents the weight matrix of the neural network in the second layer, b2 represents the bias vector of the neural network in the second layer, and f2 is the activation function used to calculate the mean in the second layer;

[0023] The calculation formula of the latent variable standard deviation σ is: in, The activation function used to calculate the standard deviation in the second layer;

[0024] In this way, the model is able to sample samples from the latent space that conform to a specific distribution, which helps capture the uncertainty and latent structure of the data during the encoding process;

[0025] The latent feature variables are obtained by sampling the latent space for the mean μ and standard deviation σ of the latent variables. The formula is: z = μ + σ⊙∈, where ⊙ represents the multiplication operation between elements. For two vectors, the result of their multiplication ⊙ is a new staggered vector, ∈~N(0, I) standard normal distribution;

[0026] Furthermore, when sampling from the distribution of latent variables, sampling is performed through the reparameterization technique, which solves the non-differentiable problem when sampling from a parameterized distribution, allowing the model to be trained through backpropagation and effectively learn the distribution of latent variables and reconstruct data. The reparameterization technique is used to sample from the distribution of latent variables, ensuring that latent variables can be reasonably generated when processing data with uncertainty. For example, when there is a certain randomness in warehouse management data (such as random fluctuations in the time of goods entering and leaving the warehouse), the reparameterization technique can better handle this uncertainty.

[0027] The steps to obtain the initial reconstructed sequence by optimizing the decoder part of the variational autoencoder based on the latent feature variables are:

[0028] The decoder part based on the potential feature variables by optimizing the variational autoencoder includes two decoder layers and an output layer;

[0029] The decoding process of the first layer decoder:

[0030] Map the latent feature variable z to a new intermediate representation The Swish activation function is introduced. The activation function formula is expressed as: F(x) = x*f(x), f(x) is the activation function of the encoding layer, x represents the input data, and the decoding formula of the first layer decoder is expressed as: in, represents the weight matrix of the first layer decoder, represents the bias term of the first layer decoder, and F represents the Swish activation function;

[0031] The decoding process of the second layer decoder:

[0032] The decoding result of the first layer decoder is further mapped to obtain the second layer decoder result, which is expressed as: in, represents the weight matrix of the second layer decoder, represents the bias term of the second layer decoder, the intermediate representation is the decoding result of the second layer decoder;

[0033] The output layer of the optimized variational autoencoder is expressed as:

[0034] The intermediate representation Perform the initial calculation through the output layer and add a loss function with a regularization parameter λ: Among them, L represents the initial loss function, Represents the output layer weight matrix The sum of all elements of , where i is the index;

[0035] The output layer weight matrix To update, use the update formula: in, Represents the output layer weight matrix The sign function of hour, when hour, when hour,

[0036] Update the output layer weight matrix through a learning rate a It is expressed as: in, is the updated weight matrix;

[0037] Update the bias of the output layer by the learning rate a It is expressed as: in, is the updated bias;

[0038] Finally, the output result of the output layer is obtained: Among them, f3 is the activation function of the output layer, is the initial reconstructed sequence output by the output layer.

[0039] Furthermore, the multi-layer architecture can abstract the features of the input data layer by layer, better capture the complex structure and potential features in the data on the encoder side, and reconstruct the data more accurately on the decoder side. Regularization helps prevent the autoencoder from overfitting. By limiting the size of the weights, the model fits the training data more smoothly. The optimized multi-layer decoder can gradually recover the data from the low-dimensional latent space. Each layer of the decoder is trying to reconstruct different aspects of the data, and ultimately converts the latent variables into reconstructed data similar to the original data. For example, when managing data in a warehouse, the decoder can reconstruct reasonable information such as the quantity, specifications, batches, and time of goods based on the latent variables, avoiding the information loss and reconstruction errors that may occur in the direct conversion from high-dimensional to low-dimensional and then to high-dimensional.

[0040] The steps of using the initial reconstructed sequence as a generator for generating a hybrid adversarial network are:

[0041] The generative adversarial hybrid network consists of a discriminator and a generator;

[0042] The initial reconstructed sequence obtained by optimizing the variational autoencoder As input to the generator in a generative adversarial hybrid network;

[0043] The generator is an n-layer generator, and the input of the first layer generator is the initial reconstruction sequence The output after transformation through a neural network layer is expressed as Where fm represents the activation function of the mth (m≤n) layer generator;

[0044] Add a residual connection in the generator, expressed as For the mth layer, the initial reconstructed sequence as the output of the generator is represented as: in, represents the generator output of the m-1 layer, i.e. the previous layer, represents the weight matrix of the mth layer of the generator, Represents the bias term of the mth layer of the generator.

[0045] Furthermore, in warehouse management data processing, due to the high complexity and diversity of the data, deep neural networks will face training difficulties. The application of residual connections can improve the training efficiency and stability of the network, thereby better processing warehouse data.

[0046] The steps of jointly training and optimizing the generative adversarial hybrid network through a gradient penalty to obtain the true reconstruction sequence are:

[0047] Define the discriminator of the generated adversarial hybrid network as D, and the input data of the discriminator D is (refers to the output of the generator), the output data is On input and output Random interpolation in is expressed as: Among them, ∈~U(0,1) is uniformly distributed;

[0048] Let the discriminator be D pairs of random interpolation The gradient of Calculate the second norm of the gradient, square the difference between it and 1, and then multiply it by the weight parameter ε. The gradient penalty term is: in, Represents input Output and the expectation of ∈;

[0049] Furthermore, the gradient The steps to obtain are:

[0050] When the discriminator gets the input and output After the discriminant result of random interpolation in the discriminator, the discriminator framework automatically records the calculation and then obtains it through back propagation The gradient of

[0051] The steps to obtain are:

[0052] For all possible inputs Output And ∈ average, multiple sampling of different inputs Output And ∈, calculate each time Then find the average value, and the final value is

[0053] Based on gradient penalty Joint training optimizes the generation of adversarial hybrid networks. The overall objective function of joint training is expressed as: in, represents the loss term of the discriminator D, ε D is the weight parameter of the loss term of the discriminator D, represents the loss term of the generator, ε H is the weight parameter of the loss term of the generator D;

[0054] Furthermore, the loss term of the generator The steps to obtain are:

[0055] Based on the loss term formula: in, Represents the data input through the generator Output probability after passing through the discriminator;

[0056] The loss term of the discriminator D The steps to obtain are:

[0057] Based on the loss term formula: in, Represents the output of the real warehouse data after passing through the discriminator, Indicates that the input data is (represents the output of the generator) The output probability after passing through the discriminator, Q is the number of data points;

[0058] The overall objective function of this joint training is used to converge the optimized generative adversarial hybrid network. The overall objective function is used to make the generative adversarial hybrid network continuously converge, and finally an optimized generative hybrid adversarial network is obtained. The initial reconstructed sequence is input into the optimized generative hybrid adversarial network. The output of the generator of the optimized generative hybrid adversarial network is the true reconstructed sequence.

[0059] Furthermore, in this process, the training of the discriminator is stabilized by introducing a gradient penalty term to prevent the discriminator from being too powerful and causing the generator to fail to converge (mode collapse). The overall objective function of the joint training comprehensively considers the losses of the generator and the discriminator. The generator and the discriminator are converged by alternating training loss terms, so that the entire network can converge to a better state and finally obtain a true reconstructed sequence. This method can effectively optimize the data reconstruction process and improve the accuracy and reliability of the data when processing warehouse management data.

[0060] The steps of a series of behavior sequences for obtaining the latest stored data are:

[0061] The warehouse behavior data is collected by sensors in the warehouse, and the behavior data of each entry into the warehouse is recorded in chronological order to form a series of behavior sequences K = (k1, k2, k3, k n ), where k n Indicates the behavior data of the nth storage operation.

[0062] In warehouse management, behavioral data includes various operational behaviors related to the warehousing of goods, such as the transportation path of goods (the route from the warehouse entrance to the storage location), the transportation equipment used (forklifts, carts, etc.), the operator's operating steps (such as scanning the goods barcode, placing goods on the shelf, etc.), etc. These data are sorted and recorded in chronological order to form a behavioral sequence.

[0063] The steps of inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model to obtain anomaly scores are:

[0064] Define the true reconstruction sequence output by the optimized generative adversarial hybrid network as

[0065] The multimodal deep Boltzmann machine model consists of a visible layer (including nodes of behavior sequences and true reconstruction sequences) and multiple hidden layers;

[0066] Furthermore, the visible layer of the multimodal deep Boltzmann machine is composed of a behavior sequence and a true reconstruction sequence, and the elements in these sequences constitute the node state V of the visible layer;

[0067] The model is trained by contrastive divergence algorithm, the visible layer node bias, hidden layer node bias and connection weight are randomly initialized, and a visible layer node V corresponding to a behavior sequence and a true reconstruction sequence is selected. x As the initial state, start from the initial state and conduct model training;

[0068] For new behavior sequences and true reconstruction sequences, calculate their probabilities under the model:

[0069] First, the hidden layer is sampled through a conditional probability: Each hidden layer node is sampled, and the collected hidden layer nodes are represented as

[0070] The gradient of the bias term and weight is calculated, and the model parameters are updated through the obtained gradient. The hidden layer sampling, gradient calculation, and parameter update are repeated for all training samples until the model converges. After the model converges, a trained multimodal deep Boltzmann machine model is obtained.

[0071] Furthermore, the gradient calculation is implemented by contrastive divergence algorithm. First, a visible layer node state corresponding to a behavior sequence and a true reconstruction sequence is selected from the training data as the initial state. The hidden layer state is sampled according to the current visible layer state, and the new visible layer state is sampled according to the sampled hidden layer state.

[0072] The anomaly score is calculated based on the energy function of the trained multimodal deep Boltzmann machine model. The calculation formula is:

[0073] Among them, the formula for obtaining the energy function is: Among them, j represents the node index of the visible layer, represents the hidden layer index, Represents the jth node V of the visible layer j With The product of the biases, summed over all visible layer nodes, is Represents the hidden layer Nodes and their biases The product of all hidden layer nodes is Indicates that the visible layer nodes are weighted With hidden layer nodes The energy term generated by the interaction is the sum of all pairwise combinations of visible and hidden layer nodes.

[0074] The anomaly score is determined by the probability obtained by the above calculation. The anomaly score is inversely proportional to the probability, that is, the lower the probability, the higher the anomaly score.

[0075] Warehouse management system, including:

[0076] Warehouse data collection module: collect warehouse management data and extract feature vectors of warehouse management data;

[0077] Initial sequence acquisition module: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables;

[0078] Real sequence acquisition module: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the real reconstructed sequence;

[0079] Sequential pattern decision module: obtains a series of behavior sequences of the latest incoming data, and obtains the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it indicates that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

[0080] The present invention has the following beneficial effects:

[0081] 1. In the present invention, starting from data collection, detailed information of goods is obtained through warehouse scanning equipment to form a feature vector, covering quantity, specifications, batches and time, etc., providing a comprehensive and detailed data basis for subsequent processing, and then encoding and reconstructing the feature vector by optimizing the variational autoencoder. Its multi-layer structure and special encoding and decoding mechanism can deeply explore the potential structure and complex relationship in the data and capture uncertainty factors. For example, when processing the data of goods in and out of the warehouse, the potential inventory change patterns between different types of goods and different batches can be found, which provides the possibility for more accurate data analysis. Then, the reconstructed data is further optimized by generating adversarial hybrid networks. The multi-layer structure and residual connection of the generator enable it to better learn data features and generate high-quality reconstructed data. The discriminator and the generator communicate with each other. Gradient penalty joint training ensures the authenticity and reliability of generated data, improves the accuracy and stability of data reconstruction, and makes the reconstructed data more in line with actual business conditions. Finally, the multimodal deep Boltzmann machine is used to fuse the behavior sequence and the real reconstruction sequence, fully considers the multimodal information, and trains the model through the contrast divergence algorithm. The probability is calculated to obtain the abnormal score, which can more accurately determine whether the input cargo data is abnormal. Whether it is human operation errors, such as data entry errors, or other complex reasons that cause mismatches in warehouse data, they can be effectively detected, greatly reducing the risk of data anomalies being missed. The entire warehouse management system forms an organic whole, and each link is closely connected and works together. From data collection to the final abnormality judgment, each step has corresponding technical means to ensure data quality and processing effects. Improve the accuracy and reliability of warehouse management. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A method step diagram of the warehouse management method and system proposed by the present invention;

[0083] Figure 2 This is a system block diagram of the warehouse management method and system proposed in the present invention. DETAILED DESCRIPTION

[0084] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0085] Embodiment 1

[0086] like Figure 1 to Figure 2 As shown, the warehouse management method proposed by the present invention includes:

[0087] S1: Collect warehouse management data and extract feature vectors of warehouse management data;

[0088] S2: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables;

[0089] S3: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the true reconstructed sequence;

[0090] S4: Obtain a series of behavior sequences of the latest incoming data, and obtain the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it means that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

[0091] The steps to obtain the feature vector of warehouse management data are as follows:

[0092] When goods enter the warehouse, the quantity and specification information of the goods are obtained through the warehouse scanning equipment, and the entry time and batch are recorded at the same time to form a complete feature vector, expressed as X=(Q, S, B, T), where Q is the quantity, S is the specification, B is the batch, and T is the time.

[0093] Furthermore, this vector form of data can effectively reflect the details of warehouse operations. Warehouse management data contains various key information about the goods entering and leaving the warehouse, including quantity, specifications, batches and time, which can be represented in vector form, providing a basis for subsequent data processing and comprehensively capturing the actual situation of warehouse operations.

[0094] The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain the potential feature variables in the following steps:

[0095] The vector feature X is used as the input of an autoencoder, and the autoencoder is optimized to a multi-layer autoencoder structure to obtain an optimized variational autoencoder;

[0096] The input data of the input layer of the optimized variational autoencoder is X;

[0097] The encoder layer of the optimized variational autoencoder consists of two encoders, and the decoder layer consists of two decoders;

[0098] The encoding process of the first layer encoder:

[0099] The input data X is input to the first layer encoder, which maps X to an intermediate representation through a neural network: Among them, W1 represents the weight matrix of the neural network in the first layer, b1 represents the bias vector of the neural network in the first layer, and f1 is the activation function of the neural network in the first layer. This intermediate representation is the encoding result of the first layer encoder;

[0100] The encoding process of the second layer encoder:

[0101] The intermediate representation Input to the second layer encoder, the second layer encoder will It is further mapped to the mean μ and standard deviation σ of the latent variable;

[0102] The calculation formula for the mean μ of the latent variable is: Wherein, W2 represents the weight matrix of the neural network in the second layer, b2 represents the bias vector of the neural network in the second layer, and f2 is the activation function used to calculate the mean in the second layer;

[0103] The calculation formula of the latent variable standard deviation σ is: in, The activation function used to calculate the standard deviation in the second layer;

[0104] In this way, the model is able to sample samples from the latent space that conform to a specific distribution, which helps capture the uncertainty and latent structure of the data during the encoding process;

[0105] The latent feature variables are obtained by sampling the latent space for the mean μ and standard deviation σ of the latent variables. The formula is: z = μ + σ⊙∈, where ⊙ represents the multiplication operation between elements. For two vectors, the result of their multiplication ⊙ is a new staggered vector, ∈~N(0, I) standard normal distribution;

[0106] Furthermore, when sampling from the distribution of latent variables, sampling is performed through the reparameterization technique, which solves the non-differentiable problem when sampling from a parameterized distribution. Through the reparameterization technique, the sampling operation is decomposed into two parts: one part is the deterministic calculation determined by the neural network parameters (calculation of the mean μ and the standard deviation σ), and the other part is sampling from a fixed standard normal distribution (sampling of ∈). The parameters in the encoder can be smoothly updated through back propagation, so that the model can be trained through back propagation and can effectively learn the distribution of latent variables and reconstruct data. The reparameterization technique is used to sample from the latent variable distribution, ensuring that the latent variables can be reasonably generated when processing data with uncertainty. For example, when there is a certain randomness in warehouse management data (such as random fluctuations in the time of goods entering and leaving the warehouse), the reparameterization technique can better handle this uncertainty.

[0107] The steps to obtain the initial reconstructed sequence by optimizing the decoder part of the variational autoencoder based on the latent feature variables are:

[0108] The decoder part of the variational autoencoder based on the latent feature variables by optimizing it includes two decoder layers and an output layer;

[0109] The decoding process of the first layer decoder:

[0110] Map the latent feature variable z to a new intermediate representation The Swish activation function is introduced. The activation function formula is expressed as: F(x) = x*f(x), f(x) is the activation function of the encoding layer, x represents the input data, and the decoding formula of the first layer decoder is expressed as: in, represents the weight matrix of the first layer decoder, represents the bias term of the first layer decoder, and F represents the Swish activation function;

[0111] The decoding process of the second layer decoder:

[0112] The decoding result of the first layer decoder is further mapped to obtain the second layer decoder result, which is expressed as: in, represents the weight matrix of the second layer decoder, represents the bias term of the second layer decoder, the intermediate representation is the decoding result of the second layer decoder;

[0113] The output layer of the optimized variational autoencoder is expressed as:

[0114] The intermediate representation Perform the initial calculation through the output layer and add a loss function with a regularization parameter λ: Among them, L represents the initial loss function, Represents the output layer weight matrix The sum of all elements of , where i is the index;

[0115] The output layer weight matrix To update, use the update formula: in, Represents the output layer weight matrix The sign function of hour, when hour, when hour,

[0116] Update the output layer weight matrix through a learning rate a It is expressed as: in, is the updated weight matrix;

[0117] Update the bias of the output layer by the learning rate a It is expressed as: in, is the updated bias;

[0118] Finally, the output result of the output layer is obtained: Among them, f3 is the activation function of the output layer, is the initial reconstructed sequence output by the output layer.

[0119] Furthermore, the multi-layer architecture can abstract the features of the input data layer by layer, better capture the complex structure and potential features in the data on the encoder side, and reconstruct the data more accurately on the decoder side. Regularization helps prevent the autoencoder from overfitting. By limiting the size of the weights, the model fits the training data more smoothly. The optimized multi-layer decoder can gradually recover the data from the low-dimensional latent space. Each layer of the decoder is trying to reconstruct different aspects of the data, and ultimately converts the latent variables into reconstructed data similar to the original data. For example, when managing data in a warehouse, the decoder can reconstruct reasonable information such as the quantity, specifications, batches, and time of goods based on the latent variables, avoiding the information loss and reconstruction errors that may occur in the direct conversion from high-dimensional to low-dimensional and then to high-dimensional.

[0120] The steps to use the initial reconstructed sequence as the generator of the Generative Adversarial Hybrid Network are:

[0121] The generative adversarial hybrid network consists of a discriminator and a generator;

[0122] The initial reconstructed sequence obtained by optimizing the variational autoencoder As input to the generator in a generative adversarial hybrid network;

[0123] The generator is an n-layer generator, and the input of the first layer generator is the initial reconstruction sequence The output after transformation through a neural network layer is expressed as Where fm represents the activation function of the mth (m≤n) layer generator;

[0124] Add a residual connection in the generator, expressed as For the mth layer, the initial reconstructed sequence as the output of the generator is represented as: in, represents the generator output of the m-1 layer, i.e. the previous layer, represents the weight matrix of the mth layer of the generator, Represents the bias term of the mth layer of the generator.

[0125] Furthermore, in warehouse management data processing, due to the high complexity and diversity of the data, deep neural networks will face training difficulties. The application of residual connections can improve the training efficiency and stability of the network, thereby better processing warehouse data.

[0126] The steps to obtain the true reconstruction sequence are as follows:

[0127] Define the discriminator of the generated adversarial hybrid network as D, and the input data of the discriminator D is (refers to the output of the generator), the output data is On input and output Random interpolation in is expressed as: Among them, ∈~U(0,1) is uniformly distributed;

[0128] Let the discriminator be D pairs of random interpolation The gradient of Calculate the second norm of the gradient, square the difference between it and 1, and then multiply it by the weight parameter ε. The gradient penalty term is: in, Represents input Output and the expectation of ∈;

[0129] Furthermore, the gradient The steps to obtain are:

[0130] When the discriminator gets the input and output After the discriminant result of random interpolation in the discriminator, the discriminator framework automatically records the calculation and then obtains it through back propagation The gradient of

[0131] The steps to obtain are:

[0132] For all possible inputs Output And ∈ average, multiple sampling of different inputs Output And ∈, calculate each time Then find the average value, and the final value is

[0133] Based on gradient penalty Joint training optimizes the generation of adversarial hybrid networks. The overall objective function of joint training is expressed as: in, represents the loss term of the discriminator D, εD is the weight parameter of the loss term of the discriminator D, represents the loss term of the generator, ε H is the weight parameter of the loss term of the generator D;

[0134] Furthermore, the loss term of the generator The steps to obtain are:

[0135] Based on the loss term formula: in, Represents the data input through the generator Output probability after passing through the discriminator;

[0136] The loss term of the discriminator D The steps to obtain are:

[0137] Based on the loss term formula: in, Represents the output of the real warehouse data after passing through the discriminator, Indicates that the input data is (represents the output of the generator) The output probability after passing through the discriminator, Q is the number of data points;

[0138] The overall objective function of this joint training is used to converge the optimized generative adversarial hybrid network. The overall objective function is used to make the generative adversarial hybrid network continuously converge, and finally an optimized generative hybrid adversarial network is obtained. The initial reconstructed sequence is input into the optimized generative hybrid adversarial network. The output of the generator of the optimized generative hybrid adversarial network is the true reconstructed sequence.

[0139] Furthermore, in this process, the training of the discriminator is stabilized by introducing a gradient penalty term to prevent the discriminator from being too powerful and causing the generator to fail to converge (mode collapse). The overall objective function of the joint training comprehensively considers the losses of the generator and the discriminator. The generator and the discriminator are converged by alternating training loss terms, so that the entire network can converge to a better state and finally obtain a true reconstructed sequence. This method can effectively optimize the data reconstruction process and improve the accuracy and reliability of the data when processing warehouse management data.

[0140] The steps for obtaining the latest data in the database are as follows:

[0141] The warehouse behavior data is collected by sensors in the warehouse, and the behavior data of each entry into the warehouse is recorded in chronological order to form a series of behavior sequences K = (k1, k2, k3, k n ), where k n Indicates the behavior data of the nth storage operation.

[0142] In warehouse management, behavioral data includes various operational behaviors related to the warehousing of goods, such as the transportation path of goods (the route from the warehouse entrance to the storage location), the transportation equipment used (forklifts, carts, etc.), the operator's operating steps (such as scanning the goods barcode, placing goods on the shelf, etc.), etc. These data are sorted and recorded in chronological order to form a behavioral sequence.

[0143] The steps of inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model to obtain the anomaly score are:

[0144] The true reconstruction sequence output by the optimized generative adversarial hybrid network is defined as

[0145] The multimodal deep Boltzmann machine model consists of a visible layer (nodes containing behavior sequences and true reconstruction sequences) and multiple hidden layers;

[0146] Furthermore, the visible layer of the multimodal deep Boltzmann machine is composed of a behavior sequence and a true reconstruction sequence, and the elements in these sequences constitute the node state V of the visible layer;

[0147] The model is trained by contrastive divergence algorithm, the visible layer node bias, hidden layer node bias and connection weight are randomly initialized, and a visible layer node V corresponding to a behavior sequence and a true reconstruction sequence is selected. x As the initial state, start from the initial state and conduct model training;

[0148] For new behavior sequences and true reconstruction sequences, calculate their probabilities under the model:

[0149] First, the hidden layer is sampled through a conditional probability: Each hidden layer node is sampled, and the collected hidden layer nodes are represented as

[0150] The gradient of the bias term and weight is calculated, and the model parameters are updated through the obtained gradient. The hidden layer sampling, gradient calculation, and parameter update are repeated for all training samples until the model converges. After the model converges, a trained multimodal deep Boltzmann machine model is obtained.

[0151] Furthermore, the gradient calculation is implemented by contrastive divergence algorithm. First, a visible layer node state corresponding to a behavior sequence and a true reconstruction sequence is selected from the training data as the initial state. The hidden layer state is sampled according to the current visible layer state, and the new visible layer state is sampled according to the sampled hidden layer state.

[0152] The anomaly score is calculated based on the energy function of the trained multimodal deep Boltzmann machine model. The calculation formula is:

[0153] Among them, the formula for obtaining the energy function is: Among them, j represents the node index of the visible layer, represents the hidden layer index, Represents the jth node V of the visible layer j With The product of the biases, summed over all visible layer nodes, is Represents the hidden layer Nodes and their biases The product of all hidden layer nodes is Indicates that the visible layer nodes are weighted With hidden layer nodes The energy term generated by the interaction is the sum of all pairwise combinations of visible and hidden layer nodes.

[0154] The anomaly score is determined by the probability obtained by the above calculation. The anomaly score is inversely proportional to the probability, that is, the lower the probability, the higher the anomaly score.

[0155] In this embodiment, firstly, the quantity, specification, batch and time information of the goods are obtained through the warehouse scanning equipment to form a feature vector. This vector form covers the key information of the goods in and out of the warehouse, and provides a comprehensive and detailed data basis for subsequent processing. Then, the data is deeply mined through the optimized variational autoencoder. The optimized variational autoencoder adopts a multi-layer structure. The encoder maps the feature vector to the latent space, calculates the mean and standard deviation of the latent variables, and uses the reparameterization technique to sample to obtain the latent feature variables, which helps to deeply mine the potential structure and complex relationships in the data and capture uncertain factors, such as discovering the inventory change patterns between different types of goods and batches. Then, the data quality is improved by generating an adversarial hybrid network. The generative adversarial hybrid network The generator uses a multi-layer structure and residual connection, takes the initial reconstruction sequence as input, learns data features, and generates high-quality reconstructed data. The discriminator and generator are jointly trained through gradient penalty to ensure the authenticity and reliability of the generated data. The multimodal deep Boltzmann machine model is used to judge anomalies. The visible layer of the multimodal deep Boltzmann machine model fuses the behavior sequence and the true reconstruction sequence, fully considering the multimodal information. The model is trained through the contrast divergence algorithm, and the probability is calculated to obtain the anomaly score. It can accurately determine whether the input cargo data is abnormal. Whether it is human operation errors (such as data entry errors) or other complex reasons that cause mismatches in inbound and outbound data, it can be effectively detected, greatly reducing the risk of missing data anomalies and improving the accuracy and reliability of warehouse management.

[0156] Embodiment 2

[0157] like Figure 1 to Figure 2 As shown, based on the first embodiment, the warehouse management system includes:

[0158] Warehouse data collection module: collect warehouse management data and extract feature vectors of warehouse management data;

[0159] Initial sequence acquisition module: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables;

[0160] Real sequence acquisition module: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the real reconstructed sequence;

[0161] Sequential pattern decision module: obtains a series of behavior sequences of the latest incoming data, and obtains the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it indicates that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

[0162] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A warehouse management method, characterized in that: include: S1: Collect warehouse management data and extract feature vectors of warehouse management data; S2: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables; S3: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the true reconstructed sequence; S4: Obtain a series of behavior sequences of the latest incoming data, and obtain the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it means that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

2. The warehouse management method according to claim 1, characterized in that: The steps for obtaining the feature vector of the warehouse management data are as follows: When goods enter the warehouse, the quantity and specification information of the goods are obtained through the warehouse scanning equipment, and the entry time and batch are recorded at the same time to form a complete feature vector, expressed as X=(Q, S, B, T), where Q is the quantity, S is the specification, B is the batch, and T is the time.

3. The warehouse management method according to claim 1, characterized in that: The step of encoding and reconstructing the collected feature vectors through an optimized variational autoencoder to obtain potential feature variables is as follows: The vector feature X is used as the input of an autoencoder, and the autoencoder is optimized to a multi-layer autoencoder structure to obtain an optimized variational autoencoder; The input data of the input layer of the optimized variational autoencoder is X; The encoder layer of the optimized variational autoencoder is composed of two encoders, and the decoder layer is composed of two decoders; The encoding process of the first layer encoder: The input data X is input to the first layer encoder, which maps X to an intermediate representation through a neural network: Among them, W1 represents the weight matrix of the neural network in the first layer, b1 represents the bias vector of the neural network in the first layer, and f1 is the activation function of the neural network in the first layer. This intermediate representation is the encoding result of the first layer encoder; The encoding process of the second layer encoder: The intermediate representation Input to the second layer encoder, the second layer encoder will It is further mapped to the mean μ and standard deviation σ of the latent variable; The calculation formula for the mean μ of the latent variable is: Wherein, W2 represents the weight matrix of the neural network in the second layer, b2 represents the bias vector of the neural network in the second layer, and f2 is the activation function used to calculate the mean in the second layer; The calculation formula of the latent variable standard deviation σ is: in, The activation function used to calculate the standard deviation in the second layer; The latent space sampling of the mean μ and standard deviation σ of the latent variable is used to obtain the latent feature variable, which is expressed as: z=μ+σ⊙∈, where ⊙ represents the multiplication operation between elements. For two vectors, the result of their multiplication ⊙ is a new interleaved vector, ∈~N(0,I) standard normal distribution.

4. The warehouse management method according to claim 1, characterized in that: The step of obtaining the initial reconstructed sequence by optimizing the decoder part of the variational autoencoder based on the potential feature variables is: The decoder part based on the potential feature variables by optimizing the variational autoencoder includes two decoder layers and an output layer; The decoding process of the first layer decoder: Map the latent feature variable z to a new intermediate representation The Swish activation function is introduced. The activation function formula is expressed as: F(x) = x*f(x), f(x) is the activation function of the encoding layer, x represents the input data, and the decoding formula of the first layer decoder is expressed as: in, represents the weight matrix of the first layer decoder, represents the bias term of the first layer decoder, and F represents the Swish activation function; The decoding process of the second layer decoder: The decoding result of the first layer decoder is further mapped to obtain the second layer decoder result, which is expressed as: in, represents the weight matrix of the second layer decoder, represents the bias term of the second layer decoder, the intermediate representation is the decoding result of the second layer decoder; The output layer of the optimized variational autoencoder is expressed as: The intermediate representation Perform the initial calculation through the output layer and add a loss function with a regularization parameter λ: Among them, L represents the initial loss function, Represents the output layer weight matrix The sum of all elements of , where i is the index; The output layer weight matrix To update, use the update formula: in, Represents the output layer weight matrix The sign function of hour, when hour, when hour, Update the output layer weight matrix through a learning rate α It is expressed as: in, is the updated weight matrix; Update the bias of the output layer by the learning rate α It is expressed as: in, is the updated bias; Finally, the output result of the output layer is obtained: Among them, f3 is the activation function of the output layer, is the initial reconstructed sequence output by the output layer.

5. The warehouse management method according to claim 1, characterized in that: The steps of using the initial reconstructed sequence as a generator for generating a hybrid adversarial network are: The generative adversarial hybrid network consists of a discriminator and a generator; The initial reconstructed sequence obtained by optimizing the variational autoencoder As input to the generator in a generative adversarial hybrid network; The generator is an n-layer generator, and the input of the first layer generator is the initial reconstruction sequence The output after transformation through a neural network layer is expressed as where f m represents the activation function of the mth (m≤n) layer generator; Add a residual connection in the generator, expressed as For the mth layer, the initial reconstructed sequence as the output of the generator is represented as: in, represents the generator output of the m-1 layer, i.e. the previous layer, represents the weight matrix of the mth layer of the generator, Represents the bias term of the mth layer of the generator.

6. The warehouse management method according to claim 1, characterized in that: The steps of jointly training and optimizing the generative adversarial hybrid network through a gradient penalty to obtain the true reconstruction sequence are: Define the discriminator of the generated adversarial hybrid network as D, and the input data of the discriminator D is (refers to the output of the generator), the output data is On input and output Random interpolation in is expressed as: Among them, ∈~U(0,1) is uniformly distributed; Let the discriminator be D pairs of random interpolation The gradient of Calculate the second norm of the gradient, square the difference between it and 1, and then multiply it by the weight parameter ε. The gradient penalty term is: in, Represents input Output and the expectation of ∈; The gradient The steps to obtain are: When the discriminator gets the input and output After the discriminant result of random interpolation in the discriminator, the discriminator framework automatically records the calculation and then obtains it through back propagation The gradient of Said The steps to obtain are: For all possible inputs Output And ∈ average, multiple sampling of different inputs Output And ∈, calculate each time Then find the average value, and the final value is Based on gradient penalty Joint training optimizes the generation of adversarial hybrid networks. The overall objective function of joint training is expressed as: in, represents the loss term of the discriminator D, ε D is the weight parameter of the loss term of the discriminator D, represents the loss term of the generator, ε H is the weight parameter of the loss term of the generator D; The loss term of the generator The steps to obtain are: Based on the loss term formula: in, Represents the data input through the generator Output probability after passing through the discriminator; The loss term of the discriminator D The steps to obtain are: Based on the loss term formula: in, Represents the output of the real warehouse data after passing through the discriminator, Indicates that the input data is (represents the output of the generator) The output probability after passing through the discriminator, Q is the number of data points; The overall objective function of this joint training is used to converge the optimized generative adversarial hybrid network. The overall objective function is used to make the generative adversarial hybrid network continuously converge, and finally an optimized generative hybrid adversarial network is obtained. The initial reconstructed sequence is input into the optimized generative hybrid adversarial network. The output of the generator of the optimized generative hybrid adversarial network is the true reconstructed sequence.

7. The warehouse management method according to claim 1, characterized in that: The steps of a series of behavior sequences for obtaining the latest stored data are: The warehouse behavior data is collected by sensors in the warehouse, and the behavior data of each entry into the warehouse is recorded in chronological order to form a series of behavior sequences K = (k1, k2, k3, k n ), where k n Indicates the behavior data of the nth storage operation.

8. The warehouse management method and system according to claim 1, characterized in that: The steps of inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model to obtain anomaly scores are: The true reconstruction sequence output by the optimized generative adversarial hybrid network is defined as The multimodal deep Boltzmann machine model consists of a visible layer and multiple hidden layers; The visible layer of the multimodal deep Boltzmann machine consists of a behavior sequence and a true reconstruction sequence. The elements in these sequences constitute the node state V of the visible layer. The model is trained by contrastive divergence algorithm, the visible layer node bias, hidden layer node bias and connection weight are randomly initialized, and a visible layer node V corresponding to a behavior sequence and a true reconstruction sequence is selected. x As the initial state, start from the initial state and conduct model training; For new behavior sequences and true reconstruction sequences, calculate their probabilities under the model: First, the hidden layer is sampled through a conditional probability: Each hidden layer node is sampled, and the collected hidden layer nodes are represented as The gradient of the bias term and weight is calculated, and the model parameters are updated through the obtained gradient. The hidden layer sampling, gradient calculation, and parameter update are repeated for all training samples until the model converges. After the model converges, a trained multimodal deep Boltzmann machine model is obtained. The anomaly score is calculated based on the energy function of the trained multimodal deep Boltzmann machine model. The calculation formula is: Among them, the formula for obtaining the energy function is: Among them, j represents the node index of the visible layer, represents the hidden layer index, Represents the jth node V of the visible layer j With The product of the biases, summed over all visible layer nodes, is Represents the hidden layer Nodes and their biases The product of all hidden layer nodes is Indicates that the visible layer nodes are weighted With hidden layer nodes The energy term generated by the interaction is the sum of all pairwise combinations of visible and hidden layer nodes. The anomaly score is determined by the probability obtained by the above calculation. The anomaly score is inversely proportional to the probability, that is, the lower the probability, the higher the anomaly score.

9. A warehouse management system, which is implemented based on the warehouse management method according to any one of claims 1 to 8, characterized in that: include: Warehouse data collection module: collect warehouse management data and extract feature vectors of warehouse management data; Initial sequence acquisition module: The collected feature vectors are encoded and reconstructed through an optimized variational autoencoder to obtain potential feature variables, and the initial reconstructed sequence is obtained by optimizing the decoder part of the variational autoencoder based on the potential feature variables; Real sequence acquisition module: Through a generative adversarial hybrid network, the initial reconstructed sequence is used as the input of the generator of the generative adversarial hybrid network, and the generative adversarial hybrid network is jointly trained and optimized through a gradient penalty to obtain the real reconstructed sequence; Sequential pattern decision module: obtains a series of behavior sequences of the latest incoming data, and obtains the anomaly score by inputting the true reconstruction sequence and the behavior sequence into a multimodal deep Boltzmann machine model. When the anomaly score is greater than the anomaly threshold, it indicates that the entered cargo data is abnormal, and the corresponding incoming data is marked and a reminder is issued.

Images