Intelligent storage adaptive identification system based on deep learning

Abstract

The invention relates to the technical field of intelligent storage recognition, and particularly discloses an intelligent storage adaptive recognition system based on deep learning. According to the method, multi-source goods storage identification data are fused, a defect area is positioned and a goods defect value is identified through processing such as image difference, the goods sealing performance is judged by using a DTW algorithm, factors such as a storage environment are integrated to predict an actual residual period of efficacy of goods, and a goods storage state discrimination vector is constructed according to the goods defect value, the sealing performance and the actual residual period of efficacy. The cumulative probability of the goods storage state grade is calculated through a multivariate ordered Logistic regression model, the category with the maximum probability is selected as the storage state grade, and intelligent judgment of the goods storage state is achieved; meanwhile, corresponding strategies are executed based on different goods identification results, intelligence and automation of goods warehouse management are achieved, and the overall efficiency and reliability of warehouse operation are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehouse identification technology, specifically to an intelligent warehouse adaptive identification system based on deep learning. Background Technology

[0002] The safety and standardization of goods warehousing are of paramount importance, directly determining the quality stability and safety of goods. With the increasing variety of goods and the continuous expansion of warehousing scale, traditional goods warehousing management models are gradually failing to meet the demands of efficient and intelligent modern warehousing. Various pain points in warehousing identification and management are becoming increasingly prominent, severely hindering the standardized development of the goods warehousing industry.

[0003] Currently, the identification and management of goods in the warehousing process relies on a combination of manual inspection and traditional automated equipment, which suffers from numerous problems such as low identification accuracy, low efficiency, poor adaptability, and an imperfect management loop. In terms of identifying defects in goods storage, traditional methods mostly employ manual visual inspection or simple image comparison techniques. Manual inspection is affected by subjective factors such as the experience and fatigue of operators, easily leading to missed or false detections, especially in its insufficient ability to identify hidden defects such as minor scratches and small tears. Simple image comparison techniques lack in-depth mining of image features, failing to extract detailed features such as texture and color of defective areas, and also unable to quantify the degree of defect analysis, making it difficult to meet the needs of goods packaging defect identification.

[0004] In terms of product sealing identification, existing technologies mostly employ contact-based detection methods such as air pressure detection and vacuum detection. These methods are cumbersome to operate, have low detection efficiency, and are prone to damaging product packaging, making them unsuitable for large-scale, continuous warehousing inspection scenarios. While some non-contact detection methods use ultrasonic technology to collect sealing-related signals, they lack effective signal analysis algorithms. They cannot cope with issues such as signal timing distortion and noise interference caused by diverse product packaging specifications and complex storage environments. They struggle to identify sealing abnormalities such as leaks and incomplete seals, leading to substandard products entering subsequent stages and causing quality risks such as moisture absorption and spoilage.

[0005] In managing the remaining shelf life of goods, existing models largely rely on manual recording of the date of goods entering the warehouse and the expiration date, and calculate the remaining shelf life by simple time difference. This ignores the impact of factors such as fluctuations in the storage environment (temperature, humidity, light intensity), defects in product packaging, and sealing on the shelf life. In reality, uneven fluctuations in the storage environment, reduced protective capabilities due to packaging defects, and environmental corrosion caused by abnormal sealing all accelerate the depletion of the product's shelf life. This leads to a significant deviation between the remaining shelf life calculated based on pure time and the actual situation, making it easy for goods to deteriorate within their expiration date or for expired goods to be failed to be handled in a timely manner, posing safety hazards and economic losses to the management of goods storage.

[0006] Therefore, there is an urgent need for a deep learning-based intelligent warehouse adaptive recognition system to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based intelligent warehouse adaptive identification system to solve the problems mentioned in the background.

[0008] The objective of this invention can be achieved through the following technical solution: a deep learning-based intelligent warehouse adaptive identification system, comprising:

[0009] The multimodal data acquisition module collects image information of goods stored in the warehouse, seal identification information, and warehouse environment status information, and sends them to the defect identification module, seal identification module, and remaining shelf life identification module, respectively.

[0010] The defect identification module identifies defective areas of goods based on warehouse image information and analyzes defect values ​​based on these defective areas.

[0011] The sealing recognition module calculates the similarity between the sequence of goods to be detected and the standard sequence of goods through a dynamic time warping algorithm, and identifies the sealing of goods based on the similarity.

[0012] The remaining shelf life identification module predicts the consumed shelf life score of the goods based on the goods' storage environment status information, goods' defect value, and sealing performance, and then analyzes to obtain the actual remaining shelf life of the goods.

[0013] The identification output module constructs a goods storage status discrimination vector based on the goods defect value, sealing performance, and actual remaining shelf life. It then identifies the goods storage status level through a multivariate ordered logistic regression model to obtain the goods storage identification result.

[0014] The identification and management module implements corresponding strategies based on the identification results of the goods storage.

[0015] The beneficial effects of this invention are:

[0016] This invention addresses the shortcomings of existing technologies in terms of insufficient information dimensions and susceptibility to environmental interference by collecting three types of heterogeneous data: goods storage image information, ultrasonic sealing recognition information, and storage environment status information. This avoids large recognition errors caused by single, incomplete data sets, thereby improving the accuracy of storage status perception. The defect recognition module constructs an automated goods storage defect recognition system through differential image preprocessing, Gaussian blur denoising, lightweight convolutional neural network heatmap output, connected component analysis, and texture and color feature fusion. Compared to existing technologies that rely on human vision or traditional machine vision, this invention's deep learning algorithm can capture minute defect features. Through heatmap visualization and multi-feature analysis, it reduces the false negative and false positive rates, meeting the real-time detection needs of large-scale goods storage.

[0017] The sealing recognition module of this invention uses the DTW algorithm to process the similarity between the product sequence to be detected and the standard sequence. Through local cost matrix construction, cumulative distance matrix optimization, and path constraint efficiency improvement, it effectively solves the sequence alignment problem caused by time axis stretching and distortion. Compared with the destructive detection or single threshold judgment methods in the prior art, the non-contact detection of this invention does not require damage to the packaging and evaluates the sealing status in real time. It improves the sensitivity to anomalies such as small leaks and poor sealing to a certain extent, ensuring the integrity of product packaging and storage safety.

[0018] The remaining shelf life identification module of this invention measures the uneven distribution of various indicators in the storage environment data through a standard deviation model. It uses a weighted model to comprehensively measure the storage environment status of goods by combining environmental discrete values ​​and environmental deviation values, thereby obtaining an environmental impact value. This enables a quantitative assessment of the impact of the storage environment on the remaining shelf life of goods. Based on the environmental impact value, the defect value of the goods, and the sealing performance, a dynamic shelf life prediction model is constructed. This model overcomes the limitations of existing technologies that rely solely on the static calculation of shelf life based on the manufacturing date. The model considers the impact of environmental fluctuations and packaging integrity on the degradation rate of goods, making the shelf life prediction more consistent with the actual storage conditions. This effectively avoids premature product failure caused by various factors such as the storage environment, making the prediction of remaining shelf life more reasonable.

[0019] This invention's identification output module constructs a goods storage status discrimination vector based on the goods' defect value, sealing performance, and actual remaining shelf life. It then calculates the cumulative probability of each goods storage status level using a multivariate ordered logistic regression model and selects the category with the highest probability as the storage status level, thus achieving intelligent discrimination of goods storage status. Simultaneously, the identification management module executes corresponding strategies based on different goods identification results and intelligently plans picking or isolation paths, improving operational efficiency, reducing operational congestion, and realizing intelligent and automated goods storage management, thereby enhancing the overall efficiency and reliability of warehouse operations. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 As shown, this invention is a deep learning-based intelligent warehousing adaptive recognition system, specifically applied in goods warehousing recognition, including:

[0024] The multimodal data acquisition module collects images of goods stored in the warehouse, seal identification information, and warehouse environment status information, and sends these to the defect identification module, seal identification module, and remaining shelf life identification module, respectively; specifically as follows:

[0025] Smart cameras are used to collect images of goods stored in the warehouse. High-frequency sound waves are emitted to the goods via an ultrasonic transmitter, and the receiver captures the reflected signals and converts them into electrical signals, obtaining the time-domain waveform sequence of the reflected waves, i.e., the amplitude curve changing over time. It should be noted that well-sealed goods have a uniform internal structure, and the reflected wave waveform exhibits a stable and regular pattern. If there are leaks, incomplete seals, or other sealing abnormalities, sudden changes in acoustic impedance at the abnormal point will cause abnormal fluctuations in the amplitude and phase of the reflected wave. Therefore, the time-domain waveform directly carries the characteristic information of the sealing state. A sensor array is used to collect warehouse environmental status information, including temperature, humidity, and light intensity, and the data is cleaned and normalized.

[0026] The defect identification module identifies defective areas in goods based on warehouse storage images and analyzes defect values ​​based on these areas; specifically as follows:

[0027] The goods storage image and a normal reference image are preprocessed, including scaling them to the same size and normalizing the pixel values. Then, pixel-level difference operations are performed between the preprocessed goods storage image and the normal reference image to obtain a difference image. This difference image highlights the differences between the two images, i.e., potential defect areas. The difference operation formula is as follows: ,in, For difference images, Images of goods stored in the warehouse. This is a normal reference image. For pixels; Gaussian blur is applied to the difference image to obtain the processed difference image, which suppresses noise and pixel-level alignment errors. The Gaussian blur formula is: ,in, The Gaussian kernel is a parameter that controls the degree of blur.

[0028] The product storage image, normal reference image, and processed differential image are stitched together along the R, B, and G channels to obtain a three-channel input tensor. The stitched three-channel input tensor contains the overall appearance information of the product and details of defect changes. The three-channel input tensor is then fed into a lightweight convolutional neural network to output a defect probability heatmap. The heatmap visually displays the probability of a defect in each pixel of the image, with higher brightness intensity indicating a greater likelihood of a defect in that area.

[0029] The defect probability heatmap is upsampled using bilinear interpolation to upgrade the low-resolution heatmap to a high-resolution one, increasing image detail and clarity. A first threshold is then applied for binarization to obtain a preliminary defect localization mask. The binarization formula is as follows: ,in, This represents the probability value of the corresponding pixel in the defect probability heatmap. Indicates the first threshold. This indicates a preliminary defect location mask. This indicates an area where defects may exist. Indicates a normal area or a background area;

[0030] Connectivity analysis is performed on the initial defect location mask to extract all connected regions. A connected region is a region consisting of adjacent pixels with a value of 1 in the initial defect location mask. Through connectivity analysis, all possible defect regions in the initial defect location mask can be identified, providing a foundation for subsequent noise filtering and accurate defect location. Each connected region is enclosed by a minimum bounding rectangle and its area is extracted. A second threshold is set. If the area of ​​a connected region is less than the preset second threshold, it is judged as noise and removed; otherwise, it is judged as a candidate defect region. Based on this, a set of candidate defect regions is obtained.

[0031] Extract the area of ​​each defect region in the candidate defect region set and sum them up to obtain the total area of ​​the candidate defect regions. Calculate the ratio of the total area of ​​the candidate defect regions to the total area of ​​the region to obtain the defect area ratio.

[0032] Texture features, including contrast, entropy, and energy, are extracted from the candidate defect region set using local binary mode. Similarly, reference texture features, including reference contrast, reference entropy, and reference energy, are extracted from the normal reference image. The ratio of the reference contrast to the contrast of the candidate defect region set is calculated to generate contrast anomalies. Similarly, entropy anomalies and energy anomalies are obtained. The contrast anomalies, entropy anomalies, and energy anomalies of the candidate defect region set are summed to obtain the texture anomaly value of the candidate defect region set. It should be noted that the larger the contrast anomaly value, entropy anomaly value, and energy anomaly value of the candidate defect region, the larger the texture anomaly value, indicating that the texture difference between the corresponding regions of the candidate defect region and the normal reference image is greater, and the greater the degree of product defect.

[0033] Obtain the color features of the candidate defect region set, including RGB values, and construct a color histogram of the candidate defect region set. Similarly, construct a color histogram of the corresponding region of the normal reference image, and then use a color anomaly detection model. The color anomaly values ​​of the candidate defect region set were calculated. 'r' represents the warehouse number of the goods. This represents the histogram corresponding to the a-th region in the candidate defect region set. This represents the histogram corresponding to the a-th region in the set of candidate defect regions corresponding to the normal reference image. This represents the number of pixels corresponding to the b-th discrete color interval within the histogram of the a-th region in the candidate defect region set. This represents the number of pixels corresponding to the b-th discrete color interval within the histogram of the a-th region in the set of candidate defect regions of the normal reference image; in the formula, The term represents the L2 norm of the color histogram vector, reflecting the overall energy of the histogram. The dot product of two color histogram vectors represents the degree of similarity between the two color histograms. It indicates that the smaller the cosine similarity, the greater the color difference between the candidate defect region and the normal region.

[0034] Based on the defect area ratio, texture outlier, and color outlier of the candidate defect region set, the formula is used to... Calculate the defect value of the goods ,in, This indicates the defect area ratio and color outlier value of the candidate defect region. These are preset weighting factors for the defect values ​​of goods corresponding to the defect area ratio, texture outlier, and color outlier; for example, .

[0035] The airtightness identification module calculates the similarity between the product's test sequence and the product's standard sequence using the Dynamic Time Warping (DTW) algorithm, and identifies the product's airtightness based on the similarity level; specifically as follows:

[0036] Extract the time-domain waveform sequence of the reflected wave, and denote it as the sequence to be detected. , where n is the number of sampling points in the sequence to be detected, corresponding to the time series length of the sequence to be detected.

[0037] Obtain the time-domain waveform sequence of the reflected wave from the goods that have passed the sealing test, and denote it as the standard sequence. m represents the number of sampling points in the standard sequence, corresponding to the time series length of the standard sequence. The product test sequence and the product standard sequence are extracted, and the sequences are denoised by moving average filtering to eliminate random noise from the sensor and preserve the overall trend of the sequence. The filtered sequence is then normalized. After normalization, the numerical ranges of the standard sequence and the test sequence are consistent, making the subsequent DTW distance calculation more comparable.

[0038] Construct a local cost matrix C with dimensions n×m, where n is the length of the sequence to be detected and m is the length of the standard sequence. Each element in the local cost matrix C... The Euclidean distance between the amplitude at time point i of the sequence to be detected and the amplitude at time point j of the standard sequence is expressed by the following formula: ,in, This represents the amplitude at the i-th time point in the sequence to be detected after filtering and normalization. This represents the amplitude at the j-th time point in the standard sequence after filtering and normalization. It should be noted that the local cost matrix transforms the point-to-point differences between the two time series into matrix elements, laying the data foundation for subsequent dynamic programming to find the time alignment path with the minimum overall difference.

[0039] Construct the cumulative distance matrix Cumulative distance matrix The size is consistent with the local cost matrix C. First, initialize the boundary conditions and establish the recursion starting point, including the initialization of the first row (horizontal constraint) and the initialization of the first column (vertical constraint). Then, initialize the cumulative distance matrix. The first line has the following recursive rule: And the initial point satisfies This ensures that the first-row path extends only horizontally, avoiding temporal backtracking; and it applies the cumulative distance matrix. The recursive rule for the first column is: And the initial point satisfies This ensures that the first column path extends only in the vertical direction, conforming to the unidirectional nature of time series.

[0040] For the cumulative distance matrix Non-boundary points (i>1 and j>1) in the range. It is obtained by adding the minimum cumulative distances in the left, top, and top-left directions to the local cost matrix C, in order to ensure the rationality of the time sequence and avoid path intersections. Its recursive formula is: ;

[0041] Set the cumulative distance matrix Path constraint window width ,in For any point If the constraints are met If the value is true, then retain that point for recursion; otherwise, remove it. Let it be set to infinity and excluded from the recursive process; it should be noted that by using path constraints, the computational efficiency of large-scale detection and recognition is greatly improved, meaningless long-distance matching is avoided, and the rationality of the path is ensured.

[0042] It should be further explained that the cumulative distance matrix is ​​constructed in order to find the optimal alignment path between two sequences, taking into account the possible scaling or distortion of the sequence time axis, so as to accurately measure the similarity between them.

[0043] The cumulative distance matrix bottom right element The DTW distance between the test sequence and the standard sequence is denoted as . The DTW distance integrates all local differences between two sequences on the time axis and takes into account the time regularization factor, which can effectively reflect the similarity in trend between the sequence to be detected and the standard sequence.

[0044] Through formula The airtightness of the goods was calculated. , This represents the preset DTW distance threshold; the smaller the DTW distance, the more similar the trend of the sequence to be detected is to the standard sequence. The closer the value is to 1, the better the sealing of the goods; conversely, the closer it is to 1, the better the sealing of the goods. The closer it is to 0, the worse the sealing of the goods.

[0045] The remaining shelf life identification module, based on information about the product's storage environment, defect values, and sealing performance, predicts the product's consumed shelf life score and then analyzes it to obtain the actual remaining shelf life of the product; specifically as follows:

[0046] Acquire data on the warehouse environment status of goods from their entry into the warehouse to their current state. Use a standard deviation model to measure and calculate the uneven distribution of various indicators in the environmental status data from goods' entry into the warehouse to their current state, thus obtaining the discrete values ​​of the warehouse environment. The standard deviation model is as follows: ,in, These represent the mean values ​​of temperature, humidity, and light intensity, respectively. This represents temperature, humidity, and light intensity data, with f representing the data index. It should be noted that the greater the environmental fluctuations, the greater the dispersion of the storage environment, and the higher the uncertainty of product degradation.

[0047] Extract the standard warehousing environment status data corresponding to the product's expiration date, compare and analyze the warehousing environment status data from product entry to the present with the standard warehousing environment status data, and calculate the environmental deviation value. The greater the environmental parameters exceed the standards, the stronger the coercive effect on the goods.

[0048] It should be noted that the statistical method for environmental deviation values ​​is as follows:

[0049] When a certain indicator in the storage environment status data is greater than its corresponding standard data maximum value, the deviation value of that indicator is calculated as the difference between the corresponding value of the indicator and its corresponding standard data maximum value. Conversely, when a certain indicator in the storage environment status data is less than its corresponding standard data minimum value, the deviation value of that indicator is calculated as the difference between the standard data minimum value of the indicator and its corresponding value. Based on this analysis, the sum of temperature deviation values, the sum of humidity deviation values, and the sum of light intensity deviation values ​​in the storage environment status data are obtained, and then the total deviation value is obtained, which is recorded as the environmental deviation value.

[0050] Environmental discrete values and environmental deviation value A weighted model is used to comprehensively measure the storage environment of goods and obtain the environmental impact value. The weighted model is as follows: ,in, These are preset proportional coefficients, and ;

[0051] Based on environmental impact value, product defect value, and sealing performance, length conversion is performed according to a preset ratio. A circle is constructed with the sealing performance length as the radius, and a sector region one is constructed within the circle with the environmental impact value length as arc length one. A sector region two is constructed within the circle with the product defect value length as arc length two. The remaining arc area values ​​outside of sector regions one and two are extracted and denoted as... Through the consumed expiration date prediction model Calculate the consumed expiration period fraction It should be noted that using the area of ​​a geometric figure to intuitively quantify the coupling risk among the environment, defects, and sealing is a good approach. The larger the remaining arc area, the greater the deviation from the ideal storage state of the shelf life. A nonlinear mapping of the remaining arc area is performed using the hyperbolic tangent function, and the output of the consumed shelf life fraction ranges from (0,1). The larger the remaining arc area S, the higher the value of the fraction. A value closer to 1 indicates a higher proportion of the product's shelf life has been consumed; the smaller S is, the higher the percentage of the product's shelf life has been consumed. A value close to 0 indicates that almost no extra duration has been consumed.

[0052] Obtain the product's expiration date and the current date, and perform a difference calculation to obtain the product's baseline remaining validity period, denoted as [missing information]. Based on the product's remaining shelf life and the percentage of expired shelf life, the formula is used to determine the product's shelf life. The actual remaining validity period was calculated. .

[0053] The identification output module constructs a goods storage status discrimination vector based on the goods defect value, sealing performance, and actual remaining shelf life. It then identifies the goods storage status level using a multivariate ordered logistic regression model, yielding the goods storage identification result. Details are as follows:

[0054] Extract product defect values, sealing performance, and actual remaining shelf life to construct a product storage status discrimination vector. ;

[0055] Define the dependent variable This indicates the storage status level of goods and is a three-category variable. Indicates level one. Indicates level two. The three levels are represented by a multivariate ordered logistic regression model. The cumulative probability of the goods storage status level is calculated. These are the intercept parameters obtained from model fitting; different levels of k correspond to different intercepts. These are the feature weight vector and discriminant vector obtained from model fitting. The dot product reflects the degree of influence of three characteristics—defect value, sealing performance, and remaining shelf life—on the storage status level.

[0056] Through formula The probabilities of each storage status level are derived, and the category with the highest probability is selected as the storage status level of the goods; for example, if If it is the maximum of the three, then the storage status level of the goods is determined to be Level 1; if If it is the maximum of the three, then the storage status level of the goods is determined to be Level 2; if If the value is the highest among the three, then the storage status level of the goods is determined to be level three;

[0057] When the storage status level of goods is Level 1, a normal storage signal is generated and the goods are recorded as Level 1 goods; when the storage status level of goods is Level 2, a priority outbound signal is generated and the goods are recorded as Level 2 goods; when the storage status level of goods is Level 3, a recall isolation signal is generated and the goods are recorded as Level 3 goods; the priority outbound signal, Level 2 goods, recall isolation signal, and Level 3 goods are recorded as the goods storage identification result.

[0058] The identification and management module executes corresponding strategies based on the goods storage identification results; details are as follows:

[0059] When the product identification result is a priority outbound signal, the storage location distribution information of the secondary products is retrieved, the responsible picking personnel for each storage location are retrieved, and a storage location positioning command is sent to the mobile terminal of the responsible picking personnel to obtain their real-time location. At the same time, the status information of available AGVs in the storage area is retrieved, including the current battery level, current task progress, and load capacity of the AGVs. After filtering all available AGVs according to the conditions of battery level ≥80% and current task progress ≤20%, the AGV with the largest load capacity is selected and marked as the target AGV. Several picking paths are generated based on the secondary product storage location, picking port, and the current location of the target AGV. The path with the shortest path length and the lowest operation congestion rate is selected as the target picking path.

[0060] It should be noted that the analysis process for the congestion rate of the picking route is as follows: obtain the number of personnel, equipment, and goods stacking volume of several picking routes; extract the preset maximum number of personnel, maximum allowable number of equipment, and maximum temporary goods stacking volume of the area; calculate the personnel saturation, equipment saturation, and goods saturation; and weight these three saturations to obtain the congestion rate. If the personnel saturation, equipment saturation, and goods saturation of a picking route are higher, then the congestion rate of that picking route is higher.

[0061] The location distribution of secondary goods, target AGV numbers, and target picking routes are sent to the mobile devices of responsible picking personnel and the AGV scheduling system. After the responsible picking personnel and target AGVs work together to complete the picking of secondary goods, they send a picking completion instruction to the identification management module via their mobile devices. When the identification management module receives the picking completion instruction, it updates the inventory status of the secondary goods in real time to "picked and awaiting priority outbound", and simultaneously pushes the goods information to the order review system and marks it as secondary priority review. After the review is completed, the secondary goods are assigned to the priority outbound platform, and a secondary goods priority outbound record is generated and pushed to the warehouse operations department terminal to ensure that the outbound flow is completed quickly.

[0062] When the product identification result is a recall isolation signal, the warehouse location distribution information of the third-level products is retrieved, the responsible warehouse personnel corresponding to each warehouse location are retrieved, and the warehouse location command is sent to the mobile terminal of the responsible warehouse personnel. The mobile terminal includes, but is not limited to, smart handheld terminals and smartwatches, to obtain the real-time location of the responsible warehouse personnel.

[0063] Retrieve information on all available isolation areas within the storage area, including the remaining capacity and location of each isolation area. Calculate the distance between each level 3 goods storage location and the responsible warehouse personnel, and the distance between each level 3 goods storage location and the available isolation area. Sum the two to obtain the interval distance. After filtering all available isolation areas based on the condition that the remaining capacity is greater than or equal to the level 3 goods inventory quantity, select the isolation area with the smallest interval distance and mark it as the target isolation area. Generate several isolation paths based on the location of the level 3 goods storage area in the target isolation area, and select the shortest path as the target isolation path.

[0064] The warehouse location distribution of Level 3 goods, the location of the target isolation area, and the target isolation path are sent to the mobile terminal of the responsible warehouse personnel. After the responsible warehouse personnel complete the transfer operation of Level 3 goods to the target isolation area, they send a transfer completion instruction to the identification management module through the mobile terminal. When the identification management module receives the transfer completion instruction, it synchronously updates the inventory status of Level 3 goods to recall isolation pending disposal, automatically generates a recall goods isolation record, and pushes it to the interface terminal of the quality management department and the corresponding supplier. At the same time, it locks all outbound permissions for Level 1 goods.

[0065] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent warehouse adaptive identification system, characterized in that, include: The multimodal data acquisition module collects image information of goods stored in the warehouse, seal identification information, and warehouse environment status information, and sends them to the defect identification module, seal identification module, and remaining shelf life identification module, respectively. The defect identification module identifies defective areas of goods based on warehouse image information, and identifies and analyzes defect values ​​based on these defective areas. The sealing recognition module calculates the similarity between the sequence of goods to be detected and the standard sequence of goods through a dynamic time warping algorithm, and identifies the sealing of goods based on the similarity. The remaining shelf life identification module predicts the consumed shelf life score of the goods based on the goods' storage environment status information, goods' defect value, and sealing performance, and then analyzes to obtain the actual remaining shelf life of the goods. The identification output module constructs a goods storage status discrimination vector based on the goods defect value, sealing status and actual remaining shelf life, and identifies the goods storage status level through a multivariate ordered logistic regression model to obtain the goods storage identification result; The identification and management module executes corresponding strategies based on the identification results of the goods storage.

2. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, The specific steps for identifying the defective areas of the goods are as follows: The goods storage image and the normal reference image are preprocessed, and pixel-level difference operation is performed on the preprocessed goods storage image and the normal reference image to obtain the difference image. Gaussian blurring is applied to the difference image to obtain the processed difference image; The goods storage image, the normal reference image, and the processed difference image are stitched together along the R, B, and G channels respectively to obtain a three-channel input tensor; the three-channel input tensor is then input into a lightweight convolutional neural network to output a defect probability heatmap. The defect probability heatmap is upsampled using bilinear interpolation and binarized using a first threshold to obtain a preliminary defect location mask. Connectivity analysis is performed on the initial defect location mask to extract all connected regions. Each connected region is enclosed by the minimum bounding rectangle and its area is extracted. A second threshold is set. If the area of ​​a connected region is less than the preset second threshold, it is judged as noise and removed. Otherwise, it is judged as a candidate defect region. Based on this, a set of candidate defect regions is obtained.

3. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, Based on the identification and analysis of the defective areas of the goods, the specific steps are as follows: Extract the area of ​​each defect region in the candidate defect region set and sum them up to obtain the total area of ​​the candidate defect regions. Calculate the ratio of the total area of ​​the candidate defect regions to the total area of ​​the region to obtain the defect area ratio. Texture features, including contrast, entropy, and energy, are extracted from the candidate defect region set using local binary mode. Similarly, reference texture features, including reference contrast, reference entropy, and reference energy, are extracted from the normal reference image. The ratio of the reference contrast to the contrast of the candidate defect region set is calculated to generate contrast anomalies. Similarly, entropy anomalies and energy anomalies are obtained. The contrast anomalies, entropy anomalies, and energy anomalies of the candidate defect region set are accumulated to obtain the texture anomalies of the candidate defect region set. The color features of the candidate defect region set, including RGB values, are obtained, and a color histogram of the candidate defect region set is constructed. Similarly, a color histogram of the corresponding region of the normal reference image is constructed. The color anomaly values ​​of the candidate defect region set are then calculated using a color anomaly detection model. ; Based on the defect area ratio, texture outlier, and color outlier of the candidate defect region set, the formula is used to... Calculate the defect value of the goods ,in, This indicates the defect area ratio and color outlier value of the candidate defect region. These are the weighting factors for the product defect values ​​corresponding to the preset defect area ratio, texture anomaly value, and color anomaly value.

4. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, The similarity between the sequence of goods to be tested and the standard sequence of goods is calculated using a dynamic time warping algorithm. The airtightness of the goods is then identified based on the similarity. The specific steps are as follows: Extract the time-domain waveform sequence of the reflected wave, and denote it as the sequence to be detected. n is the number of sampling points in the sequence to be detected, corresponding to the time series length of the sequence to be detected; Obtain the time-domain waveform sequence of the reflected wave from the goods that have passed the sealing test, and denote it as the standard sequence. m is the number of sampling points in the standard sequence, corresponding to the time series length of the standard sequence; extract the product sequence to be detected and the product standard sequence, denoise the sequence by moving average filtering, and normalize the filtered sequence; Construct a local cost matrix C with dimensions n×m, where n is the length of the sequence to be detected and m is the length of the standard sequence. Each element in the local cost matrix C... This represents the Euclidean distance between the amplitude of the i-th time point of the sequence to be detected and the amplitude of the j-th time point of the standard sequence; Construct the cumulative distance matrix Cumulative distance matrix The size is consistent with the local cost matrix C. First, initialize the boundary conditions and establish the recursion starting point, including the initialization of the first row and the first column. Then, initialize the cumulative distance matrix. The first line has the following recursive rule: And the initial point satisfies For the cumulative distance matrix The recursive rule for the first column is: And the initial point satisfies ; For the cumulative distance matrix Non-boundary points in It is obtained by adding the minimum cumulative distances in the left, top, and top-left directions to the local cost matrix C, and its recursive formula is: ; Set the cumulative distance matrix Path constraint window width ,in For any point If the constraints are met If the value is true, then retain that point for recursion; otherwise, remove it. Let it be infinity, and exclude it from the recursive process; The cumulative distance matrix bottom right element The DTW distance between the test sequence and the standard sequence is denoted as . ; through formula The airtightness of the goods was calculated. , This indicates the preset DTW distance threshold.

5. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, Based on the information regarding the storage environment, defect values, and sealing of the goods, the consumed shelf life fraction of the goods is predicted, and then the actual remaining shelf life of the goods is analyzed. The specific steps are as follows: Acquire the warehousing environment status data of goods from warehousing to the present, and use the standard deviation model to measure and calculate the uneven distribution of various indicators in the warehousing environment status data from warehousing to the present, to obtain the warehousing environment discrete value; extract the warehousing standard environment status data corresponding to the product's expiration date, and compare and analyze the warehousing environment status data from warehousing to the present with the warehousing standard environment status data to calculate the environmental deviation value. The environmental discrete values ​​and environmental deviation values ​​are input into the weighted model to obtain the environmental impact value; Based on environmental impact value, product defect value, and sealing performance, length conversion is performed according to a preset ratio. A circle is constructed with the sealing performance length as the radius, and a sector region one is constructed within the circle with the environmental impact value length as arc length one. A sector region two is constructed within the circle with the product defect value length as arc length two. The remaining arc area values ​​outside of sector regions one and two are extracted and denoted as... Through the consumed expiration date prediction model Calculate the consumed expiration period fraction ; Obtain the product's expiration date and the current date, and perform a difference calculation to obtain the product's baseline remaining validity period, denoted as [missing information]. Based on the product's remaining shelf life and the percentage of expired shelf life, the formula is used to determine the product's shelf life. The actual remaining validity period was calculated. .

6. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, Based on the product defect value, sealing performance, and actual remaining shelf life, a product storage status discrimination vector is constructed. A multivariate ordered logistic regression model is used to identify the product storage status level, yielding the product storage identification result. The specific steps are as follows: Extract product defect values, sealing performance, and actual remaining shelf life to construct a product storage status discrimination vector. ; Define the dependent variable This indicates the storage status level of goods and is a three-category variable. Indicates level one. Indicates level two. The three levels are represented by a multivariate ordered logistic regression model. The cumulative probability of the goods storage status level is calculated. These are the intercept parameters obtained from model fitting; different levels of k correspond to different intercepts. These are the feature weight vector and discriminant vector obtained from model fitting. The dot product reflects the degree of influence of three characteristics—defect value, sealing performance, and remaining shelf life—on the storage status level. Through formula The probabilities of each storage status level are derived, and the category with the highest probability is selected as the storage status level of the goods. When the storage status level of goods is Level 1, a normal storage signal is generated and the goods are recorded as Level 1 goods; when the storage status level of goods is Level 2, a priority outbound signal is generated and the goods are recorded as Level 2 goods; when the storage status level of goods is Level 3, a recall isolation signal is generated and the goods are recorded as Level 3 goods; the priority outbound signal, Level 2 goods, recall isolation signal, and Level 3 goods are recorded as the goods storage identification result.

7. The intelligent warehousing adaptive identification system based on deep learning according to claim 1, characterized in that, The specific steps for implementing the corresponding strategy based on the goods storage identification results are as follows: When the product identification result is a priority outbound signal, the storage location distribution information of the secondary products is retrieved, the responsible picking personnel for each storage location are retrieved, and a storage location positioning command is sent to the mobile terminal of the responsible picking personnel to obtain their real-time location. At the same time, the status information of available AGVs in the storage area is retrieved, including the current battery level, current task progress, and load capacity of the AGVs. After filtering all available AGVs according to the conditions of battery level ≥80% and current task progress ≤20%, the AGV with the largest load capacity is selected and marked as the target AGV. Several picking paths are generated based on the secondary product storage location, picking port, and the current location of the target AGV. The path with the shortest path length and the lowest operation congestion rate is selected as the target picking path. The location distribution of secondary goods, target AGV number, and target picking route are sent to the mobile terminal of the responsible picking personnel and the AGV scheduling system. When the product identification result is a recall isolation signal, the warehouse location distribution information of the third-level products is retrieved, the responsible warehouse personnel corresponding to each warehouse location are retrieved, and the warehouse location command is sent to the mobile terminal of the responsible warehouse personnel to obtain the real-time location of the responsible warehouse personnel. Retrieve information on all available isolation areas within the storage area, including the remaining capacity and location of each isolation area. Calculate the distance between each level 3 goods storage location and the responsible warehouse personnel, and the distance between each level 3 goods storage location and the available isolation area. Sum the two to obtain the interval distance. After filtering all available isolation areas based on the condition that the remaining capacity is greater than or equal to the level 3 goods inventory quantity, select the isolation area with the smallest interval distance and mark it as the target isolation area. Generate several isolation paths based on the location of the level 3 goods storage area in the target isolation area, and select the shortest path as the target isolation path. The location distribution of the three-level goods, the location of the target isolation area, and the target isolation path are sent to the mobile device of the responsible warehouse personnel.

Images